FE-KFormer: A keypoint-informed transformer model for quantifying the nonlinear cooling effects of urban green space

The nonlinear mechanisms underlying the cooling efficiency of urban green spaces (UGS) are poorly characterized, creating a critical need for predictive models to quantify these relationships. To bridge this gap, this paper develops the FE-KFormer, a novel Keypoint-informed Transformer model with feature enhancement, to accurately predict both current and future UGS cooling efficiency based on landscape metrics.. We generate a high-resolution thermal dataset by fusing MODIS and Landsat 8 land surface temperature products, which is then used to train FE-KFormer. The FE-KFormer model demonstrates markedly superior predictive performance compared to traditional machine learning models (e.g., Random Forest, LightGBM) and deep learning model (i.e., CNN), achieving an R² of approximately 0.80 for cooling density prediction and substantially lower mean absolute and root mean squared errors, with reductions of up to ∼85 %. Evaluation results indicate synergistic and antagonistic interactions among the cooling effects of different landscape metrics. The cooling effect of UGS shows pronounced seasonal variability, being stronger in the rainy season (April-September) than in the dry season (October-March). By varying the landscape metric features, prediction results from the FE-KFormer model reveal a positive correlation between area-related and connectivity-related metrics and UGS cooling effects, while shape-related and density-related metrics show a negative correlation. Cooling intensity is more sensitive to changes in the landscape matrices than cooling distance. The FE-KFormer model provides a decision-support tool for urban planners to evaluate UGS cooling performance and assess the potential impact of different planning strategies through the lens of landscape metrics.