Work

Long-term large-scale winter wheat yield estimation: A Bayesian temporal convolutional network

Year: 2025

Type: article

Abstract: Long-term and large-scale crop yield forecasting plays a pivotal role in formulating seasonal crop management strategies and maintaining global food security. Although studies worldwide have integrated climate and remote sensing data to predict yields at regional and national scales, the long-term and large-scale winter wheat yield forecasting in China remains scarce. For this purpose, we propose a novel Bayesian-Temporal Convolutional Network (BTCN) which integrates phenological aggregation to assess the contributions and interferences of multi-source data on yield estimation across large spatial and temporal scales. Initially, the study takes the yield estimation of winter wheat in China from 2001 to 2019 as a case study and temporally aggregates multi-source predictors based on phenological stages. Secondly, the county-level temporal feature imagery is employed as input for the BTCN to predict crop yields, benchmarking its performance against three other widely utilized machine learning models. Finally, the study conducts a spatiotemporal analysis of feature importance and uncertainty associated with multi-source data in yield estimation. In winter wheat yield estimation across China, our model excels with an average R 2 of 0.741, surpassing the other three typical methods. The spatiotemporal analysis highlights geographical yield predictors disparities across China, with EVI and water factors in the north, and early-stage water and heat factors in the south. At the same time, the phenological stages also differ by region, and an earlier maturation period in the Yangtze River basin. In addition, our results indicate that overabundant predictors can introduce noise and uncertainty, affecting yield estimation reliability. In short, the novel BTCN framework not only improves the accuracy of large-scale winter wheat yield prediction but also largely enhances our understanding of its uncertainties and responds to spatiotemporal features, making it highly significant for sustainable agricultural development.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2025.101632

Authors: Shuyuan Wang, Xiaobin Xu, Yunhui Tan, Ting Hu, Xiong Xv, Hongyan Zhang