MACHINE LEARNING IN REMOTE SENSING DATA PROCESSING

Gustavo Camps-Valls

Image Processing Laboratory (IPL), Universitat de València, Spain

Catedrático A. Escardino - 46980 Paterna, València, Spain
[email protected], http://www.uv.es/gcamps

ABSTRACT
Remote sensing data processing deals with real-life applica-
tions with great societal values. For instance urban monitor-
ing, fire detection or flood prediction from remotely sensed
multispectral or radar images have a great impact on eco-
nomical and environmental issues. To treat efficiently the ac- ฀

quired data and provide accurate products, remote sensing has

evolved into a multidisciplinary field, where machine learn-
ing and signal processing algorithms play an important role

฀
nowadays. This paper serves as a survey of methods and ap-
O
plications, and reviews the latest methodological advances in
machine learning for remote sensing data analysis. ฀ ฀ ฀ ฀ ฀
฀ ฀
฀ ฀ ฀
฀ ฀ ฀
1. INTRODUCTION ฀
฀ ฀ ฀
฀ ฀ ฀ ฀

฀ ฀ ฀

Remote sensing is the field of science studying and modeling

the processes occurring on the Earth’s surface and their inter- Fig. 1. Principle of imaging spectroscopy.
action with the atmosphere [1]. Earth observation at local and
global scales is nowadays an increasing need. By monitor-
ing urban growth, estimating temperature or ocean salinity, known as spectrum and is used to identify materials in the
and identifying objects on the surface, remote sensing pro- scene. Figure 1 shows the principle of imaging spectroscopy
vides valuable information for policy and decision makers, as to perform satellite remote sensing. The resulting multispec-
well as for tourism or defense applications. These objectives tral image consists of a simultaneous acquisition of spatially
are possible because materials in a scene reflect, absorb, and coregistered images, in several, spectrally contiguous bands
emit electromagnetic radiation in a different way depending from a remotely operated platform [1, 2].
of their molecular composition and shape. Remote sensing The diversity of objectives and the special characteris-
exploits this physical fact and deals with the acquisition of in- tics of the data give rise to the use of a wide range of ma-
formation about a scene (or specific object) at a short, medium chine learning and signal processing algorithms. The statis-
or long distance. tical characterization of remote sensing images turns to be
According to the type of energy resources involved in difficult because of pixel’s high dimensionality, presence of
the data acquisition, remote sensing imaging instruments can different kinds of noise sources and uncertainty, their inher-
be passive or active. In this paper we will focus on pas- ent non-linear nature, and the high spatial and spectral redun-
sive sensors which have experienced a great evolution in the dancy. Machine learning has been successfully applied in re-
last decades, and pose challenging problems for the machine mote sensing for classification, regression, clustering, coding,
learning and signal processing communities. Passive systems or source separation. However, we feel that promising new
exploit solar radiation to capture the emergent radiation, which learning paradigms, such as transfer, active, structured, rein-
is acquired by an airborne or satellite spectrometer at differ- forcement, semisupervised or manifold learning, have been
ent wavelengths. The acquired signal or spectral signature is payed little or no attention. This paper reviews both tradi-
tional and new trends in machine learning for remote sensing
This work was partially supported by projects from the Spanish Ministry
of Science TEC2006-13845/TCM, CSD2007-00018 and AYA2008-05965- data processing with the main goal of estimulating research
C04-03. and development in both directions.
 2. TRADITIONAL MACHINE LEARNING FOR of images and the interest in monitoring Earth’s changes at lo-
REMOTE SENSING cal and global scales. On the one hand, many multi-temporal
supervised methods have been used during the last years, such
In this section, we review the traditional machine learning ap- as evidence reasoning [31], generalized least squares [32],
proaches to remote sensing applications. Only the most rele- neural networks [33] or support vector machines (SVMs) [34,
vant applications are revised: classification, feature selection 35]. Hidden Markov random fields [36] and fuzzy-based ap-
and extraction, regression and unmixing. proaches [37] have been also used to link time-varying statis-
tics. On the other hand, change detection approaches typically
use image subtraction or ratioing, change vector analysis, or
2.1. Image classification
cross-correlation analysis [1]. Recently, neural networks [38]
Classification maps are the main product of remote sensing and kernel methods [9] have been used. Composite kernels
image processing. In the last years, data-driven approaches have been specifically designed for the combination of mul-
have gained relevance in the remote sensing community. In titemporal, multisensor and multisource information [9, 39].
particular, non-parametric methods have demonstrated good Recent advances focus on the reduction of the user interven-
performance. Supervised and unsupervised are revised here. tion, either by using semi- or unsupervised methods [40, 41].

