Land Cover Classification Using IRS LISS III Image and DEM in a Rugged Terrain: A Case Study in Himalayas

Transcription

1 Land Cover Classification Using IRS LISS III Image and DEM in a Rugged Terrain: A Case Study in Himalayas A. K. Saha Department of Earth Sciences, Indian Institute of Technology Roorkee, Roorkee , India ashisksaha@gmail.com M. K. Arora Department of Civil Engineering, Indian Institute of Technology Roorkee, Roorkee , India manojfce@iitr.ernet.in E. Csaplovics Institute of Photogrammetry and Remote Sensing, TU- Dresden, Dresden, D-01069, Germany csaplovi@rcs.urz.tu-dresden.de R. P. Gupta Department of Earth Sciences, Indian Institute of Technology Roorkee, Roorkee , India rpgesfes@iitr.ernet.in Abstract Digital image classification is generally performed to produce land cover maps from remote sensing data, particularly for large areas. The performance of image classifiers that utilize only the remote sensing data may deteriorate, especially in mountainous regions, due to the presence of shadows of high peaks. In this study, a multisource classification approach to map land cover in Himalayan region with high mountain peaks having elevations up to 4785 m above mean sea level has been adopted. Remote sensing data from IRS LISS III image along with NDVI and DEM data layers have been used to perform multi-source classification using maximum likelihood classifier. The results show a substantial improvement in accuracy of classification on incorporation of NDVI and DEM as ancillary data over the classification performed solely on the basis of remote sensing data. Introduction The knowledge of spatial land cover information is essential for proper management, planning and monitoring of natural resources (Zhu, 1997). For example, it is a desired input for many agricultural, geological, hydrological and ecological models. Also, any natural hazard study such as landslide hazard zonation (e.g. Gupta et al., 1999; Saha et al., 2002) highly depends on the availability of accurate and up-to-date land cover information. Due to synoptic view, map like format and repetitive coverage, satellite remote sensing imagery is a viable source of gathering quality land cover information at local, regional and global scales (Csaplovics, 1998; Foody, 2002). Moreover, remote sensing data are, in particular, useful for land cover mapping in mountainous regions such as the Himalayas, since these areas are generally inaccessible due to high altitudes and ruggedness of the terrain. Over the years, a number of studies to map land cover using remote sensing data in high mountain areas have been reported with varying degrees of accuracy. This may be due to a large number of factors that influence the remote sensing process. These include the presence of shadows due to high altitude of the terrain, the cloud cover, deep narrow valleys and ravines, low sun angles, steep slopes and differential vegetation cover. Therefore, due to changes in environmental conditions, spectral characteristics also change from one region to the other (Arora and Mathur, 2001). Hence, classification only on the basis of spectral data from a remote sensing sensor alone may not be sufficient to gather effective land cover information. A classification approach that incorporates data from other sources may therefore be more effective than that is based solely upon the multispectral data from a single remote sensing sensor. The ancillary data from Geocarto International, Vol. 20, No. 2, June 2005 Published by Geocarto International Centre, G.P.O. Box 4122, Hong Kong. geocarto@geocarto.com Website: 33

2 other sources may be acquired from topographic maps (Bruzzone et al., 1997), geological (Gong, 1996) and other maps. The most useful information that can be obtained from topographic maps is the elevation contours, which can be digitized to produce a raster Digital Elevation Models (DEMs). DEMs along with their derivatives such as slope and aspect provide the basis for multi-source classification (Jones et al., 1988; Frank, 1988; Janssen et al., 1990). Data from different remote sensing sensors may also be combined to produce multi-sensor classification (Michelson et al., 2000). Moreover, a number of derivatives of multispectral images such as Principal Components Analysis (PCA) and Normalized Difference Vegetation Index (NDVI) may also be incorporated in the classification process to enhance the quality of land cover information from remote sensing data in mountainous regions (Eiumnoh and Shrestha, 2000). In mountainous regions such as the Himalayas, shadows are the major source of confusion in extracting land cover information from remote sensing data. Although, there is no suitable method to completely remove the effect of shadows, several alternatives exist to minimize their effect in order to improve classification from remotely sensed data. Many of the methods are based on shaded relief models that are produced from DEM. However, some of the studies (e.g. Kawata et al., 1988, Civco, 1989, Colby, 1991, Curran and Foody, 1994) have