Instant Ebook Access, One Click Away – Begin at ebookgate.

com

Image Analysis Classification and Change Detection

in Remote Sensing With Algorithms for ENVI IDL and
Python 3rd Edition Morton John Canty

https://ebookgate.com/product/image-analysis-classification-
and-change-detection-in-remote-sensing-with-algorithms-for-
envi-idl-and-python-3rd-edition-morton-john-canty/

OR CLICK BUTTON

DOWLOAD EBOOK

Get Instant Ebook Downloads – Browse at https://ebookgate.com

Click here to visit ebookgate.com and download ebook now
Instant digital products (PDF, ePub, MOBI) available
Download now and explore formats that suit you...

Image Registration for Remote Sensing 1st Edition

Jacqueline Le Moigne

https://ebookgate.com/product/image-registration-for-remote-
sensing-1st-edition-jacqueline-le-moigne/

ebookgate.com

Spectral Spatial Classification of Hyperspectral Remote

Sensing Images 1st Edition Jon Atli Benediktsson

https://ebookgate.com/product/spectral-spatial-classification-of-
hyperspectral-remote-sensing-images-1st-edition-jon-atli-benediktsson/

ebookgate.com

Kernel Methods for Remote Sensing Data Analysis 1st

Edition Gustavo Camps-Valls

https://ebookgate.com/product/kernel-methods-for-remote-sensing-data-
analysis-1st-edition-gustavo-camps-valls/

ebookgate.com

Remote Sensing of Impervious Surfaces Taylor Francis

Series in Remote Sensing Applications 1st Edition Qihao
Weng
https://ebookgate.com/product/remote-sensing-of-impervious-surfaces-
taylor-francis-series-in-remote-sensing-applications-1st-edition-
qihao-weng/
ebookgate.com
Detection Algorithms for Wireless Communications With
Applications to Wired and Storage Systems 1st Edition
Gianluigi Ferrari
https://ebookgate.com/product/detection-algorithms-for-wireless-
communications-with-applications-to-wired-and-storage-systems-1st-
edition-gianluigi-ferrari/
ebookgate.com

Datums and Map Projections For Remote Sensing GIS and

Surveying 2Rev Ed Edition Iliffe

https://ebookgate.com/product/datums-and-map-projections-for-remote-
sensing-gis-and-surveying-2rev-ed-edition-iliffe/

ebookgate.com

Scale Issues in Remote Sensing 1st Edition Qihao Weng

https://ebookgate.com/product/scale-issues-in-remote-sensing-1st-
edition-qihao-weng/

ebookgate.com

Remote Sensing Imagery 1st Edition Florence Tupin

https://ebookgate.com/product/remote-sensing-imagery-1st-edition-
florence-tupin/

ebookgate.com

Regression Analysis with Python Luca Massaron

https://ebookgate.com/product/regression-analysis-with-python-luca-
massaron/

ebookgate.com
 IMAGE ANALYSIS,
CLASSIFICATION AND
CHANGE DETECTION
IN REMOTE SENSING
With Algorithms for ENVI/IDL and Python

THIRD EDITION
This page intentionally left blank
 IMAGE ANALYSIS,
CLASSIFICATION AND
CHANGE DETECTION
IN REMOTE SENSING
With Algorithms for ENVI/IDL and Python

THIRD EDITION

Morton J. Canty
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2014 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20140206

International Standard Book Number-13: 978-1-4665-7038-2 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize to
copyright holders if permission to publish in this form has not been obtained. If any copyright material has
not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmit-
ted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented,
including photocopying, microfilming, and recording, or in any information storage or retrieval system,
without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.
com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and
registration for a variety of users. For organizations that have been granted a photocopy license by the CCC,
a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used
only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
Contents

Preface to the First Edition xi

Preface to the Second Edition xiii

Preface to the Third Edition xv

List of Figures xvii

Program Listings xxiii

1 Images, Arrays, and Matrices 1

1.1 Multispectral satellite images . . . . . . . . . . . . . . . . . . 4
1.2 Synthetic aperture radar images . . . . . . . . . . . . . . . . . 8
1.3 Algebra of vectors and matrices . . . . . . . . . . . . . . . . . 11
1.3.1 Elementary properties . . . . . . . . . . . . . . . . . . 11
1.3.2 Square matrices . . . . . . . . . . . . . . . . . . . . . . 14
1.3.3 Singular matrices . . . . . . . . . . . . . . . . . . . . . 16
1.3.4 Symmetric, positive definite matrices . . . . . . . . . . 16
1.3.5 Linear dependence and vector spaces . . . . . . . . . . 17
1.4 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . 18
1.5 Singular value decomposition . . . . . . . . . . . . . . . . . . 21
1.6 Finding minima and maxima . . . . . . . . . . . . . . . . . . 23
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2 Image Statistics 35
2.1 Random variables . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.1 Discrete random variables . . . . . . . . . . . . . . . . 36
2.1.2 Continuous random variables . . . . . . . . . . . . . . 37
2.1.3 Random vectors . . . . . . . . . . . . . . . . . . . . . 40
2.1.4 The normal distribution . . . . . . . . . . . . . . . . . 42
2.1.5 The gamma distribution and its derivatives . . . . . . 45
2.2 Parameter estimation . . . . . . . . . . . . . . . . . . . . . . . 47
2.2.1 Random samples . . . . . . . . . . . . . . . . . . . . . 47
2.2.2 Sample distributions and interval estimators . . . . . . 50
2.3 Multivariate distributions . . . . . . . . . . . . . . . . . . . . 53
2.3.1 Vector sample functions and the data matrix . . . . . 54
2.3.2 Provisional means . . . . . . . . . . . . . . . . . . . . 55

v
vi Image Analysis, Classification, and Change Detection in Remote Sensing

2.3.3 Real and complex multivariate sample distributions . . 59

2.4 Bayes’ Theorem, likelihood and classification . . . . . . . . . 61
2.5 Hypothesis testing . . . . . . . . . . . . . . . . . . . . . . . . 64
2.6 Ordinary linear regression . . . . . . . . . . . . . . . . . . . . 69
2.6.1 One independent variable . . . . . . . . . . . . . . . . 69
2.6.2 Coefficient of determination (R2 ) . . . . . . . . . . . . 71
2.6.3 More than one independent variable . . . . . . . . . . 72
2.6.4 Regularization, duality and the Gram matrix . . . . . 74
2.7 Entropy and information . . . . . . . . . . . . . . . . . . . . . 76
2.7.1 Kullback–Leibler divergence . . . . . . . . . . . . . . . 78
2.7.2 Mutual information . . . . . . . . . . . . . . . . . . . . 78
2.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3 Transformations 85
3.1 The discrete Fourier transform . . . . . . . . . . . . . . . . . 85
3.2 The discrete wavelet transform . . . . . . . . . . . . . . . . . 91
3.2.1 Haar wavelets . . . . . . . . . . . . . . . . . . . . . . . 91
3.2.2 Image compression . . . . . . . . . . . . . . . . . . . . 95
3.2.3 Multiresolution analysis . . . . . . . . . . . . . . . . . 98
3.3 Principal components . . . . . . . . . . . . . . . . . . . . . . 105
3.3.1 Image compression and reconstruction . . . . . . . . . 107
3.3.2 Primal solution . . . . . . . . . . . . . . . . . . . . . . 108
3.3.3 Dual solution . . . . . . . . . . . . . . . . . . . . . . . 110
3.4 Minimum noise fraction . . . . . . . . . . . . . . . . . . . . . 111
3.4.1 Additive noise . . . . . . . . . . . . . . . . . . . . . . . 112
3.4.2 Minimum noise fraction in ENVI . . . . . . . . . . . . 115
3.5 Spatial correlation . . . . . . . . . . . . . . . . . . . . . . . . 117
3.5.1 Maximum autocorrelation factor . . . . . . . . . . . . 117
3.5.2 Noise estimation . . . . . . . . . . . . . . . . . . . . . 119
3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

4 Filters, Kernels and Fields 127

4.1 The Convolution Theorem . . . . . . . . . . . . . . . . . . . . 127
4.2 Linear filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.3 Wavelets and filter banks . . . . . . . . . . . . . . . . . . . . 135
4.3.1 One-dimensional arrays . . . . . . . . . . . . . . . . . 135
4.3.2 Two-dimensional arrays . . . . . . . . . . . . . . . . . 141
4.4 Kernel methods . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.4.1 Valid kernels . . . . . . . . . . . . . . . . . . . . . . . 146
4.4.2 Kernel PCA . . . . . . . . . . . . . . . . . . . . . . . . 150
4.5 Gibbs–Markov random fields . . . . . . . . . . . . . . . . . . 152
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Contents vii

5 Image Enhancement and Correction 159

5.1 Lookup tables and histogram functions . . . . . . . . . . . . . 159
5.2 High-pass spatial filtering and feature extraction . . . . . . . 161
5.2.1 Sobel filter . . . . . . . . . . . . . . . . . . . . . . . . 161
5.2.2 Laplacian-of-Gaussian filter . . . . . . . . . . . . . . . 163
5.2.3 OpenCV functions . . . . . . . . . . . . . . . . . . . . 166
5.2.4 Invariant moments . . . . . . . . . . . . . . . . . . . . 170
5.3 Panchromatic sharpening . . . . . . . . . . . . . . . . . . . . 176
5.3.1 HSV fusion . . . . . . . . . . . . . . . . . . . . . . . . 176
5.3.2 Brovey fusion . . . . . . . . . . . . . . . . . . . . . . . 179
5.3.3 PCA fusion . . . . . . . . . . . . . . . . . . . . . . . . 179
5.3.4 DWT fusion . . . . . . . . . . . . . . . . . . . . . . . . 180
5.3.5 À trous fusion . . . . . . . . . . . . . . . . . . . . . . . 181
5.3.6 A quality index . . . . . . . . . . . . . . . . . . . . . . 184
5.4 Radiometric correction of polarimetric SAR imagery . . . . . 185
5.4.1 Speckle statistics . . . . . . . . . . . . . . . . . . . . . 185
5.4.2 Multi-look data . . . . . . . . . . . . . . . . . . . . . . 188
5.4.3 Speckle filtering . . . . . . . . . . . . . . . . . . . . . . 191
5.5 Topographic correction . . . . . . . . . . . . . . . . . . . . . . 196
5.5.1 Rotation, scaling and translation . . . . . . . . . . . . 196
5.5.2 Imaging transformations . . . . . . . . . . . . . . . . . 197
5.5.3 Camera models and RFM approximations . . . . . . . 199
5.5.4 Stereo imaging and digital elevation models . . . . . . 201
5.5.5 Slope and aspect . . . . . . . . . . . . . . . . . . . . . 207
5.5.6 Illumination correction . . . . . . . . . . . . . . . . . . 208
5.6 Image–image registration . . . . . . . . . . . . . . . . . . . . 212
5.6.1 Frequency domain registration . . . . . . . . . . . . . 212
5.6.2 Feature matching . . . . . . . . . . . . . . . . . . . . . 217
5.6.3 Re-sampling with ground control points . . . . . . . . 222
5.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

6 Supervised Classification Part 1 231

6.1 Maximizing the a posteriori probability . . . . . . . . . . . . 232
6.2 Training data and separability . . . . . . . . . . . . . . . . . . 233
6.3 Maximum likelihood classification . . . . . . . . . . . . . . . . 238
6.3.1 ENVI’s maximum likelihood classifier . . . . . . . . . 239
6.3.2 A modified classifier for ENVI and a Python script . . 241
6.4 Gaussian kernel classification . . . . . . . . . . . . . . . . . . 243
6.5 Neural networks . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.5.1 The neural network classifier . . . . . . . . . . . . . . 252
6.5.2 Cost functions . . . . . . . . . . . . . . . . . . . . . . 257
6.5.3 Backpropagation . . . . . . . . . . . . . . . . . . . . . 258
6.5.4 Overfitting and generalization . . . . . . . . . . . . . . 264
6.6 Support vector machines . . . . . . . . . . . . . . . . . . . . . 266
6.6.1 Linearly separable classes . . . . . . . . . . . . . . . . 267
viii Image Analysis, Classification, and Change Detection in Remote Sensing

6.6.2 Overlapping classes . . . . . . . . . . . . . . . . . . . . 272

6.6.3 Solution with sequential minimal optimization . . . . . 274
6.6.4 Multiclass SVMs . . . . . . . . . . . . . . . . . . . . . 275
6.6.5 Kernel substitution . . . . . . . . . . . . . . . . . . . . 277
6.6.6 A modified SVM classifier . . . . . . . . . . . . . . . . 278
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280

7 Supervised Classification Part 2 285

7.1 Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . 285
7.1.1 Majority filtering . . . . . . . . . . . . . . . . . . . . . 286
7.1.2 Probabilistic label relaxation . . . . . . . . . . . . . . 286
7.2 Evaluation and comparison of classification accuracy . . . . . 288
7.2.1 Accuracy assessment . . . . . . . . . . . . . . . . . . . 289
7.2.2 Cross-validation on the cloud . . . . . . . . . . . . . . 294
7.2.3 Model comparison . . . . . . . . . . . . . . . . . . . . 296
7.3 Adaptive boosting . . . . . . . . . . . . . . . . . . . . . . . . 300
7.4 Classification of polarimetric SAR imagery . . . . . . . . . . . 305
7.5 Hyperspectral image analysis . . . . . . . . . . . . . . . . . . 307
7.5.1 Spectral mixture modeling . . . . . . . . . . . . . . . . 307
7.5.2 Unconstrained linear unmixing . . . . . . . . . . . . . 310
7.5.3 Intrinsic end-members and pixel purity . . . . . . . . . 311
7.5.4 Anomaly detection: The RX algorithm . . . . . . . . . 313
7.5.5 Anomaly detection: The kernel RX algorithm . . . . . 315
7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320

8 Unsupervised Classification 323

8.1 Simple cost functions . . . . . . . . . . . . . . . . . . . . . . . 324
8.2 Algorithms that minimize the simple cost functions . . . . . . 326
8.2.1 K-means clustering . . . . . . . . . . . . . . . . . . . . 327
8.2.2 Kernel K-means clustering . . . . . . . . . . . . . . . . 329
8.2.3 Extended K-means clustering . . . . . . . . . . . . . . 332
8.2.4 Agglomerative hierarchical clustering . . . . . . . . . . 335
8.2.5 Fuzzy K-means clustering . . . . . . . . . . . . . . . . 337
8.3 Gaussian mixture clustering . . . . . . . . . . . . . . . . . . . 339
8.3.1 Expectation maximization . . . . . . . . . . . . . . . . 340
8.3.2 Simulated annealing . . . . . . . . . . . . . . . . . . . 343
8.3.3 Partition density . . . . . . . . . . . . . . . . . . . . . 343
8.3.4 Implementation notes . . . . . . . . . . . . . . . . . . 344
8.4 Including spatial information . . . . . . . . . . . . . . . . . . 346
8.4.1 Multiresolution clustering . . . . . . . . . . . . . . . . 346
8.4.2 Spatial clustering . . . . . . . . . . . . . . . . . . . . . 348
8.5 A benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
8.6 The Kohonen self-organizing map . . . . . . . . . . . . . . . . 353
8.7 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . 357
8.7.1 Segmenting a classified image . . . . . . . . . . . . . . 358
Contents ix

8.7.2 Object-based classification . . . . . . . . . . . . . . . . 360

8.7.3 The mean shift . . . . . . . . . . . . . . . . . . . . . . 361
8.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363