2.1.1. Supervised Image Classification 2.2. Feature Selection and Extraction

These methods use labeled information about class member- A critical issue when working with high dimensional datasets,
ship of single pixels (labeled by expert users) to build a model such as hyperspectral images, is that the computational time
able to generalize to the whole image (or set of images). At is increased and the high collinearity and presence of noisy
present, the most successful methods are neural networks [3, bands can degrade the quality of the model. But maybe more
4] and support vector machines [5]. The latter have been ap- important is the study of the relative relevance of the acquired
plied to both multispectral [6,7] and hyperspectral [5,8,9] data bands to perform a given task. Remember that spectral bands
in a wide range of domains, including object recognition [10], have a physical meaning and can be related to the properties
landcover and multi-temporal classification [9,11,12], and ur- of the elements to be identified or modeled.
ban monitoring [13] to name a few. Another field of growing Feature selection has been studied in remote sensing un-
interest is that of classifier ensembles [14, 15] and boosting der classical discriminative criteria [42]. Lately, advanced
methods [16]. machine learning methods have been used, such as genetic
algorithms [43], or SVM-based recursive feature elimination
[44]. Recently more attention has been focused on feature ex-
2.1.2. Unsupervised Image Classification
traction methods. Even though the use of linear methods such
Unsupervised classification of remote sensing images is a crit- as PCA or PLS is quite common, recent advances to cope
ical problem in many applications, either for visualization and with nonlinearities in the data based on multivariate kernel
monitoring of similar areas in the scene or as a pre-processing machines have been presented [45].
step for supervised classifiers. Many clustering methods have
been designed, such as rule-based [17], neural networks [18–
2.3. Signal Unmixing
20], or based on SVMs [21]. Three approaches dominate the
field. First, fuzzy clustering has been used alone [22, 23] or Pixels are invariably a mixture of the signatures of the var-
combined with multiobjective optimization [24] for exploit- ious materials found within the spatial extent of the ground
ing spatially membership relations. Second, fusion of mul- instantaneous field view. An important problem in remote
tisource information has been conducted either with graph- sensing is the development of automatic extraction methods
cuts [25], projection pursuit [26], hierarchical clustering [27], of the spectral pure pixels (known as endmembers) directly
or Markov random fields [28] for contextual regularization. from the image. These pure pixels are the basis to express
Also, multicomponent image segmentation with self-organizing all pixels as a linear (or non-linear) combination of them, and
maps (SOM) and hybrid genetic algorithms [20] have been this, in turn, allows subpixel detection [2] or mineral map-
proposed. Finally, it is worth mentioning the use of dynamic ping [46]. Some classical techniques for this purpose include
clustering strategies for spatio-temporal reasoning [29] and the N-FINDR algorithm [47], the vertex component algorithm
visualization [30]. (VCA) in [48], and an orthogonal subspace projection (OSP)
technique in [49], among others [50]. Selection of the free pa-
2.1.3. Multitemporal Classification and Change Detection rameters and inclusion of spatial information in the unmixing
process are key issues nowadays [51]. Recently support vec-
Multitemporal and change detection problems are very active tor domain description (SVDD) has been also used to select
because of the increasing availability of complete time series the pure pixels [45].
2.4. Regression and Model Inversion (typically few) labeled data and the wealth of unlabeled sam-
ples to model the manifold data structure. In remote sens-
Robust, fast and accurate regression tools are a critical de-
ing, several methods have been developed, either generative
mand in remote sensing. The estimation of biophysical pa-
or discriminative. The estimation of the conditional density
rameters is of special relevance in order to better understand
to be included in generative models have been extensively ex-
the environment dynamics at local and global scales [1]. The
ploited [64]. Recently, many graph-based methods have been
inversion of analytical models introduces a higher level of
developed for classification [61, 65], regression [55], and tar-
complexity, induces an important computational burden, and
get detection [66, 67]. Also, the design of cluster and bagged
sensitivity to noise becomes an important issue. In the recent
kernels have been successfully presented [68]. Also the trans-
years, nevertheless, the use of empirical models adjusted to
ductive SVM has been applied for image classification [69,
learn the relationship between the acquired spectra and ac-
70] and change detection [40]. In [71], a semisupervised ker-
tual ground measurements has become very attractive. Para-
nel Fisher discriminant classifier was proposed. These meth-
metric models have some important drawbacks, which typ-
ods, however, cannot readily applicable to large scale prob-
ically lead to poor prediction results on unseen (test) data.
lems with millions of unlabeled samples, as is often the case.
