Deep learning-based prediction and analysis of urban heat island dynamics using multispectral indices and precipitation data
TL;DR
The findings demonstrate the effectiveness of deep learning for 1 km LST prediction and provide a transferable framework for urban climate monitoring and sustainable land-use planning.
Abstract
Infographic summarizing deep learning models, land surface temperature prediction, urban heat island analysis, key findings, and study area in Peshawar District. Rapid urbanization has altered the thermal environment of Peshawar District, Pakistan, intensifying the urban heat island (UHI) effect through increased impervious surfaces, vegetation loss, and reduced surface moisture. This study developed a comparative deep learning framework for land surface temperature (LST) prediction using multispectral remote sensing indices (NDVI, NDWI, NDBI, and NDBaI) and precipitation data. Four models were evaluated: Pix2Pix generative adversarial network (GAN), convolutional neural network–long short-term memory (CNN–LSTM), U-Net, and vision transformer (ViT). Using MODIS LST and CHIRPS precipitation data (2000–2020), model performance was assessed with root mean squared error (RMSE), mean absolute error (MAE), R², structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). Mean LST and UHI intensity increased throughout the study period, with LST showing strong positive correlations with NDBI and NDBaI (r = 0.94) and strong negative correlations with NDVI (r = –0.89) and NDWI (r = –0.91). Pix2Pix GAN achieved the highest predictive performance (R² = 0.94–0.96, RMSE = 1.0–1.2 °C, SSIM ≈ 0.90), outperforming ViT, CNN–LSTM, and U-Net. These findings demonstrate the effectiveness of deep learning for 1 km LST prediction and provide a transferable framework for urban climate monitoring and sustainable land-use planning.