Physics-Guided Spatiotemporal Learning for Urban Landscape Thermal Comfort Prediction and Green Infrastructure Optimization

Abstract

Urban landscape design significantly influences outdoor thermal comfort, particularly in densely developed areas experiencing increasing heat exposure and climate variability. Nevertheless, accurately predicting microclimatic conditions remains challenging due to complex interactions among vegetation coverage, surface materials, urban geometry, and meteorological dynamics. This study proposes a physics-guided spatiotemporal learning framework for predicting outdoor thermal comfort and optimizing urban green infrastructure configurations. The proposed framework combines graph-based spatial representations, temporal attention mechanisms, and physically informed constraints to model interactions among urban vegetation, land surface temperature, solar radiation, humidity, and wind conditions. A multi-objective optimization module is introduced to evaluate alternative landscape design scenarios, including tree canopy expansion, vegetation redistribution, and permeable surface allocation. The framework is designed for evaluation using approximately 25,000 hourly environmental observations collected from urban meteorological stations, satellite-derived surface temperature products, and geospatial vegetation datasets. Comparative experiments will include Random Forest, XGBoost, LSTM, and conventional spatiotemporal neural networks. Performance will be assessed using RMSE, MAE, thermal comfort classification accuracy, and simulated temperature reduction under alternative landscape configurations. The proposed framework aims to provide a scalable decision-support tool for climate-responsive landscape architecture and sustainable urban green infrastructure planning.