As a consequence, non-parametric and potentially non-linear
regression techniques have been effectively introduced, such
as neural networks [52, 53], support vector regression (SVR) 3.3. Transfer Learning
[54, 55], relevance vector machines (RVM) [56], or Gaussian A common problem in remote sensing is that of updating
Processes (GP) [57]. Even been more accurate than analyti- land-cover maps by classifying temporal series of images when
cal models, they lack interpretability and rely on training data only training samples collected at one time are available. This
from the observed scene, which limits its extensive use. is known as transfer learning or domain adaptation. The prob-
lem was initially tackled with partially unsupervised classi-
3. NEW TRENDS IN MACHINE LEARNING FOR fiers, under parametric formalisms [72] and neural networks
REMOTE SENSING [73]. The approach was then successfully extended to domain
adaptation SVM (DASVM) [74]. A related problem is also
The special characteristics of the acquired data motivates the that of classifying an image with samples from different im-
continuous research in machine learning methods for tackling ages, which induces the sample selection bias or covariance
particular remote sensing problems. In this section, we sum- shift problems. These problems have been recently presented
marize some promising machine learning paradigms of recent by defining proper kernel machines [75].
application in remote sensing.
3.4. Active Learning
3.1. Manifold Learning
In remote sensing, application of active learning methods that
Recently the field of manifold learning has appeared as a select the most relevant samples for training is quite recent.
powerful framework to analyze nonlinearities in the data. The A SVM method for object-oriented classification was pro-
field is related to that of dimensionality reduction and non- posed in [76], while maximum likelihood classifiers for pixel-
linear feature extraction, which is scattered throughout com- based classification was presented in [77]. Recently, this ap-
puter science, machine learning, image processing and cyber- proach was extended in [78] by proposing boosting to iter-
netics. The main goal in manifold learning is to map high atively weight the selected pixels. In [79, 80] information-
dimensional data into a lower dimension while preserving the based active learning was proposed for target detection, and
main features of the original data for better analysis. In this in [81], a model-independent active learning method was pro-
way, visualization and understanding of high-dimensional data posed for very-high resolution satellite images.
becomes feasible. Traditional linear dimensionality reduc-
tion methods fail in describing the inherent structure of re- 3.5. Structured Learning
mote sensing data. Consequently some preliminary works us-
ing nonlinear transforms have been presented, such as Isomap Most of the techniques revised so far assume a simple set of
[58, 59], Laplacian methods [60, 61] or Local Linear Embed- outputs Y, for instance binary labels Y = {−1, 1}. How-
ding [59, 62]. Besides, some algorithms that analyze the in- ever, more complex output spaces can be imagined, e.g. pre-
trinsic dimensionality of hyperspectral images can be men- dicting multiple labels (land use and land cover simultane-
tioned [63]. ously), multi-temporal image sequences, or abundance frac-
tions. Such complex output spaces are the topic of struc-
tured learning, one of the most recent developments in ma-
3.2. Semi-supervised Learning
chine learning. Only a computer vision application [82] and
A related field to manifold learning is semi-supervised learn- the preliminary results in [83] have been presented for image
ing, which is concerned in developing models that exploit the processing.
