TFSF: Topological and Feature Space Fusion with Spatio-Temporal Modeling for Crop Yield Prediction

The interactions between climate change and geographic conditions have imposed great challenges to agriculture researchers on crop yield predictions. Traditional machine learning algorithms like Lasso and Gradient Boosting Machine often fall short in terms of accuracy. Deep learning has emerged as a promising approach in agriculture modeling: many studies have used convolutional(CNN) and recurrent neural networks(RNN) to effectively capture the nonlinear relationship between yield and various factors such as climate, soil and management. How-ever, these approaches often neglect the spatial relations among different prediction units. In this paper, we introduce Graph Neural Networks(GNNs) to incorporate spatial knowledge in crop yield forecasting. Additionally, a Topology-Feature Space Fusion Graph Neural Network(TFSF-GNN) is proposed to address the limitations of topological graphs based on geographical distance which include only the static spatial information. The network is designed to compute the similarity of meteorological and environmental characteristics in different regions and generates graph structures of the feature space. Multiple graph convolutional networks are then used to extract information from the topological space, feature space, and common space. Extensive experiments on the benchmark dataset demonstrate that our proposed approach outperforms existing network structures on county-level yield prediction tasks.