Enhancing crop yield simulations under extreme climate events using a hybrid model

Crop models currently have a limited capacity to accurately simulate the impacts of extreme climate events (ECEs), and there is considerable uncertainty across different models. Consequently, the assessment of food security risks from future climate extremes based on existing frameworks is less reliable. To address this issue at global scale, we are developing an advanced hybrid model that integrates process-based crop models with information on the occurrence of extreme climate events and a deep learning framework. Specifically, our model uses outputs from multiple crop models provided by the third round of the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP 3a) as the initial input. The second input will consist of the daily occurrence of four types of extreme events: two related to temperature (heatwaves and coldwaves) and two related to precipitation (droughts and pluvials). We employ a Long Short-Term Memory (LSTM) network with an attention mechanism designed to dynamically capture the varying impacts of ECEs at different crop growth stages. The results are expected to offer a more precise simulation and deeper understanding of how ECEs affect food security. This study highlights the potential of AI-hybrid modeling to enhance the accuracy of crop impact assessments under climate change.