 4. CONCLUSIONS [10] J. Inglada, “Automatic recognition of man-made objects in
high resolution optical remote sensing images by SVM classi-
The field that machine learning occupies in remote sensing fication of geometric image features,” ISPRS J. Photogramm.
has been summarized in this paper. Attention has been paid Rem. Sens., vol. 62, 236–248, 2007.
not only to the standard machine learning paradigms (classi- [11] G. Camps-Valls, L. Gomez-Chova, J. Muñoz Marı́, J. Vila-
fication, regression and feature extraction/selection), but also Francés, and J. Calpe-Maravilla, “Composite kernels for hy-
to recently and promising ones, such as manifold, semisuper- perspectral image classification,” IEEE Geosci. Remote Sens.
vised, active, transfer and structured learning. The special Lett., vol. 3, 93–97, 2006.
peculiarities of the images open the field for research and de- [12] M. Chi, R. Feng, and L. Bruzzone, “Classification of hyper-
velopment of new methods. And viceversa, the new learning spectral remote-sensing data with primal SVM for small-sized
paradigms available offer new ways of looking at old, yet un- training dataset problem,” Adv. Space Res., vol. 41 (11), 1793–
solved, problems. 1799, 2008.
[13] M. Fauvel, J. A. Benediktsson, J. Chanussot, and J. R. Sveins-
son, “Spectral and spatial classification of hyperspectral data
5. ACKNOWLEDGMENTS
using SVMs and morphological profiles,” IEEE Trans. Geosci.
The author would like to thank Mr. Devis Tuia from the University of Remote Sens., vol. 46 (11), 3804 – 3814, 2008.
Lausanne (Switzerland) for his valuable pieces of advice and useful [14] G. J. Briem, J. A. Benediktsson, and J. R. Sveinsson, “Multiple
discussion and comments about this work. classifiers applied to multisource remote sensing data,” IEEE
Trans. Geosc. Rem. Sens., vol. 40, no. 10, 2291–2299, 2002.
6. REFERENCES [15] J. Ham, Y. C. Chen, M. M. Crawford, and J. Ghosh, “Inves-
tigation of the random forest framework for classification of
[1] T. M. Lillesand, R. W. Kiefer, and J. Chipman, Rem. Sens. and hyperspectral data,” IEEE Trans. Geosc. Rem. Sens., vol. 43,
Image Interpretation, J. Wiley & Sons, NY, 2008. no. 3, 492–501, 2005.
[2] G. Shaw and D. Manolakis, “Signal processing for hyperspec- [16] G. Camps-Valls and A. Rodrigo-González, “Classification of
tral image exploitation,” IEEE Signal Processing Magazine, satellite images with regularized AdaBoosting of RBF neu-
vol. 50, 12–16, Jan 2002. ral networks,” in Speech, Audio, Image and Biomedical Sig-
nal Processing using Neural Networks, Bhanu Prasad and
[3] J.A. Benediktsson, J. A. Palmason, and J. R. Sveinsson, “Clas- S. R. Mahadeva Prasanna, Eds., pp. 309–326. Springer-Verlag,
sification of hyperspectral data from urban areas based on ex- Germany, 2008.
tended morphological profiles,” IEEE Trans. Geosci. Remote
Sens., vol. 43, 480–490, 2005. [17] A. Baraldi, V. Puzzolo, P. Blonda, L. Bruzzone, and
C. Tarantino, “Automatic spectral rule-based preliminary map-
[4] F. Del Frate, F. Pacifici, G. Schiavon, and D. Solimini, “Use ping of calibrated landsat TM and ETM+ images,” IEEE Trans.
of neural networks for automatic classification from high- Geosc. Rem. Sens., vol. 44, no. 9, 2563–2586, Sept. 2006.
resolution images,” IEEE Trans. Geosci. Remote Sens., vol.
[18] Yanfei Zhong, Liangpei Zhang, Bo Huang, and Pingxi-
45 (4), 800–809, 2007.
ang Li, “An unsupervised artificial immune classifier for
[5] G. Camps-Valls and L. Bruzzone, “Kernel-based methods for multi/hyperspectral remote sensing imagery,” IEEE Trans.
hyperspectral image classification,” IEEE Trans. Geosci. Re- Geosc. Rem. Sens., vol. 44, no. 2, 420–431, Feb. 2006.
mote Sens., vol. 43, 1351–1362, 2005.
[19] A. Baraldi and F. Parmiggiani, “A neural network for unsuper-
[6] C. Huang, L.S. Davis, and J.R.G. Townshend, “An assessment vised categorization of multivalued input patterns: an applica-
of support vector machines for land cover classification,” Int. tion to satellite image clustering,” IEEE Trans. Geosc. Rem.
