Spatiotemporal Analysis and Deep-Learning Based Forecasting of Land Surface Temperature in the UNESCO World Heritage Sinharaja Rainforest

• Developed a deep learning framework using Convolutional Long Short-Term Memory (ConvLSTM) to forecast temperature dynamics in the Sinharaja Rainforest, a critical UNESCO World Heritage Site and biodiversity hotspot. • Integrated two decades (2000–2021) of high-resolution satellite-derived Land Surface Temperature (LST) data to capture complex, nonlinear patterns with high predictive accuracy (RMSE = 1.11°C). • Identified a subtle but persistent long-term warming signal through time series decomposition, providing essential spatiotemporal insights for climate adaptation and the management of data-scarce tropical ecosystems. Currently, climate change is a global issue receiving significant attention and it leads to unpredictable temperature patterns that threaten the biodiversity and ecological stability of forests. While accurate temperature prediction in such ecosystems is challenging due to sparse ground observations and strong spatial heterogeneity. To fill this gap, this study develops and compares multiple deep learning architectures for land surface temperature (LST) prediction in the Sinharaja Forest Reserve, Sri Lanka’s last remaining major tropical rainforest and a UNESCO World Heritage Site. Monthly MODIS LST data from 2001 to 2021 were used to analyze historical temperature dynamics and develop spatiotemporal forecasting models. The nonparametric Mann-Kendall test did not indicate a statistically significant trend. However, the time series decomposition method revealed a subtle long-term warming signal in the region. Three deep learning architectures, including ConvLSTM, CNN-LSTM, and PredRNN, were evaluated using a hindcast validation framework, where models were trained using historical observations and tested against independent satellite observations from the later years. The ConvLSTM model demonstrated the highest predictive accuracy, achieving a root mean square error (RMSE) of 1.55°C, a mean absolute error (MAE) of 1.14°C, a coefficient of determination (R²) of 0.45, and a mean absolute percentage error (MAPE) of 4.36%, demonstrating improved capability in capturing nonlinear spatiotemporal temperature patterns compared to CNN-LSTM and PredRNN models. The findings provide valuable insights to support future climate adaptation and conservation policy decisions for the globally significant Sinharaja Forest Reserve.