Enhancing Drought Forecasting Using GNN-LSTM with Attention Mechanism: A study of Jaisalmer district, Rajasthan

Droughts severely effect food production, water supply, and environment, demanding accurate prediction for mitigation. This study addresses the critical challenge of drought prediction by leveraging advanced deep learning methods to improve forecast accuracy. The research focuses on developing a Graph Neural Network-Long Short-Term Memory (GNN-LSTM) model integrated with an attention mechanism to analyze the relationship between precipitation data and drought indicators of VCI, TCI, VHI, and SPI in Jaisalmer, Rajasthan, India, from 1991 to 2023. The study acknowledges the best mapping of precipitation to compute SPI, and investigates the spatial correlation with VCI, TCI, VHI, and SPI drought indices. The proposed model GNN-LSTM with attention, scored spatial dependency in dynamic way using graph and improves long range temporal learning rate with LSTM architecture via attention to lead better accuracy and robustness in drought prediction in comparison to developed model architecture of Artificial Neural Network with Multi-Point Aggregation (ANN-MPA), and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) etc. The GNN-LSTM model outperformed in distinguishing and predicting drought and non-drought regions with Mean Absolute Error (MAE) of ± 0.1906, Mean Squared Error (MSE) of ± 0.0982, Root Mean Squared Error (RMSE) of ± 0.03285, and an R-squared score of ± 0.8239. The main aim of this research is to contribute to decision making framework to mitigate the adverse effects of climate variability, enhanced agricultural planning and helps in early warning systems, allowing for proactive measures to manage and adapt to drought conditions, ultimately contributing to a sustainable environment.