J. Remote Sens., vol. 23:4, 725–749, 2002. Sens., vol. 33, no. 2, 305–316, 1995.
[7] G. M. Foody and A. Mathur, “Toward intelligent training of [20] M. Awad, K. Chehdi, and A. Nasri, “Multicomponent image
supervised image classifications: directing training data acqui- segmentation using a genetic algorithm and artificial neural
sition for SVM classification,” Remote Sens. Environ., vol. 93, network,” IEEE Geosc. Rem. Sens. Lett., vol. 4, no. 4, pp.
107–117, 2004. 571–575, Oct. 2007.
[8] M. Fauvel, J. Chanussot, and J. A. Benediktsson, “Evaluation [21] A. Mukhopadhyay and U. Maulik, “Unsupervised pixel classi-
of kernels for multiclass classification of hyperspectral remote fication in satellite imagery using multiobjective fuzzy cluster-
sensing data,” in IEEE ICASSP - International conference on ing combined with SVM classifier,” IEEE Trans. Geosc. Rem.
Acoustics, Speech and Signal Processing, Toulouse, France, Sens., vol. 47, no. 4, 1132–1138, April 2009.
2006, II–813–II–816. [22] G. Bilgin, S. Erturk, and T. Yildirim, “Unsupervised classifi-
[9] G. Camps-Valls, L. Gómez-Chova, J. Mu noz Marı́, J. L. Rojo- cation of hyperspectral-image data using fuzzy approaches that
Álvarez, and M. Martı́nez-Ramón, “Kernel-based framework spatially exploit membership relations,” IEEE Geosc. Rem.
for multi-temporal and multi-source remote sensing data clas- Sens. Lett., vol. 5, no. 4, pp. 673–677, Oct. 2008.
sification and change detection,” IEEE Trans. Geosci. Remote [23] R.L. Cannon, J.V. Dave, J.C. Bezdek, and M.M. Trivedi, “Seg-
Sens., vol. 46 (6), pp. 1822–1835, 2008. mentation of a thematic mapper image using the fuzzy c-means
 clustering algorithm,” IEEE Trans. Geosc. Rem. Sens., vol. [38] L. Bruzzone, R. Cossu, and G. Vernazza, “Detection of land-
GE-24, no. 3, 400–408, 1986. cover transitions by combining multidate classifiers,” Pattern
[24] S. Bandyopadhyay, U. Maulik, and A. Mukhopadhyay, “Mul- Recognition for Remote Sensing (PRRS 2002), vol. 25(13),
tiobjective genetic clustering for pixel classification in remote 1491–1500, 2004.
sensing imagery,” IEEE Trans. Geosc. Rem. Sens., vol. 45, no. [39] D. Tuia, F. Ratle, A. Pozdnoukhov, and G. Camps-Valls,
5, 1506–1511, 2007. “Multi-source composite kernels for urban image classifica-
[25] M. Tyagi, F. Bovolo, A.K. Mehra, S. Chaudhuri, and L. Bruz- tion,” IEEE Geosc. Rem. Sens. Lett., 2008, Accepted.
zone, “A context-sensitive clustering technique based on [40] F. Bovolo, L. Bruzzone, and M. Marconcini, “A novel ap-
graph-cut initialization and expectation-maximization algo- proach to unsupervised change detection based on a semi-
rithm,” IEEE Geosc. Rem. Sens. Lett., vol. 5, no. 1, pp. 21–25, supervised SVM and a similarity measure,” IEEE Trans.
2008. Geosc. Rem. Sens., vol. 46(7), 2070–2082, 2008.
[26] C.M. Bachmann, T.F. Donato, G.M. Lamela, W.J. Rhea, M.H.
[41] S. Ghosh, L. Bruzzone, S. Patra, F. Bovolo, and A. Ghosh, “A
Bettenhausen, R.A. Fusina, K.R. Du Bois, J.H. Porter, and B.R.
novel context-sensitive technique for unsupervised change de-
Truitt, “Automatic classification of land cover on smith island,
tection based on hopfield type neural networks,” IEEE Trans.
va, using hymap imagery,” IEEE Trans. Geosc. Rem. Sens.,
Geosc. Rem. Sens., vol. 45(3), 778–789, 2007.
vol. 40, no. 10, 2313–2330, Oct 2002.
