GraphCrossFormer A Graph-Enhanced Cross-Attention Transformer for Crop Yield Prediction

Accurate forecasting of agricultural yields is essential to ensuring global food security and developing effective climate change adaptation strategies. As global environmental changes make weather patterns increasingly unpredictable and shift agricultural zones, traditional forecasting models often fall short. These models struggle to capture the complex interplay between crop characteristics and dynamic environmental factors that impact yields. To address this, we introduce GraphCrossFormer—an advanced transformer-based architecture designed specifically to improve crop yield predictions. Its key innovations include: (1) A GraphFormer module that models feature dependencies through message-passing in fully connected graphs, and (2) A cross-attention mechanism that dynamically integrates scenario-level context like emission trajectories and adaptation strategies.Together, these components learn spatial, semantic, and contextual relationships within a unified, interpretable framework, significantly boosting prediction accuracy. This thesis utilizes a global dataset covering wheat, rice, corn and soybeans, and predicts their impacts on yields based on various meteorological data,and GraphCrossFormer consistently outperformed alternatives. It achieves improved prediction accuracy and reduced errors for different crop types and climate scenarios. Furthermore, interpretability analyses using attention-basedand gradient-based methods reveal the relative importance of critical climate variables and management factors. Ultimately, the proposed model offers an effective, powerful, and interpretable tool for reliable crop yield forecasts in the context of climate-resilient agricultural planning.