9 Change Detection 369

9.1 Algebraic methods . . . . . . . . . . . . . . . . . . . . . . . . 370
9.2 Postclassification comparison . . . . . . . . . . . . . . . . . . 371
9.3 Principal components analysis (PCA) . . . . . . . . . . . . . 371
9.3.1 Iterated PCA . . . . . . . . . . . . . . . . . . . . . . . 374
9.3.2 Kernel PCA . . . . . . . . . . . . . . . . . . . . . . . . 376
9.4 Multivariate alteration detection (MAD) . . . . . . . . . . . . 377
9.4.1 Canonical correlation analysis (CCA) . . . . . . . . . . 379
9.4.2 Orthogonality properties . . . . . . . . . . . . . . . . . 381
9.4.3 Scale invariance . . . . . . . . . . . . . . . . . . . . . . 384
9.4.4 Iteratively re-weighted MAD . . . . . . . . . . . . . . 384
9.4.5 Correlation with the original observations . . . . . . . 387
9.4.6 Regularization . . . . . . . . . . . . . . . . . . . . . . 387
9.4.7 Postprocessing . . . . . . . . . . . . . . . . . . . . . . 391
9.5 Decision thresholds . . . . . . . . . . . . . . . . . . . . . . . . 391
9.6 Unsupervised change classification . . . . . . . . . . . . . . . 395
9.7 Change detection with polarimetric SAR imagery . . . . . . . 398
9.7.1 Single polarimetry: The gamma distribution . . . . . . 398
9.7.2 Quad polarimetry: The complex Wishart distribution 400
9.8 Radiometric normalization of multispectral imagery . . . . . 403
9.8.1 Scatterplot matching . . . . . . . . . . . . . . . . . . . 404
9.8.2 IR-MAD normalization . . . . . . . . . . . . . . . . . 406
9.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411

A Mathematical Tools 415

A.1 Cholesky decomposition . . . . . . . . . . . . . . . . . . . . . 415
A.2 Vector and inner product spaces . . . . . . . . . . . . . . . . 417
A.3 Complex numbers, vectors and matrices . . . . . . . . . . . . 418
A.4 Least squares procedures . . . . . . . . . . . . . . . . . . . . . 420
A.4.1 Recursive linear regression . . . . . . . . . . . . . . . . 420
A.4.2 Orthogonal linear regression . . . . . . . . . . . . . . . 422
A.5 Proof of Theorem 7.1 . . . . . . . . . . . . . . . . . . . . . . . 425

B Efficient Neural Network Training Algorithms 429

B.1 The Hessian matrix . . . . . . . . . . . . . . . . . . . . . . . . 429
B.1.1 The R-operator . . . . . . . . . . . . . . . . . . . . . . 430
B.1.2 Calculating the Hessian . . . . . . . . . . . . . . . . . 433
B.2 Scaled conjugate gradient training . . . . . . . . . . . . . . . 434
B.2.1 Conjugate directions . . . . . . . . . . . . . . . . . . . 434
B.2.2 Minimizing a quadratic function . . . . . . . . . . . . 437
B.2.3 The algorithm . . . . . . . . . . . . . . . . . . . . . . . 440
x Image Analysis, Classification, and Change Detection in Remote Sensing

B.3 Kalman filter training . . . . . . . . . . . . . . . . . . . . . . 444

B.3.1 Linearization . . . . . . . . . . . . . . . . . . . . . . . 444
B.3.2 The algorithm . . . . . . . . . . . . . . . . . . . . . . . 446
B.4 A neural network classifier with hybrid training . . . . . . . . 452

C ENVI Extensions in IDL 453

C.1 Installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
C.2 Documentation . . . . . . . . . . . . . . . . . . . . . . . . . . 454
C.2.1 ENVI extensions for Chapter 4 . . . . . . . . . . . . . 455
C.2.2 ENVI extensions for Chapter 5 . . . . . . . . . . . . . 455
C.2.3 ENVI extensions for Chapter 6 . . . . . . . . . . . . . 460
C.2.4 ENVI extensions for Chapter 7 . . . . . . . . . . . . . 461
C.2.5 ENVI extensions for Chapter 8 . . . . . . . . . . . . . 463
C.2.6 ENVI extensions for Chapter 9 . . . . . . . . . . . . . 466

D Python Scripts 471

D.1 Installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
D.1.1 Required packages . . . . . . . . . . . . . . . . . . . . 471
D.1.2 Eclipse and Pydev . . . . . . . . . . . . . . . . . . . . 472
D.2 Documentation . . . . . . . . . . . . . . . . . . . . . . . . . . 473
D.2.1 Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . 473
D.2.2 Scripts for Chapter 1 . . . . . . . . . . . . . . . . . . . 473
D.2.3 Scripts for Chapter 4 . . . . . . . . . . . . . . . . . . . 474
D.2.4 Scripts for Chapter 5 . . . . . . . . . . . . . . . . . . . 474
D.2.5 Scripts for Chapter 6 . . . . . . . . . . . . . . . . . . . 476
D.2.6 Scripts for Chapter 7 . . . . . . . . . . . . . . . . . . . 477
D.2.7 Scripts for Chapter 8 . . . . . . . . . . . . . . . . . . . 478
D.2.8 Scripts for Chapter 9 . . . . . . . . . . . . . . . . . . . 479

Mathematical Notation 481

References 483

Index 495
Preface to the First Edition

This textbook had its beginnings as a set of notes to accompany seminars and
lectures conducted at the Geographical Institute of Bonn University and at its
associated Center for Remote Sensing of Land Cover. Lecture notes typically
continue to be refined and polished over the years until the question inevitably
poses itself: “Why not have them published?” The answer of course is “By
all means, if they contribute something new and useful.”
So what is “new and useful” here? This is a book about remote sensing
image analysis with a distinctly mathematical-algorithmic-computer-oriented
flavor, intended for graduate-level teaching and with, to borrow from the
remote sensing jargon, a rather restricted FOV. It does not attempt to match
the wider fields of view of existing texts on the subject, such as Schowengerdt
(1997), Richards and Jia (2006), Jensen (2005) and others. However, the
topics that are covered are dealt with in considerable depth, and I believe that
this coverage fills an important gap. Many aspects of the analysis of remote
sensing data are quite technical and tend to be intimidating to students with
moderate mathematical backgrounds. At the same time, one often witnesses
a desire on the part of students to apply advanced methods and to modify
them to fit their particular research problems. Fulfilling the latter wish, in
particular, requires more than superficial understanding of the material.
The focus of the book is on pixel-oriented analysis of visual/infrared Earth
observation satellite imagery. Among the topics that get the most attention
are the discrete wavelet transform, image fusion, supervised classification with
neural networks, clustering algorithms and statistical change detection meth-
ods. The first two chapters introduce the mathematical and statistical tools
necessary in order to follow later developments. Chapters 3 and 4 deal with
spatial/spectral transformations, convolutions and filtering of multispectral
image arrays. Chapter 5 treats image enhancement and some of the pre-
processing steps that precede classification and change detection. Chapters
6 and 7 are concerned, respectively, with supervised and unsupervised land
cover classification. The last chapter is about change detection with heavy
emphasis on the use of canonical correlation analysis. Each of the 8 chap-
ters concludes with exercises, some of which are small programming projects,
intended to illustrate or justify the foregoing development. Solutions to the
exercises are included in a separate booklet. Appendix A provides some ad-
ditional mathematical/statistical background and Appendix B develops two
efficient training algorithms for neural networks. Finally, Appendix C de-
scribes the installation and use of the many computer programs introduced

xi
xii Image Analysis, Classification and Change Detection in Remote Sensing

in the course of the book.

I’ve made considerable effort to maintain a consistent, clear mathemati-
cal style throughout. Although the developments in the text are admittedly
uncompromising, there is nothing that, given a little perseverance, cannot
be followed by a reader who has grasped the elementary matrix algebra and
statistical concepts explained in the first two chapters. If the student has am-
bitions to write his or her own image analysis programs, then he or she must
be prepared to “get the maths right” beforehand. There are, heaven knows,
enough pitfalls to worry about thereafter.
All of the illustrations and applications in the text are programmed in RSI’s
ENVI/IDL. The software is available for download at the publisher’s website:
http://www.crcpress.com/e products/downloads/default.asp
Given the plethora of image analysis and geographic information system (GIS)
software systems on the market or available under open source license, one
might think that the choice of computer environment would have been diffi-
cult. It wasn’t. IDL is an extremely powerful, array- and graphics-oriented,
universal programming language with a versatile interface (ENVI) for im-
porting and analyzing remote sensing data — a peerless combination for my
purposes. Extending the ENVI interface in IDL in order to implement new
methods and algorithms of arbitrary sophistication is both easy and fun.
So, apart from some exposure to elementary calculus (and the aforesaid
perseverance), the only other prerequisites for the book are a little familiarity
with the ENVI environment and the basic knowledge of IDL imparted by such
excellent introductions as Fanning (2000) or Gumley (2002). For everyday
problems with IDL at any level from “newbie” on upward, help and solace
are available at the newsgroup
comp.lang.idl-pvwave
frequented by some of the friendliest and most competent gurus on the net.
I would like to express my thanks to Rudolf Avenhaus and Allan Nielsen
for their many comments and suggestions for improvement of the manuscript
and to CRC Press for competent assistance in its preparation. Part of the
software documented in the text was developed within the Global Monitor-
ing for Security and Stability (GMOSS) network of excellence funded by the
European Commission.

Morton Canty
Preface to the Second Edition

Shortly after the manuscript for the first edition of this book went to the pub-
lisher, ENVI 4.3 appeared along with, among other new features, a support
vector machine classifier. Although my decision not to include the SVM in
the original text was a conscious one (I balked at the thought of writing my
own IDL implementation), this event did point to a rather glaring omission
in a book purporting to be partly about land use/land cover classification.
So, almost immediately, I began to dream of a Revised Second Edition and
to pester CRC Press for a contract. This was happily forthcoming and the
present edition now has a fairly long section on supervised classification with
support vector machines.
The SVM is just one example of so-called kernel methods for nonlinear
data analysis, and I decided to make kernelization one of the themes of the
revised text. The treatment begins with a dual formulation for ridge regres-
sion in Chapter 2 and continues through kernel principal components analysis
in Chapters 3 and 4, support vector machines in Chapter 6, kernel K-means
clustering in Chapter 8 and nonlinear change detection in Chapter 9. Other
new topics include entropy and mutual information (Chapter 1), adaptive
boosting (Chapter 7) and image segmentation (Chapter 8). In order to ac-
commodate the extended material on supervised classification, discussion is
now spread over the two Chapters 6 and 7. The exercises at the end of each
chapter have been extended and re-worked and, as for the first edition, a
solutions manual is provided.
I have written several additional IDL extensions to ENVI to accompany the
new themes, which are available, together with updated versions of previous
programs, for download on the Internet. In order to accelerate some of the
more computationally intensive routines for users with access to CUDA (par-
allel processing on NVIDIA graphics processors), code is included which can
make use of the IDL bindings to CUDA provided by Tech-X Corporation in
their GPULib product:
http://gpulib.txcorp.com
Notwithstanding the revisions, the present edition remains a monograph
on pixel-oriented analysis of intermediate-resolution remote sensing imagery
with emphasis on the development and programming of statistically moti-
vated, data-driven algorithms. Important topics such as object-based feature
analysis (for high-resolution imagery), or the physics of the radiation/surface
interaction (for example, in connection with hyperspectral sensing) are only

xiii
xiv Image Analysis, Classification and Change Detection in Remote Sensing

touched upon briefly, and the huge field of radar remote sensing is left out
completely. Nevertheless, I hope that the in-depth focus on the topics covered
will continue to be of use both to practitioners as well as to teachers.
I would like to express my appreciation to Peter Reinartz and the German
Aerospace Center for permission to use the traffic scene images in Chapter 9
and to NASA’s Land Processes Distributed Active Archive Center for free and
uncomplicated access to archived ASTER imagery. Thanks also go to Peter
Messmer and Michael Galloy, Tech-X Corp., for their prompt responses to my
many cries for help with GPULib. I am especially grateful to my colleagues
Harry Vereecken and Allan Nielsen, the former for generously providing me
with the environment and resources needed to complete this book, the latter
for the continuing inspiration of our friendship and long-time collaboration.

Morton Canty
Preface to the Third Edition

A main incentive for me to write a third edition of this book stemmed from my
increasing enthusiasm for the Python programming language. I began to see
the advantage of illustrating the many image processing algorithms covered
in earlier editions of the text not only in the powerful and convenient, but
not inexpensive, ENVI/IDL world, but also on a widely available open source
platform. Python, together with the Numpy and Scipy packages, can hold its
own with any commercial array processing software system. Furthermore, the
Geospatial Data Abstraction Library (GDAL) and its Python wrappers allow
for great versatility and convenience in reading, writing and manipulating dif-
ferent image formats. This was enough to get me going on a revised textbook,
one which I hope will have appeal beyond the ENVI/IDL community.
Another incentive for a new edition was hinted at in the preface to the
previous edition, namely the lack of any discussion of the vast and increas-
ingly important field of radar remote sensing. Obviously this would be a topic
for (at least) a whole new book, so I have included material only on a very
special aspect of particular interest to me, namely multivariate statistical clas-
sification and change detection algorithms applied to polarimetric synthetic
aperture radar (polSAR) data. Up until recently, not many researchers or
practitioners have had access to this kind of data. However with the advent
of several spaceborne polarimetric SAR instruments such as the Japanese
ALOS, the Canadian Radarsat-2, the German TerraSAR-X and the Italian
COSMO-SkyMed missions, the situation has greatly improved. Chapters 5, 7
and 9 now include treatments of speckle filtering, image co-registration, super-
vised classification and multivariate change detection with multi-look polSAR
data.
The software associated with the present edition includes, along with the
ENVI/IDL extensions, Python scripts for all of the main processing, classifica-
tion and change detection algorithms. In addition, many examples discussed
in the text are illustrated with Python scripts as well as in IDL. The Appen-
dices C and D separately document the installation and use of the ENVI/IDL
and Python code. For readers who wish to use the Eclipse/Pydev develop-
ment environment (something which I highly recommend), the Python scripts
are provided in the form of a Pydev project.
What is missing in the Python world, of course, is the slick GUI provided
by ENVI. I have made no attempt to mimic an ENVI graphical environment
in Python, and the scripts provided content themselves with reading imagery
from, and writing results to, the file system. A rudimentary command line

xv
xvi Image Analysis, Classification and Change Detection in Remote Sensing

script for RGB displays of multispectral band combinations in different his-

togram enhancement modes is included.
For an excellent introduction to scientific computing in Python see Langtan-
gen (2009). The book by Westra (2013) provides valuable tips on geospatial
development in Python, including GDAL programming. The definitive refer-
ence on IDL is now certainly Galloy (2011), an absolute must for anyone who
uses the language professionally.
With version 5.0, a new ENVI graphics environment and associated API has
appeared which has a very different look and feel to the old “ENVI Classic”
environment, as it is officially referred to. Fortunately the classic environment
is still available and, for reasons of compatibility with previous versions, the
IDL programming examples in the text use the classic interface and its asso-
ciated syntax. Most of the ENVI/IDL extensions as documented in Appendix
C are provided both for the new as well as for the classic GUI/API.
I would like to express my appreciation to the German Aerospace Center
for permission to use images from the TerraSAR-X platform and to Henning
Skriver, DTU Space Denmark, for allowing me to use his EMISAR polari-
metric data. My special thanks go to Allan Nielsen and Frank Thonfeld for
acquainting me with SAR imagery analysis and to Rudolf Avenhaus for his
many helpful suggestions in matters statistical.

