A causal deep learning framework for simulating the thermal impact of urban land use changes
Document Type
Article
Publication Date
2026
Department/School
Geography and Geology
Publication Title
Sustainable Cities and Society
Abstract
As climate change and urbanization intensify urban heat, planners lack scalable tools to predict the thermal outcomes of specific landscape interventions. This research presents a causal deep learning framework to simulate the counterfactual thermal impacts of urban land-use changes. Using Detroit, Michigan, as a case study, a U-Net Convolutional Neural Network (CNN) was developed to estimate Land Surface Temperature (LST) by integrating 2D spatial proxies, including multi-spectral indices, albedo, and building density. Model performance was validated through a series of “natural experiments” using a Difference-in-Differences (DiD) logic, comparing simulated impacts with real-world land-use transitions observed between 2019 and 2025. Results indicate high ex-post predictive fidelity (R2 of 0.95 ± 0.005, Mean Absolute Error (MAE) of 0.404 ± 0.030 K, Root Mean Square Error (RMSE) of 0.509 ± 0.025 K) and nearly 100% accuracy in predicting the direction of thermal change across all test cases. Physical interpretability analyses via Integrated Gradients confirmed the framework learned plausible, non-linear relationships, including vegetation cooling thresholds and urban canyon shading. To validate the model as a proactive planning tool, we developed an empirically calibrated Treatment Library. In 'blind' ex-ante simulations of nine historical interventions, the framework successfully predicted the direction of thermal change in 100% of cases with a Mean Absolute Error of 0.619 K, providing a highly accurate 'what-if' simulation engine for policymakers. While the model successfully isolates landscape-driven thermal signals, it exhibits a conservative bias, especially in industrial areas where anthropogenic heat is a dominant factor. Ultimately, this data-efficient framework offers a globally scalable solution for cities to proactively evaluate planning scenarios and advance climate equity.
Link to Published Version
Recommended Citation
Gijon Van Linden, L. G., & Xie, Y. (2026). A causal deep learning framework for simulating the thermal impact of urban land use changes. Sustainable Cities and Society, 148, 107657. https://doi.org/10.1016/j.scs.2026.107657
Comments
Y. Xie is a faculty member in EMU's Department of Geography and Geology.
*L. G. Gijon Van Linden is an EMU student.