Unmanned Aerial Systems (UAS) are expected to play a key role in the next techno-social transformation, reshaping logistics, aerial transport, and surveillance sectors. However, to ensure the safe and scalable integration of UAS into urban airspaces, effective Unmanned Traffic Management (UTM) systems are needed, particularly to address the risks associated with low-altitude flights over complex urban environments. This work proposes a methodology to estimate overflight risk based on the presence of physical, social, and infrastructural attributes. Using high-resolution satellite and geospatial data, attribute-based risk indices are defined and used to train Convolutional Neural Networks (CNNs) capable of generating continuous-valued heatmaps that highlight safer or riskier areas for UAS operations. The CNNs were trained on a dataset derived from the central region of São Paulo, Brazil, using U-Net-like architectures adapted for regression tasks. Experimental results show that the trained models could accurately predict risk levels across diverse urban scenes and demonstrated generalisation when applied to aerial imagery from other major cities worldwide. The findings support the potential of CNN-based approaches in enabling real-time, onboard risk assessment, and contributing to more efficient and adaptive UTM path planning in urban contexts.
Andrade et al. (Thu,) studied this question.