shown that due to the presence of errors in creating DEMs and the way the DEM is applied in the classification process, the correction for shaded slopes may get over-estimated. To counteract this problem of overcorrection, the use of NDVI image as an additional layer has been recommended, since the band ratio derivatives may help in nullifying the topographic component to some extent (Holben and Justice, 1981; Apan, 1997). It may however be mentioned that the NDVI alone may not completely remove the shadow effect. Recently, Eiumnoh and Shrestha (2000) exploited the advantages of incorporating both NDVI and DEM in the classification process and showed an improvement in the classification accuracy of the order of 10 to 20%. Motivated by the success of multisource classification in neutralizing the effect of shadows, the aim of this study is to present a case study to derive accurate land cover map using remote sensing and ancillary data in an area of high elevation and rugged terrain in the Himalayas, where shadows are the major problem and thus the implementation of multi-source classification will be of great value. The work reported in this paper is accomplished in part of a major on-going GIS based research to assess the impact of landslide related activities on route planning in Himalayas, where accurate land cover information forms a key data input. The multispectral image from IRS-LISS-III sensor has been used as the primary data with NDVI and DEM as the additional data layers to implement multi-source landcover classification using the logical channel approach (Tso and Mather, 2001). Separability analysis on the basis of transformed divergence has also been performed to examine the relative importance of various spectral bands and ancillary data layers in the classification process. The classification has been performed using the most widely used Maximum Likelihood Classifier (MLC). The Study Area, Data and Field Work The study area of about 730 km 2, situated in the Himalayas, covers a portion of Rudraprayag and Chamoli Districts of the newly formed Uttaranchal State in India (Fig. 1). The terrain is highly rugged with elevations varying from 880 m to 4785 m above mean sea level. The rivers Alaknanda and Madhyamaheshwar Ganga flow through the northwest and southeast parts of the area respectively. A large number of streams and tributaries of both the rivers form the drainage network of the region. The area is also dominated by dense forest cover (>50%). Terrace cultivation is associated with agricultural and fallow fields. Gopeshwar and Chamoli are the major towns surrounded by many small villages spread all over the region. Barren land and high mountain areas covered with snow are situated in the northeastern part of the region. From geological point of view, the region comprises of the Lesser Himalayas and the Higher Himalayas (Valdiya, 1980) and consists of a range of lithological units namely sandstones, limestones to granite-augen gneiss. Structurally, the region is complex due to the presence of various thrusts and faults. Consequently, the earthquake and landslide activities in the region often result in local change of land cover characteristics. Therefore, regular mapping and monitoring of land cover is desired. The present study is based on mapping land cover from IRS-1C remote sensing data. The LISS III multispectral Figure 1 Location map of the study area 34

3 image (23.5 m spatial resolution) (Fig. 2) has been used as the primary data to produce land cover classification, whereas the PAN image (5.8 m spatial resolution) (Fig. 3), due to its fine spatial resolution, has been used as reference data for creation of training and testing data sets. Several other studies (e.g., Fisher and Pathirana, 1990, Foody and Arora, 1996; Shalan et al., 2003) have also used finer resolution images for this purpose in the absence of insufficient field survey data. Nevertheless, in this study, the preparation of reference data was ably assisted with field surveys conducted in earlier years and in December, The season of the last field visit coincided with the date of acquisition of LISS III and PAN images to allow for same atmospheric and environmental conditions. Since the terrain is inadequately road networked and thus inaccessible due to high elevations and ruggedness, the information on existing land cover was collected only along the accessible roads during the field surveys. The Global Positioning System (GPS) receivers were also used during the field survey to obtain accurate location of land cover classes for their easy demarcation on geo-rectified LISS III and PAN images. The ancillary information in the form of DEM was derived from topographic maps. More description of these data sets is provided in Table 1. Figure 2 IRS 1C LISS III colour infrared composite (NIR, Red, Green RGB, date of acquisition: ) Methodology A number of data processing steps are involved in performing multi-source classification. These include preprocessing of LISS III image to correct for atmospheric