Authors: Assistant Professor B.Naresh, D. Swasthitha
Abstract: Semantic segmentation of satellite imagery has become a fundamental task in remote sensing because it enables accurate identification of land-use patterns for applications such as urban planning, environmental monitoring, disaster management, and resource assessment. However, the complex spatial characteristics of satellite images, including varying resolutions, atmospheric interference, and class imbalance, make precise pixel-level classification a challenging problem. This study presents a modified U-Net-based deep learning framework for multi-class semantic segmentation of satellite images using the Dubai Satellite Imagery Dataset. The proposed model classifies six distinct land-use categories, namely buildings, roads, vegetation, water bodies, barren land, and miscellaneous regions. The preprocessing pipeline incorporates image patch generation, MinMax normalization, one-hot encoding, and extensive data augmentation to improve training efficiency and enhance model generalization. To address the imbalance among land-cover classes, the framework combines Dice Loss with Categorical Focal Loss, enabling improved segmentation performance for both majority and minority classes. Model optimization is performed using the Adam optimizer, while segmentation quality is evaluated through Accuracy, Intersection over Union (IoU), and Jaccard Index. Experimental results demonstrate that the proposed approach achieves reliable segmentation performance across diverse satellite scenes, making it suitable for practical remote sensing applications. The study further discusses strategies for handling spatial resolution challenges and improving segmentation robustness in complex urban environments.
