Deep learning-driven forecasting of nighttime urban heat patterns in Kolkata: a spatiotemporal approach for climate-resilient city planning

Urbanization has intensified land surface temperature (LST) patterns, exacerbated urban heat island (UHI) effects, and posed significant challenges for sustainable city planning. This study investigates long-term LST dynamics (2000–2022) and develops deep learning models to forecast future night-time LST over Kolkata, India. Analysis of the 23-year record reveals a consistent warming trend, with annual mean LST ranging from 18.58°C to 20.32°C and an overall average of 19.48°C. Urban areas exhibited higher mean LST (19.88°C) compared to rural surroundings (17.60°C), confirming persistent UHI intensity. High-temperature zones (22–24°C) initially concentrated in central Kolkata in 2000 progressively expanded towards peripheral regions by 2022, particularly along north – south and north – east growth corridors, indicating spatial amplification of urban thermal stress. To model and predict future LST, Long Short-Term Memory (LSTM) and Convolutional Long Short-Term Memory (ConvLSTM) architectures were developed using integrated predictors including NDVI, NDBI, MNDWI, DBU, DBS, aerosol optical depth (AOD), and topographic variables. Models were trained using 2000–2021 data and validated against 2022 observations. The models effectively captured complex spatiotemporal dynamics influencing urban thermal environments. After hyperparameter tuning, ConvLSTM achieved superior performance (R2 = 0.94; RMSE = 0.25°C), outperforming LSTM (R2 = 0.89; RMSE = 0.36°C), with 99.68% of pixels exhibiting prediction errors within ±1°C. Long-term forecasts for 2031 revealed continued intensification of urban heat, with mean LST projected to rise from 18.78°C to 19.45°C in Kolkata, while urban cores are expected to reach mean temperatures of 20.62°C in Kolkata. Seasonal predictions highlighted pronounced summer thermal stress, with mean urban LSTs predicted to exceed 24.34°C in Kolkata by 2031. These findings underscore the value of ConvLSTM-based frameworks in proactively identifying emerging heat hotspots and supporting climate-resilient urban planning strategies in rapidly urbanizing regions.