Morton Canty
List of Figures

1.1 ASTER color composite image over Jülich, Germany. . . . . . . 2

1.2 A spatial subset of the HH polarization band of a TerraSAR-X
image acquired over the Rhine river, Germany. . . . . . . . . . 3
1.3 A vector with two components. . . . . . . . . . . . . . . . . . . 11
1.4 Illustrating the inner product. . . . . . . . . . . . . . . . . . . . 12
1.5 A function of one variable with a minimum at x∗ . . . . . . . . . 24
1.6 First principal component of the VNIR spectral bands for the
Jülich scene. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.1 Histogram of sums of 12 uniformly distributed random num-

bers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2 Plots of the chi-square probability density. . . . . . . . . . . . . 47
2.3 Spectral band 3N of the ASTER image over Jülich with regions
of interest (ROIs) covering areas of mixed forest. . . . . . . . . 48
2.4 Histogram of the pixels in Figure 2.3 under the ROIs. . . . . . 49
2.5 Critical region for rejecting the hypothesis µ = µ0 . . . . . . . . 67

3.1 Logarithm of the power spectrum for the 3N band of the Jülich
ASTER image. . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.2 The Haar scaling function. . . . . . . . . . . . . . . . . . . . . . 91
3.3 A row of c pixel intensities represented as a piecewise constant
function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.4 Basis functions Cn for function space Vn . . . . . . . . . . . . . 92
3.5 The Haar mother wavelet. . . . . . . . . . . . . . . . . . . . . . 93
3.6 Spatial subset of the Jülich ASTER image before and after
wavelet compression. . . . . . . . . . . . . . . . . . . . . . . . . 97
3.7 Haar wavelet coefficients for the image of Figure 3.6. . . . . . . 98
3.8 Approximation to the Haar scaling function. . . . . . . . . . . . 100
3.9 Approximation to the Daubechies D4 scaling function. . . . . . 104
3.10 RGB color composites of principal components of a LANDSAT
7 ETM+ image over Jülich. . . . . . . . . . . . . . . . . . . . . 107
3.11 Eigenvalues of the minimum noise fraction (MNF) transforma-
tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.12 Two classes of observations in a two-dimensional feature space. 125

4.1 Illustrating convolution in the spatial and frequency domains. . 130

xvii
xviii Image Analysis, Classification, and Change Detection in Remote Sensing

4.2 Radar ranging. . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

4.3 Gaussian filter in the frequency domain. . . . . . . . . . . . . . 134
4.4 The 3N band of the Jülich ASTER image after low-pass filter-
ing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.5 The 3N band of the Jülich ASTER image after high-pass fil-
tering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.6 Schematic representation of Equation (4.12). . . . . . . . . . . 138
4.7 Schematic representation of the filter bank H, G. . . . . . . . . 139
4.8 Schematic representation of the filter H ∗ . . . . . . . . . . . . . 140
4.9 Schematic representation of the synthesis bank H ∗ , G∗ . . . . . 140
4.10 Wavelet filter bank. . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.11 Result of application of the filter bank of Figure 4.10 to the 3N
band of the Jülich ASTER image. . . . . . . . . . . . . . . . . 143
4.12 Recursive application of the filter bank of Figure 4.10. . . . . . 143
4.13 Ehe eigenvalues for kernel PCA on the image of Figure 3.10. . . 151
4.14 Pixel neighborhoods. . . . . . . . . . . . . . . . . . . . . . . . . 153
4.15 Cliques for a 4-neighborhood and 8-neighborhood. . . . . . . . 153

5.1 Power spectrum image of the Sobel filter. . . . . . . . . . . . . 161

5.2 Sobel edge detection on the 3N band of the Jülich ASTER
image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.3 Power spectrum image of the Laplacian filter. . . . . . . . . . . 163
5.4 Laplacian-of-Gaussian filter. . . . . . . . . . . . . . . . . . . . . 164
5.5 Image contours from the 3N band of the Jülich ASTER image. 166
5.6 Corners and edges detected from a small spatial subset of Fig-
ure 1.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.7 An “aircraft” feature translated, scaled and rotated. . . . . . . 172
5.8 Edge contours for the first spectral band of the LANDSAT 7
ETM+ image of Figure 3.10. . . . . . . . . . . . . . . . . . . . 173
5.9 Histogram of the logarithms of the first Hu moments of the
contours of Figure 5.8. . . . . . . . . . . . . . . . . . . . . . . . 175
5.10 The RGB color cube. . . . . . . . . . . . . . . . . . . . . . . . . 176
5.11 Panchromatic fusion with the principal components transfor-
mation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.12 Panchromatic sharpening of an IKONOS multispectral image
with the DWT filter bank. . . . . . . . . . . . . . . . . . . . . . 181
5.13 À trous image sharpening scheme for a multispectral to panchro-
matic resolution ratio of two. . . . . . . . . . . . . . . . . . . . 182
5.14 Panchromatic sharpening with the à trous wavelet transform. . 183
5.15 RGB composites of a TerraSAR-X image acquired over the city
of Mannheim, Germany. Left: unfiltered; right: after adaptive
filtering with the MMSE filter. . . . . . . . . . . . . . . . . . . 193
5.16 RGB composites of a TerraSAR-X image acquired over the city
of Mannheim, Germany. Left: unfiltered; right: after adaptive
filtering with the gamma MAP filter. . . . . . . . . . . . . . . . 194
List of Figures xix

5.17 Basic imaging process. . . . . . . . . . . . . . . . . . . . . . . . 198

5.18 Determining the height of a vertical structure with an RFM. . 200
5.19 The stereo imaging process. . . . . . . . . . . . . . . . . . . . . 202
5.20 ASTER along-track stereo acquisition geometry. . . . . . . . . 203
5.21 ASTER 3N band (nadir camera) over a hilly region in North
Korea. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
5.22 ASTER 3B band (back-looking camera) registered to Figure
5.21. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
5.23 A rudimentary digital elevation model (DEM) from the ASTER
stereo pair of Figures 5.21 and 5.22. . . . . . . . . . . . . . . . 205
5.24 DEM generated with the commercial product AsterDTM. . . . 205
5.25 Pixel elevations in an 8-neighborhood. . . . . . . . . . . . . . . 207
5.26 Slope image calculated with the DEM of Figure 5.24. . . . . . . 209
5.27 Aspect image calculated with the DEM of Figure 5.24. . . . . . 209
5.28 Angles involved in computation of local solar incidence. . . . . 210
5.29 Cosine of local solar incidence angle γi stretched across the
DEM of Figure 5.24. . . . . . . . . . . . . . . . . . . . . . . . . 210
5.30 Application of the C-correction method. . . . . . . . . . . . . . 211
5.31 Phase correlation of two identical images shifted relative to one
another. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
5.32 Frequency domain registration of quad polarimetric SAR im-
agery, RGB composites (HH,HV,VV). Left: TerraSAR-X, right:
Radarsat-2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
5.33 Image-image registration with contour matching. . . . . . . . . 218
5.34 Closed contours derived from an ASTER image over Nevada. . 219
5.35 Tie-points obtained by contour matching. . . . . . . . . . . . . 220
5.36 Tie-points obtained with ENVI’s feature-based matching pro-
cedure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

6.1 RGB color composite of the first three principal components of

the nine nonthermal bands of the Jülich ASTER scene. . . . . 232
6.2 Ground reference data for four land cover categories. . . . . . . 235
6.3 ROIs for supervised classification. . . . . . . . . . . . . . . . . . 236
6.4 Supervised classification of a portion of the Jülich ASTER
scene. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
6.5 Training the Gaussian kernel classifier. . . . . . . . . . . . . . . 248
6.6 A linear discriminant for two classes. . . . . . . . . . . . . . . . 249
6.7 An artificial neuron representing Equation (6.24). . . . . . . . . 250
6.8 The logistic activation function. . . . . . . . . . . . . . . . . . . 250
6.9 A single-layer neural network. . . . . . . . . . . . . . . . . . . . 252
6.10 A two-layer feed-forward neural network with L hidden neu-
rons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
6.11 An example of overfitting during neural network training. . . . 265
6.12 Distance d to the separating hyperplane. . . . . . . . . . . . . . 267
xx Image Analysis, Classification, and Change Detection in Remote Sensing

6.13 A maximal margin hyperplane for linearly separable training

data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
6.14 Slack variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
6.15 Two nonlinearly separable classes. . . . . . . . . . . . . . . . . 277
6.16 Linearly separable classes in a three-dimensional, nonlinear fea-
ture space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
6.17 Principal components analysis with a neural network. . . . . . 283

7.1 An example of post-classification processing. . . . . . . . . . . . 289

7.2 The binomial distribution for n = 2000 and θ = 0.1. . . . . . . 291
7.3 Adaptive boost training of a neural network. . . . . . . . . . . 304
7.4 Maximum likelihood classification of an EMISAR L-band quad
polarimetric SAR image acquired over a test agricultural area
in Denmark. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
7.5 AVIRIS hyperspectral image cube over the Santa Monica Moun-
tains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
7.6 AVIRIS spectrum at one pixel location. . . . . . . . . . . . . . 308
7.7 The first 50 eigenvalues of the MNF transformation. . . . . . . 311
7.8 The n-D visualizer displaying pixel purity indices. . . . . . . . . 312
7.9 Anomaly detection for an ASTER image using the RX and
kernel RX algorithms. . . . . . . . . . . . . . . . . . . . . . . . 320

8.1 Unsupervised classifications of the ASTER image over Jülich. . 331

8.2 Gaussian mixture clustering of the first five principal compo-
nents of the Jülich ASTER scene. . . . . . . . . . . . . . . . . . 350
8.3 Classification of a toy image. . . . . . . . . . . . . . . . . . . . 352
8.4 The Kohonen self-organizing map in two dimensions. . . . . . . 353
8.5 Movement of a synaptic weight vector in the direction of a
training vector. . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
8.6 Kohonen self-organizing map of the nine spectral bands of the
Jülich ASTER scene. . . . . . . . . . . . . . . . . . . . . . . . . 355
8.7 The 63 = 216 SOM neurons in the feature space of the Jülich
ASTER scene. . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
8.8 Post-classification segmentation of the cluster associated with
green vegetation in the Jülich ASTER scene. . . . . . . . . . . 357
8.9 Classification of the segments in Figure 8.8 on the basis of their
invariant moments. . . . . . . . . . . . . . . . . . . . . . . . . . 360
8.10 Standardized logarithms of the first three Hu moments of the
classified segments of Figure 8.9. . . . . . . . . . . . . . . . . . 361
8.11 Mean shift segmentation of a spatial subset of Figure 6.1. . . . 362
8.12 Two clusters which are not linearly separable. . . . . . . . . . . 364

9.1 LANDSAT 5 TM images over a water reservoir in Hindustan. . 369

9.2 Scatterplot of the two spectral bands of Figure 9.1. . . . . . . . 372
9.3 Iterated principal components analysis. . . . . . . . . . . . . . . 372
List of Figures xxi

9.4 The change probability for the bitemporal image in Figure 9.1
after five iterations of principal components analysis. . . . . . . 375
9.5 Two traffic scenes taken 0.7 seconds apart. . . . . . . . . . . . . 376
9.6 Change detection with linear and nonlinear PCA. . . . . . . . . 377
9.7 The fifth kernel principal component for the nonlinear bitem-
poral traffic scene. . . . . . . . . . . . . . . . . . . . . . . . . . 378
9.8 Multivariate alteration detection (MAD) using the six nonther-
mal bands of the bitemporal image of Figure 9.1. . . . . . . . . 382
9.9 Scatterplots of V6 vs. U6 and of V1 vs. U1 from Figure 9.8. . . . 382
9.10 The sixth MAD variate for the bitemporal image in Figure 9.1. 385
9.11 The canonical correlations ρi , i = 1 . . . 6, under 20 iterations of
the MAD transformation of the image in Figure 9.1. . . . . . . 386
9.12 Correlation of the fourth iteratively reweighted MAD (IR-MAD)
component with the seven spectral bands for the scene in Fig-
ure 9.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
9.13 The fourth IR-MAD component. . . . . . . . . . . . . . . . . . 388
9.14 Gaussian mixture fit to a MAD/MNF component. . . . . . . . 393
9.15 Color composite of MAD/MNF components 1, 2, and 3. . . . . 394
9.16 RGB composite of an IR-MAD image. . . . . . . . . . . . . . . 395
9.17 Four clusters in MAD feature space projected onto the 5,6
plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
9.18 Overlay of the harvested grain cluster onto spectral band 4 of
the August 2001 LANDSAT ETM+ image. . . . . . . . . . . . 397
9.19 Wishart change detection with TerraSAR-X data. . . . . . . . . 402
9.20 Radiometric normalization of bands 2 and 3 for two ASTER
images using scatterplot matching. . . . . . . . . . . . . . . . . 405
9.21 Principle of scatterplot matching. . . . . . . . . . . . . . . . . . 406
9.22 Mosaic of two LANDSAT 7 ETM+ images over Morocco. . . . 408
9.23 Regressions of the December 1999 reference scene on the Octo-
ber 2000 target scene. . . . . . . . . . . . . . . . . . . . . . . . 409
9.24 As Figure 9.22 after automatic radiometric normalization. . . . 410

A.1 Illustrating Inequality (A.18). . . . . . . . . . . . . . . . . . . . 426

B.1 Search directions in weight space. . . . . . . . . . . . . . . . . . 436

B.2 An isolated output neuron. . . . . . . . . . . . . . . . . . . . . 444
B.3 Determination of the synaptic weights for an isolated neuron
with the Kalman filter. . . . . . . . . . . . . . . . . . . . . . . . 446
B.4 Training with backpropagation, Kalman filter, and scaled con-
jugate gradient. . . . . . . . . . . . . . . . . . . . . . . . . . . . 450

C.1 ENL determination in ENVI Classic. . . . . . . . . . . . . . . . 458

C.2 Radiometric normalization. . . . . . . . . . . . . . . . . . . . . 468
This page intentionally left blank
Program Listings

1.1 Reading and displaying a multispectral image in ENVI/IDL. . 6

1.2 Reading and displaying a multispectral image in Python. . . . 7
1.3 Reading and displaying a polarimetric SAR image in IDL. . . . 8
1.4 Reading and displaying a polarimetric SAR image in Python. . 9
1.5 Determining the covariance matrix of an image in ENVI/IDL. . 17
1.6 Diagonalization of the covariance matrix in ENVI/IDL. . . . . 20
1.7 Principal components analysis in Python. . . . . . . . . . . . . 27
1.8 Principal components analysis in Python (continued). . . . . . 28
1.9 Principal components analysis in ENVI/IDL. . . . . . . . . . . 29
1.10 Principal components analysis with the new ENVI API. . . . . 31
2.1 Illustrating the Central Limit Theorem. . . . . . . . . . . . . . 44
2.2 An IDL object class for the method of provisional means. . . . 56
2.3 Calculation of the covariance matrix of an image with the pro-
visional means method and ENVI tiling. . . . . . . . . . . . . . 57
2.4 A Python object class for the method of provisional means. . . 58
2.5 Mutual information using IDL histogram functions. . . . . . . . 79
3.1 Displaying the power spectrum of an image band in ENVI. . . 88
3.2 Displaying the power spectrum of an image band in Python. . . 89
3.3 Image compression with the Haar wavelet transform. . . . . . . 96
3.4 Cascade algorithm approximation to the scaling function. . . . 101
3.5 Image reconstruction from principal components. . . . . . . . . 109
3.6 Solving the generalized eigenvalue problem in IDL by Cholesky
decomposition. . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.7 Solving the generalized eigenvalue problem in Python by Cholesky
decomposition. . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.8 Estimation of the noise covariance matrix in IDL from the dif-
ference of one-pixel shifts. . . . . . . . . . . . . . . . . . . . . . 120
3.9 Estimation of the noise covariance matrix in Python from the
difference of one-pixel shifts. . . . . . . . . . . . . . . . . . . . . 121
3.10 Simulation of two classes of observations. . . . . . . . . . . . . 124
4.1 Illustrating convolution in the spatial and frequency domains. . 129
4.2 A Python program to illustrate radar ranging. . . . . . . . . . 131
4.3 Illustrating filtering in the frequency domain. . . . . . . . . . . 133
4.4 Application of two iterations of the discrete wavelet transform
to an image band with the object class DWT. . . . . . . . . . . . 142
4.5 The FILTER method for the IDL object class DWT. . . . . . . . . 144

xxiii
xxiv Image Analysis, Classification, and Change Detection in Remote Sensing

