Remote sensing-based winter wheat yield estimation integrating machine learning and crop growth multi-scenario simulations
Xin Du, Jiong Zhu, Jingyuan Xu, Qiangzi Li, Zui Tao, Yuan Zhang, Hongyan Wang, Haoxuan Hu
Abstract
Accurate and timely winter wheat yield prediction is critical for effective agricultural management and food security. This study used the World Food Studies (WOFOST) model, a widely adopted crop growth simulation model, to dynamically simulate winter wheat yield under various growth scenarios to produce a simulated dataset. Based on this dataset, custom yield estimation models were developed based on available remote sensing data. Validation with field-measured and county-level statistics demonstrated a robust and spatially extensive capability for accurate yield estimation, with R2, RMSE, and MRE values of 0.57, 424.80 kg/ha, and 6.57% at the plot level, and 0.58, 345.53 kg/ha, and 4.93% at the county level, notably improving on traditional field-based methods (R2 = 0.03–0.46) that primarily rely on limited field surveys and statistical models. Model simplification showed that accuracy decreased when fewer remote sensing images were used, yet achieved reasonable estimates (two temporal phases: R2 of 0.41/0.40 at plot/county level). Findings highlighted that data collection during key growth stages is essential for accuracy, and that a dataset of at least 5,000 records suffices for reliable results. This study offers important insights and direction for enhancing yield prediction with efficient data acquisition and modeling strategies in large-scale applications.