Biophysical Segmentation Neural Model: An Automatic Algorithm for Urban Environment Classification. Case of Study for the Metropolitan Zone of the Valley of Mexico

Urban environments exhibit complex land cover spatial distribution and interactions driven by rapid anthropogenic expansion, requiring robust and physically interpretable classification methodologies. This study presents a biophysically driven neural segmentation model for automated multi-class land cover classification in urban environments, integrating principles of urban remote sensing, spectral decomposition, and deep learning. The proposed framework is sustained in the Vegetation–Impervious–Soil (VIS) conceptual model and the Biophysical Composition Index (BCI), providing a biophysically basis for semantic segmentation beyond purely data-driven approaches. Multispectral imagery from the Landsat 9 OLI-TIRS sensor was employed to characterize the Metropolitan Zone of the Valley of Mexico (MZVM), a highly heterogeneous and densely urbanized region. Reflectance data were transformed through a Gram–Schmidt orthogonalization procedure to derive the tasseled cap components of brightness, greenness, and wetness, which capture the fundamental radiative responses of urban surfaces, vegetation, soil, and water. These biophysical features were used both to generate supervised label masks via BCI-based reclassification and to augment the neural network input space. The segmentation model is implemented as a 2D convolutional encoder–decoder architecture, designed to balance architectural depth and computational efficiency under limited hardware resources. Two experimental configurations were evaluated: a baseline model using only NIR–Red–Green reflectance bands and an augmented model incorporating the biophysical brightness–greenness–wetness components. The dataset was strategically subsampled to mitigate class imbalance, and model performance was assessed using accuracy, precision, recall, F1-score, and Matthew’s Correlation Coefficient (MCC), the latter provides a robust evaluation for imbalanced, pixel-wise classification tasks. Results demonstrate that biophysical feature augmentation consistently improves segmentation performance, yielding higher predictive performance and superior evaluation metrics across most land cover classes. The optimal configuration achieved reliable discrimination among seven urban and natural surface classes, including impervious materials, vegetation states, soils, and water bodies. Model generalization was further validated using an independent urban dataset from Guadalajara, Mexico, confirming the scalability of the approach. Overall, this work highlights the value of embedding physically interpretable biophysical parameters within deep learning architectures, offering a robust and transferable framework for urban land cover mapping, environmental monitoring, and sustainable urban planning applications.