Abstract
Accurate segmentation of land cover using satellite imagery is essential for efficient Earth resource management. This work introduces a refined U-Net a type of CNN (Convolutional neural network) designed specifically for the DeepGlobe a Land Cover Classification Challenge dataset. The derived dataset consists of high-resolution images showcasing diverse landscapes, including urban zones, agricultural regions, forests, and bodies of water. To tackle the issue of class imbalance, a focal loss function is applied during training, giving more weight to difficult-to-classify regions and minimizing the influence of easily recognizable features. Furthermore, temperature scaling is utilized to enhance model confidence calibration. The experimental findings reveal substantial improvements in both segmentation accuracy and the overall quality of land cover mapping, particularly for minority classes. This approach offers a reliable solution for advancing land cover categorization in remote sensing applications.