errors, registration of LISS III and PAN images, generation of ancillary data layers, image classification and accuracy assessment. All the processing has been done on ERDAS Imagine, Arc GIS and ILWIS software. The processing steps are briefly described now. Pre-processing of LISS III image The atmospheric scattering is common in remote sensing data and is generally more pronounced in the shorter wavelength regions (e.g., blue). The effect of atmospheric scattering is to contribute some additional spectral values to the ground reflectance (Gupta, 2003; Jensen, 1986). In this Figure 3 A portion of IRS PAN data showing various landuse/landcover classes (Df Dense Forest; Sv Sparse Vegetation; Ag Agriculture; F1 Fallow; St Settlements; Fs Fresh Sediments; Wb Water Body, Date of acquisition: ) Table 1 Characteristics of remote sensing and other data used in the study Data 1. IRS 1C LISS III image (sun angle 36.31º, sun azimuth º) in 4 bands (Green: µm, Red: µm, NIR: µm and SWIR: µm) 2. IRS 1C PAN image (sun angle 40.26º, sun azimuth º) µm 3. Topographic maps (Sheet Number 53 N /2,3,6,7; scale 1:50,000) 4. Field data on land use/land cover Source National Remote Sensing Agency (NRSA), India NRSA, India Survey of India Ground truth collected during the study Date of Acquisition During December,

4 study, the LISS III image was corrected for atmospheric path radiance using dark object subtraction method (Chavez, 1988). The method is fast and easy, as it does not require information on atmospheric conditions at the time of image acquisition. To implement this method, the pixel (associated with the dark object) having minimum brightness value in the Near Infra Red (NIR) band was detected and the corresponding pixel values in all other bands were subtracted from the specific raw bands. This will result in an image that is corrected for atmospheric scattering. Geometric registration of images Accurate geometric registration of images is a pre-requisite to perform a multi-source classification. First, the LISS III image was geometrically corrected using 39 well-distributed Ground Control Points (GCPs) extracted from the topographical map. Due to non-existence of sharp, welldefined, stable and prominent features in the image, the GCPs were acquired mainly from the intersection of the drainage lines. Owing to the steep topography and narrow valleys in Himalayas, it was assumed that there was no change in drainage network between the year of topographic map generation and the date of acquisition of LISS III image. The registration was performed to sub-pixel accuracy using first order polynomial transformation and nearest neighbour resampling method, thus letting the brightness values unchanged. The PAN image was also registered with the LISS III image to a sub-pixel accuracy using 60 welldistributed GCPs. The registration of the PAN image was necessary as this image was used as reference data for accurate demarcation of training and testing areas in the LISS III image. Generation of ancillary data DEM and NDVI data layers were used as additional bands (referred as ancillary data) to perform multi-source classification. a) Production of DEM In high altitudes and rugged terrain, a major variation in the brightness values of pixels can be found due the presence of shadows. This may lead to erroneous classification. Therefore, the DEM was used as ancillary data in the classification process primarily to reduce misclassification of shadowed areas to water bodies. Moreover, the elevation information from DEM may also act as a logical rule to eliminate the presence or absence of certain classes in some elevation zones. For example, fallow land is not expected to exist at higher elevations that are covered with snow since climatic conditions do not allow for any agricultural activity at such high elevations. These areas thus should be categorized as barren land. Therefore, any presence of fallow land in the neighborhood of snow-covered areas may represent a misclassification, which can be reduced by including a DEM in the remote sensing classification process. The DEM was produced by digitizing contours from topographic map at a scale of 1:50,000 with 40 m contour interval. Triangulated Irregular Network (TIN) model was used to produce a raster DEM at 23.5m spatial resolution to match with that of LISS-III image. b) NDVI layer As the study area is dominated by different types of vegetation, NDVI was used as an ancillary data layer in the classification process to enhance the separability among various vegetation classes and also to reduce the shadow effect due to variations in topography. The NDVI data layer was generated from Red and NIR bands of LISS-III image and is defined as, NDVI=(NIR-Red)/(NIR+Red) (1) The pixel values of the NDVI data layer range from -1 to +1 and are scaled from 0 to 255 respectively. The higher NDVI values indicate increase in biomass per unit area and vice versa. The NDVI data layer draped over DEM is presented in Fig. 4. In this figure, the NDVI values vary from to The positive values