Elimination Of Shadow Using Topographic Normalization Technique for Landcover Mapping Nizam-ud-Din The Urban Unit Gov.of Punjab, Lahore Pakistan E-mail:nizam_space@hotmail.com, Introduction Satellite imagery has become a powerful tool for landcover identification and area estimation with reasonably good accuracy {Jensen, 1996 }. It is, therefore, widely used for natural resource conservation and for many other applications such as mineral explorations, forest identification, wetland mapping, urban planning, environmental change detection etc. Satellite image acquisition, processing and interpretation have become an important practice in scientific society. When satellite sensor captures a view of the earth surface, some shadow also appears in the image along with different landcover types, which is due to the sun’s angle and topography of the terrain. The shadow is more visible in the imagery of mountainous regions. The extent of shadow over a particular terrain depends on season and time of the day when the image is captured. Shadow hides some landcover types on the image and can be problematic in obtaining appropriate measurements of the landcover types. If shadow is removed / reduced from the image the objects underneath the shadow can be identified with relatively more precision and better results of area estimation can be obtained. The study area for this research is Palas valley. This is one of the inner valleys of northern Himalayas that have been isolated from the rest the world due to its socio-cultural reasons, its difficult terrain and tribal culture. It lies in the transition zone between regions of Monsoon and Mediterranean climate {Usman, 2004}. The terrain of Palas valley is highly mountainous which causes large shadows in the image and this research is focused on “Elimination of shadow using Topographic Normalization Technique for landcover mapping”. Topographic normalization is a technique, first proposed by Tiellet, used to reduce the shadow in the image {Jensen, 1996}. the image. This makes the classification more effective for various landcover classes {Sheraz, 2002}. In this research Landsat enhanced thematic mapper plus (ETM+) sensor data with spatial resolution 30m are used. The area of interest (boundary of Palas valley) is obtained by creating a subset of the data in ERDAS Imagine 8.7 and classification of this data is achieved using supervised classification method. For supervised classification, training stage is very important and for this study training sets were mostly based on the identification of spectral signatures for different landcover classes. Spectral signatures for different landcover types were obtained after detailed discussions about tonal and textural pattern, brightness values and aspect variations. Some ancillary data such as photographs were used to have an idea of the slope/aspect variations of landcover types. The results of classification identified the different landcover types showing shadow as a separate landcover class. After this classification the topographic normalization technique was applied to remove the shadow effect in the original image. For this technique following inputs were used.
Finally unsupervised classification of the normalized data was performed and classes of the classified data were recoded to get final output. Objectives General Objective Generally objective of this study was to compliment the results of spectral classification of Palas valley by applying the topographic normalization process. This enabled us to make more accurate estimation of the area covered by different landcover classes. This can be helpful for image analysis and mapping of different landcovers. Specific Objective
Study Area: Palas valley is located between 34o 52’N to 35o16’N Latitudes and 72o52’E to 73o35’E longitudes at the left bank of the river Indus falling under the jurisdiction of tehsil Palas and Kohistan forest division. It is bounded on the north and northeast by Jalakot valley ,on the east by Kaghan ,on the south by Allai and on the west by river Indus.  Figure 1: Drain pattern of Palas valley  Figure 2:Palas valley location Valley covers an area of approximately 1300 km2 and is sub-watershed of the river Indus. The area of Palas valley is drained by two main nullahs; Musha’ga and Sherakot which flow into the river Indus near Keyal and Patten respectively. The entire valley is a series of rugged mountains with elevations ranging from 600 to 5151 meters above mean sea level. The topography of the area is rough with bare out-cropes and deserted precipitous slopes breaking the continuity of the forests {Usman, 2004}. The thematic map of palas valley is shown in figure 3  Figure 3:Thematic map of palas valley Materials: Satellite data: In this research Landsat-7 ETM+ data used with spatial resolution of 30m. The Landsat data preferred because of its easy availability and was free of cost. The image was already rectified with datum and spheroid WGS-84 and for UTM zone 43. The image was captured on 7th October 2001.  Figure 4:Satellite image map with palas valley in red polygon Digital Elevation Model: A digital elevation model (DEM) is a digital file consisting of terrain elevations for ground positions at regularly spaced horizontal intervals. For this study DEM was obtained form the courtesy of GIS lab, WWF Lahore, Pakistan. It was extracted by digitizing contour maps having 40m contour interval. The brightest tone in the figure 5 is representing the highest elevation and the darkest tone is representing the deepest value. The continuous variation in the brightness value is indicating height variations of the Palas valley.  