See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228543190

A comparison of land use and land cover change detection methods

Conference Paper · April 2002

CITATIONS READS
112 3,209

5 authors, including:

Daniel Civco James D Hurd

University of Connecticut University of Connecticut
111 PUBLICATIONS   4,292 CITATIONS    36 PUBLICATIONS   935 CITATIONS   

SEE PROFILE SEE PROFILE

Emily Hoffhine Wilson Mingjun Song

University of Connecticut Microsoft
18 PUBLICATIONS   1,116 CITATIONS    4 PUBLICATIONS   358 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Daniel Civco on 04 June 2015.

The user has requested enhancement of the downloaded file.

 A COMPARISON OF LAND USE AND
LAND COVER CHANGE DETECTION METHODS

Daniel L. Civco1 , James D. Hurd2 , Emily H. Wilson 3 , Mingjun Song 4 , Zhenkui Zhang 5

Laboratory for Earth Resources Information Systems

Department of Natural Resources Management & Engineering
The University of Connecticut
U-4087, Room 308, 1376 Storrs Road
Storrs, CT 06269-4087
860-486-2840 (phone)
860-486-5408 (fax)

ABSTRACT

The principal objective of this project was to compare the results of different land use and land cover
change detection approaches: traditional post-classification cross-tabulation, cross-correlation analysis,
neural networks, knowledge-based expert systems, and image segmentation and object-oriented
classification. A combination of both direct T1 to T2 change detection as well as post-classification
analysis was employed. The test sites, located in the Stony Brook Millstone River Watershed in New
Jersey, consisted of two 5122 image blocks representative of the range of cover types and changes in the
watershed. Nine la nd use and land cover classes were selected for analysis: Dense Urban, Residential,
Turf & Grass, Agriculture, Deciduous Forest, Coniferous Forest, Water, Wetland, and Barren Land.
Twenty-three possible change and no-change classes were identified. Landsat Thematic Mapper data
from March 27, 1989 and September 3, 1989 represented conditions at T1 , and Landsat Enhanced
Thematic Mapper data from May 4, 2000 and September 23, 1999 were used for T2 . It was observed that
there are merits to each of the four methods examined, and that, at this point of this research, no single
approach can solve the land use change detection problem. This paper overviews the procedures used and
presents some of the results of the change detection experiment.

INTRODUCTION

Advances in remote sensing science, and in our ability to analyze temporal changes in our landscape,
hold great promise for putting to rest any questions of the relevancy of remote sensing to local land use
decisions. This assumption was the foundation for the formation of the “NAUTILUS” (Northeast
Applications of Useable Technology In Land planning for Urban Sprawl) Regional Earth Sciences
Application Center (RESAC) at UConn, one of nine RESACs designated by NASA in 1999. The
RESAC system was created with the goal of applying remote sensing research to pressing regional
problems. In the case of the Northeast RESAC, this translated to a variety of landscape characterization
techniques focused on providing information on the Northeast’s urbanizing landscape to local decision
makers.

Previous and on-going efforts of NAUTILUS and NEMO (Nonpoint Education for Municipal
Officials) investigators have demonstrated that 30-meter pixel land cover data derived from Landsat

1
Associate Professor and Director, [email protected]
2
Research Assistant, [email protected]
3
Research Assistant, [email protected]
4
Graduate Research Assistant, [email protected]
5
Graduate Research Assistant, [email protected]

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
Thematic Mapper (TM) imagery can be used for educational and planning applications at the local level.
But more accurate land cover information is needed if we are to move beyond first generation impacts of
educational programs and provide local end-users with information and products that can be easily and
directly incorporated into land use plans and policies. These considerations are at the core of the
NAUTILUS Project's research. Since a focus of this program is urban growth, it is clear that accurate land
use and land cover change data are required.

