Remote Sensing Based Land Cover Classification Using Machine Learning and Deep Learning: A Comprehensive Survey

Abstract

Since the 1990s, remote sensing images have been used for land cover classification combined with MachineLearning algorithms. The traditional land surveying method only works well in places that are hard to get to, likehigh mountain regions, arid and semi-arid land, and densely forested areas. As the satellites and airborne sensorspass over a specific point of land surface periodically, it is possible to assess the change in land cover over a longtime. With the advent of ML methods, automated land cover classification has been at the center of researchfor the last few decades. From 2015 forward, a technical shift has been noticed with the emergence of severalbranches of Neural Networks (NN) and Deep Learning (DL). This paper examines current practices, problems,and trends in satellite image processing. This formal review focused on the summarization of major classificationapproaches from 1995. Two dominant research trends have been noticed in automated land cover classification,e.g., per pixel and subpixel analysis. Classical machine learning algorithms and deep learning methods are mainlyused for per-pixel analysis, whereas Fuzzy algorithms are used for sub-pixel analysis. The current article includesthe research gap in automated land cover classification to provide comprehensive guidance for subsequent researchdirection.