Figure 5: DEM of Palas valley. (Courtesy, GIS LAB WWF Pakistan) Ground Truthing: Before going through the process of classification, photographs taken from helicopter and from the ground were analyzed for having a good understanding of forest and various other landcovers in the study area, specially the forest cover (i.e. classes of conifer and broadleaf forests). This pictorial data helped to analyze the forest landcover association with slope and aspect variations such as conifers and broadleaves. It also provided confidence in taking decisions about deceptive classes (such as mixed forests). Following are some of the photographs used for ground truth analysis. Methodology: The flow chart describes the complete process of this study as follows;  Figure 6:Work flow diagram Originally the image was downloaded from the link www.glcf.umiacs.umd.edu/portal/geocover/ in “.Tiff” format and was transformed into “.img " format for further processing in ERDAS Imagine 8.7. For this purpose image was imported into “.img” format. Truncation of the Study Area: Landsat image has a span of 185×185 km and in this study; area of interest was only Palas valley therefore a subset of the image was developed by using shape file of the study area. This subset of Palas valley was developed in ERDAS Imagine 8.7 using subset utility of the software. Band combinations: To identify different landcovers different false color combinations were used. For example, 542 as RGB were used for better identification of vegetation and soil. Similarly 742 as RGB was used for identification of snow, moraine and water. Other band combinations such as 321, 541, as RGB were also analyzed to make accurate decisions about different landcover variations. Supervised Classification: Supervised classification was performed in order to compare the results of classification after topographic normalization. Following processes have been performed for this purpose. Training Stage The analyst defines on the image a small area, called a training site, which is representative of each terrain category, or class {Floyed, 1999}. Training areas for this study were defined by the signature identification, using feature space and validation by the ground truth photographs. Also the primitive knowledge of photographs helped us in defining the classes. Eight major classes were identified with shadow as a separate landcover class. For major landcover classes, 223 representative signatures were identified. While assigning a class to a particular signature, extra care was taken in order to avoid mixing of the classes. Classification Stage: For this study supervised classification was performed after identifying 223 spectral signatures from the study area. For classification process Maximum Likelihood algorithm was used. Output Stage: After classification, recoding of 223 representative signatures into 8 distinct landcover classes was performed. Shadow was identified as a separate landcover class so that its extent and area occupied by it could be measured. Table 1 is representing number of representative signatures for each major landcover class. These representative classes were recoded as a major landcover class.
Table 1: No. of representative signatures for each landcover type. Topographic Normalization of the data: For normalization process both image data and DEM of the study area should be in same projection system {Jensen, 1996}. For normalization, solar azimuth and solar elevation angles were used that were measured at the time of image acquisition. Solar azimuth and solar elevation angle can easily be obtained from the meta data file of the image. After topographic normalization the original brightness values of the pixels were changed and most of the pixels which were representing shadow. In original non-normalized image, became brighter in tone. However the pixels representing sheer shadow changed very little in brightness and they remained dark. Unsupervised Classification of Topographically Normalized data: After applying normalization process unsupervised classification was performed with different number of classes to check its results. It was concluded after analyzing different results that better results can be obtained with large numbers of classes. After different experiments unsupervised classification with 250 classes was performed and then recoded in 9 major landcover classes. Previously landcover was identified into 8 landcover classes, where as an additional class as confusion class is also identified using unsupervised classification of normalized data because no certain decision could be taken for this class to be identified into any other landcover class. Results of Supervised Classification Supervised classification of the study area produced major landcover classes like conifer, broadleaf, water, snow, moraines, soil, shrubs/grasses and shadow. These are the major landcover types in the study area as shown in figure 7. Left window in the figure 7 shows attributes of supervised classification of the study area along with class name, area (in hectares) and specific color. The window at right is representing the classified thematic image with different colors assigned to each class.  