Improved semantic segmentation for urban mapping with a combined loss function approach

Accurate semantic segmentation of buildings from aerial imagery is vital for urban planning but challenging due to factors like class imbalance and edge definition. This research introduces a novel-combined loss function integrating Binary Cross-Entropy (BCE), Tversky, and Edge-Weighted Hinge loss within a U-Net architecture. The primary innovation lies in leveraging this specific combination to achieve a synergistic balance across key segmentation metrics (Accuracy, Precision, Dice Coefficient, and Recall), addressing multiple challenges simultaneously. The approach was rigorously evaluated on three diverse datasets: a newly developed Isfahan dataset capturing unique Middle Eastern urban features, alongside the standard INRIA and Massachusetts Buildings datasets. Experimental results highlight the effectiveness of the combined loss, achieving 96.53% Accuracy and a 92.35% Dice Coefficient on the Isfahan dataset, while crucially demonstrating superior overall balance and robustness across all three datasets compared to individual or simpler combined losses. This work shows that strategically combining loss functions can significantly improve segmentation performance, providing a powerful approach for urban mapping and related remote sensing applications like urban planning and industrial inspection.