Land use and land cover mapping has long been an area of research focus for UConn investigators,
having explored a wide range of alternative approaches including expert systems (Civco, 1989), neural
networks (Civco, 1993; Wang and Civco, 1996), evidential reasoning (Wang and Civco, 1992), and other
techniques with roots in the field of artificial intelligence (AI). Area-wide land use and land cover
mapping projects (Civco and Hurd, 1999) have utilized multitemporal and multiresolution remote sensing
data (Zhou and Civco, 1998). Recent efforts have addressed improved methods for LULC change
detection (Hurd et al., 2001), and hierarchical image segmentation and object-oriented LULC
classification. Several techniques for determining land cover change have been evaluated by NAUTILUS
investigators. These have included post classification change detection (Hurd et al., 1992), multidate
classification change detection, cross-correlation analysis (Hurd et al., 2001), multidate principal
components analysis and RGB-NDVI color composite change detection (Hoffhine, 2000).

Computer-assisted production of spatially-detailed and thematically-accurate land use and land cover
information from satellite image data continues to be a challenge for the remote sensing research
community. The problems and limitations associated with single -date land cover information extraction
are compounded when attempting to produce land use change information using multitemporal data.

The principal objective of this project was to compare the results, qualitatively and quantitatively, of
different land use and land cover change detection approaches, including traditional post-classification
cross-tabulation, cross-correlation analysis, neural networks, knowledge-based expert systems, and image
segmentation and object-oriented classification. There were nine land use and land cover types of interest
{dense urban, residential, turf & grass, agriculture, deciduous forest, coniferous forest, water, non-
forested wetland, and barren}. These investigators placed a constraint on which land use changes were
plausible, or possible, and disallowed those there were deemed infeasible, or highly unlikley (i.e., water
becoming developed, or a residential area becoming agriculture). Only the 23 change and no change
classifications as shown in Table 1 were allowed. All subsequent classification change detection and
mapping methods adhered to this prescription.

Table 1. Allowable Change and No-change Land Cover Combinations

FromèTo Urban Residntl Turf&Grass Agric Decid Conif Water Wetland Barren
Urban
Residential
Turf&Grass
Agriculture
Deciduous
Coniferous
Water
Wetland
Barren

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
 METHODS
Study Area

The areas chosen for analysis were two sites within the Stony Brook Millstone Watershed located in
New Jersey, and one of NAUTILUS’s four study watersheds (Figure 1). The Stony Brook Millstone
River Watershed encompasses a 265-square mile region between Philadelphia and the New York
metropolitan area. The watershed includes 26 separate municipalities and is, as Stony Brook Millstone
Watershed Association6 Executive Director George Hawkins put it, "in the center of a development
storm". Geospatial information for this watershed can be found on NAUTILUS’s Stony Brook Millstone
River Maps & Data website 7 .

ME

VT
NH
NY
MA

CT

NJ

Figure 1. Location of Stony Brook Millstone Watershed in New Jersey and USGS MRLC Land Use and
Land Cover

Data

Two 512 pixel square subimages were selected from within the watershed for analysis. These were
chosen on the basis of the nature and magnitude of land use changes observed in the watershed. Landsat
Thematic Mapper data from March 27, 1989 and September 3, 1989 represented conditions at T1 , and
Landsat Enhanced Thematic Mapper data from May 4, 2000 and September 23, 1999 were used for T2.
All four dates for the two sites are depicted in Figure 2 (Site 1) and Figure 3 (Site 2).

Change Detection

As stated in the Introduction section of this paper, the research was designed to conduct a comparison
of four different land cover and land use change detection methods: Traditional Post-classification
(Lunetta and Elvidge, 1999), Cross-correlation Analysis (Koeln and Bissonnette, 2000), Neural Networks
(Dai and Khorram, 1999), and Image Segmentation and Object-oriented Classification (Niemeyer and

6
http://www.thewatershed.org/
7
http://resac.uconn.edu/maps_data/study_watersheds/maps_datanj.html

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
Canty, 2001). The research design attempted to standardize, as much as each method would permit, the
procedures so that results would enable a systematic, objective comparison among the methods. Due to
the inherent differences among the four approaches, however, some variation in each of the
methodologies was necessary. This became evident as the research progressed.

Traditional Post-classification.