4.6 The filter method for the Python object DWTArray (excerpt
from auxil.py). . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.7 Calculating a kernel matrix. . . . . . . . . . . . . . . . . . . . . 148
5.1 Histogram equalization in Python (excerpt from auxil.py) . . . 160
5.2 Calculating the power spectrum of the filter h1 in Equation
(5.5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.3 LoG filtering with sign change detection in IDL. . . . . . . . . 165
5.4 Corner detection with OpenCV. . . . . . . . . . . . . . . . . . . 168
5.5 Illustrating the Hu invariant moments. . . . . . . . . . . . . . . 171
5.6 Hu invariant moments of image contours. . . . . . . . . . . . . 174
5.7 HSV panchromatic sharpening with ENVI batch procedures
and ENVI Classic. . . . . . . . . . . . . . . . . . . . . . . . . . 177
5.8 HSV panchromatic sharpening with ENVI batch procedures
and the new ENVI GUI. . . . . . . . . . . . . . . . . . . . . . . 178
5.9 An IDL routine for the Wang-Bovik quality index. . . . . . . . 184
5.10 Estimation of the equivalent number of looks. . . . . . . . . . . 190
5.11 Calculation of a DEM with cross-correlation. . . . . . . . . . . 206
5.12 Image matching by phase correlation. . . . . . . . . . . . . . . 214
5.13 Polarimetric SAR image registration. . . . . . . . . . . . . . . . 216
5.14 Polarimetric SAR image registration (continued). . . . . . . . . 217
5.15 Image–image registration with the similarity transform. . . . . 224
5.16 Image–image registration with the similarity transform (con-
tinued). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
6.1 Calculating class membership probabilities from rule images. . 240
6.2 Excerpt from the program module MAXLIKE_RUN.PRO. . . . . . . 241
6.3 Excerpt from the Python module supervisedclass.py. . . . . . 242
6.4 Excerpt from the program module KERNEL_RUN.PRO. . . . . . . . 246
6.5 Excerpt from the program module KERNEL_RUN.PRO. . . . . . . . 247
6.6 Excerpt from the object class FFN. . . . . . . . . . . . . . . . . 255
6.7 Excerpt from the Python module supervisedclass.py. . . . . . 256
6.8 Excerpt from the object class FFNBP. . . . . . . . . . . . . . . . 263
6.9 Excerpt from the Python module supervisedclass.py. . . . . . 264
6.10 Excerpt from the Python module supervisedclass.py. . . . . . 279
7.1 Cross-validation on the cloud. . . . . . . . . . . . . . . . . . . . 295
7.2 Excerpt from the script rx.py . . . . . . . . . . . . . . . . . . . 314
7.3 Excerpt from the ENVI/IDL extension KRX_RUN.PRO. . . . . . . 318
7.4 Excerpt from the ENVI/IDL extension KRX_RUN.PRO. . . . . . . 318
8.1 K-means clustering of a multispectral image with Python. . . . 328
8.2 Excerpt from the program KKMEANS_RUN.PRO. . . . . . . . . . . . 330
8.3 Extended K-means clustering. . . . . . . . . . . . . . . . . . . . 334
8.4 Extended K-means clustering (continued). . . . . . . . . . . . . 335
8.5 Excerpt from the IDL program module EM.PRO. . . . . . . . . . 344
8.6 Excerpt from the Python script em.py. . . . . . . . . . . . . . . 345
8.7 Excerpt from the program module EM_RUN.PRO. . . . . . . . . . 347
8.8 Excerpt from the program module EM.PRO. . . . . . . . . . . . . 349
Program Listings xxv

8.9 Generating a toy image. . . . . . . . . . . . . . . . . . . . . . . 351

8.10 Excerpt from the program SEGMENT_CLASS_RUN.PRO. . . . . . . . 358
8.11 Clustering segments on the basis of their invariant moments. . 359
8.12 Generating another toy image. . . . . . . . . . . . . . . . . . . 364
9.1 Iterated principal components. . . . . . . . . . . . . . . . . . . 373
9.2 Iterated principal components (continued). . . . . . . . . . . . 374
9.3 Excerpt from the Python script imad.py. . . . . . . . . . . . . . 389
9.4 Scatterplot matching. Excerpt from BSLFCPNORM RUN.PRO. . . . 407
9.5 Copying a spatial subset between images. . . . . . . . . . . . . 412
A.1 An IDL procedure for orthogonal linear regression. . . . . . . . 424
B.1 Excerpt from the object class FFNCG. . . . . . . . . . . . . . . . 435
B.2 Excerpt from the object class FFNCG. . . . . . . . . . . . . . . . 442
B.3 Excerpt from the object class FFNCG (continued). . . . . . . . . 443
B.4 Excerpt from the object class FFNKAL. . . . . . . . . . . . . . . . 449
B.5 Excerpt from the object class FFNKAL (continued). . . . . . . . . 451
This page intentionally left blank
1
Images, Arrays, and Matrices

There are many Earth observation satellite-based sensors, both active and
passive, currently in orbit or planned for the near future. Representative of
these, we describe briefly the multispectral ASTER system (Abrams et al.,
1999) and the TerraSAR-X synthetic aperture radar satellite (Pitz and Miller,
2010). See Jensen (2005), Richards and Jia (2006) and Mather and Koch
(2010) for overviews of remote sensing satellite platforms.
The Advanced Spaceborne Thermal Emission and Reflectance Radiometer
(ASTER) instrument was launched in December 1999 on the Terra space-
craft. It is being used to obtain detailed maps of land surface temperature,
reflectance and elevation and consists of sensors to measure reflected solar
radiance and thermal emission in three spectral intervals:

• VNIR: Visible and near-infrared bands 1, 2, 3N, and 3B, in the spectral
region between 0.52 and 0.86 µm (four arrays of charge-coupled detectors
(CCDs) in pushbroom scanning mode).

• SWIR: Short wavelength infrared bands 4 to 9 in the region between

1.60 and 2.43 µm (six cooled PtSi-Si Schottky barrier arrays, pushbroom
scanning).

• TIR: Thermal infrared bands 10 to 14 covering a spectral range from 8.13

to 11.65 µm (cooled HgCdTe detector arrays, whiskbroom scanning).

The altitude of the spacecraft is 705 km. The across- and in-track ground
sample distances (GSDs), i.e., the detector widths projected through the sys-
tem optics onto the Earth’s surface, are 15 m (VNIR), 30 m (SWIR) and
90 m (TIR).∗ The telescope associated with the 3B sensors is back-looking at
an angle of 27.6o to provide, together with the 3N sensors, along-track stereo
image pairs. In addition, the VNIR camera can be rotated from straight down
(nadir) to ± 24o across-track. The SWIR and TIR instrument mirrors can be
pointed to ± 8.5o across-track. Like most platforms in this ground resolution
category, the orbit is near polar, sun-synchronous. Quantization levels are 8
bits for VNIR and SWIR and 12 bits for TIR. The sensor systems have an
average duty cycle of 8% per orbit (about 650 scenes per day, each 60×60

∗ At the time of writing, the SWIR sensor is no longer functioning although both the

VNIR and TIR systems are still producing good data.

1
2 Images, Arrays, and Matrices

FIGURE 1.1
ASTER color composite image (1000 × 1000 pixels) of VNIR bands 1 (blue),
2 (green), and 3N (red) over the town of Jülich in Germany, acquired on May
1, 2007. The bright areas are open cast coal mines. (See color insert.)

km2 in area) with revisit times between 4 and 16 days. Figure 1.1 shows a
spatial/spectral subset of an ASTER scene. The image is a UTM (Universal
Transverse Mercator) projection oriented along the satellite path (rotated ap-
proximately 16.4o from north) and orthorectified using a digital terrain model
generated from the stereo bands.
Unlike passive multi- and hyperspectral imaging sensors, which measure
reflected solar energy or the Earth’s thermal radiation, synthetic aperture
radar (SAR) airborne and satellite platforms supply their own microwave
radiation source, allowing observations which are independent of time of day
or cloud cover. The radar antenna on the TerraSAR-X satellite, launched in
 3

FIGURE 1.2
A 5000×5000-pixel spatial subset of the HH polarimetric band of a TerraSAR-
X quad polarimetric image acquired over the Rhine River, Germany, in so-
called Stripmap mode. The data are slant-range, single-look, complex. The
gray-scale values correspond to the magnitudes of the complex pixel values.

June, 2007, emits and receives X-band radar (9.65 GHz) in both horizontal and
vertical polarizations to provide surface imaging with a geometric resolution
from about 18 m (scanSAR mode, 10 km×150 km swath) down to 1 m (high-
resolution Spotlight mode, 10 km×5 km swath). It flies in a sun-synchronous,
near-polar orbit at an altitude of 514 km with a revisit time for points on
the equator of 11 days. Figure 1.2 shows a TerraSAR-X HH polarimetric
band (horizontally polarized radiation emitted and detected) acquired over
the Rhine River, Germany, in April, 2010. The data are at the “single-look,
slant-range complex” processing level, and are not map-projected.
4 Images, Arrays, and Matrices

1.1 Multispectral satellite images

A multispectral image such as that shown in Figure 1.1 may be represented
as a three-dimensional array of gray-scale values or pixel intensities

gk (i, j), 1 ≤ i ≤ c, 1 ≤ j ≤ r, 1 ≤ k ≤ N,

where c is the number of pixel columns (also called samples) and r is the
number of pixel rows (or lines). The index k denotes the spectral band, of
which there are N in all. For data at an early processing stage a pixel may be
stored as a digital number (DN), often in a single byte, so that 0 ≤ gk ≤ 255.
This is the case for the ASTER VNIR and SWIR bands at processing level L1A
(unprocessed reconstructed instrument data), whereas the L1A TIR data are
quantized to 12 bits (as unsigned integers) and thus stored as digital numbers
from 0 to 212 − 1 = 4095. Processed image data may of course be stored in
byte, integer or floating point format and can have negative or even complex
values.
The gray-scale values in the various bands encode measurements of the
radiance L∆λ (x, y) in wavelength interval ∆λ due to sunlight reflected from
some point (x, y) on the Earth’s surface, or due to thermal emission from
that surface, and focused by the instrument’s optical system along the array
of sensors. Ignoring all absorption and scattering effects of the intervening
atmosphere, the at-sensor radiance available for measurement from reflected
sunlight from a horizontal, Lambertian surface, i.e., a surface which scatters
reflected radiation uniformly in all directions, is given by

L∆λ (x, y) = E∆λ · cos θz · R∆λ (x, y)/π. (1.1)

The units are [W/(m2 · sr · µm)], E∆λ is the average spectral solar irradiance
in the spectral band ∆λ, θz is the solar zenith angle, R∆λ (x, y) is the surface
reflectance at coordinates (x, y), a number between 0 and 1, and π accounts
for the upper hemisphere of solid angle. The conversion between DN and
at-sensor radiance is determined by the sensor calibration as measured (and
maintained) by the satellite image provider. For example, for ASTER VNIR
and SWIR L1A data,

L∆λ (x, y) = A · DN/G + D.

The quantities A (linear coefficient), G (gain), and D (offset) are tabulated

for each of the detectors in the arrays and included with each acquisition.
Atmospheric scattering and absorption models may be used to deduce at-
surface radiance, surface temperature and emissivity or surface reflectance
from the observed radiance at the sensor. Reflectance and emissivity are
directly related to the physical properties of the surface being imaged. See
Multispectral satellite images 5

Schowengerdt (1997) for a thorough discussion of atmospheric effects and their

correction.
Various conventions are used for storing the image array gk (i, j) in computer
memory or other storage media. In band interleaved by pixel (BIP) format,
for example, a two-channel, 3 × 3 pixel image would be stored as

g1 (1, 1) g2 (1, 1) g1 (2, 1) g2 (2, 1) g1 (3, 1) g2 (3, 1)

g1 (1, 2) g2 (1, 2) g1 (2, 2) g2 (2, 2) g1 (3, 2) g2 (3, 2)
g1 (1, 3) g2 (1, 3) g1 (2, 3) g2 (2, 3) g1 (3, 3) g2 (3, 3),

whereas in band interleaved by line (BIL) it would be stored as

g1 (1, 1) g1 (2, 1) g1 (3, 1) g2 (1, 1) g2 (2, 1) g2 (3, 1)

g1 (1, 2) g1 (2, 2) g1 (3, 2) g2 (1, 2) g2 (2, 2) g2 (3, 2)
g1 (1, 3) g1 (2, 3) g1 (3, 3) g2 (1, 3) g2 (2, 3) g2 (3, 3),

and in band sequential (BSQ) format as

g1 (1, 1) g1 (2, 1) g1 (3, 1)

g1 (1, 2) g1 (2, 2) g1 (3, 2)
g1 (1, 3) g1 (2, 3) g1 (3, 3)
g2 (1, 1) g2 (2, 1) g2 (3, 1)
g2 (1, 2) g2 (2, 2) g2 (3, 2)
g2 (1, 3) g2 (2, 3) g2 (3, 3).

In the computer language IDL, so-called column major indexing is used for
arrays (Gumley, 2002) and the elements in an array are numbered from zero.
This means that, if a gray-scale image g is assigned to an IDL array variable
G, then the intensity value g(i, j) is addressed as G[i-1,j-1]. An N -band
multispectral image is stored in BIP format as an N × c × r array in IDL, in
BIL format as a c × N × r and in BSQ format as a c × r × N array. So, for
example, in BIP format the value gk (i, j) is stored at G[k-1,i-1,j-1].
In numerical Python (Python with the Numpy package) arrays are also
numbered from zero, but the indexing is row major. The array shapes are
thus reversed relative to IDL. For example, a BSQ image has shape N × r × c.
Auxiliary information, such as image acquisition parameters and georef-
erencing, is sometimes included with the image data on the same file, and
the format may or may not make use of compression algorithms. Examples
are the GeoTIFF∗ file format used, for instance, by Space Imaging Inc. for
distributing Carterra(c) imagery and which includes lossless compression, the
HDF-EOS (Hierarchical Data Format-Earth Observing System) files in which
ASTER images are distributed, and the PCIDSK format employed by PCI

∗ GeoTIFF is an open source specification and refers to TIFF files which have geographic

(or cartographic) data embedded as tags within the file. The geographic data can be used
to position the image in the correct location and geometry on the screen of a geographic
information display.
6 Images, Arrays, and Matrices

Listing 1.1: Reading and displaying a multispectral image in ENVI/IDL.

1 PRO EX1_1
2
3 envi_select , title = ’ Choose multispectral image ’ , $
4 fid = fid , dims = dims , pos = pos
5 IF ( fid EQ -1) THEN BEGIN
6 PRINT , ’ cancelled ’
7 RETURN
8 ENDIF
9
10 envi_file_query , fid , fname = fname
11
12 ; image dimensions
13 cols = dims [2] - dims [1]+1
14 rows = dims [4] - dims [3]+1
15 bands = n_elements ( pos )
16
17 ; BSQ array
18 mage = fltarr ( cols , rows , bands )
19
20 FOR i =0 , bands -1 DO image [* ,* , i ] = $
21 envi_get_dat a ( fid = fid , dims = dims , pos = pos [ i ])
22
23 ; display first band
24 window , 11 , xsize = cols , ysize = rows , title = fname
25 tvscl , image [* ,* ,0]
26
27 END

Geomatics(c) with its image processing software, in which auxiliary informa-

tion is in plain ASCII and the image data are not compressed. ENVI uses a
simple “flat binary” file structure with an additional ASCII header file.
The IDL program shown in Listing 1.1 makes use of the ENVI procedures
ENVI_SELECT , ENVI_FILE_QUERY and the ENVI function ENVI_GET_DATA() to read
a multispectral image into an IDL array with BSQ interleave format and then
display the first spectral band with the IDL procedures WINDOW and TVSCL.
This script uses what is now referred to as the “ENVI Classic” interface to
IDL, and in order to execute it, an ENVI Classic (or simply ENVI for pre 5.0
versions) interactive session should be up and running. With ENVI Version 5,
a new graphical user interface and associated object-oriented application pro-
gramming interface (API) was introduced. An example of its use is given in
Listing 1.10 at the end of this chapter.
Listing 1.2 is a simple and fairly self-explanatory Python script which does
the same thing with the aid of GDAL (the Geospatial Data Abstraction Li-
brary) and the matplotlib.pyplot package. It also uses the select_infile(),
Multispectral satellite images 7

Listing 1.2: Reading and displaying a multispectral image in Python.

