Domain Adaptation for Performance Enhancement of Deep Learning Models for Remote Sensing Scenes Classification

Abstract

Deep neural networks and in particular Convolutional Neural Networks (CNNs) are considered to be the state-of-the-art tools for scene classification. However, training deep CNN models requires huge amounts of labeled data to achieve excellent classification accuracy. Thus, an important goal in deep learning is how to reduce the data labelling burden. Domain Adaptation (DA) is the main technique in this regard. The goal is to classify the target domain correctly by learning from the source domain. This chapter examines the basic concepts required to understand RS. Then, it proceeds to describe in detail a method for multi-source semi-supervised domain adaptation in remote sensing scene classification called Semi-Supervised Domain Adaptation Network (SSDAN). Performance results in terms of overall accuracy and Kappa coefficient values obtained when conducting experiments using single-source, two-source, and three-source scenarios are also provided. The achieved results of these two metrics reached values of more than 99%, demonstrating the efficacy of the SSDAN method.