represent different types of vegetation classes, whereas near zero and negative values indicate non-vegetation classes, such as water, snow and barren land (Fig. 4). Image Classification A series of image classification operations were performed to produce land cover map from LISS III image and ancillary data layers, which are described in the following, Selection of a land cover classification scheme A classification scheme defines the land cover classes to be considered for remote sensing image classification. Sometimes a standard classification scheme such as Anderson s land use land cover classification system (Anderson et al., 1976) is used, while at other times the number of land cover classes are chosen according to the requirements of the specific application. In this study, based on Anderson s classification system, nine land cover classes were defined. The detail description of these classes along with their interpretative characteristics both on the False Colour Composite (FCC) of LISS-III image and PAN image is provided in Table 2. Formation of training dataset Training data extraction is a critical step in a supervised image classification process. As the success of a classification highly depends on the quality of the training data, these must be selected from the regions representative of the land cover classes under investigation. Data should thus be collected from relatively homogeneous areas consisting of those classes. The collection of training data is generally a time consuming and tedious process, as it involves strenuous field surveys and accumulation of reference data from various sources. Therefore, the size of the training data set is kept 36

5 Table 2 Characteristics of land cover classes Land Cover Class Description Characteristics on LISS-III FCC Characteristics on PAN image Dense forest Tall dense trees Low vegetation density with exposed ground surface Crops on hill terraces as step cultivation Dark red with rough texture Dark tone with rough texture Sparse vegetation Dull red to pinkish Light tone with dark patches Agriculture Dull red and smooth appearance Step like arrangement of fields Fallow Agricultural fields without crops Bluish/greenish grey with smooth texture Yellowish Bright tone with smooth texture Barren Exposed rocks without vegetation Towns and villages; block like appearance Fresh landslide debris and river sediments on the bank Very bright tone Typical blocky appearance with light tone Settlements Bluish Fresh sediments Cyanish Bright tone Water body Rivers and lakes Cyanish blue to blue according to the depth of water and sediment content Dark tone Snow Snow covered areas on high altitude mountains Bright white Very bright tone small. Nevertheless, the number of pixels constituting the training data set must be large enough to accurately characterize the land cover classes. As a rule of thumb, the number of training pixels for each class may be kept as 30 times the number of bands under consideration (Mather, 1999). In this study, the training data set consisted of about 1% of the total pixels in the LISS III image. The number of training samples for each class (Table 3) were chosen in proportion to the area covered by the respective classes on the ground. Similar to other studies, the fine spatial resolution PAN image and topographic map were used as reference data (ground truth) to delineate the training pixels on the LISS III image. First, the PAN image was visually interpreted on screen based on the characteristics defined in Table 2 and the previous knowledge of the study area to delineate nine land cover classes. Wherever there appeared to be confusion in identifying the classes, these were verified in the field. The PAN image derived land cover information was used to demarcate training areas on LISS III image. The quality of training areas, thus identified, was evaluated through histogram plots. Majority of training areas were normally distributed having single peak, which is a requirement of the maximum likelihood classifier used in this study. Separability analysis The dataset for multi-source classification consisted of six data layers (four bands of multi-spectral LISS III image, two ancillary data sources - NDVI and the DEM). For convenience, Green band, Red band, NIR band, Short Wave IR (SWIR) band, NDVI and DEM data layers have been numbered as 1, 2, 3, 4, 5 and 6 respectively. A separability analysis was performed using the training dataset, selected earlier, to identify the combination of bands that shows the highest distinction between the land cover classes. Table 3 Number of training pixels for each land cover class used in classification Land Cover Class Number of Training Pixels 1 Dense Forest Sparse vegetation Agriculture Fallow Barren Settlements Fresh sediments Water body Snow 1642 Total Separability is a statistical measure devised on the basis of spectral distances computed for a combination of bands. From a number of separability measures, the Transformed Divergence (TD) has been used in this study (Jensen, 1986). The TD values range from 0 to A value close to 2000 indicates the best separability. The values between 1800 and 2000 are generally considered adequate for the selection of appropriate band combinations. Since, the focus of the present study is on the inclusion of ancillary data in the classification process, the average TD values of various band combinations that included ancillary data, were computed. Various fiveband combinations that produced average TD values near to 2000 were considered appropriate for classification (Table 4). Although all the five-band combinations are equally good, the band combination 1, 2, 3, 4 and 6 resulted in the highest average TD value. This illustrates that LISS III 37