[27] A.R.S. Marcal and L. Castro, “Hierarchical clustering of mul- [42] L. Bruzzone and S.B. Serpico, “A technique for features se-
tispectral images using combined spectral and spatial criteria,” lection in multiclass problems,” Int. J. Rem. Sens., vol. 21 (3),
IEEE Geosc. Rem. Sens. Lett., vol. 2, no. 1, pp. 59–63, 2005. 549–563, 2000.
[28] A. Sarkar, M.K. Biswas, B. Kartikeyan, V. Kumar, K.L. Ma- [43] M. Pal, “Support vector machine-based feature selection for
jumder, and D.K. Pal, “A MRF model-based segmentation land cover classification: a case study with DAIS hyperspectral
approach to classification for multispectral imagery,” IEEE data,” Int. J. Rem. Sens., vol. 27 (14), 2877–2894, 2006.
Trans. Geosc. Rem. Sens., vol. 40, no. 5, 1102–1113, 2002. [44] R. Archibald and G. Fann, “Feature selection and classification
[29] P. Heas and M. Datcu, “Modeling trajectory of dynamic clus- of hyperspectral images with support vector machines,” IEEE
ters in image time-series for spatio-temporal reasoning,” IEEE Geosc. and Rem. Sens. Lett., vol. 4 (4), 674–679, 2007.
Trans. Geosc. Rem. Sens., vol. 43, no. 7, 1635–1647, 2005.
[45] G. Camps-Valls and L. Bruzzone, Kernel Methods for Remote
[30] Xiuping Jia and J.A. Richards, “Cluster-space representation Sensing Data Analysis, J. Wiley & Sons, 2009.
for hyperspectral data classification,” IEEE Trans. Geosc. Rem.
Sens., vol. 40, no. 3, 593–598, 2002. [46] N. Keshava and J. F. Mustard, “Spectral unmixing,” IEEE
Signal Processing Magazine, vol. 19, no. 1, 44–57, 2002.
[31] F. Wang, “A knowledge-based vision system for detecting land
changes at urban fringes,” IEEE Trans. Geosc. Rem. Sens., vol. [47] M. E. Winter, “N-FINDR: an algorithm for fast autonomous
31, no. 1, 136–145, 1993. spectral end-member determination in hyperspectral data,”
[32] J. T. Morisette and S. Khorram, “An introduction to using gen- Proc. SPIE Im. Spectr. V, vol. 3753, 266–277, 2003.
eralized linear models to enhance satellite-based change de- [48] J. M. P. Nascimento and J. M. Bioucas-Dias, “Vertex Com-
tection,” in Proceedings of the International Conference on ponent Analysis: A Fast Algorithm to Unmix Hyperspectral
Geoscience and Remote Sensing, 1997, 1282–1284. Data,” IEEE Trans. Geosc. Rem. Sens., vol. 43, no. 4, 898–
[33] D. Kushardono, K. Fukue, H. Shimoda, and T. Sakata, “Com- 910, 2005.
parison of multi-temporal image classification methods,” in [49] J. C. Harsanyi and C.-I Chang, “Hyperspectral image classifi-
Proceedings of the International Conference on Geoscience cation and dimensionality reduction: An orthogonal subspace
and Remote Sensing, 1995, 1282–1284. projection,” IEEE Trans. Geosc. Rem. Sens., vol. 32, no. 4,
[34] J. Li and R. M. Narayanan, “A shape-based approach to change 779–785, 1994.
detection of lakes using time series remote sensing images,”
[50] A. Plaza, P. Martinez, R. Perez, and J. Plaza, “A quantitative
IEEE Trans. Geosc. Rem. Sens., vol. 41, no. 11, 2466–2477,
and comparative analysis of endmember extraction algorithms
2003.
from hyperspectral data,” IEEE Trans. Geosc. Rem. Sens., vol.
[35] D. Liu, M. Kelly, and P. Gong, “Classifying multi-temporal 42, no. 3, 650–663, 2004.
Landsat TM imagery using Markov random fields and support
vector machines,” in 3rd International Workshop on the Anal- [51] A. Plaza and C.-I Chang, “Impact of initialization on design of
ysis of Multi-temporal Remote Sensing Images, 2005. endmember extraction algorithms,” IEEE Trans. Geosc. Rem.
Sens., vol. 44, no. 11, 3397–3407, 2006.