Figure 7: Results of supervised classification. The map after supervised classification is given as above in quantitative measures. Confusion class was occurred due the similar brightness values over different landcovers. Classified image, representing the different landcovers in different colors is shown in right hand window.  Figure 8: Thematic map after supervised classification. Landcovers obtained by supervised classification are shown in the following bar graph. This graph represents the quantitative measure of different landcover types.  Figure: 9: Bar graph of different landcovers. Results of Classification after Topographic Normalization Un-supervised classification of the normalized data also produced major landcover classes like supervised classification such as conifers, broad leaf, soil, water, snow, moraines, shrubs, shadow and a confusion class which is shown in left-hand. The results obtained after normalization of the data show that the area occupied by shadow is decreased significantly. As we can see in the figure 10 that the area occupied by shadow is 3211(approx) hectors.  Figure 10: Results of unsupervised classification of topographically normalized data. The bar graph of the results obtained after normalization of the data is shown below. The graph is representing different landcover types along with their area. As indicated by graphs of the results, extent of shadow is decreased significantly from 10090 hectares to 3211 hectares.  Figure 11: Bar graph of different landcovers after normalization The map after topographic normalization is;  Figure 12: Thematic map of palas valley after topographic normalization. Comparisons and Conclusions Results of classification obtained before and after topographic normalization of the data are shown in the graph below. Black buildings are representing landcovers obtained before topographic normalization of the data while white buildings are representing the landcover types after normalization of the data.  Figure 13:Graph between results before and after normalization By the analysis of results we can see that before normalization of the data shadow was a significantly large landcover class as shown in the figure 13.This shadow was occupying different landcover types and restricted us to make a decision about the landcover types underneath it. In the figure 14 pie chart can better represent the shadow before and after normalization and the amount of shadow eliminated. As shown in the graph, 68% of the total shadow has been eliminated and landcovers under the shadow have been assigned to their respective classes.  Figure 14: Shadow percentage graph From the bar graph it can be seen that landcovers are improved quantitatively showing an increase in the area. This class as well shows that the area under the shadow has been classified. Figure 15 is representing the percentage increase in areas of different landcover classes after applying topographic normalization. As shown in the graph, areas occupied by broadleaf, snow, shrubs/grass and moraine increased by 12.65%, 1.697%, 0.818% and 0.715% respectively. This increase in extents of landcovers shows that after topographic normalization shadow is classified as different landcover classes.  Figure 15: Graph is representing the percentage increase in areas of different landcover classes. During the recoding process of topographically normalized data a confusion class was formed. It is the landcover about which no decision could be taken with confidence. The reason was that, some signatures obtained after topographic normalization were quite confusing and their pattern was not matching with any single landcover class. Further Analysis of the landcover type in the confusion class with reference data showed that most of the landcover could be conifers, soil or snow. Suggestions For further improvement in results we can make certain measures such as given below:
Jensen, J. R. (1996). Introductory Digital Image Processing: NewJersey: Prentice-Hall. Sheraz, S. (2002). Vegetation mapping of Margalla Hills National Park Islamabad. Department of Space Sciences. Lahore, University of the Punjab. Usman, A. (2004). Study of Habitat Zones in Moist Temperate Forest Using Satellite Data of Palas Valley. Department of Space Sciences. Lahore, University of the Punjab. Lillesand, T. M., and R. W. Kiefer (1987). Remote Sensing and Image Interpretation. New York, John Wiley & Sons, Inc Floyed, F. S. (1999). Remote sensing principles and interpretation. New York. Hammad, G. (2006). Extraction of digital elevation model from stereo satellite data. Department of Space Sciences. Lahore, University of the Punjab. Chang, K. T. Introduction to Geographic Information System, International Addition. Congalton, R. (1991). A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data. Floyed, F. S. (1999). Remote sensing principles and interpretation. New York. Huurneman, L. L. F. J. a. G. C. (2001). Principles of Remote Sensing. The Netherlands, ITC Educational Textbook Series. Kloer, B. R. (1994). Hybrid Parametric/Non-Parametric Image Classification. ACSM-ASPRS. annual convention Riano, D., Chuvieco E., Salas J., and Aguado I (2003). Assessment of different Topographic Corrections in Landsat-TM Data for Mapping Vegetation Types. Schowengerdt, R. A. (1980). Reconstruction of Multispatial, Multispectral Image Data Using Spatial Frequency Content. zafeer Saqib (2006). Modelling potential Distribution of Taxus Wallichiana in Palas Valley. M.Dare, P. (2005). Shadow analysis in High-Resolution Satellite Imagery of Urban Areas. |