A 14-band multiseasonal composite was created from the March 27, 1989 and September 3, 1989 Landsat
TM data. Unsupervised classification using the ISODATA algorithm was applied, and 75 spectral clusters
were generated, of which 47 were labeled into land cover classes and 28 were labeled as unknown. The
image pixels of unknown class membership were extracted for subsequent unsupervised classification
into 58 clusters, of which 33 were labeled into known classes and 25 were labele d as unknown. This
process of cluster busting was iterated a third time until all areas (classes) of the image could be
accounted for satisfactorily. The separate clustering and labeling results were overlain to form a single
composite classification map for T1 consisting of the nine classes of interest. The same repeated clustering
procedure was applied to the May 4, 2000 and September 23, 1999 ETM+ imagery to produce a final land
cover classification map for T2 .

Cross-correlation Analysis

Cross-correlation Analysis (CCA) overcomes many of the limitations of conventional change detection
methods. Cross-correlation Analysis is a change detection method developed by Earthsat, Inc. and
measures the differences between an existing land cover image and a recent single date multispectral
image (Koeln and Bissonnette, 2000). The benefit of this technique is that it eliminates the problems
associated with radiometric and phenological differences that are so readily experienced when performing
change detection.

Cross-correlation works by using the class boundaries from the base land cover image to derive an
expected class average spectral response. This information is used to derive a Z-statistic for each pixel
falling within a given land cover type. The Z-statistic describes how close a pixel's response is to the
expected spectral response of its corresponding class value in the land cover image. Pixels that have
undergone change between the T1 land cover image and the T2 multispectral image will produce high Z-
statistic values while pixels that have not changed will produce low Z-statistic values. The equation is:

2
r −µ 
n  ijk ic
jk

Z = ∑  
i = 1 σ ic 
jk

 jk 
where,

Zjk is the Z score for a pixel of a given class

i is the band number in the multispectral image
n is the number of bands
cjk is the thematic class being analyzed
rijk is the reflectance in band i for pixels in a given class
µic is the mean reflectance value in band i of all pixels in a given class
σic is the standard deviation of the reflectance value in band i of all pixels in a given class

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
 Spring Summer

1989 1989

Site 1

2000 1999

Figure 2. Landsat TM Imagery for Site 1. Bands 4, 5, and 3 {R, G, B, respectively}

1989 1989

Site 2

2000 1999

Figure 3. Landsat ETM+ Imagery for Site 2 Bands 4, 5, and 3 {R, G, B, respectively}

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
 In this experiment, the classification produced from the March 27, 1989 and September 3, 1989
Landsat TM data served as the baseline (T1 ) for determining areas of change in the T2 ETM data (May 4,
2000 and September 23, 1999). Graphical models created with ERDAS Spatial Modeler were developed
to perform CCA. Figure 4 is an illustration of the nature of CCA change detection. Additional detail on
cross-correlation analysis is provided in Koeln and Bissonnette (2000) and Hurd et al. (2001).

(a) T1 land cover (b) T1 TM (c) Changed forest (d) T2 ETM (e) T2 land cover
Figure 4. Example of residential growth detection using cross-correlation analysis (from Hurd et al.,
2001). Deciduous forest area in T1 (a) was used to extract the class’ spectral properties (b) and CCA
applied to T2 image data (d) to detect areas of significant difference (c), which were determined
through a user-defined threshold. These areas (c) were extracted and land cover for T2 was determined.

Neural Networks

Change detection was performed using a combination of tools, principally the NAUTILUS Image
Processing System (NIPS) and NeuralSIM8 . The former is a program developed by two of this paper’s
authors and consists of a suite of tools for performing supervised training site selection, conducting
classification using backpropagating neural networks, support vector machines, as well as the maximum
likelihood algorithm, deriving texture metrics from Haralick’s gray level co-occurrence matrix, and
several other specialized image processing functions. The latter, now known as NeuralWorks Predict® is
an integrated tool for creating and deploying prediction and classification neural network applications,
and operates within a Microsoft® Excel environment.