[36] F. Melgani and S. B. Serpico, “A Markov random field
approach to spatiotemporal contextual image classification,” [52] P. Cipollini, G. Corsini, M. Diani, and R. Grass, “Retrieval of
IEEE Trans. Geosci. Remote Sens., vol. 41, no. 11, pp. 2478– sea water optically active parameters from hyperspectral data
2487, 2003. by means of generalized radial basis function neural networks,”
IEEE Trans. On Geosc. Rem. Sens., vol. 39, 1508–1524, 2001.
[37] F. Melgani, “Classification of multitemporal remote-sensing
images by a fuzzy fusion of spectral and spatio-temporal con- [53] B. Dzwonkowski and X.-H. Yan, “Development and appli-
textual information,” Intl. Journal of Pattern Recognition and cation of a neural network based colour algorithm in coastal
Artificial Intelligence, vol. 18, no. 2, 143–156, 2004. waters,” Intl. J. Rem. Sens., vol. 26, no. 6, 1175–1200, 2005.
[54] G. Camps-Valls, L. Bruzzone, J.L. Rojo-Álvarez, and F. Mel- [69] L. Bruzzone, M. Chi, and M. Marconcini, “A novel transduc-
gani, “Robust support vector regression for biophysical param- tive SVM for semisupervised classification of remote sensing
eter estimation from remotely sensed images,” IEEE Geosc. images,” IEEE Trans. Geosci. Remote Sens., vol. 44 (11), pp.
Rem. Sens. Lett., vol. 3, no. 3, 339–343, 2006. 3363–3373, 2006.
[55] G. Camps-Valls, J. Muñoz-Marı́, L. Gómez-Chova, K. Richter, [70] M. Chi and L. Bruzzone, “Semi-supervised classification of
and J. Calpe-Maravilla, “Biophysical parameter estimation hyperspectral images by SVMs optimized in the primal,” IEEE
with a semi-supervised support vector machine,” IEEE Geosc. Trans. Geosc. Rem. Sens., vol. 45(4), 1870–1880, 2007.
Rem. Sens. Lett., 2009, Accepted. [71] M. Dundar and A. Langrebe, “A cost-effective semisupervised
[56] G. Camps-Valls, L. Gomez-Chova, J. Vila-Francés, J. Amorós- classifier approach with kernels,” IEEE Trans. Geosci. Remote
López, J. Muñoz-Marı́, and J. Calpe-Maravilla, “Retrieval of Sens., vol. 42, no. 1, pp. 264–270, 2004.
oceanic chlorophyll concentration with relevance vector ma- [72] L. Bruzzone and D.F. Prieto, “Unsupervised retraining of a
chines,” Rem. Sens. Envir., vol. 105, no. 1, 23–33, Nov 2006. maximum likelihood classifier for the analysis of multitempo-
[57] L. Pasolli, F. Melgani, and E. Blanzieri, “Estimating biophysi- ral remote sensing images,” IEEE Trans. Geosc. Rem. Sens.,
cal parameters from remotely sensed imagery with Gaussian vol. 39, no. 2, 456–460, Feb 2001.
Processes,” in IEEE International Geoscience and Remote [73] L. Bruzzone and R. Cossu, “A multiple-cascade-classifier sys-
Sensing Symposium, IGARSS’08, Boston, USA, 2008. tem for a robust and partially unsupervised updating of land-
[58] C.M. Bachman, T.L. Ainsworth, and R.A. Fusina, “Exploit- cover maps,” IEEE Trans. Geosc. Rem. Sens., vol. 40, no. 9,
ing manifold geometry in hyperspectral imagery,” IEEE Trans. 1984–1996, Sep 2002.
Geosci. Remote Sens., vol. 43 (3), 441–454, 2005. [74] L. Bruzzone and M. Marconcini, “Toward the automatic updat-
[59] C.M. Bachman, T.L. Ainsworth, and R.A. Fusina, “Improved ing of land-cover maps by a domain-adaptation SVM classifier
manifold coordinate representations of large-scale hyperspec- and a circular validation strategy,” IEEE Trans. Geosc. Rem.
tral scenes,” IEEE Trans. Geosci. Remote Sens., vol. 44 (10), Sens., vol. 47, no. 4, 1108–1122, April 2009.
pp. 2786–2803, 2006. [75] L. Gómez-Chova, G. Camps-Valls, L. Bruzzone, and J. Calpe-
[60] W. Kim, M. Crawford, J. Tilton, and J. Ghosh, “Spatially Maravilla, “Semi-supervised remote sensing image classifi-
adapted manifold learning for classification of hyperspectral cation based on clustering and the kernel mean map,” in
imagery with insufficient labeled data,” in IEEE Geosc. Rem. IEEE International Geoscience and Remote Sensing Sympo-
Sens. Symp. IGARSS, 2008. sium, IGARSS’08, Boston, USA, 2008.