6 image together with DEM data layer has produced the best separability among various pairs of land cover classes. However, since the average TD values for different band combinations did not show any variation, all the five band combinations as well as the complete data set (1,2,3,4,5,6) were used to perform classification. Maximum likelihood classification (MLC) Over the years, a number of image classifiers have been developed. MLC has been found to be the most accurate and commonly used classifier, when data distributional assumptions are met. This classifier is based on the decision rule that the pixels of unknown class membership are allocated to those classes with which they have the highest likelihood of membership (Foody et al., 1992). The detailed formulation of this classifier may be found in Richards and Jia (1999) and has not been provided here. MLC has been used here to produce a number of land cover maps using different band combinations as described in the previous section. Classification accuracy assessment No image classification is said to be complete unless its accuracy has been assessed. To determine the accuracy of classification, a sample of testing pixels is selected on the classified image and their class identity is compared with the reference data (ground truth). The choice of a suitable sampling scheme and the determination of an appropriate sample size for testing data plays a key role in the assessment of classification accuracy (Arora and Agarwal, 2002). The pixels of agreement and disagreement are generally compiled in the form of an error matrix. It is a c x c matrix (c is the number of classes), the elements of which indicate the number of pixels in the testing data. The columns of the matrix depict the number of pixels per class for the reference data, and the rows show the number of pixels per class for the classified image. From this error matrix, a number of accuracy measures such as overall accuracy, user s and producer s accuracy, may be determined (Congalton, 1991). The overall accuracy is used to indicate the accuracy of whole classification (i.e. number of correctly classified pixels divided by the total number of pixels in the error matrix), whereas the other two measures indicate the accuracy of individual classes. User s accuracy is regarded as the probability that a pixel classified on the map actually represents that class on the ground or reference data, whereas producer s accuracy represents the probability that a pixel on reference data has been correctly classified. In this study, PAN image derived land cover information together with previous knowledge of the area through earlier field visits was used as reference data to generate testing data set. A total of 200 testing pixels for each class were randomly selected, which are significantly larger than the sample size of 75 to 100 pixels per class as recommended by Congalton (1991) for accuracy assessment purposes. For effective comparison, the same testing dataset was used to determine the overall and producer s accuracy from different land cover classifications. Results and Discussions The aim of this study is to implement a multisource classification approach to produce an accurate land cover map for its subsequent use in GIS based application on landslide hazard zonation. The accuracy of land cover maps obtained from multisource classification of dataset using different band combinations is shown in Table 5. The classification based only on spectral data of LISS III image produced an accuracy of 86.94%, which is more than the minimum accuracy criterion of 85% overall accuracy as reported in Anderson s land use land cover classification Table 4 Table 5 Figure 4 Various band combinations and their average TD values (Bands 1,2,3,4: LISS III bands; Band 5: NDVI; Band 6: DEM) Band Combinations Average TD 1,2,3,4, ,2,3,4, ,2,3,5, ,2,4,5, ,3,4,5, ,3,4,5, The overall accuracies for land cover classifications produced from various band combinations. Bands 1, 2, 3, 4, 5, and 6 have been defined in the text. The bold figure corresponds to the highest accuracy Band Combination Overall Accuracy 1, 2,3, 4,5, ,2,3,4, ,2,3,4, ,2,4,5, ,3,4,5, ,2,3,5, ,3,4,5, ,2,3, Normalized Difference Vegetation Index (NDVI) image from LISS III imagery draped over Digital Elevation Model. 38

