Self-Supervised Techniques for Satellite Imagery: A Novel Approach to Land Cover Classification

This study investigates the use of self-supervised learning (SSL) methods for land cover classification through satellite images, emphasizing clustering algorithms and logistic regression. The research uses an unsupervised method, applying K-Means, Mean Shift, and Gaussian Mixture Model (GMM) clustering techniques for feature extraction from unlabeled data, and then classifies via logistic regression. To assess the effectiveness of these techniques, the Silhouette Score was utilized, with Mean Shift clustering attaining the top score, signifying better cluster cohesion and separation. The findings emphasize the promise of self-supervised methods in addressing the issues posed by restricted labeled datasets for land cover classification, indicating that Mean Shift clustering is the most efficient technique for feature extraction. The results highlight the significance of choosing suitable clustering techniques to enhance classification precision in satellite image assessment, aiding the progress of self-supervised learning in remote sensing fields. This study introduces an innovative and efficient method for land cover classification that does not depend on large amounts of labeled data, showcasing considerable promise for extensive use in environmental monitoring and land-use planning.