What if transformers revolutionize geospatial forecasting? ConvLSTM-Transformer-ARIMA framework for LST forecasting

The present study introduces a novel integrated framework that merges Convolutional Long Short-Term Memory (ConvLSTM) networks, Transformer architecture, and Autoregressive Integrated Moving Average (ARIMA) models to predict the dynamics of Land Surface Temperature (LST) in China’s Sichuan-Chongqing region. The region’s complex topography and rapid urbanization present substantial difficulties for accurate LST forecasting. To tackle these challenges, we employed comprehensive LST raster data from 2001 to 2020, coupled with 19 influencing factors. Extensive collinearity and redundancy analyses were conducted to identify 11 critical drivers, such as Carbon Fixation Potential (CFP), Thermal Circulation Potential (TCP), Net Primary Production (NPP), Artificial Light Heat Index (ALHI), and factors related to urban expansion. The hybrid model capitalizes on the strengths of ConvLSTM in capturing spatiotemporal dependencies, Transformer in enhancing global context modeling, and ARIMA in short-term trend forecasting. Compared to conventional CNN and LSTM models, the proposed framework achieves an $R^2$ of 0.9054, representing an approximately 10% improvement in explained variance, while maintaining comparable MAE levels. This indicates that the hybrid architecture enhances the ability to capture complex spatiotemporal dynamics and improves the stability of predictions. Meanwhile, the model’s predictive accuracy was verified using observed LST data from 2021 to 2023, showing high precision with low Mean Absolute Error (MAE) and Mean Squared Error (MSE), as well as high Pearson Correlation Coefficient (PCC) and Coefficient of Determination ($R^2$). Moreover, SHAP (SHapley Additive exPlanations) analysis pinpointed key factors influencing LST variations, with vegetation productivity (GPP, NPP) and meteorological parameters (WDSP, STP) emerging as dominant contributors. This research underscores the efficacy of the hybrid ConvLSTM-Transformer-ARIMA model in capturing complex spatiotemporal LST dynamics and offers actionable insights for urban planning and climate adaptation strategies in regions with challenging environmental conditions.