Optimized Land Cover Classification: A Lightweight Approach to High-Resolution Remote Sensing

Land cover classification is a critical challenge in remote sensing, essential for monitoring physical and anthropogenic phenomena on a large scale. While extensive research has focused on multi-class classification, where each pixel is assigned to a single class, these traditional approaches need help with the complexities of high-resolution remote sensing images in multilabel classification tasks. Convolutional neural networks (CNNs) based on deep learning have shown promising results, but their many trainable parameters often lead to overfitting. To address this issue, we propose a lightweight convolutional neural network specifically designed to reduce model complexity. Our experiments, conducted on the widely recognized UC Merced Land Use dataset, demonstrate that the proposed model achieves higher precision, recall, and overall accuracy while reducing the number of trainable parameters by more than 39% compared to existing models. This reduction mitigates overfitting and significantly decreases training and testing time, highlighting the efficiency and effectiveness of our approach for land cover classification in remote sensing.