7 system (Anderson, 1976). On inclusion of NDVI data layer with spectral data, this accuracy marginally dropped to 84.78%, however it increased remarkably to 91.00% when both DEM and NDVI data layers were included in the classification process. The highest accuracy of 92.04% was obtained with the five-band combination (i.e., band combination 1, 2, 4, 5, 6) that contains both NDVI and DEM data layers, thus producing a considerable increase of the order of 5% in accuracy. To assess the accuracy of individual land cover classes, Table 6 Producer s accuracy of individual classes derived from classifications using 1,2,3,4 band combination vis-a-vis 1,2,4,5,6 band combination. 1, 2, 3, 4, 5, and 6 bands have been defined in the text Classes Producer s Accuracy (%) 1,2,3,4 1,2,4,5,6 Dense Forest Sparse Vegetation Agriculture Fallow Barren land Settlements Fresh sediments Water body Snow cover producer s accuracies were also determined for the classification that provided the highest overall accuracy (i.e., the classification obtained by using band combination 1, 2, 4, 5, 6). These accuracy values were also compared with those obtained from the classification produced by using only LISS III spectral data (Table 6). A glance at producer s accuracy values shows that the accuracy of most of the classes has increased when NDVI and DEM data layers are added in the classification process. This illustrates that the misclassifications between the classes have been reduced. In particular, the classes namely sparse vegetation, agriculture, fallow and barren land, settlements and snow showed a substantial increase in accuracy ranging from 1.5% to 20%. Two explicit reasons may be stated for this increase in accuracy. First, the class barren land was considerably misclassified with the classes settlements and water body when only spectral data were used. Since, at high elevations, the presence of these classes is scarce, addition of DEM data layer reduced this misclassification. Secondly, due to the presence of shadows in the region, the classification using only spectral data showed misclassification of fallow land and sparse vegetation to the class forest. The addition of NDVI and DEM data layers reduced the shadow effect and hence the misclassification. On visual comparison of FCC (Fig. 2) and the land cover classification of highest accuracy (Fig. 5), it can further be observed that the addition of DEM and NDVI data layers resulted in the correct classification of shadowed areas to their corresponding vegetation classes, which was not the case when only spectral data was used for classification. Thus, this study clearly demonstrates the utility of incorporating NDVI and DEM in the remote sensing classification process. The land cover classification of highest accuracy had some stray pixels over the whole image. To remove these stray pixels so as to produce smooth land cover classification, a 30 majority filter was applied over the classification with the highest accuracy. The resulting product was considered as the final land cover map to be used as input for subsequent GIS based study. Figure 5 The land cover classification with the highest accuracy (i.e., 92.06%), produced from the band combination 1, 2, 4, 5, 6 (i.e., Green, Red and SWIR bands of IRS LISS III image, NDVI image and DEM) Conclusions Remote sensing data are attractive for land cover classification, particularly in the high mountainous regions, where most of the areas are inaccessible due to the rugged terrain. However, classification just on the basis of the reflectance characteristics of remote sensing data may not be appropriate due the presence of shadows in these areas. Therefore, the use of ancillary datasets in addition to remote sensing data has been recommended. The case study presented in this paper also showed a remarkable increase in accuracy of land cover classification on incorporation of DEM and NDVI data layers with IRS-LISS-III image. Classification accuracies of the order of 90% were obtained for the majority of the nine land cover classes to be mapped. The addition of the ancillary data substantially reduced the misclassifications 39

8 incurred due to the effect of shadow in the image and also due to the similarity in spectral characteristics of some classes in high elevation areas. The present study thus highlights the effectiveness of integrating DEM and NDVI data layers with the spectral data to enhance the quality of land cover classifications in mountainous regions such as the Himalayas. Acknowledgments AKS is grateful to Council of Scientific and Industrial Research (CSIR), New Delhi, India for Senior Research Fellowship. He is also thankful to German Academic Exchange Service (DAAD), Bonn for the award of DAAD Sandwich Model Fellowship, during which a major portion of this work was carried out at Institute of Photogrammetry and Remote Sensing, Dresden University of Technology, Dresden, Germany. Assistance provided by Pallov Pal during field surveys is also acknowledged. References Anderson, J M, Hardy, E E, Roach, J T and Witmert, R E A Land