Automatic Extraction of Built-up from SAR Imagery

Urban landscape is dynamic in nature. Extent of urban area changes frequently, therefore it is necessary to monitor the urban land cover maps regularly. Remote sensing technique is an efficient tool for urban analysis. Synthetic Aperture Radar has capability to acquire data independent of solar illumination and weather conditions. Independent imaging quality of synthetic aperture radar gives periodic images of earth without any interruption. The present study analyses the potential of SAR imagery in built up area extraction. Advantage of combining two different aspect (look direction) images has been studied. TerraSAR-X images in Strip map mode with HH polarization acquired on of 27th may 2010 as ascending right look and 30th August 2014 as descending right look have been used. Study area comprises Jodhpur city and surrounding area with spatial extent of latitude 26 degrees 10' 15 '' N to 26 degrees 19' 33 '' N and longitude 72 degrees 51' 06 '' E to 73 degrees 5' 10 '' E. Images have been multi looked by 5*4 looks with 8.17042*9.82941 meters resolution in range*azimuth direction respectively to reduce speckles and salt and pepper noise. To make perfect square pixels, resampling of 1010 meters in range*azimuth direction has been done. sigma degrees Calibration and geocoding has been done to calibrate the data using orbital files and reference SRTM 90m Digital Elevation Model. Backscattering co-efficient variation has been studied for both the imageries by drawing samples for built-up. Backscattering variation in both images found to be highly variable for all samples. Built-up area has been extracted from individual image by using grey level thresholding. Some part of Hilly area and built-up got mix-up in grey level thresholding. To avoid such mix-up and to get overall built-up, layer stacking of ascending right look image of 27th may 2010 and descending right look image of 30th August 2014 have been done. Training samples for built-up area have been drawn. Supervised classification with Support Vector Machine classifier has been carried out. Radial Basis Function with 0.5 kernel with 0.2 probability threshold has been used to classify the built-up from stacked image. The classified output has been compared with output generated from Resourcesat-2 LISS-IV data and found a better matching. To validate the output, random samples were drawn on classified output. 912 out of 1000 samples were found to be of built-up class.