[61] L. Gómez-Chova, G. Camps-Valls, J. Muñoz Marı́, and [76] P. Mitra, B. Uma Shankar, and S.K. Pal, “Segmentation of mul-
J. Calpe, “Semi-supervised image classification with Lapla- tispectral remote sensing images using active support vector
cian support vector machines,” IEEE Geosc. Rem. Sens. Lett., machines,” Pattern Recogn. Lett., vol. 25, 1067–1074, 2004.
vol. 5, pp. 336–340, 2008. [77] S. Rajan, J. Ghosh, and M. M. Crawford, “An active learn-
[62] T. Han and D. Goodenough, “Nonlinear feature extraction of ing approach to hyperspectral data classification,” IEEE Trans.
hyperspectral data based on locally linear embedding (LLE),” Geosci. Remote Sens., vol. 46 (4), pp. 1231–1242, 2008.
in IEEE Geosc. Rem. Sens. Symp. IGARSS, 2005. [78] G. Jun and J. Ghosh, “An efficient active learning algo-
[63] C.-I Chang and Q. Du, “Estimation of number of spectrally rithm with knowledge transfer for hyperspectral remote sens-
distinct signal sources in hyperspectral imagery,” IEEE Trans. ing data,” in IEEE Geosc. Rem. Sens. Symp. IGARSS, 2008.
Geosc. Rem. Sens., vol. 42, no. 3, 608–619, 2004. [79] C. Zhang, S. Franklin, and M.A. Wulder, “Geostatistical and
[64] Q. Jackson and D.A. Landgrebe, “An adaptive classifier design texture analysis of airborne-acquired images used in forest
for high-dimensional data analysis with a limited training data classification,” Int. J. Remote Sens., vol. 25, 859–865, 2004.
set,” IEEE Trans. Geosc. Rem. Sens., pp. 2664–2679, Dec. [80] Q. Liu, X. Liao, and L. Carin, “Detection of unexploded ord-
2001. nance via efficient semisupervised and active learning,” IEEE
[65] G. Camps-Valls, T. Bandos, and D. Zhou, “Semi-supervised Trans. Geosci. Remote Sens., vol. 46 (9), pp. 2558–2567, 2008.
graph-based hyperspectral image classification,” IEEE Trans. [81] D. Tuia, F. Ratle, F. Pacifici, M. Kanevski, and W.J. Emery,
Geosc. Rem. Sens., vol. 45, no. 10, 3044–3054, Oct 2007. “Active learning methods for remote sensing image classifica-
[66] J. Muñoz-Marı́, L. Gómez-Chova, G. Camps-Valls, and tion,” IEEE Trans. Geosci. Remote Sens., in press.
J. J. Calpe-Maravilla, “Image classification with semi- [82] M.B. Blaschko and C.H. Lampert, “Learning to localize ob-
supervised one-class support vector machine,” in SPIE Remote jects with structured output regression,” in Computer Vision:
Sensing Conference, Cardiff, Wales, UK, Sep 2008. ECCV 2008., D. Forsyth, P. Torr, and A. Zisserman, Eds. 2008,
[67] L. Capobianco and G. Camps-Valls, “Hyperspectral target de- 2–15, Springer.
tection with a semi-supervised kernel orthogonal subspace pro- [83] D. Tuia, M. Kanevski, J Muñoz Marı́, and G. Camps-Valls,
jection,” IEEE Trans. Geosc. Rem. Sens., 2009, Accepted. “Structured SVM for remote sensing image classification,”
[68] D. Tuia and G. Camps-Valls, “Semi-supervised remote sens- in IEEE Workshop Machine Learning and Signal Processing,
ing image classification with cluster kernels,” IEEE Geosci. MLSP09, Grenoble, France, 2009.
Remote Sens. Lett., vol. 6 (1), 224–228, 2009.