1 # !/ usr / bin / env python
2 # Name : ex1_1 . py
3 IMPORT auxil . auxil as auxil
4 FROM numpy IMPORT *
5 FROM osgeo IMPORT gdal
6 FROM osgeo . gdalconst IMPORT GA_ReadOnly
7 IMPORT matplotlib . pyplot as plt
8
9 DEF main ():
10
11 gdal . AllRegister ()
12 infile = auxil . select_infil e ()
13 IF infile :
14 inDataset = gdal . Open ( infile , GA_ReadOnly )
15 cols = inDataset . RasterXSize
16 rows = inDataset . RasterYSize
17 bands = inDataset . RasterCount
18 ELSE :
19 RETURN
20
21 # spectral and spatial subsets
22 pos = auxil . select_pos ( bands )
23 x0 , y0 , rows , cols = auxil . select_dims ([0 ,0 , rows , cols ])
24
25 # BSQ array
26 image = zeros (( LEN ( pos ) , rows , cols ))
27 k = 0
28 FOR b IN pos :
29 band = inDataset . GetRasterBand ( b )
30 image [k ,: ,:]= band . ReadAsArray ( x0 , y0 , cols , rows )\
31 . astype ( FLOAT )
32 k += 1
33 inDataset = None
34
35 # display first band
36 band0 = image [0 ,: ,:]
37 mn = amin ( band0 )
38 mx = amax ( band0 )
39 plt . imshow (( band0 - mn )/( mx - mn ) , cmap = ’ gray ’ )
40 plt . show ()
41
42 IF __name__ == ’ __main__ ’:
43 main ()
8 Images, Arrays, and Matrices

select_pos() and select_dims() functions defined in the auxil.py module

which accompanies this book. The script is run from the command prompt in
Windows or from a console window on Unix-like systems with the command
python ex1_1.py. In the latter case, the “Shebang” #! in the first line allows it
to be run simply by typing the filename, assuming the path to the env utility
is /usr/BIN.

1.2 Synthetic aperture radar images

Synthetic aperture radar (SAR) systems differ significantly from optical/in-
frared sensor-based platforms. Richards (2009) and Oliver and Quegan (2004)
provide thorough introductions to SAR remote sensing, SAR image statistics,
image analysis and interpretation.
The power received by a radar transmitting/receiving antenna reflected
from a distributed (as opposed to point) target a distance D from the antenna
is given by (Richards, 2009)

PT GT GR λ2 σ o ∆a ∆r
PR = [W ], (1.2)
(4π)3 D4

where PT is the transmitted power [W · m−2 ], λ is the operating wave-

Listing 1.3: Reading and displaying a polarimetric SAR image in IDL.

1 PRO ex1_2
2 COMPILE_OPT IDL2
3
4 fname = dialog_pickf i le ( filter = ’ *. xml ’ , $
5 title = ’ Select COSAR image ’)
6 IF fname EQ ’ ’ THEN RETURN
7 tdx_openfile , / no_envi , filename = fname , $
8 saInfo = saInfo
9 data = tdx_readspat ia l ( band =[0] , $
10 xs =1000 , $
11 xe =5999 , $
12 ys =1000 , $
13 ye =5999 , $
14 saInfo = saInfo )
15 tdx_closefile , saInfo
16 window , 11 , xsize =5000 , ysize =5000
17 tvscl , abs ( data )
18
19 END
Synthetic aperture radar images 9

Listing 1.4: Reading and displaying a polarimetric SAR image in Python.

1 # !/ usr / bin / env python
2 # Name : ex1_2 . py
3 IMPORT auxil . auxil as auxil
4 FROM numpy IMPORT *
5 FROM osgeo IMPORT gdal
6 FROM osgeo . gdalconst IMPORT GA_ReadOnly
7 IMPORT matplotlib . pyplot as plt
8
9 DEF main ():
10 gdal . AllRegister ()
11 infile = auxil . select_infil e ( filt = ’ *. xml ’)
12 IF infile :
13 inDataset = gdal . Open ( infile , GA_ReadOnly )
14 cols = inDataset . RasterXSize
15 rows = inDataset . RasterYSize
16 bands = inDataset . RasterCount
17 ELSE :
18 RETURN
19 pos = auxil . select_pos ( bands )
20 x0 , y0 , rows , cols = auxil . select_dims ([0 ,0 , rows , cols ])
21 # BSQ array
22 image = zeros (( LEN ( pos ) , rows , cols ) , dtype = complex64 )
23 k = 0
24 FOR b IN pos :
25 band = inDataset . GetRasterBand ( b )
26 image [k ,: ,:]= band . ReadAsArray ( x0 , y0 , cols , rows )\
27 . astype ( COMPLEX )
28 k += 1
29 inDataset = None
30 # display magnitude in linear 2% stretch
31 band0 = ABS ( image [0 ,: ,:])
32 band0 = auxil . lin2pcstr ( band0 )
33 mn = amin ( band0 )
34 mx = amax ( band0 )
35 plt . imshow (( band0 - mn )/( mx - mn ) , cmap = ’ gray ’)
36 plt . show ()
37
38 IF __name__ == ’ __main__ ’:
39 main ()

length [m], GT (GR ) is the transmitting (receiving) antenna gain, ∆a (∆r )

is the azimuth (ground range) resolution [m] and σ o is the unitless scattering
coefficient (referred to as the radar cross section) of the target surface. The
scattering coefficient is related to the (bio)physical properties of the surface
being irradiated, notably its water content.
10 Images, Arrays, and Matrices

In later chapters we will be concerned primarily with fully and partially

polarimetric SAR data. A full, or quad, polarimetric SAR measures a 2 × 2
scattering matrix S at each resolution cell on the ground. The scattering
matrix relates the incident and the backscattered electric fields E i and E b
according to
 b   i 
b i Eh shh shv Eh
E = SE or = . (1.3)
Evb svh svv Evi
i(b) i(b)
Here Eh and Ev denote the horizontal and vertical components of the
incident (backscattered) oscillating electric fields directly at the target. These
can be deduced from the transmitted and received radar signals via the so-
called far-field approximations; see Richards (2009). If both horizontally and
vertically polarized radar pulses are emitted and discriminated, then they
determine, from Equation (1.3), the four complex scattering matrix elements.
Complex numbers provide a convenient representation for the amplitude E
of an electric field:
 
E = |E|cos(ωt + φ) = Re |E|ei(ωt+φ) ,

where ω = 2πf and φ are the angular frequency and phase of the radiation
and Re denotes “real part.”∗ It is usually convenient to work exclusively
with complex amplitudes |E|ei(ωt+φ) , bearing in mind that only the real part
is physically significant. When the oscillating electric fields are described by
complex numbers in this way, the scattering matrix elements are also complex.
A full polarimetric, or quad polarimetric SAR image then consists of four
complex bands shh , shv , svh and svv , one for each pixel-wise determination of
an element of the scattering matrix. So-called reciprocity (Richards, 2009),
which normally applies to natural targets, implies that shv = svh . The squared
amplitudes of the scattering coefficients, i.e., |shh |2 , |shv |2 , etc., constitute the
radar cross sections for each polarization combination. These in turn replace
σ o in Equation (1.2) and determine the received power in each polarization
channel.
Listing 1.3 gives a basic IDL script for reading polarimetric TerraSAR-X
imagery in COSAR format, an annotated binary matrix format with each
sample (pixel) held in 32 bytes (16 bytes real, 16 bytes imaginary) stored
with the most significant byte first (big endian format). The script makes use
of the freely available TerraSAR-X/TanDEM-X SSC/CoSSC reader for IDL
and ENVI (Exelis/DLR, 2012). After opening the data file, a 500 × 500 pixel
subset of the first (HH) band is read into the IDL variable data. Then the
magnitudes of the complex pixel values are displayed in an IDL window as in
Figure 1.2.
Listing 1.4 shows a similar script in Python which makes use of GDAL and
the matplotlib library as well as the auxil package to access and plot the
data.

∗A brief introduction to complex numbers is given in Appendix A.

Algebra of vectors and matrices 11

1.3 Algebra of vectors and matrices

It is very convenient, in fact essential, to use a vector representation for multi-
spectral or SAR image pixels. In this book, a pixel will be represented in the
form  
g1 (i, j)

g(i, j) =  .. 
. , (1.4)
gN (i, j)
which is understood to be a column vector of spectral intensities or gray-scale
values at the image position (i, j). It can be thought of as a point in N -
dimensional Euclidean space, commonly referred to as input space or feature
space. In the case of SAR images, the vector components may be complex
scattering amplitudes.
Since we will be making extensive use of the vector notation of Equation
(1.4), some of the basic properties of vectors, and of matrices which generalize
them, will be reviewed here. One can illustrate — and the reader can easily
verify — many of these properties for 2-component vectors and 2 × 2 matrices.
A list of frequently used mathematical symbols is given at the end of the book.

1.3.1 Elementary properties

The transpose of the two-component
column vector
 
x1
x x= , (1.5)
x2
✒
shown in Figure 1.3, is the row vector
x2
x⊤ = (x1 , x2 ). (1.6)

The sum of two column vectors is

x1 given by
     
FIGURE 1.3 x1 y x1 + y1
A vector with two components. x+y = + 1 = ,
x2 y2 x2 + y2
(1.7)
and their inner product by
 
y1
x⊤ y = (x1 , x2 ) = x1 y1 + x2 y2 . (1.8)
y2
The length or Euclidean norm of the vector x is
q √
kxk = x21 + x22 = x⊤ x . (1.9)
12 Images, Arrays, and Matrices

The programming language IDL is especially good at manipulating array

structures like vectors and matrices. In the following dialog, a column vector
is assigned to the variable X and then printed together with its transposed
form:
1 IDL > X =[[1] ,[2]]
2 IDL > PRINT , X
3 1
4 2
5 IDL > PRINT , transpose ( X )
6 1 2
The inner product of x and y can
be expressed in terms of the vector
❃ lengths and the angle θ between the
x two vectors:

x⊤ y = kxkkyk cos θ; (1.10)

✿
θ y see Figure 1.4 and Exercise 1. If
x cos θ θ = 90o, the vectors are said to be
orthogonal, in which case x⊤ y = 0.
FIGURE 1.4 Any vector can be expressed in
Illustrating the inner product. terms of orthogonal unit vectors, e.g.,

     
x1 1 0
x= = x1 + x2 = x1 i + x2 j , (1.11)
x2 0 1
where the symbols i and j denote vectors of unit length along the x and y
directions, respectively.
A 2 × 2 matrix is written in the form
 
a11 a12
A= . (1.12)
a21 a22

The first index of a matrix element indicates its row, the second its column.
While numerical Python indexes two-dimensional array and matrix objects
in this way, the indices of the elements in Equation (1.12) do not correspond
to the IDL convention. The first index of a 2-dimensional IDL array denotes
the column, the second the row. Thus, IDL may be said to be array-oriented
but not matrix-oriented. This can cause confusion and must be kept in mind
when writing programs in IDL. Depending on the context, we will refer to an
n × m IDL array and an m × n matrix as one and the same object.
When a matrix is multiplied with a vector, the result is another vector, e.g.,
        
a11 a12 x1 a11 x1 + a12 x2 a11 a11
Ax = = = x1 + x2 .
a21 a22 x2 a21 x1 + a22 x2 a21 a22
Algebra of vectors and matrices 13

In general, for A = (a1 , a2 . . . aN ), where the vectors ai are the columns of A,

Ax = x1 a1 + x2 a2 + · · · + xN aN . (1.13)
The IDL operator for matrix and vector multiplication is ##, for example:
1 IDL > A =[[1 ,2] ,[3 ,4]]
2 IDL > PRINT , A
3 1 2
4 3 4
5 IDL > PRINT , A ## X
6 5
7 11
Numerical Python is similarly efficient in manipulating arrays, vectors and
matrices. The scalar multiplication operator ∗ is interpreted as matrix multi-
plication for matrix operands:
1 >>> IMPORT numpy
2 >>> X = numpy . mat ([[1] ,[2]])
3 >>> A = numpy . mat ([[1 ,2] ,[3 ,4]])
4 >>> PRINT A
5 [[1 2]
6 [3 4]]
7 >>> PRINT A * X
8 [[ 5]
9 [11]]
The product of two 2 × 2 matrices is given by
    
a11 a12 b11 b12 a11 b11 + a12 b21 a11 b12 + a12 b22
AB = =
a21 a22 b21 b22 a21 b11 + a22 b21 a21 b12 + a22 b22
and is another matrix. More generally, the matrix product AB is allowed
whenever A has the same number of columns as B has rows. So if A has
dimension ℓ × m and B has dimension m × n, then AB is ℓ × n with elements
m
X
(AB)ij = aik bkj i = 1 . . . ℓ, j = 1 . . . n. (1.14)
k=1

Matrix multiplication is not commutative, i.e., AB 6= BA in general. However

it is associative:
(AB)C = A(BC) (1.15)
so we can write, for example,
(AB)C = A(BC) = ABC (1.16)
without ambiguity. The outer product of two vectors of equal length, written
xy ⊤ , is a matrix, e.g.,
      
⊤ x1 x1 0 y1 y2 x1 y1 x1 y2
xy = (y1 , y2 ) = = . (1.17)
x2 x2 0 0 0 x2 y1 x2 y2
14 Images, Arrays, and Matrices

Matrices, like vectors, have a transposed form, obtained by interchanging

their rows and columns:
 
⊤ a11 a21
A = . (1.18)
a12 a22

Transposition has the properties

(A + B)⊤ = A⊤ + B ⊤
(1.19)
(AB)⊤ = B ⊤ A⊤ .

The analogous operation to transposition for complex vectors and matrices

is conjugate transposition (see Appendix A), for which aij is replaced by a∗ji ,
where the asterisk denotes complex conjugation. (The complex conjugate of
a + ib is a − ib.) We will write the conjugate transpose of A as A† . Conjugate
transposition has the same properties as given above for ordinary transposi-
tion.

1.3.2 Square matrices

A square matrix A has equal numbers of rows and columns. The determinant
of a p × p square matrix is written |A| and defined as
X
|A| = (−1)f (j1 ...jp ) a1j1 a2j2 · · · apjp . (1.20)
(j1 ...jp )

The sum is taken over all permutations (j1 . . . jp ) of the integers (1 . . . p) and
f (j1 . . . jp ) is the number of transpositions (interchanges of two integers) re-
quired to change (1 . . . p) into (j1 . . . jp ). The determinant of a 2 × 2 matrix,
for example, is given by

|A| = a11 a22 − a12 a21 .

The determinant has the properties

|AB| = |A||B|
(1.21)
|A⊤ | = |A|.

The identity matrix is a square matrix with ones along its diagonal and
zeroes everywhere else. For example
 
1 0
I= ,
0 1

and for any A,

IA = AI = A. (1.22)
Algebra of vectors and matrices 15

The matrix inverse A−1 of a square matrix A is defined in terms of the

identity matrix by the requirements
A−1 A = AA−1 = I. (1.23)
For example, it is easy to verify that a 2 × 2 matrix has inverse
 
−1 1 a22 −a12
A = .
|A| −a21 a11
In IDL, continuing the previous dialog,
1 IDL > PRINT , determ ( A )
2 -2.00000
3 IDL > PRINT , invert ( A )
4 -2.00000 1.00000
5 1.50000 -0.500000
6 IDL > PRINT , A # # invert ( A )
7 1.00000 0.000000
8 0.000000 1.00000
The matrix inverse has the properties
(AB)−1 = B −1 A−1
(1.24)
(A−1 )⊤ = (A⊤ )−1 .
If the transpose of a square matrix is its inverse, i.e., if
A⊤ A = I, (1.25)
then it is referred to as an orthonormal matrix.
A system of n linear equations of the form
n
X
yi = aj xj (i), i = 1 . . . n, (1.26)
j=1

can be written in matrix notation as

y = Aa, (1.27)
where y = (y1 . . . yn )⊤ , a = (a1 . . . an )⊤ and Aij = xj (i). Provided A is
nonsingular (see below), the solution for the parameter vector a is given by
a = A−1 y. (1.28)
The trace of a square matrix is the sum of its diagonal elements, e.g., for a
2 × 2 matrix,
tr(A) = a11 + a22 . (1.29)
The trace has the properties
tr(A + B) = trA + trB
(1.30)
tr(AB) = tr(BA).
16 Images, Arrays, and Matrices

1.3.3 Singular matrices

If |A| = 0, then A has no inverse and is said to be a singular matrix. If A is
nonsingular, then the equation

Ax = 0 (1.31)

only has the so-called trivial solution x = 0. To see this, multiply from the
left with A−1 . Then A−1 Ax = Ix = x = 0.
If A is singular, Equation (1.31) has at least one nontrivial solution x 6= 0.
This, again, is easy to see for a 2 × 2 matrix. Suppose |A| = 0. Writing
Equation (1.31) out fully:

a11 x1 + a12 x2 = 0
a21 x1 + a22 x2 = 0.

