Integrated Knowledge Guided Graph Attention Network with Mixture Density Network for Crop Yield Prediction

In recent years, the accurately and timely predict the crop yield has become hard to achieve because crop growth has complex and changing patterns across space and time. However, the existing model struggle to capture dynamic spatial temporal, nonlinear and external environmental factors such as temperature, soil conditions and precipitation, due to its graph performance. To overcome this problem, this paper proposed a Knowledge guided Graph Attention network with a Mixture Density Network (KGAT-MDN) to predict crop yield with more flexibility. The proposed KGAT-MDN first uses a 3D Convolutional Neural Network (CNN) to extracts important features. Next, the graph attention network used to model, the proposed method as a node-level regression problem on a graph built from the distribution system. After, mixture density networks are used to predict multiple possible outcomes at once, along with their importance, which are represented as Gaussian distributions. At last, location-aware Spatial Attention Graph Network is used, which uses geospatial knowledge to combine the features of nearby areas used for the final result. The experiments results obtained 0.50 for RMSE, 0.95 for R2 and 10.11 for MAPE on Crop yield prediction dataset by outperforming the existing KSTAGE in predicting crop yield in agriculture.