A deep learning framework combining CNN and GRU for improving wheat yield estimates using time series remotely sensed multi-variables

Accurate and timely crop yield estimation is crucial for crop market planning and food security. Combining remotely sensed big data with deep learning for yield estimation has attracted extensive attention. However, it is still challenging to understand and quantify the time cumulative effects of crop growth over time for crop yield estimation. In this study, we combined the powerful feature extraction capability of the convolutional neural network (CNN) and the advantage of time series memory of the gated recurrent unit (GRU) network to develop a novel deep learning model called CNN-GRU for estimating county-level winter wheat yields in the Guanzhong Plain using three remotely sensed variables, vegetation temperature condition index (VTCI), leaf area index (LAI) and fraction of photosynthetically active radiation (FPAR). The CNN-GRU model was able to extract features related to yield from the input variables and the accuracy of the proposed model (R2 = 0.64, RMSE = 462.56 kg/ha, MRE = 8.90 %) was higher than that of the GRU model (R2 = 0.62, RMSE = 479.79 kg/ha, MRE = 9.34 %), and the CNN-GRU model’s reliability and robustness were confirmed by applying the leave-one-year-out cross-validation. Furthermore, we applied the proposed CNN-GRU model to simulate the wheat yields in the Plain pixel by pixel and examined the spatiotemporal patterns of the estimated yields. The distribution of yields presented the spatial characteristics of low yields in the east and high yields in the west, and the inter-annual variation characteristics of overall stability and steady increase. Additionally, we explored the possibility of timely prediction of winter wheat yield and the contribution of the multi-variables at different growth stages to yield estimation based on the ability of deep learning to reveal cumulative effects and non-linear relationships between influencing factors and yield. It was found that the information reflected by the multi-variables from late March to late April was important for yield estimation and the best prediction could be achieved approximately 20 days before the harvest of winter wheat. Our study demonstrated that combining CNN and GRU was an efficient and promising approach to improve yield estimation, offering great promise for global crop yield estimation.