PSNet: a knowledge guided deep learning approach for county-level corn yield estimation

Accurate crop yield estimation is important for global food security. Data-driven deep learning approaches have shown great potential for agricultural system monitoring, but are limited by their out-of-sample prediction failure and low interpretability. How to embed knowledge into deep learning models to address the above challenges remains an open question. In this study, we developed a deep learning model named PSNet following the concept of hierarchical yield levels to estimate county-level crop yield. The PSNet model mainly consists of PotentialNet and StressNet to capture the interactions among crop, environment, and technological trend. The PotentialNet is developed to capture the spatiotemporal pattern of the rice yield potential based on environmental and local technological conditions. The StressNet is designed to capture the negative impact of climate stresses, which caused the yield gap between yield potential and actual yield. We applied the model to analyze the county-level rainfed corn yield in the US Corn Belt from 2006 to 2020. The Random Forest (RF) and Long Short-term Memory (LSTM) models were chosen as baselines. The results showed that the PSNet model achieved better yield estimation accuracy than baselines under the normal (R² = 0.82) and stressful climate conditions (R² = 0.77). The ablation results indicated that PotentialNet contributed to the yield estimation under normal climate conditions, while the StressNet was better at capturing the yield losses under climate stresses. This study provided a promising approach to extract the pattern of yield potential and stress impact to achieve good estimation performance across various growth conditions.