Mapping Deprived Urban Areas with an Optimized Loss-weight Feature-guided Deep Learning Model

The growth of deprived urban areas (DUA), often associated with slums or informal settlements, is one of the consequences of rapid urbanization. Earth Observation (EO) data provides valuable information for mapping and monitoring such urban areas to assess the Sustainable Development Goal (SDG) indicator 11.1.1. Previous studies show that building density is one of the most informative morphometric variables to map DUA. However, building density when available are often mono-temporal and lack information about the exact date it was assessed. To address this gap, we present a deep learning-based approach that integrates building density regression from EO data to guide the learning process of a pixel-wise classification network. Our methodology optimizes the combined loss function of a dual-output semantic segmentation model, balancing classification and regression tasks, using Sentinel-1 and Sentinel-2 as input. This balance improves the accuracy of building density predictions, which, in turn, enhances the detection of DUAs and model interpretability. We evaluated our approach in Salvador (Brazil) and Nairobi (Kenya), achieving improvements of 9.75% and 0.60%, respectively, compared to previous studies.