Biophysical Index-Driven Automated U-Net Training for Building Extraction Across Diverse Regions

Automated building extraction from high-resolution remote sensing imagery has advanced substantially with the integration of deep learning approaches. Accurate and efficient building detection from satellite imagery is essential for environmental monitoring, urban planning, and geospatial analysis. This study presents a comparative performance analysis of automated building detection across four regions with different urban structures and climatic conditions. These regions were selected to evaluate the robustness and generalization of the proposed approach under diverse urban and environmental contexts. Specifically, we integrated ten biophysical indices into the U-Net architecture to assess multi-class land cover classification, with a particular focus on fully automated building extraction. The selected indices include three building indices (i.e., NDBI, IBI2, and BAEI), three vegetation indices (i.e., MSAVI, NNIR, and NDVI_RE), two water indices (i.e., MNDWI and GVMI), and two soil indices (i.e., BSI_INV and MSI). The model was trained to segment four land-cover classes: buildings, vegetation, water, and soil. Sentinel-2 30% cloud-free multispectral imagery was used for Bochum, Germany; Granada, Spain; Adelaide, Australia; and Nador, Morocco/Melilla, Spain, acquired respectively on August 11, 2025; August 20, 2025; September 14, 2025; and October 23, 2025. We implemented two experimental configurations with fully automated training pipelines: i) standard U-Net using spectral bands, and ii) enhanced U-Net integrating ten biophysical indices. Model predictions were validated against OpenStreetMap (OSM) building data, and performance was evaluated using F1-score and intersection over union (IoU). The comparative results across the four study regions demonstrate that incorporating biophysical indices consistently improved the performance of U-Net architecture. Weighted F1-score increased from baseline values ranging from 0.10 to 0.48 to enhanced values of 0.31–0.57, while the weighted IoU improved from 0.05–0.32 to 0.18–0.40. Across the four regions, the enhanced model achieved a mean weighted F1-score of 0.429 ± 0.116 compared to 0.343 ± 0.172 for the base model, while for weighted IoU, the enhanced model achieved a mean of 0.281 ± 0.097 versus 0.218 ± 0.119 for the base model. These consistent improvements across both evaluation metrics demonstrate that integrating biophysical indices into the standard U-Net model yields the best overall performance, highlighting the value of spectral-index information for fully automated building detection.