Biophysical Indices Impact on Cognitive Building Detection

Accurate detection of buildings from satellite imagery presents a significant challenge in complex urban environments where traditional spectral and geometric approaches often fail to distinguish built structures from surrounding vegetation and natural features. This research introduces an enhanced U-Net deep learning architecture that integrates multispectral satellite data with biophysical indices to improve building detection performance. We have developed and validated our approach using 12 high-resolution multispectral images collected from Sentinel-2 satellite imagery between January and December 2024 over Nador-Morocco and Melilla-Spain. This area was selected due to its critical status as a Mediterranean biodiversity hotspot, and to its diverse landscape composition, including dense urban areas and varied vegetation cover, providing an ideal urban-rural interface for evaluating fully automated building detection algorithms under varying environmental conditions. The methodology incorporates 10 biophysical indices (i.e., NDVI, NNIR, EVI, GEMI, SVI, MSAVI, NDBI, BU, UI, IBI2) as additional input channels to the standard U-Net architecture. The enhanced U-Net model achieved superior performance with approximately 5% higher accuracy compared to the standard U-Net architecture. The inclusion of 10 biophysical indices enhanced the model's ability to differentiate between built-up areas and dense vegetation, addressing key challenges in building detection from satellite imagery.