NIPS was used to select training sites for each of the nine land cover classes, doing so on a pixel-by-pixel
basis for each of the four dates on Landsat data. Training was performed independently for T1 and T2
imagery. The data for Sites 1 and 2 were pooled into a combined dataset, one for each time period of
Landsat data. A routine in NIPS permitted the cross-combination of selected from-to change classes, as
well as no change classes. Preliminary analyses revealed that the spectral heterogeneity of static and
change classes was compounded by different appearances of the same land use type in different years but
in the same season (e.g., a field covered by grass in spring T1 but bare soil in spring T2 , causing a
difference in reflectance due to phenology rather than real change). Therefore, subsequent analyses were
restricted to the summertime Landsat imagery only. A total of 133 pixels was selected for the nine classes
of interest from the September 3, 1989 TM imagery and 133 from the September 23, 1999 ETM imagery.
NIPS cross-combined the two independent training data sets into one in which every T1 pixel’s brightness
vector was paired with every T2 pixel’s brightness vector producing a dataset of 13,140 observations,
each with the brightness values for summer 1989 and summer 1999. This dataset was imported into Excel
and processed with NeuralSIM. The C++ code containing the parameters for the final network was
compiled and executed with NIPS to produce the final classification maps in an ERDAS Imagine format.

8
NeuralWare, 230 East Main Street, Suite 200, Carnegie, PA 15106-2700, USA

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
Image Segmentation and Object-oriented Classification

The software system named eCognition9 is based object-oriented processing and classification of
remote sensing imagery. As opposed to most other pattern recognition algorithms which operate on a
pixel-by-pixel basis, eCognition segments a multispectral image into homogeneous objects, or regions,
based on neighboring pixels’ spectral and spatial properties. Image segmentation can be performed at
different levels of resolution, or granularity. A knowledge-based approach is used to classify objects into
information categories, using Fuzzy Logic based on attributes of image objects and their mutual relations.
This classification can be performed at different levels of the classification hierarchy.

Numerous trials (and errors) for image segmentation and classification on T1 and T2 Landsat data
were attempted. In the final model, TM/ETM+ Bands 3, 4, and 5 from both the spring and summer T1 TM
data were used for segmentation, as were they for segmentation of the T2 ETM+ data. Four hierarchical
levels of segmentation were performed using different settings for scale (resolution), spectral
homogeneity, and shape (smoothness and compactness). Further, vegetation indices were calculated and
added to the image layer stack, and were used in the knowledge-base classifier to assist in removing some
of the erroneous change detected between differing field conditions of agriculture and residential.

After each date of imagery was satisfactorily segmented and classified with eCognition, the results
were combined into a T1 è T2 change image using ERDAS Imagine’s Knowledge-based Engineer
classification procedure.

RESULTS AND DISCUSSION

Figures 5 and 6 depict the classification results for the four methods for Sites 1 and 2, respectively.
Even a casual inspection of the results reveals some similarities as well as some extreme differences.
Traditional post classification change determination and cross-correlation analysis produced comparable
results overall. In both these cases, the T1 classifications were the same. For traditional post classification,
an independent T2 land cover map is cross-tabulated with T1 to identify change. For cross-correlation,
areas of change to be classified are based on spectral departure of class means defined by land cover areas
in the T1 classification. Accordingly, some similarities are expected.

Whereas the neural networks produced acceptably high accuracy for land cover change detection for
the training and test data sets (approximately 92% each), the final area-wide results were less satisfactory
overall. Many areas of omission and commission are evident. The problem is not with the neural network
approach itself, but how results of the trained network are turned into mapped information. In this
experiment, the three neurons (classes) with the highest output values were saved as image files, as were
the corresponding weighted non-linear sums, ranging from 0.0 to 1.0. In some instances, a pixel clearly
belonged to the winning (1st ) class, having an output value of 0.99, for instance, while 2nd and 3rd class
values were substantially lower. In other instances, the highest ranked class (output neuron) may not have
had differed at that much from the output values of competing classes (for example, 0.92 versus 0.91
versus 0.89). And in other instances, the magnitude of the winning output neuron was not very high,
sometimes lower than 0.5. The original intent of outputting the three highest classes, either change or no-
change, was to enable the possibility of performing mixed pixel analysis. At this point in the research,
however, that has not been done. In this experiment, only the class value from the highest output neuron
was used, only if that value exceeded 0.95. If it did not, then the classification from the T1 land cover
maps developed for traditional post classification and cross-correlation analysis was used. Clearly, this
rule is suboptimal, resulting in a number of land cover change omissions. A better way in which to

