Satellite Image Segmentation using Deep Learning

This paper offers a deep learning approach for semantic satellite imagery segmentation utilizing a modified U-Net architecture. Leveraging a Dubai satellite imagery dataset, the model classifies six distinct land-use categories, including buildings, roads, vegetation, and water bodies. Preprocessing methods including patch creation, normalization, and class balance are incorporated into the implementation through augmenting data. Combining categorical focal loss and dice loss, which addresses class imbalance and improves segmentation accuracy, is one of the major advances. The study demonstrates the model's robustness in different fields of environmental monitoring and urban planning applications. Insights into spatial resolution challenges and model optimization strategies are also discussed.