IRS LISS – III multispectral data semantic segmentation using deep

Abstract

Identification of land use/land cover and measurement of area under cultivation for various classes is a very significant job. Indeed, the methods and tools used for land use and land cover analysis often require significant human effort, consume substantial time, and can be very expensive. emote Sensing (RS) plays a pivotal in addressing these challenges efficiently and effectively. Remote sensing provides significant and distinguished data. The current investigation performs semantic segmentation of multispectral remote sensing images using a fully convolutional network (FCN). A multispectral image captures image data within specific wavelength ranges across the electromagnetic spectrum (EM). It has more than 100 nm resolution and less the 10 bands. Semantic Segmentation aims at a pixel-level classification of RS images where every individual pixel is categorized into a distinct class. A total of 4 classes were successfully identified in the present study. Identified classes are water bodies, uncultivated land, residential area and vegetation. Deep learning algorithms U-Net and DeepLabv3+ were used to perform classification on 2 datasets of different sizes and seasons (Dataset – 1:1470 images, Dataset – 2:13500 images). U-Net outperforms DeeplabV3+ in terms of performance. U-net achieved an accuracy of 81 % for dataset - 1, and 84% for dataset - 2 respectively whereas Deeplabv3+ achieved an accuracy of 36% for dataset -1, and 29% for dataset - 2.