Enhancing U-Net, SegFormer, Swin U-Net, and DeepLabV3+ for Automated Building Detection Using Biophysical Indices

Recent advances in deep learning have significantly enhanced the ability to automatically extract and map urban features from Sentinel-2 multispectral imagery with a spatial resolution of 10 m. Accurate and automated building detection from remote sensing imagery is essential for urban planning, environmental monitoring, and sustainable development. This study presents a comparative analysis of four deep learning architectures for automated building detection within a multi-class semantic segmentation framework. We evaluated U-Net, SegFormer, Swin-UNet, and DeepLabV3+ architectures using Sentinel-2 multispectral imagery of Bochum, Germany, acquired in August 2025. Indeed, ten biophysical indices were integrated into each architecture, providing complementary spectral and biophysical data to better disentangle buildings from spectrally similar land-cover classes. The models were trained to segment four land-cover classes: buildings, vegetation, water, and soil. The impact of the biophysical index was evaluated across two experimental, fully automated training pipeline configurations: i) standard models using spectral bands, and ii) enhanced models integrating ten biophysical indices. The selected indices are NDBI, IBI2, BAEI, MSAVI, NNIR, NDVI_RE, MNDWI, GVMI, BSI_INV, and MSI. Results showed that integrating biophysical indices consistently improved the performance of all investigated architectures. Weighted F1-scores increased from baseline values ranging from 0.39-0.61 to enhanced values of 0.56-0.69, while overall weighted IoU values improved from 0.26–0.44 to 0.40–0.53. For building class specifically, F1-score increased from 0.4237 for baseline U-Net to 0.5435 for U-Net enhanced with biophysical indices, while the building IoU increased from 0.2688 to 0.3670. These class-specific results provide direct evidence of the improvement achieved in building detection through the integration of biophysical indices. Overall, findings demonstrate that incorporating biophysical information into deep learning frameworks can enhance semantic segmentation performance and provide complementary information for automated building detection.