9
Definiens AG, Trappentreustr. 1, 80339 München, Germany

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
perform the change classification would be to use rules that vary by class and that take into consideration
the classes and their output neuron value for the 2nd and 3rd output neurons. Nonetheless, given the
fundamental logic behind this approach to discover changes automatically, neural networks deserve
further examination.

The image segmentation and object-oriented classification method for change detection holds much
promise. Initially, an attempt was made to perform a multitemporal segmentation, in which change
objects (regions) would have emerged somewhat automatically. However, given the spectral
heterogeneity of the Landsat scenes used, coupled with differences that are more phonologically-related
rather than real change, the resulting segmented image is nearly impossible to interpret. The knowledge-
base that would have to be developed would be infinitely complex. These and other investigators have
observed a similar level of complexity with the results from multitemporal cluster analysis. The approach
then resorted to one of performing multiresolution image segmentation for T1 and T2 image data
independently, and to develop a rule -base for determining regions (objects) of change. In one sense, this
is somewhat analogous to the traditional post-classification change detection. A major difference,
however, is the classification and change detection of objects, rather than simply pixels. This is evident
from the less-pixelated appearance of the change maps in Figures 5 and 6 compared with those of the
other three methods.

Figure 7 shows the results for a small portion of Site 2, in which development has occurred, along
with the corresponding Landsat images for summer T1 and summer T2 . Several noteworthy observations
can be made. There is substantial agreement among three of the methods, neural networks being the one
most dissimilar. Whereas traditional post-classification and cross-correlation analysis revealed the
residential growth in the center of the image (the agricultural field), they at the same time falsely-
committed already-built areas, just below center, an area that seemed to be under development at the time
of the T1 Landsat data. In the object-oriented classification, these two areas were correctly classified, as
agriculture è urban and pre-existing urban, respectively. Much of this change area was not detected
using the neural network, but again, it is believed to be a problem in establishing appropriate thresholds
rather than with the process itself. In this small area, however, there are a couple instances in which the
neural network did out-perform the other two methods. In the upper left-hand corner of the image, there is
clearly an area having undergone a barren è urban change (purple) between T1 and T2 . The neural
network correctly identified this area and type of change. Whereas the traditional post-classification and
cross-correlation analysis methods classified this area as change, as well, it was falsely labeled as
agriculture è urban. In the case of object-oriented classification, the area was labeled as urban in T1 .
Also, in the middle -left of the image, the object-oriented classification identified urban growth in an area
that had already been developed in T1. For the most part, however, the image segmentation and object-
oriented classification and change detection appeared to have produced better overall results, especially in
terms not only detecting and characterizing the nature of change, but also in minimizing the salt-and-
pepper effect caused by isolated and non-contiguous pixels.

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
 (a) Traditional Post-classification (b) Cross-correlation Analysis

(c) Neural Networks (d) Object-oriented Classification

Figure 5. Change Detection Results for Site 1

Urban Agriculture Forest Water Barren Agr to Forest to Barren to

Urban Urban Urban

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
 (a) Traditional Post-classification (b) Cross-correlation Analysis

(c) Neural Networks (d) Object-oriented Classification

Figure 6. Change Detection Results for Site 2

Urban Agriculture Forest Water Barren Agr to Forest to Barren to

Urban Urban Urban

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
Traditional Post-classification September 3, 1989 Cross-correlation Analysis

