Spatio-Temporal Temperature Forecasting in Eastern India Using GCN-LSTM Model

Forecasting temperature trends in geographically diverse regions remains a major challenge due to the combined spatial and temporal complexities involved. Located in eastern India, the state of Odisha boasts diverse geographical features such as hills, forests, and coastal regions, where weather patterns significantly influence daily life, agriculture, and disaster management. This study presents a spatio-temporal temperature forecasting model based on a hybrid architecture which combines Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) networks. The proposed GCNLSTM model productively captures both spatial dependencies between locations and temporal trends in temperature data. The model is trained on historical daily maximum temperature data spanning from 1990 to 2014 and evaluated over the period 2015 to 2024 using three real-world datasets-IMD, NASA POWER, and ERA5-Land for the Odisha region. Experimental results demonstrate that the proposed model outperforms other models across all key metrics, achieving lower MAE and RMSE, higher $R^{2}$, and improved accuracy. These findings highlight the model's simplicity, robustness, and potential as a trustworthy method for long-term temperature forecasting and climate planning in data-sensitive regions like Odisha that can support for heatwave warnings, crop planning, and climate-related decisions.