Efficiency analysis of attention modules and combined loss functions for urban road segmentation

The main objective of this study is enhancing the accuracy and robustness of road detection from high-resolution remote sensing imagery by developing improved UNet3+ encoder–decoder architecture. While existing UNet3+ have demonstrated promising results, their effectiveness remains limited in complex urban environments. To address these limitations, this research proposes a modified UNet3+ model that simultaneously integrates pyramid pooling, spatial attention, and channel attention modules, along with a hybrid loss function combining Binary Cross Entropy and Soft Dice losses. The model was evaluated on two benchmark datasets composed of DeepGlobe and Massachusetts under diverse environmental conditions. The results demonstrate that the proposed model achieves consistent improvements over baseline networks. In dense urban areas, the proposed model improved F1-score by 7.1% and IoU by 7.6% compared with the original UNet3+. In areas with soil and vegetation cover with less environmental interference indicated the improvements of 11.2 and 15.4 from the proposed model.