Neural Networks September 23, 1999 Object-oriented Classification

Figure 7. Detail of Change Detection for Portion of Site 2

CONCLUSIONS

The results of this research reveal that there is merit to each of the several land use change detection
methods studied, and that there appears to be no single best way in which to perform change analysis. It is
apparent that, for any method using T1 and T2 land cover maps to determine change, the accuracy of those
methods can be no better than that of each of the input maps, and is often quite lower. Methods such as
multidate clustering, image differencing, multidate principal components analysis, or RGB-NDVI color
composite change detection obviate the need for a high degree of a priori knowledge, but require
substantial a posteriori interpretation. The methods addressed in this research each explicitly identify a
priori the types and natures of land use change to be expected to occur within multitemporal remote
sensing data. The most significant conclusion of this study is that much research remains to be done to
improve upon the results of land use and land cover change detection. These investigators firmly believe
that an approach based on image-segmentation and rule -based classification is potentially such an
improved methodology, and accordingly intend on pursuing the avenues of neural network and object-
oriented classification change detection, perhaps in an integrated approach.

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002
 ACKNOWLEDGMENTS

This material is based upon work supported by the National Aeronautics and Space
Administration under Grant NAG13-99001/NRA-98-OES-08 RESAC-NAUTILUS, “Better Land Use
Planning for the Urbanizing Northeast:Creating a Network of Value-Added Geospatial Information,
Tools, and Education for Land Use Decision Makers". CLEAR Publication Number 020115.1 SAES
Scientific Contribution Number 2079.

LITERATURE CITED

Civco, D.L. (1989). Knowledge-based land use and land cover mapping. in Proc. of the 1989 Annual Meeting of the
American Society for Photogrammetry and Remote Sensing, Baltimore, MD. pp. 276-291.

Civco, D.L. (1993). Artificial neural networks for land cover classification and mapping. International Journal of
Geographic Information Systems 7(2):173-186.

Civco, D.L. and J.D. Hurd. (1999). A hierarchical approach to land use and land cover mapping using multiple
image types. Proc. 1999 ASPRS Annual Convention, Portland, OR. pp. 687-698.

Dai, X and S. Khorram. (1999). Remotely Sensed Change Detection Based on Artificial Neural Networks
Photogrammetric Engineering and Remote Sensing 65(10):1187-1194.

Hoffhine, E. (2000). Forest Change Detection, Harvest and Cover Type Distinction Using Multiple Dates of Landsat
Thematic Mapper. M.S.-degree Thesis, University of Maine, Orono.

Hurd, J.D., D.L. Civco, C. LaBash, and P. August. (1992). Coastal wetland mapping and change detection in the
northeast United States. in Proc. 1992 ASPRS/ACSM/RT'92 Convention, Washington, D.C. 1:130-139

Hurd, J.D., E.H. Wilson, S. Lammey and D.L. Civco. (2001). Characterization of Forest Fragmentation and Urban
Sprawl using Time Sequential Landsat Imagery. Proc. 2001 ASPRS Annual Convention, St. Louis, MO. 13 p.

Koeln, G. and J. Bissonnette. (2000). Cross-correlation analysis: mapping landcover change with a historic
landcover database and a recent, single-date multispectral image. in Proc. 2000 ASPRS Annual Convention,
Washington, D.C. 8 p.

Lunetta, R. and C. Elvidge. (1999). Remote Sensing Change Detection. Taylor & Francis, 320 p.

Niemeyer, I. and M.J. Canty. (2001): Object-Oriented Post-Classification of Change Images. In: Proceedings of
SPIE’s International Symposium on Remote Sensing, Toulouse, France, 17-21 September 2001, SPIE Vol. 4545. 9
p.

Wang, Y. and D.L. Civco (1992). Evidential reasoning and post-classification of misclassified pixels: a GIS
approach for improving the classification accuracy of remotely sensed data. in Proc. 1992 ASPRS/ACSM/RT'92
Convention, Washington, D.C. 4:160-170.

Zhou, J. and D.L. Civco. (1998). A wavelet transform method to merge Landsat TM and SPOT Panchromatic Data.
International Journal of Remote Sensing 19(4):743-757.

Wang, Y. and D.L. Civco. (1996). Three artificial neural network paradigms in high dimensional multisource spatial
data classification. The Association of Chinese Professionals in GIS Geographic Information Science 1(2):73-87

2002 ASPRS-ACSM Annual Conference and FIG XXII Congress April 22-26, 2002

View publication stats