Spatio-temporal modeling of climate change impacts on farming using GNN-LSTM with attention and ensemble learning

Abstract

Climate change poses a great danger to agriculture today, which in turn affects crop yields, quality, and food security; thus, efforts to create simulation tools that would help in the assessment of these effects are paramount. This work proposes a model of Graph Neural Networks (GNN) and Long Short-Term Memory with attention (LSTMatt) networks. The GNN learns the spatial configuration and interconnection between climate and agricultural data, and the LSTMatt learns the temporal patterns. The model is a combination of outputs from LSTMatt and GNN by Random Forest-based bagging, which thereby improves accuracy and robustness of predictions. This integrated approach provides a more holistic view to ensure a solution to some of the challenges affecting the policymakers, farmers, urban rooftop planners, and smart city developers to make better decisions on suitable options for practices in agriculture to mitigate the effects of climate change.