To get a nontrivial solution, assume without loss of generality that a12 6= 0.

Just choose x1 = 1. Then the above two equations imply that
a11 a11
x2 = − and a21 − a22 = 0.
a12 a12
The latter equality is satisfied because |A| = a11 a22 − a12 a21 = 0.

1.3.4 Symmetric, positive definite matrices

The variance–covariance matrix, which we shall meet in the next chapter and
which plays a central role in digital image analysis, is both symmetric and
positive definite.

DEFINITION 1.1 A square matrix is said to be symmetric if A⊤ = A.

The p × p matrix A is positive definite if

x⊤ Ax > 0 (1.32)

for all p-dimensional vectors x 6= 0.

The expression x⊤ Ax in the above definition is called a quadratic form. If

x⊤ Ax ≥ 0 for all x then A is positive semi-definite. Definition 1.1 can be
generalized to complex matrices; see Exercise 9.
Listing 1.5 shows how to extract the covariance matrix from an ENVI im-
age. In line 21 the image is read into an array G, in which each pixel vector
constitutes a row of the array. We will refer to such an array as a data design
matrix or simply data matrix; see Section 2.3.1. Many of our IDL and Python
programs will involve manipulations of the data matrix. Here, in line 22, it is
passed to the IDL function CORRELATE() with keyword COVARIANCE set, which
returns the covariance matrix. The result for the three VNIR bands 1, 2, and
3N of the Jülich image in Figure 1.1 is a symmetric 3 × 3 matrix:
Algebra of vectors and matrices 17

Listing 1.5: Determining the covariance matrix of an image in ENVI/IDL.

1 PRO EX1_3
2
3 envi_select , title = ’ Choose multispectral image ’ , $
4 fid = fid , dims = dims , pos = pos
5 IF ( fid EQ -1) THEN BEGIN
6 PRINT , ’ cancelled ’
7 RETURN
8 ENDIF
9
10 envi_file_query , fid , fname = fname
11
12 cols = dims [2] - dims [1]+1
13 rows = dims [4] - dims [3]+1
14 bands = n_elements ( pos )
15 pixels = cols * rows
16
17 ; array with pixel vectors as rows
18 G = fltarr ( bands , pixels )
19
20 FOR i =0 , bands -1 DO $
21 G [i ,*]= envi_get_data ( fid = fid , dims = dims , pos = pos [ i ])
22 C = correlate (G ,/ covariance ,/ double )
23
24 PRINT , ’ Covariance matrix for image ’+ fname
25 PRINT , C
26
27 END

1 ENVI > EX1_3

2 ’ Covariance matrix for image E :\ satellite images \ juelich ’
3 87.117627 92.394480 -10.518867
4 92.394480 115.83875 -77.420886
5 -10.518867 -77.420886 374.61514

We will see how to show that this matrix is positive definite in Section 1.4.

1.3.5 Linear dependence and vector spaces

Vectors are said to be linearly dependent when any one can be expressed as a
linear combination of the others. Here is a formal definition:

DEFINITION 1.2 A set S of vectors x1 . . . xr is said to be linearly

dependent
Pr if there exist scalars c1 . . . cr , not all of which are zero, such that
i=1 c i x i = 0. Otherwise they are linearly independent.
18 Images, Arrays, and Matrices

A matrix is said to have rank r if the maximum number of linearly inde-

pendent columns is r. If the p × p matrix A = (a1 . . . ap ), where ai is its ith
column, is nonsingular, then it has full rank p. If this were not the case, then
there must exist a set of scalars c1 . . . cp , not all of which are zero, for which
c1 a1 + . . . + cp ap = Ac = 0.
In other words, there would be a nontrivial solution to Ac = 0, contradicting
the fact that A is nonsingular.
The set S in Definition 1.2 is said to constitute a basis for a vector space
V , comprising all vectors that can be expressed as a linear combination of the
vectors in S. The number r of vectors in the basis is called the dimension of
V . The vector space V is also an inner product space by virtue of the inner
product definition Equation (1.8) in Subsection 1.3.1. Inner product spaces
are elaborated upon in Appendix A.

1.4 Eigenvalues and eigenvectors

In image analysis, it is frequently necessary to solve an eigenvalue problem. In
the simplest case, and the one which will concern us primarily, the eigenvalue
problem consists of finding eigenvectors u and eigenvalues λ that satisfy the
matrix equation
Au = λu, (1.33)
where A is both symmetric and positive definite. Equation (1.33) can be
written equivalently as
(A − λI)u = 0, (1.34)
so for a nontrivial solution we must have
|A − λI| = 0. (1.35)
This is known as the characteristic equation for the matrix A. For instance,
in the case of a 2 × 2 matrix eigenvalue problem,
    
a11 a12 u1 u1
=λ , (1.36)
a21 a22 u2 u2
the characteristic equation is
(a11 − λ)(a22 − λ) − a212 = 0,
which is a quadratic equation in λ with solutions
 q 
(1) 1 2 2
λ = a11 + a22 + (a11 + a22 ) − 4(a11 a22 − a12 )
2
 q  (1.37)
(2) 1 2 2
λ = a11 + a22 − (a11 + a22 ) − 4(a11 a22 − a12 ) .
2
Discovering Diverse Content Through
Random Scribd Documents
earnestly. Besides, at home they would be hungry for news; how
could he help saying yes.
He walked beside her, but confidences were impossible in the
presence of Sir Algernon, although that gentleman made himself
exceedingly agreeable according to his wont. Still, Hugh could look
at Sydney and hear her speak, and that was something.
They reached the hotel all too soon. Lady Frederica was looking
out for them and the introduction was made. She was civil, but by
no means cordial, and conveyed an accent of disapproval into her
polite surprise at seeing Mr. Chichester so far from town.
Sydney explained eagerly, but Lady Frederica’s “Indeed!” was
discouraging, and there was a pause. Hugh felt he was expected to
take his leave, and took it.
“Good-bye, Sydney, I’m—awfully glad to have seen you.”
“Good-bye! Good-bye, Hugh—my love to them at home, a great
deal of love, you know, Hugh. Good-bye!”
Oh, dear! how much there was that Sydney wanted to say to him!
If only Lady Frederica would have left them for a little time alone! If
only Sir Algernon had not been there when they met! She wanted—
oh, so much!—to hear the little things that letters never tell; those
little items of everyday home news for which she felt so sick with
longing suddenly. Why hadn’t she asked this, that, and the other?
She seemed to have said nothing but good-bye. She was very quiet
upon the homeward drive, so quiet that Sir Algernon looked
curiously at her more than once. And when they reached the castle,
and the girl had gone up to the school-room, he went into the library
to St. Quentin.
“Got any views for that little girl, Quin?” he asked carelessly, when
he had answered his host’s inquiries as to the conditions of the
roads, the “pace of the greys,” and other details of their day.
“Possibly, but none that I need your advice upon, thanks,” was the
answer.
 “Don’t get riled, old man, I wasn’t offering it.” Sir Algernon lit a
cigarette with great care and sat down by the fire. “It strikes me that
she has views of her own, as well,” he concluded.
“Suppose we leave Sydney out of the conversation, altogether!”
said St. Quentin.
“Oh, just as you please, of course. Do you want the people who
brought her up—the Chichesters—to be a tabooed subject as well?”
“What of them?”
“Oh, a son is at Donisbro’, that’s all.”
“One of the Chichesters?”
“Yes; she called him Hugh.”
Sir Algernon leaned back luxuriously in his chair, stretching out his
feet to the cheerful blaze.
“You don’t mean to say that my aunt allowed the child to enter
into conversation with him?” St. Quentin’s tone was very sharp; Sir
Algernon laughed lightly.
“Don’t look so fierce, old chap. I was the guilty party, I’m afraid. I
was escorting her back to Lady Frederica after leaving a parcel with
some girl or other, when we ran across this young chemist’s
assistant, or whatever he is. They fairly rushed into each other’s
arms. I couldn’t interfere very well, you see, though I did venture to
suggest, after a lengthy period of patient freezing, that there was a
limit to the time he ought to keep her standing in the street. He
walked with us to the hotel, and there Lady Frederica choked him
off. You needn’t look so furious, Quin, there wasn’t much harm
done; only I fancy Miss Sydney isn’t quite the pliable little wax saint
you think her, she——”
“Leave her name alone, please!”
“Oh, very well! You’ve grown uncommonly stand-offish of late, my
dear chap; you’ll be showing me the door next, eh?” His laugh was
not particularly pleasant.
 St. Quentin was frowning heavily. “You might leave me quiet a
bit,” he said. “I’m not in the best of humours, to-night.”
“Don’t mention it,” said Sir Algernon, rising and flinging his
cigarette away; “it’s quite unnecessary, I assure you.” And he went
to Lady Frederica in the drawing-room.
“Would you go to his lordship in the library, please, ma’am, if quite
convenient,” a footman said, a little later, coming to the school-room,
where Sydney and Miss Osric, undeterred by the approach of dinner,
were thoroughly enjoying a very late tea.
Sydney put down her cup and got up at once.
“Are you quite rested now, dear?” asked Miss Osric. “You looked
tired when you came in, and I am sure, if you are tired still, Lord St.
Quentin would excuse you.”
“I don’t think I’m tired,” Sydney said, and went down the wide
stairs and across the hall to the library.
St. Quentin was alone, but she knew Sir Algernon had been there
by the smell of smoke. Her cousin’s eyebrows were drawn close
together, and there was a look upon his face which was new to her.
He seemed to have forgotten to smile at her entrance to-day.
“Come here, Sydney,” he said sharply. “I have something to say to
you. I hear you met that young Chichester this afternoon.” His
contemptuous tone made the colour flame into her face.
“Yes, I did,” she said a little bit defiantly; “of course I was going to
tell you about it.”
“Were you?” said St. Quentin. “Now, Sydney, we had better
understand each other. The Chichesters brought you up, and of
course you owe a debt of gratitude to them in consequence. I have
no objection whatsoever to your paying it—in any reasonable way. I
spoke to Braemuir on the subject when he was staying here, and he
promised me to use his influence towards getting some of those
boys a start in life. I don’t suppose you know that, though the estate
is by no means as unencumbered as I could wish, I offered to refund
your doctor what he spent on you in your childhood, and——”
“He said ‘No,’ of course!” Sydney cried, with flashing eyes. “Why, I
was father’s child—of course he wouldn’t be paid for keeping me!”
“Don’t indulge in heroics, please; they bore me,” St. Quentin
observed drily. “Yes, Dr. Chichester—try to drop the expression
‘father,’ please, in speaking of him; it only makes you sound
ridiculous—Dr. Chichester, I say, refused my offer with some heat.
Like you, he appeared to consider it insulting. Tastes differ; mine is,
as you know, for common sense. Now, I should be obliged if you
would kindly give me your attention for five minutes. You are going
to occupy a great position, and I do not intend to have those
Chichesters hanging round you. Those brother-and-sister friendships
are charming in theory, but they don’t work. I know what they lead
to. I should be obliged if you would correspond less frequently with
the doctor’s family, and shall request Aunt Rica to see to it. And I
distinctly forbid you to have anything to do with that young man
when next he happens to be staying in these parts. Do you
understand me?”
“Do you mean you want me to forget mother and father, and all
the rest of them at home?” Sydney cried. There was an odd
expression on St. Quentin’s face, as he watched the growing
indignation upon hers.
“Well, something like it—you won’t find it very difficult in time, I
assure you,” was his answer.
“I don’t mean to do it!” she said with a trembling voice. “I shall
have to obey you about not writing so often, or speaking to Hugh if I
meet him, but I can’t and I won’t forget them! I hate this place! I
wish I had never come, and when you talk like that I hate you!... I
was beginning to care about you, but I don’t now at all!” She was
fighting to keep back her sobs. “Do you forget the people you have
cared for, that you want me to?” she asked him fiercely, and went
quickly out.
St. Quentin turned his head and looked after her.
“Do I forget?” he muttered; “no, I wish I did!”
 CHAPTER XI
ON THE CHURCH TOWER

