Integration of Crop Growth Model and Random Forest for Winter Wheat Yield Estimation From UAV Hyperspectral Imagery

Accurate estimation of crop yield is critical for sustainable development. Machine learning (ML) has been introduced to estimate crop yield; however, the shortages of field measurements have troubled researchers. In this study, the CW-RF model, a new wheat yield estimation model suitable for the North China Plain, was established using the random forest (RF) algorithm, and the CERES-Wheat model was chosen to simulate training samples for RF at field plot scale. According to CERES-Wheat model simulation, the leaf area index (LAI) and leaf nitrogen content (LNC) at the jointing and heading stages were selected as the most sensitive parameters, and were retrieved from unmanned aerial vehicle (UAV) hyperspectral imagery using the directional second derivative (DSD) and angular insensitivity vegetation index (AIVI) methods respectively. Then the retrieved LAI and LNC were input into the CW-RF model to estimate wheat yield. The field validation in Luohe, Henan showed that the root mean squared error (RMSE) of retrieved LAI and LNC were 6.27% and 12.17% at jointing stages, 9.21% and 13.64% at heading stages, respectively. The RMSE of the estimated yield was 1,008.08 kg/ha, and the mean absolute percent error (MAPE) of the estimated yield was 9.36% compared with the measured yield, demonstrating the available of the CW-RF model in wheat yield estimation. This study showed that data simulated by the CERES-Wheat model can be important data source for ML-based wheat yield estimation model, and the UAV is a reliable platform for wheat growth monitoring and yield estimation.