NortheastChinaSoybeanYield20m: an annual soybean yield dataset at 20 m in Northeast China from 2019 to 2023

Abstract

Accurate monitoring of crop yield is critical for ensuring food security. While various yield datasets covering Northeast China exist, they were produced at a coarse spatial resolution and remain inadequate for capturing small-scale spatial heterogeneity. Current yield estimation methods, such as machine learning models and the assimilation of remotely sensed biophysical variables into crop growth models, are heavily reliant on ground observations and are computationally expensive. To address these limitations, we propose a hybrid framework that couples the World Food Studies Simulation Model (WOFOST) and a Gated Recurrent Unit (GRU) model to generate a high-resolution (20 m) soybean yield dataset in Northeast China from 2019 to 2023 (NortheastChinaSoybeanYield20m). First, to generate a comprehensive training dataset, WOFOST was employed to simulate diverse soybean growth scenarios by accounting for variations in climate, crop varieties, soil types and agro-management practices. The GRU model was then trained to establish the relationships between model-simulated leaf area index (LAI) and soybean yield. The trained model was applied to estimate soybean yield in Northeast China using two stage-averaged LAI variables derived from Sentinel-2, which were validated as a feasible alternative to time-series LAI. The accuracy of estimates was evaluated using in situ measurements and government statistical data. The overall root mean squared error (RMSE) was 287.44 and 272.36 kgha-1 at the field and regional scales, respectively. The model exhibited consistent interannual stability, with mean relative errors (MREs) averaging 11.46 % and 7.94 % at the municipal and provincial scales, respectively. The dataset effectively captured spatiotemporal yield variability, offering potential for optimizing soybean production, guiding precision agriculture practices, and informing agricultural policy.