On the morning following the expedition to Donisbro’, Lady

Frederica received an apologetic note from Herr Felsbaden, Sydney’s
music-master, regretting his inability to give Miss Lisle her lesson
that day, owing to a severe cold. If convenient to Lady Frederica and
Miss Lisle, he would come to the Castle on Friday afternoon instead.
The note was sent in to Miss Osric, when Lady Frederica had
glanced through it over her early cup of tea, and governess and
pupil read it together.
Sydney was looking pale and heavy-eyed this morning, Miss Osric
saw, and guessed that Lord St. Quentin had said something to
distress the girl. It was a bright sunny morning, with that
exhilaration in the air which only a perfect winter’s day has the
power to give.
“Suppose, as you have no master coming this morning, we go out
for a walk as soon as we have read a little, Sydney dear?” Miss Osric
suggested. “It is such a lovely morning, and you look tired. I think
the air would do you good.”
“I have a little headache,” Sydney owned, and they set out for
their walk at about 10.30.
The frost was thick in the park, and every little twig upon the
great bare trees outlined clearly against a sky of pale cloudless blue.
Sydney wondered why she did not feel the old exhilaration that a
morning such as this would have once awakened in her, even in
smoky London.
But if she could not enjoy the perfect morning, they soon met
somebody who could!
 As they passed the gate of the Vicarage, Mr. Seaton came out,
holding Pauly by the hand. The child was in a state of absolutely wild
delight, dancing and jumping by his father’s side, and his eyes
glittering like two stars under the tangle of red hair.
“Going up the great big ’normous tower!” he informed Sydney, as
she stooped to kiss him. “Going to walk miles and miles and miles up
ladders, almost to the sky!”
The Vicar laughed and shook hands with both the girls.
“I have to give some orders about new bell-ropes; ours were
rotten, and I’ve had them taken down,” he explained. “And it was an
old promise I should take this monkey up the tower next time I had
to go there. Do you two feel inclined, I wonder, to come with us, and
walk ‘miles and miles and miles up ladders, almost to the sky’?”
Sydney looked at the tower, standing grey and tall outlined sharply
on the blue, and then at Miss Osric. “Should you like it? It would be
lovely, I think.”
“We should like to go up very much indeed, if Mr. Seaton doesn’t
mind the bother of us,” said Miss Osric, and the four went on
together to Lislehurst Church at the farther end of the village.
The church itself had been rebuilt in the eighteenth century, when
the black oak panelling had been removed as “dirty-looking” and
replaced by whitewash, and relieved at intervals by the St. Quentin
Arms painted on it in the gaudiest colours. At the same time, the few
bits of exquisite stained glass which had survived a visit from the
“root and branch” men of the Commonwealth days had been taken
away to make room for a complete set of crudely coloured windows,
which vexed the soul of Mr. Seaton whenever his eyes fell upon
them. But the old tower had been left intact, and was considered by
the learned to be one of the finest specimens of fourteenth century
architecture left in England.
There was a tradition that the saintly Bishop Ken had once
climbed it, and had pronounced the view from the top to be “a
foretaste of Heaven.”
 Sydney, when she saw the perpendicular ladders tied together,
which those who went beyond the belfry chamber were compelled to
climb, doubted privately the probability of anyone so old and frail as
the non-juring Bishop had grown when he came to Blankshire,
having strength or breath to reach the summit!
“You are not frightened, are you?” asked the Vicar, when he had
given his orders to the man awaiting him in the belfry chamber, now
emptied of its dangling ropes. “Don’t try it, if you feel in the least bit
nervous, for it is a stiffish climb!”
To be quite honest, Sydney did not particularly like the look of the
many ladders to be scaled, but she would have died sooner than
own her fears.
After all, this was not so very much more difficult than going up
the ladders in that oast-house in Kent, where they had gone to see
the men stamp out a hop-pocket, when the whole family had spent
that happy fortnight in a Kentish farm-house last summer. Only then
Hugh had been there to help her, and pull her up that awkward step
where two rungs had gone from the ladder. Her back was to the
Vicar, but Miss Osric saw the sudden wistfulness in the girl’s grey
eyes.
“Well, come on, if you really don’t feel nervous,” Mr. Seaton said.
“Oh, Hiram,” as the old clerk came stumbling down the ladders at
the sound of their voices, “you here? That’s just as well. Now you
can go up in front and get the little tower door open for the ladies.”
“Gentleman up the tower now, sir,” Hiram said, touching his
battered hat.
“All right; he won’t interfere with us,” the Vicar said. “Now, Miss
Lisle, will you go first, and take Hiram’s hand where the ladders
cross. Miss Osric, you next. Then Pauly. Hold tight, you little monkey,
or I’ll take you down again! I’ll bring up the rear, and then if
anybody slips, I’ll catch them.”
The procession started, Mr. Seaton keeping a firm grip of his small
son’s blouse the whole time, and calling at intervals directions to the
others.
Up, up they went, clinging to the ladders set perpendicularly
against the rough grey walls, worn with the lapse of time. Higher
and higher still they went, till Sydney and Miss Osric felt as though
they had been climbing for hours instead of minutes.
The elders had no breath for speech, but little Pauly chattered
unceasingly. “Did these funny stairs go right up into Heaven? Would
there be angels at the top of the tower? Would there be stars?
Would there be at least a hole through which Pauly might look into
Heaven when he came so near it?”
Sydney could hear his shrill little voice talking on, and his father’s
grave tones answering him now and then. As they came higher the
echoes caught up the two voices and made the old tower ring with
them in a way that sounded strange and very eerie, Sydney thought.
“Getting tired, Miss Lisle?” called the Vicar cheerily, as she set foot
on the highest ladder.
His words must have been heard by “the gentleman” of whom old
Hiram had spoken, for a square of blue and sunshine opened
suddenly above her, and, as she toiled up the final rungs, a hand,
whose touch was certainly familiar, grasped hers, and swung her
over that last awkward step, where she seemed to hang over a
yawning black gulf for a moment, before landing upon terra firma
outside the tower.
“Hugh!” She had forgotten everything for the moment, except the
joy of seeing him again, but in an instant, like a bitter wind, her
cousin’s words swept back upon her—“I forbid you to have anything
to do with that young man.”
Hugh could not think why she withdrew her hand, and went back
to the little low tower door with a cloud on the face that had been so
bright a minute since. “How slow the others are in getting up!” she
said.
 Hugh watched her uneasily, as she gave her hand to Miss Osric
and helped her through the doorway; then proceeded to the same
office for little Pauly. Surely it was very unlike Sydney to have
nothing to say to him, to be absorbed in these comparative
strangers, when he was at her elbow. Surely her manner had
changed with extraordinary speed since yesterday.
She on her part had been rapidly considering the situation. It was
plainly impossible to go down the tower again the very minute after
she had come up it. What excuse could she make that had the
slightest sound of reason? None, she was quite aware. Plainly the
only thing that she could do was to obey her cousin’s order in the
spirit though not in the letter.
She was rather pale, but her voice was steady as she bent over
little Pauly, devoting herself to answering his many questions.
Mr. Seaton talked to Miss Osric and to Hugh, who answered him a
little absently. His eyes were fixed on Sydney. The Vicar looked from
one to the other in a rather puzzled way from time to time, as he did
the honours of the splendid view that lay before them.
Glimpses of the Castle showed through its encircling trees, but in
summer, Mr. Seaton said, when all the leaves were out, it was
completely hidden.
He pointed out in succession the quaint little villages, dotted at
intervals about the valley, with some interesting comment upon
each. There was Loam, which boasted the finest chancel-screen in
the county. Miss Lisle and Miss Osric ought to see it one of these
bright days: it was most distinctly worth the trouble of a visit. That
tiny church, with a tower that looked as though some giant had sat
upon it long ago, was Marston. Did Mr. Chichester remember a
humorous account in the papers two or three years back, of a
famous “kill” which had taken place in Marston churchyard, when the
fox had taken refuge in one of the old stone box tombs, and held
the narrow entry, worn by age and weather in the stone, for full an
hour?
 Styles and Hurstleigh lay out yonder; it was in Hurstleigh that the
Manor stood, which a loyal lady of the Civil Wars had defended
against General Ireton, till relieved by her husband just as the little
garrison were reduced to the last straits.
At another time Sydney would have been immensely interested in
the story, but to-day somehow she could not care even to see the
place where Madam Courtenay caught the first glimpse of the scarlet
mantled horsemen, riding to her succour only just in time.
She could not put herself to-day into the place of the cavalier lady
and rejoice with her; she could only feel herself, Sydney Lisle,
behaving in a horrid, stiff, unkind way to the brother Hugh, who kept
looking at her with those troubled, questioning eyes.
Miss Osric was the only member of the party who really enjoyed
Mr. Seaton’s explanations, for little Pauly thought them dull to the
last degree. He wanted to know several things, and no one would
attend to his shrill questions. Sydney was looking where Mr. Seaton
pointed, with unseeing eyes, and his father took no notice of various
impatient tugs at his hand. Pauly wanted dreadfully to know why the
sky had gone away again, instead of being quite near as he had
expected, and whether mother and the angels would hear him if he
were to call up to them very loudly, now this minute, and whether a
big man, who was big enough to lean over the stone parapet of the
tower which his own head barely reached, could see “In Memory of
Rose” on the white marble cross in the churchyard down below.
Hugh, to pacify him, looked over, and pronounced that he could
see “no end of crosses.”
But this by no means satisfied Pauly. Hugh must see that special
grave where Daddy took him every Sunday, after service.
“Tell me where to look,” Hugh said; “but you keep still, young
man, if you please. Don’t you go trying to lean over!”
He stared down. “Is your cross a tall one, near a tree?” he asked
presently. Pauly gave a bound of delight.
 “Yes, that’s where ‘In memorwy of Wose’ is. Do you see the lovely
holly on the grave? I stuck lots in the tin, I did weally, and my
fingers was all bleedy after. I didn’t mind. Boys don’t mind being
bleedy. ’Spect that big girl that you keep on looking at would mind.
Girls cwy when they’re bleedy, don’t they? Do you cwy? I s’pose
not,’cause you’re a big man. Did you see my lovely holly? No, you
won’t see where you are. Oh, look! You can see my lovely holly this
side of the tower as well.”
“I say—stand still!” Hugh said sharply, turning his head round.
Pauly, in a state of wild excitement, was climbing up the three-foot
parapet as nimbly as a cat. “Get down!” Hugh shouted, springing to
his feet, and darting over to the child. He spoke too late.
Pauly had reached the top, and was kneeling on it, peering down
upon his “lovely holly.” “Oh, I can see it! I can see my holly!” he
screamed joyfully, clinging and laughing.
Whether the height turned him giddy, or he lost his balance by
leaning too far, no one knew. There was only time for a cry of horror,
and a frantic grasp into emptiness upon Hugh’s part. The child had
fallen from the parapet!
The poor father staggered backward, his hand to his head—the
two girls clung together, speechless; only Hugh was able to look
over. The next moment he was shaking Mr. Seaton fiercely by the
shoulder.
“Quick, sir! Down and cut the belfry ropes. Please God, we’ll save
him yet!”
The Vicar, scarcely able to believe his ears, looked over.
Some nine feet down the tower, at each corner, a large projecting
gargoyle served the purpose of a water-spout, and it was on one of
these little Pauly had fallen—the creature’s stone ear having caught
his blouse as he bumped against it in his fall. He was lying on his
back across the gargoyle’s neck, his legs and head swinging into
space, his frock hitched half across the hideous head. He was still at
the moment, but how long would he remain so? Below him was a
drop of seventy feet.
Hugh flung off his coat, and put his leg over the parapet. “Hurry
with the ropes; I’ll go to him.”
“No, no, not you!” the Vicar cried. “I must.”
But Hugh was already letting himself down. “Quick with the
ropes!” was all he said.
 “‘Quick with the ropes!’ was all he said.”

(Page 128)

Sydney and Miss Osric looked at one another. “The belfry ropes
are gone!”
Before they had finished speaking, Mr. Seaton was tearing in a
neck-or-nothing fashion down the ladders. It was well for him that
he was forced to act, and not wait to think. Ropes must be got, and
immediately, for what ladder would be long enough? He did not even
cast one glance back at the tower as he rushed through the
churchyard in search of a rope.
There was nothing that Miss Osric and Sydney had the power to
do but wait and pray. They clung to one another silently, with set,
white faces, as Hugh commenced his difficult and dangerous
descent, with one eye on the little figure, which might move and be
dashed from its precarious resting-place at any moment. Was the
child stunned? Hugh almost hoped he might be. Any movement
must almost certainly be fatal to his balance.
But as the young man felt carefully his third step in that perilous
climb, there was a quiver in the dark blue bundle on the gargoyle,
and a scared little face was uplifted to his. The hearts of the girls
above stood still.
Hugh was struggling desperately for a foothold which it seemed
impossible to find. Would the child move, or look down? Should he
do so, nothing could save him.
“It’s all right, old chap!” Hugh called in his cheeriest tone. “You
just keep still where you are. Yes, that’s right; now look at me. I’m
coming down to take you up again. No, don’t try and sit up—you can
see me splendidly from where you are.”
His voice broke off, as he all but lost both hold and footing. He
regained it with a frantic struggle and descended another step.
“Look at me, Pauly!”
Pauly’s round eyes gazed up wonderingly. Hugh neared the
gargoyle, and set his teeth for a mighty effort.
Pauly was a particularly large and strong boy for not quite five
years old, and, even on firm ground, would be no joke to lift in one
hand. But the thing must be done. Hugh strengthened his hold with
his right hand, and took an anxious downward glance. Some of the
village men were trying to join ladders, but they were far too short.
Mr. Seaton was running frantically up the road beyond the
churchyard, with a coil of rope on his arm. In the clear air Hugh
could see his upturned face, dead white, with eyes staring wildly.
He could not possibly get through the churchyard and up the
tower in less than ten minutes—Hugh thought he would probably
take longer. It was not therefore possible to risk leaving Pauly on the
gargoyle till he himself should have the help of a rope.
He took the firmest grip he could of the roughened stonework of
the tower with feet and right hand, and loosed cautiously the other,
reaching with it towards the blue bundle on the gargoyle. “Steady,
Pauly, keep quite still, old chap!”
With a struggle that brought beads of perspiration out upon his
forehead and nearly sent him flying into space, he grasped the child,
and raised him slowly from his resting-place; then stepped down on
to the gargoyle, and stood there, clasping Pauly closely, and leaning
back against the wall with closed eyes.
He was too physically exhausted with the terrible anxiety and
effort of the last few minutes to make any further movement then.
Besides, it was now a necessity to wait for the rope. The upward
climb would be impossible when burdened by the well-grown boy.
He had to concentrate all his powers on keeping steady on the
slender foothold, which was all the gargoyle afforded, and waiting
for the help which Mr. Seaton would bring.
It seemed hours before a shout from above came down cheerily to
him, and a rope end struck him on the shoulder. “Now, Pauly,” he
said, “hold on round my neck for all you’re worth, there’s a good
little chap!”
He took a firm grip of the child’s blouse in his teeth, and,
steadying himself with infinite difficulty, fastened the rope beneath
his own arms, in the strongest knots that he could make. Then,
using his hands as a trumpet, he called “Ready!”
His left arm was round Pauly, his right grasped the rope above his
head. “Now hold tight, little chap, and don’t be frightened!”
 Pauly carried out this order by taking as good a grip as the hair-
cutter allowed of Hugh’s head, and it was in this position that the
two were at length hauled over the parapet by the united strength of
the Vicar, Hiram, and the Vicarage gardener, whom Mr. Seaton had
met while searching for a rope.
Mr. Seaton wrung Hugh’s hand in silence, and held his son to him,
in silence also. No one seemed to have much voice for speech just
then; even Pauly was subdued and shaken by his fall, though he had
escaped with nothing worse than grazed knees.
The descent from the tower was very quiet and sober. A strong
shudder went through the party as they passed the belfry chamber
and thought about the awful moment when they had realised that
the ropes were gone.
His father carried Pauly, and Hugh went in front of Sydney and
Miss Osric, and gave them his hand where the ladders turned. He
and Sydney never spoke the whole way down.
They were in the churchyard at last, and Pauly was demanding to
be shown “the funny little step where me and him was standing.”
The Vicar, shivering, hushed him, and turned to Hugh. “You’ll come
in and lunch with me?” he said, a little huskily, his hand upon the
young man’s shoulder.
“Thank you, I will,” Hugh answered gravely.
“And, Sydney, we must hurry back,” Miss Osric suggested. “I am
sure it is getting late.”
Sydney moved a step away; then took a sudden resolution.
She went to Hugh and held out her hand. “Good-bye, Hugh.
Please understand,” she said very low.
Hugh took the little gloved hand in his, and read rightly the
trouble in her eyes.
“It’s all right—don’t you bother, Syd,” he said. “I understand.”
 CHAPTER XII
MERRY CHRISTMAS

“What a lot of times I seem to have said ‘Merry Christmas’ this

afternoon!” Sydney remarked as she and Miss Osric went round the
village in Sydney’s little pony carriage with the pair of lovely little bay
ponies she so much enjoyed driving. “And the sad thing is, that
nobody here seems to feel particularly happy,” she went on. “Mrs.
Andrews, to whom I took that crossover just now, said—‘It was hard
enough to feel joyful when her man was bent double with
rheumatism from the dampness of his cottage!’ Miss Osric, are the
cottages in very bad repair here? Lord Braemuir seemed to think so,
and so do the people who live in them. But when I asked Lady
Frederica she said—‘Poor people always grumbled; if it wasn’t one
thing, it was sure to be another!’ What do you think?”
Miss Osric hesitated for a little while before replying.
“Well, Sydney,” she said at length, “I don’t know whether I ought
to tell you this, but it seems to me right you should know something
of the cottages on the estate. It will be your business to know by-
and-by. You know my father is chaplain to the hospital at Donisbro’,
and he has often told me that the amount of cases coming from the
cottages on this estate is appalling. People have been brought to the
hospital from Loam and Lislehurst, and even Styles, where the
ground is higher, simply crippled with rheumatism, and off and on
there have been a good many cases of diphtheria and fever. That
doesn’t speak well for the cottages, you know.”
Sydney pulled up the ponies in the middle of the road.
“I shall ask Mr. Fenton,” she said slowly; “I don’t think I could ask
St. Quentin.”
 “I think asking Mr. Fenton is not at all a bad idea,” Miss Osric said
cordially; “but, my dear Sydney, we mustn’t dawdle here in the cold
even to discuss points of duty. Have you any more presents to
distribute?”
“Just one for Pauly at the Vicarage,” the girl said, gathering up the
reins again; “that is the parcel underneath the seat that you said
took up as much room as we did. It’s a horse and waggon—a horse
with real hair—and I think Pauly will be able to get himself into the
waggon if he tucks his legs up. I’m sure he will be pleased—the
darling!”
“I wonder how long that quarter’s allowance is going to last,”
laughed Miss Osric, as they turned the ponies’ heads up the drive to
the Vicarage. “You’ve been so lavish over Christmas presents,
Sydney; that parcel for London alone must have nearly ruined you!”
“I am rather near bankruptcy,” owned Sydney. “It is shocking to
confess, but I never had such a lot of money to spend in my life, and
I went and spent it. But I am not a bit sorry,” she concluded, “for,
just for once, they will have at home exactly what they wanted.”
Pauly had seen them coming from the window of his father’s study,
against which he was flattening his small round nose till it looked
exactly like a white button. He flew to the door and cast himself
upon them in the hall with a shriek of delight.
“Oh, do you know, it’s going to be Chwistmas Day to-morrow!” he
exclaimed, “and I am going to church in the morning like a big man,
and Santa Claus is coming in the night, daddy finks, to put fings in
my stocking, ’cause I’ve been a very good boy for years and not
runned away or been lostened!”
The Vicar, too, was not behindhand in his welcome, though he
was not quite so conversational as his little son.
“Come into the study, both of you,” he said; “we’ve got a real Yule
log there, haven’t we, Pauly?—such a monster!—and I’m sure you
must be frozen.”
 The Sydney of six weeks ago would have accepted Mr. Seaton’s
offer, but the Sydney of to-day had learned to think what would
annoy her cousin and Lady Frederica.
“I am afraid we must hurry back, mustn’t we, Miss Osric?” she
said. “We shall be rather late as it is. We have been all round the
village, wishing ever so many people a happy Christmas, so we must
only just wish the same to you, and ask you to tell Santa Claus to
see if he can’t find a rather large, knobby parcel in the corner of the
hall for Pauly, when he comes to visit you to-night.”
“It’s very good of you,” said the Vicar. “Pauly, don’t tear Miss Lisle’s
clothes to pieces in your joy. You spoil him, you know, Miss Lisle, if
you will allow me to say so. Well, if you must go, a very happy
Christmas to you both! You are going the right way to make it a
happy one, I think.”
“Mr. Seaton, one thing,” Sydney asked as they went through the
hall together. “Are the people miserable here because their cottages
want rebuilding?”
Mr. Seaton looked at the earnest face beside him, and wondered if
the wish to help her poorer neighbours would continue when she
had the power.
“Yes,” he said, “I am sorry to own that most of the cottages here
are in a very neglected condition. But landlords have no easy time of
it, I know, and often lack the means to do all they want.”
“Thank you,” said Sydney, and then she kissed little Pauly, and she
and Miss Osric got into the carriage and drove away, the Vicar
watching them, with his small son, riotous and conversational, on his
shoulder, till they turned out into the road again.
“I don’t think I ever knew anybody more devoted to a child than
that man is,” said Miss Osric, as they reached the lodge gates. “What
would he have done if he had lost him the other day?”
“Oh, don’t talk about that dreadful morning!” said Sydney with a
shiver.
 Lady Frederica had no love for Christmas.
“One is expected to be so aggressively cheerful and social,” she
complained, “when one is really feeling bored to extinction! And now
St. Quentin’s illness casts a gloom over everything; it is most absurd
to attempt any feeling of festivity. He wouldn’t like it at all.”
“Did Cousin St. Quentin care for Christmas when he was well?”
Sydney asked a little wistfully.
“Well, I remember one year, when both his father and mother
were alive, they had the regular old-fashioned sort of Christmas, and
he certainly seemed to enjoy it. The Dean of Donisbro’ and his
daughter Katharine were here, I remember. The Dean had slipped
upon a slide some tiresome boy had made when he came over to
dine here the week before Christmas, and he fell and sprained his
ankle. Of course Dr. Lorry wouldn’t let him travel, so St. Quentin got
poor dear Alicia, his mother, to go to Donisbro’ herself and bring
back Miss Morrell to spend Christmas with her father. There were
only those two, you see. My dear, Katharine Morrell was a pretty girl
in those days! You’ve seen her, haven’t you? but she has gone off a
good deal. I fancy St. Quentin admired her rather, but it didn’t come
to anything, though we all thought it would that Christmas-time. But
she was a good deal too strait-laced for him, I expect; not that he
was worse than other young man, but he ran through a lot of money
on cards and racing, and annoyed his poor father very much. Oh! Sir
Algernon, is that you?” (Sir Algernon had entered at the moment). “I
was telling Sydney of that Christmas when the Dean and Miss
Morrell were here. I forget if you have met Katharine Morrell?”
Sydney saw a strange expression cross the handsome face for a
moment. But in a second he had answered in his usual rather
languid accents, “Yes, I know her slightly; very slightly.”
Christmas Day dawned clear and sunny and Sydney, as she stood
beside Lady Frederica in the Castle pew at Lislehurst Church, felt
something of the joy of Christmas coming to her, even in this
strange place. She smiled across at little Pauly, who, standing beside
Mr. Seaton’s housekeeper, was singing, “Hark! the herald angels
sing” with all his might, and to a time and tune quite his own.
Mr. Seaton’s sermon was very short; he said he thought the
Christmas hymns and carols preached a better sermon than he had
the power to do. He only asked his people to remember that next to
God’s glory, the angels had set peace and goodwill upon earth. The
second followed on the first. He wanted all those who had to-day
been glorifying God for His great Christmas Gift, to see to it that
peace and goodwill was not lacking in that small part of God’s earth
that concerned each—his or her own home.
Sydney had not seen her cousin since her outburst on the subject
of the Chichesters, and her conscience pricked her. It was true that
St. Quentin had expressed no wish to see her, but she had made no
attempt to find out if he had one unexpressed. Surely the first move
towards that peace and goodwill of which Mr. Seaton spoke should
come from her!
She and Lady Frederica drove home together; Sydney full of
eagerness for the post, which would have come while they were at
church.
Lady Frederica laughed, and said Sydney was “the most childish
girl for her age she had ever known”; but when they reached the
Castle, she fastened a dainty little pearl brooch into the collar of the
girl’s frock, with a “There, my dear, is a Christmas present for you!”
Sydney was a good deal touched by this kindness from one who
generally seemed dissatisfied with her, but still she was undoubtedly
relieved when Lady Frederica told her that she might take her
parcels and letters to her rooms and amuse herself as she liked till
luncheon. Lady Frederica, it appeared, was going to rest after the
tremendous exertion of getting up sufficiently early to attend eleven
o’clock service!
Sydney and Miss Osric spent a blissful hour over the letters and
presents. I think Sydney cried a little over those with the London
post-mark, for Christmas-time with its associations had made her
more homesick than she knew.
They had all written to the absent one, and there were presents
from everybody. No one had forgotten her, from old nurse down to
Prissie. Sydney and Miss Osric undid parcels and munched home-
made toffee with a noble disregard for the spoiling of their appetites,
until the luncheon gong sounded, by which time the morning-room
where they were sitting looked exactly like a Christmas bazaar.
But Sydney had not forgotten her morning’s resolution, and when
lunch was over and Lady Frederica, exhausted, doubtless, by her
unaccustomed early rising, had fallen asleep in her chair, Sydney got
up and moved softly from the gold drawing-room, crossed the hall,
and tapped lightly at the door of the library.
“Come in,” said St. Quentin’s voice.
Sir Algernon was with his host, and both men looked up as she
entered. The excitement of the home letters had brought a flush to
her face, and her eyes were very bright. Sir Algernon let his cigarette
drop from between his fingers as he looked at her. “By Jove!” he
muttered.
“I didn’t mean to interrupt you,” said Sydney, flushing under his
cool survey. “I only”—with an unconsciously appealing glance in the
direction of the sofa—“I only came to give my Christmas wishes to
you, Cousin St. Quentin.”
“Thanks,” said St. Quentin, holding out his hand to her. “You’re
going for a stroll in the park, aren’t you, Bridge?”
“Ah, yes, of course I am,” his friend answered. “Have a look round
at the timber, eh, Quin? Miss Lisle, I hope you made my humble
apologies to the Vicar for not attending church this morning. Oh, all
right!” in answer to a rather impatient sound from the sofa. “I’m off,
old man. Ta-ta!”
He lounged out, and Sydney felt relieved by his absence.
“You don’t like Bridge?” her cousin asked her quickly.
 Sydney was uncompromising in her views at all times. “Not at all,”
she said.
If she had been looking at St. Quentin at the moment she would
have seen an expression of relief on his face at her answer. But she
was looking round the room, which certainly was rather untidy.
“Wouldn’t you like the hearth swept, and these cards put away in
their case, and the papers in a drawer?” she asked her cousin. “I
don’t believe Dickson has been in here since this morning, has he?”
“No, Bridge and I were talking private business.”
“Shall I put away the papers, Cousin St. Quentin?”
“Yes, in the second drawer of the writing-table, left hand side.
Lock it, please, and give me the key.”
She obeyed him, then swept up the hearth, regardless of his “Ring
for Dickson!” and finally sat down in the great brown leather chair by
the fireside.
“Cousin St. Quentin, may I ask you one or two questions?”
“Yes.”
“Must you do business with Sir Algernon? I am sure it can’t be
very good for you. You are looking much more ill. I don’t think Dr.
Lorry would like it.”
He smiled a little at her grandmotherly tone.
“Is it to do with money?” she asked, with a remembrance of a
certain pucker on father’s brow, which Christmas bills brought with
them.
“Partly; not all. Let’s talk of something else, instead of boring you
with my affairs,” her cousin said.
“They don’t bore me. Of course I care to know your bothers!” she
declared.
He raised his eyebrows and looked at her in a considering kind of
way. “Do you? I wonder why?” He laughed a little. “Go ahead and
talk to me,” he said. “Tell me what you’ve done to-day. I suppose
you had letters by the ream from your beloved Chichesters?”
Sydney reddened, remembering their last interview upon that
subject. Her cousin seemed to recollect it too.
“Has it ever struck you that you’ll have a much better time of it
when I’m gone?” he said. “As long as you look pretty and walk into a
room the right way, Aunt Rica won’t interfere with you much.”
“How can you?” the girl cried, with hot indignation. “I hate to hear
you talk like that! Why, you’ve been very kind to me—except about
the Chichesters!”
“And that’s a rather big exception, isn’t it?” St. Quentin said. “You
haven’t got much cause to like me, Sydney.”
Something in the sadness of his tone appealed to her pity.
“I do care about you!” she said. “You say those horrid things
about the Chichesters just because you don’t understand, that’s all.
Some day, perhaps, you will know that one couldn’t give up loving
people, even if one tried. But I do care about you, really! I think you
are the very bravest person that I ever met!”
St. Quentin did not answer for a minute, and when he spoke,
though it was lightly, his voice was not quite so steady as usual.
“Is it very rude to suggest to a lady, who is going to reach the
advanced age of eighteen in a few days’ time, that her experience of
life may possibly be limited?” he said. “My dear child, I regret to say
you’re out in your conception of my character. I am a coward. Of
course, I hope one is enough of a man not to make a fuss over the
inevitable, by which I mean the consequences of my motor-smash.
What is, is, and only fools whine over it. But for all that, I’m a
coward. There, let’s talk of something else!” He leaned back and
closed his eyes. “Tell me what you like.”
And Sydney told him about Lady Frederica and her present; about
Pauly and the hymn; and everything else she could think of that
might amuse or interest him.
 She told of the knobby parcels they had taken round the village in
the pony-carriage yesterday, and of the fright of one old woman
when a rolled-up pair of thick stockings had slipped from Sydney’s
over-laden arms, and gone rolling across the kitchen floor to her
very feet.
Suddenly she stopped her merry talk, and her eyes took a
thoughtful expression.
“What are you thinking of?” her cousin asked, looking across at
the creamy-gowned figure in the brown chair.
“I was thinking of the cottages,” she answered. “They are so
wretched and so damp, St. Quentin, and the people told me there
could be no ‘Merry Christmas’ for them!”
“That meddling parson has been putting you up to that idea, I
suppose!” he said sharply.
“No, I saw the cottages for myself. Oh, St. Quentin, can’t
something be done?”
“Nothing!”
She looked at him with troubled eyes. “I expect I cost a good deal
of money. Couldn’t I have fewer frocks and things of that kind? Or
perhaps,” with an effort, “we might sell Bessie: keeping a horse is so
expensive, I’ve heard father say.”
St. Quentin’s voice was stern as he stopped her. “Don’t talk of
what you do not understand. I can do nothing for the cottages at
present. If it’s any consolation to you, I will tell you this—I wish I
could. There; talk of something else, for goodness’ sake!”
She talked on, though feeling little in the mood for conversation,
and was rewarded by his exclamation of astonishment on learning
the lateness of the hour when Dickson came in to light the lamp.
“Why, I’ve kept you here two mortal hours, forgetting all about the
time; you must be sick of me! A nice way to make you spend your
Christmas Day! However, you’ve made mine a bit more cheerful.”
 As the girl passed his sofa on the way to the door, he took her
hand, saying, “Have you forgiven me for what I said about the
Chichesters the other day?”
And Sydney, remembering that morning’s sermon, said “Yes,” with
all her heart.
 CHAPTER XIII
HUGH’S BATTLE

“What’s the matter, Hugh?”

Dr. Chichester flung the question suddenly into the deep silence
which had fallen on himself and his son, as they sat together by the
study fire on a cold night shortly after Christmas.
They had done a little talking.
Dr. Chichester had said it was a bitter night, and Hugh had
assented. The doctor had remarked that a fire and a book were
wonderfully soothing after a long day’s work, and Hugh had owned
the fact. The doctor had opined that if the frost lasted, there would
shortly be skating on the Serpentine. Hugh had agreed to that as
well, but in so absent and spiritless a manner that his father plainly
saw he took no interest whatever in the skating prospects at the
present moment.
And after these attempts at conversation, silence had fallen on
them, and the doctor, forgetful of the book upon his knee, closely
scrutinised the young face before him, with its dark, sad eyes fixed
on the glowing fire.
Hugh had been curiously silent ever since that visit to Donisbro’,
his father thought to himself.
And yet, how pleased he had been at being singled out by Sir
Anthony to go with him! And he had come back, having done
everything required of him successfully enough, so far as his father
could make out. But he had been very uncommunicative over his
adventures in the quaint cathedral city.
It had been left for Sir Anthony to catch the doctor on the
staircase of Blue-friars’ Hospital, and ask him if “the boy had
remembered to tell his father that Sir Anthony had said he was a
credit to the medical profession.” Hugh had not even mentioned the
great man’s rare commendation.
What had he said about that visit? The doctor went over in his
own mind the rather bald account which the united efforts of the
family had with difficulty pumped out.
Yes, Hugh had seen Sydney. She was looking very well—this in
answer to a question from Mrs. Chichester. She had sent her love to
them all. There hadn’t been much time; Lady Frederica had been in
a great hurry to be off. There was a man with Sydney, a Sir Algernon
Bridge. Was he nice?—a query from Dolly. Well, Hugh hadn’t asked
him, but considered that he looked a sneery brute, although not
wishing to say anything against him. Yes, he had seen Sydney again:
she was up the church tower with the Vicar, who seemed a good
sort, and his boy, a jolly little chap. The incident of Pauly’s rescue
somehow failed to transpire at all. No, he hadn’t been to the Castle
—this in reply to some excited inquiries on the subject of merry-go-
rounds from Fred and Prissie. He had lunched with the Vicar, who
had said that Sydney was interested in the cottages, and took the
people soup and things. Hugh didn’t think anything much else had
happened. Oh, how was Sydney dressed? He didn’t know—
something blue, he thought. No, something red, and fur—a lot of fur.
Was she looking pretty? How should he know?
Hugh had become a little irritable at this point, his father
recollected: a circumstance almost as unprecedented as his gravity
and silence.
What was wrong with the boy?
The keen-eyed doctor noted his dejected attitude, and the
wistfulness of the gaze turned so persistently upon the fire. If Hugh
was reading his future there it certainly was not a bright one.
Dr. Chichester watched in silence for full another ten minutes, then
repeated his question with a hand upon the young man’s knee.
“Hugh, what’s the matter?”
Welcome to Our Bookstore - The Ultimate Destination for Book Lovers
Are you passionate about books and eager to explore new worlds of
knowledge? At our website, we offer a vast collection of books that
cater to every interest and age group. From classic literature to
specialized publications, self-help books, and children’s stories, we
have it all! Each book is a gateway to new adventures, helping you
expand your knowledge and nourish your soul
Experience Convenient and Enjoyable Book Shopping Our website is more
than just an online bookstore—it’s a bridge connecting readers to the
timeless values of culture and wisdom. With a sleek and user-friendly
interface and a smart search system, you can find your favorite books
quickly and easily. Enjoy special promotions, fast home delivery, and
a seamless shopping experience that saves you time and enhances your
love for reading.
Let us accompany you on the journey of exploring knowledge and
personal growth!

ebookgate.com