Smart Urban Design with Physics-Informed Neural Networks: Quantifying Temperature Reductions from Green Infrastructure Using Satellite Thermal Data

Dung Thi Vu, Rafeeque Ahmed Nizamani, Abdul Ghafoor Nizamani

Urban Heat Islands (UHIs), characterised by elevated temperatures in densely built environments, pose critical challenges to urban sustainability, public health, and energy resilience. Mitigating UHIs requires precise quantification of the cooling effects of green infrastructure; however, existing models often fail to integrate high-resolution geospatial data with physical laws. This study presents a novel Physics-Informed Neural Network (PINN) framework that addresses this gap by unifying satellite-derived data, including land surface temperature (LST), normalised difference vegetation index (NDVI), and emissivity, with vector-based urban features within a single predictive model. The framework innovatively embeds thermodynamic principles, such as heat diffusion equations, directly into the neural network's loss function, while a seamless GeoJSON (Geographic JavaScript Object Notation) to tensor format enables the integration of heterogeneous geospatial datasets. Applied to case studies in Bologna, Italy, and Washington, D.C., USA, the model accurately identified optimal locations for green infrastructure. A targeted intervention prioritising green roofs on buildings near railways in identified LST hotspots achieved a 99% reduction in the spatial extent of identified hotspots in Bologna and 98.7% in D.C., with an area-weighted average temperature decrease of 1.31°C and 1.30°C, and maximum localised cooling of 6.13°C and 8.48°C, respectively. A more comprehensive scenario incorporating green roofs near railways and new trees along main roads yielded even greater cooling, with hotspot reductions of 97.4% and 98.3%, and area-weighted average temperature drops of 1.51°C and 1.80°C for the two cities, respectively. Validation against Landsat 8 thermal imagery confirmed high predictive accuracy, with R-squared values of 0.926 and 0.937, demonstrating performance superior to conventional machine learning baselines. By bridging physics-constrained machine learning with geospatial analytics, this work provides a scalable, data-driven tool for smart urban design, offering actionable insights to combat UHIs through targeted, climate-adaptive planning.