Use Classification System for Use with Remote Sensing Data. U.S. Geological Survey Professional Paper, No. 964, Washington D. C.: Government Printing Office. Apan, A A Land cover mapping for tropical forest rehabilitation planning using remotely-sensed data. International Journal of Remote Sensing, 18(5), Arora, M K and Agarwal, K A program for sampling design for image classification accuracy assessment. Photogrammetry Journal of Finland, 18(l), Arora, M K and Mathur, S Multi-source Classification Using Artificial Neural Network in a Rugged Terrain. Geocarto International, 16(3), Bruzzone, L, Conese, C, Maselli, F and Roli, F Multi-source classification of complex rural areas by statistical and neural-network approaches. Photogrammetric Engineering & Remote Sensing, 63(5), Chavez, P S Jr An improved dark object subtraction technique for atmospheric correction of multispectral data. Remote Sensing of Environment, 24, pp Civco, D L Topographic Normalization of Landsat Thematic Mapper Digital Imagery. Photogrammetric Engineering & Remote Sensing, 55(9), Colby, J D Topographic Normalization in rugged terrain. Photogrammetric Engineering & Remote Sensing, 57(5), Congalton, R G A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment, 3 7, Csaplovics, E High Resolution space imagery for regional environmental monitoring status quo and future trends. International Archives of Photogrammetry and Remote Sensing, 32(7), Curran, P J and Foody, G M The use of remote sensing to characterize the regenerative states of tropical forests. In Environmental remote Sensing from Regional to Global Scales, edited by G. Foody and P. Curran (Chichester: John Wiley and Sons), Eiumnoh, A and Shrestha, P Application of DEM Data to Landsat Image Classification: Evaluation in a Tropical Wet-Dry Landscape of Thailand. Photogrammetric Engineering & Remote Sensing, 66(3), Fisher, P F and Pathirana, S The Evaluation of Fuzzy Membership of Land Cover Classes in the Suburban Zone. Remote Sensing of Environment, 34(2), Foody, G M Status of Land Cover Classification Accuracy Assessment, Remote Sensing of Environment, 80, Foody, G M and Arora, M K. 1996, Incorporating Mixed Pixels in the Training, Allocation and Testing Stages of Supervised Classifications, Pattern Recognition letters, 17, Foody, G M, Campbell, N A, Trodd, N M and Wood, T F Derivation and applications of probabilistic measures of class membership from maximum likelihood classification. Photogrammetric Engineering and Remote Sensing, 5 8, Frank, T D Mapping Dominant Vegetation Communities in the Colorado Rocky Mountain Front Range with LANDSAT Thematic Mapper and Digital Terrain Data. Photogrammetric Engineering & Remote Sensing, 54(12), Gong, P Integrated analysis of spatial data from multisources using evidential reasoning and artificial neural-network techniques for geological mapping. Photogrammetric Engineering & Remote Sensing, 62, Gupta, R P Remote Sensing Geology, 2 nd Edition (Heidelberg: Springer-Verlag). Gupta, R P, Saha, A K, Arora, M K and Kumar A Landslide Hazard Zonation in a part of the Bhagirathi Valley, Garhwal Himalayas, using integrated remote sensing -GIS, Himalayan Geology, 20(2), Holben, B and Justice, C An examination of spectral band ratioing to reduce the topographic effect on remotely sensed data. International Journal of Remote Sensing, 2(2), Janssen, L F, Jaarsma, J and Linder, E van der Integrating Topographic Data with Remote Sensing for Land-Cover Classification. Photogrammetric Engineering & Remote Sensing, 48(l), Jensen, J R Introductory Digital Image Processing: A Remote Sensing Perspective (New Jersey: Prentice-Hall). Jones, A R, Settle, J J and Wyatt, B K Use of Digital Terrain Data in the Interpretation of SPOT-1 HRV Multispectral Imagery. International Journal of Remote Sensing, 9(4), Kawata, Y, Ueno, S and Kusaka, T Radiometric correction for atmospheric and topographic effects on Landsat MSS images. International Journal of Remote Sensing, 9, Mather, P M Computer processing of remotely sensed images: an introduction, 2 nd edition, Wiley, Chichester. Michelson, D B, Liljeberg, B M and Pilesjo, P Comparison of algorithms for classifying Swedish landcover using landsat TM and ERS-1 SAR data. Remote Sensing of Environment, 71(l), Richards, J A and Jia, X Remote Sensing Digital Image Analysis: An Introduction. IIIrd Edition, Springer-Verlag, Heidelberg, Germany. Saha, A K, Gupta, R P and Arora, M K GIS-based Landslide Hazard Zonation in the Bhagirathi (Ganga) Valley, Himalayas. International Journal of Remote Sensing, 23(2), Shalan, M A, Arora, M K and Ghosh, S K An evaluation of fuzzy classifications from IRS I C LISS III data. International Journal of Remote Sensing, 24(15), Tso, B and Mather, P M Classification Methods for Remotely Sensed Data. Taylor and Francis, London, UK. Valdiya, K S Geology of Kumaun Lesser Himalaya (Dehra Dun: Wadia Institute of Himalayan Geology). Zhu, A X Measuring uncertainty in class assignment for natural resource maps under fuzzy logic. Photogrammetric Engineering and Remote Sensing, 63 (10),