Land Cover Segmentation from Satellite Imagery Using U-Net with Custom Loss and Morphological Postprocessing
Land cover classification is a fundamental task in remote sensing with direct applications in environmental monitoring, urban planning, and disaster management. In this project, we design and implement a semantic segmentation pipeline using a modified U-Net architecture from scratch for pixel-level classification of satellite images. The network was trained on the DeepGlobe Land Cover Classification dataset with significant attention to model generalization and class imbalance through the integration of a custom composite loss function that combines Focal Tversky Loss and Weighted Categorical Crossentropy. Data augmentation techniques and per-class weighting were used to address class imbalance, especially in underrepresented classes such as barren land and rangeland. Postprocessing techniques like morphological operations and class-wise median filtering further refined the segmentation outputs. The model achieved a mean Intersection-over-Union (IoU) of 0.67, with notable performance across dominant land types. A full Streamlit-based web interface was developed to enable real-time visualization and interaction with the model’s predictions. Additionally, we initiated a fine-tuning phase targeting Nepal- specific satellite imagery to adapt the model to the country's unique geography. A major ongoing challenge involves curating a clean and labeled dataset of Nepalese satellite patches, which is critical for domain adaptation and achieving optimal segmentation performance on local data. The project demonstrates a full end-to-end pipeline from model design to web deployment and sets a foundation for scalable geospatial AI applications tailored for Nepal.