Enhancing Urban Heat Island Analysis Through Multisensor Data Fusion and GRU-Based Deep Learning Approaches for Climate Modeling

Abstract

Rapid urbanization and land-use changes have exacerbated the urban heat island (UHI) effect, threatening urban sustainability and climate resilience. This study uses a novel gated recurrent unit (GRU)-based deep learning model in addition to the Mann–Kendall trend, Pearson correlation, and continuous wavelet to investigate the UHI phenomenon in Multan city of Pakistan. The approach utilizes the normalized difference vegetation index (NDVI) and normalized difference built-up index (NDBI) as essential variables to forecast UHI accurately using a GRU-based deep learning model using a monthly Landsat dataset from 2001 to 2023. The results from the Mann–Kendall test indicated a minor increase in monthly UHI values, accompanied by notable seasonal fluctuations with a substantial decrease in winter (Tau = −3.486), whereas a notable increase is observed in the summer season (Tau = 0.158). The NDVI exhibited a notable annual increase (Tau = 3.43), suggesting enhanced vegetation health. Conversely, NDBI showed a significant decrease (Tau = −0.907). The result of Pearson's correlation study showed that UHI is significantly negatively correlated with NDVI and positively with NDBI, with a correlation coefficient of −0.540 and 0.344, respectively. Wavelet coherence analysis revealed considerable seasonal and annual relationships between UHI, NDVI, and NDBI. The GRU-based model achieved a coefficient of determination (R²) of 0.90 with an RMSE value of 0.09, indicating robust predictive performance. The SHAP (SHapley Additive explanations) analysis revealed that NDVI is the predictor with the most significant influence. The adopted approach emphasizes vegetation's crucial function in reducing UHI's effects and offers valuable insights for urban planning and measures to mitigate climate change.