Multi-scale features assisted knowledge distillation vision transformer for land cover segmentation and classification

Abstract

The most significant problem in remote sensing interpretation is semantic segmentation, which attempts to give each pixel in the image a particular class. This research work follows the various steps, such as pre-processing, segmentation, and classification. Initially, high spatial resolution remote sensing images (RSI) are collected from the open-source dataset. In the pre processing stage, an improved guided filter (Imp-GF) is used to remove various noises from images. Next, the segmentation is done by using a knowledge distillation-based vision transformer approach integrated with an atrous spatial multi-scale pyramidal module (KD-MuViTPy). Based on the segmented image, land cover classes such as vegetation, urban areas, forest, water bodies, and roads are classified. The proposed method outperformed the Bhuvan satellite dataset, achieving better accuracy, precision, recall, F1 score, Dice score, intersection over union (IoU), and Kappa score at values of 98.01%, 98.99%, 97.49%, 98.23%, 98.23%, 96.55%, and 95.91%, respectively.