ABSTRACT
India today, as in most developing countries of the world, shows a rapid increase in urbanization. As there is an increased demand for land for urban expansion to cater to the migrant service population, there is also an increased demand for agricultural land for meeting the food requirements. All of this is leading to a drastic change in the land use/land cover pattern, as well as posing impacts for nature and resources more broadly. Due to their dynamic nature, urban areas can be difficult to map easily, but remote sensing is an effective tool capable of solving this problem. However, optical remote sensing has its constraints, like paucity of all-weather data; in addition, urban and fallow land areas often tend to merge since they both produce a cyan tone in a standard false colour composite (FCC) optical image. This makes their delineation and classification difficult. This study aimed to consider and utilize the different polarimetric decomposition models and to examine the backscatter based on the man-made and natural features detected in each. The multifrequency co-registered slant-range single look complex (co-SSC) of the X-band PolSAR dataset from TerraSAR-X was analysed, and various techniques, to fuse optical and PolSAR images, were compared. Out of these, the HSV (hue saturation value) fusion technique gave the best results in classification. It was found that PolSAR, due to its various modelling approaches and polarimetric combinations, has an advantage over both SAR and optical datasets and is found to minimize the ambiguity of merging urban built-up with fallow land features in the optical dataset. In this study, both quad and dual PolSAR datasets of TerraSAR-X were taken. The datasets were corrected for interior orientation angle and the speckle was removed using the refined Lee filter. Thereafter, coherent (Pauli’s and Sinclair’s decomposition) and incoherent decompositions (H-A-α, Yamaguchi, Freeman–Durden, and Van-Zyl) were applied on quad PolSAR datasets. The dual PolSAR datasets were band rationed to generate a suitable RGB for classification. Suitable decomposition and band-rationed products were chosen based on backscatter separability for fusion with optical LISS-IV data. Based on the high spectral separability of the HSV-fused product, it was chosen for further classification using pixel- and object-based image classification. The different machine learning classification results were compared based on the relative classification accuracies achieved.
