AUTOMATED LAND COVER MAPPING FROM LISS-III MULTISPECTRAL IMAGES USING SUPERVISED CLASSIFICATION WITH DEEP LEARNING

Remote sensing (RS) is the technique of finding and understanding information from a long distance or remote location using sensors. Land use and land cover mapping are fundamental tasks for planning and management. The deep neural network was used to perform the study. This study shows the semantic segmentation of LISS-III multispectral image using a fully convolutional network (FCN): U-net and Tiramisu. We present an innovative dataset, based on these LISS-III images that contained 4 different spectral bands (Band – 2 (Blue), Band-3 (Green), Band-4(Red), and Band-5 (Nearly Infrared), the false color composite (FCC) images and the ground truth mask images to classify total 4 classes (Surface water , Flora , Idle Land, and Housing zones) which holds 1470 labeled images where 1,255 images are used to train the model while 215 images are reserved for validation and evaluation. An FCN-based U-Net model with skip connections was trained on 256 × 256 × 3 input images to produce 256 × 256 × 4 one-hot encoded segmentation masks, representing four distinct land cover classes. Results confirmed that the U-Net architecture performs exceptionally well in semantic segmentation of LISS-III multispectral imagery for land use/land cover classification. The U-Net model successfully classified four land cover classes with an overall accuracy of 84%, significantly outperforming the Tiramisu architecture, which achieved a comparatively moderate accuracy of 52%.