Performance of Deep Learning in Land Use Land Cover Classification of Indian Remote Sensing (IRS) LISS – III Multispectral Data

Abstract

Identification of land use land cover is a very important task. However, methods existing for the above mention purpose are labor incentives, time-consuming, and costly. Remote sensing plays very important role in the mappings. classification of land cover features and offers very noteworthy and sensed information. The present study shows the semantic segmentation of Indian remote sensing (IRS) LISS-III multispectral image and the comparison of three algorithms U-Net, Deeplabv3+and Tiramisu. The deep neural network was used to perform the study. We present total 3 innovative datasets, built on these LISS-III images that has 4 different spectral bands (Band – 2 (Blue), Band-3 (Green), Band-4(Red), and Band-5 (Nearly Infrared), FCC (false color composite) images and the ground truth mask images. Dataset has 13500 labelled images. A fully-convolutional network (FCN) with skip connections is trained to take an input image of size 128 X 128 X 3 and outputs a matrix of shape 128 X 128 X 4 i.e., a one-hot encoded version of the mask. The experiment identifies 4 classes successfully (Water Bodies, Vegetation, Uncultivated Land, and Residential areas). The experiment showed that the U-Net algorithm has a very good capability for the classification of LISS -III images for land use land cover class detection then Tiramisu and Deeplabv3+. U-Net achieved accuracy 84%, Deelabv3+ achieved 29% whereas Tiramisu achieved accuracy 33%.