A self-supervised deep-driven model for automatic weather classification from remote sensing images

Abstract

Listen

Translate article

ABSTRACT The increasing use of Remote Sensing (RS) technology has resulted in the availability of large volumes of satellite imaging data, necessitating efficient, and scalable solutions for rapid analysis and classification in various interdisciplinary fields. This research focuses on addressing these challenges by examining different training and optimization techniques to achieve the highest classification accuracy on both labelled and unlabelled data from the Large-Scale Cloud Photos Dataset for Meteorology Research (LSCIDMR), which contains 10 high-resolution classes. A key aspect of this study is the implementation of a unique deep learning model based on the ResNet101 architecture, designed to handle the inter-class similarity problem through architectural improvements and imbalanced class distributions using data augmentation techniques. Furthermore, this model integrates a novel approach that combines self-learning with semi-supervised learning to process large amounts of unlabelled data. An iterative self-learning process is employed for continuous refinement, using pseudo-labels generated from the unlabelled data. To ensure higher quality and relevance, confidence-based selection of pseudo-labels is applied. Our model demonstrates superior performance in classification compared to recently reported deep learning algorithms, offering a robust solution for the automated categorization of satellite images. In this study, our ResNet101 model achieved a mean accuracy of 0.917 (±0.02), a mean precision of 0.917 (±0.01), with a mean F1 score of 0.9300 (±0.015) over five folds.