Urban growth-induced surface thermal dynamics: an integrated deep learning assessment for Colombo, Sri Lanka

Abstract

Abstract Urban growth is a major driver of land surface temperature (LST) change, yet most LST forecasting studies do not explicitly incorporate future urban expansion. To address this gap, we develop an integrated deep learning (DL) framework that couples urban growth simulation with LST prediction for Colombo, Sri Lanka, and benchmark our proposed ResNet+Transformer model against six comparison models (four DL architectures and two machine learning methods) across both tasks. For urban growth simulation, the ResNet+Transformer achieves the highest accuracy (0.908), F1 score (0.8691), Figure of Merit (0.7684), precision (0.9152), and recall (0.8274), while also yielding the lowest quantity and allocation disagreement. The simulated urban growth maps for 2035 and 2045 were then incorporated as dynamic drivers in LST prediction. In this second task, the ResNet+Transformer again outperformed all benchmark models, achieving the lowest RMSE (0.951 $$^\circ $$ C) and MAE (0.651 $$^\circ $$ C), and the highest $$R^2$$ (0.7412). The projections indicate a 2.22 $$^\circ $$ C increase in mean LST between 2025 and 2045, while surface urban heat island (SUHI) intensity rises from 4.73 $$^\circ $$ C in 2035 to 5.06 $$^\circ $$ C in 2045. Hotspot analysis reveals strong spatial clustering of warming ( $$r>0.90$$ ), and sensitivity analysis identifies urban growth as the dominant driver of future LST (2.357–2.396 $$^\circ $$ C), far exceeding other environmental factors. The proposed framework provides a robust tool for identifying future thermal risk and supporting climate-responsive urban planning.