This is the authors' abstract. We don't add key points for this paper.
The increasing population and food demand require a timely and accurate large-scale rice yield estimation method to ensure food security and the development of global food trade. However, traditional methods suffer from insufficient spatio-temporal feature extraction and inability of county- and pixel-scale yield estimation in large-area yield prediction, limiting the accuracy of yield estimation. Taking Northeast China as our study area, we propose a deep learning model, Zoned CNN-Transformer (ZCT) combining a Convolutional Neural Network (CNN) with a Transformer module. We divided Northeast China into four environmentally similar regions through cluster analysis. In each region, the features in remote sensing data, meteorology, and soil properties were extracted and modeled. We compared the estimation accuracy of ZCT with that of CNN-Transformer (CT), Zoned CNN-ALSTM (ZCAL), and Zoned ANN-Transformer (ZAT). The results show that the ZCT model performs the best in the end-of-season model performance assessment, achieving an R² of 0.89 and RMSE of 617.34 kg/ha at the county scale, and an R² of 0.87 and RMSE of 678.04 kg/ha at the pixel scale. Additionally, the ZCT model also performs well in the mid-season dynamics assessment, in which August is a critical period. We also suggest that normalized difference vegetation index, enhanced vegetation index, solar induced fluorescence, organic carbon content, and cation exchange capacity are the five most important feature variables in yield estimation through feature importance analysis. In conclusion, our ZCT model not only extracts features related to rice yield, but also reduces the impact of environmental differences on performance of the model in large regions through zonal modeling. This model realizes yield estimation at both county and pixel scales, providing an efficient and accurate solution for rice yield estimation in large regions and offering technical support for food security.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: