ABSTRACT
88Agricultural crop classification is obligatory to realize the physical and climatic requirements of diverse crops. Resourcesat -2 satellite is a key data source for the regional to global cropland other land feature classification studies. Indian Remote Sensing satellite sensor, linear imaging self-scanning (LISS-IV) data acquired on April 6, 2013, in Rabi season-derived information, was assessed for the diverse crop and other land features classification in Varanasi district, Uttar Pradesh, India. The most popular algorithms—support vector machines (SVMs), artificial neural network (ANN) and random forest (RF)—were analyzed for the land features classification. The spectral separability analysis was completed before the classification to check the separation between 16 different classes. The separability analysis using the transformed divergence (TD) method has shown high separation between the classes compared with the Jefferies Matusita (JM) distance method. The classification results derived from the kernel-based SVMs were found better than the RF and ANN algorithms except for SVMs with sigmoid kernel. The classification accuracy results were enhanced after postprocessing using different filters by pixel window size of 3 × 3 and compared. The performance analysis was also done on the basis of marginal rates, F-measure, Jaccard’s coefficient of community (JCC) and classification success index (CSI).The statistical significance in the classification accuracy results between two algorithms was analyzed using Z-test.
