Application of deep learning and remote sensing satellite data to assess glacier retreat for the past three decades in Himachal 

Abstract

Conventional glacier and glacier lake mapping change analysis using satellite data involves techniques based on band combinations and image processing. The results from computer processing of satellite data are sometimes affected by segmentation and thresholding techniques, due to which water bodies, shadows, rock surfaces, and other features are classified as glaciers and glacier lakes. Artificial intelligence (AI) based Deep learning (DL) algorithms for image classification, object detection, segmentation, and feature extraction have proven to be very efficient and accurate in recent years. Fully convolutional neural (FCN) networks-based U-Net architecture, which involves a gradual integration of superficial visual characteristics and semantic information extracted from images to segment small objects effectively, is used in the present study to extract and map glaciers and assess the change in glacier mass for past three decades, i.e. from 1994 to 2021. Deep Learning-based algorithms can automate the glacial mapping process with an identification accuracy of 95 percent and are less time-consuming. The glaciated area in Himachal Pradesh decreased at a rate of 67.84 km2 per annum from 4020.6 km2 in 1994 to 2198.5 km2 in 2021. It is observed that from 1994 to 2021, the glacier area decreased approximately by a percentage loss of 1.678 per annum, with decreasing decadal trend from 2.31% in 1994–2001 to 1.398 in 2011–2021.