Maize yield estimation based on UAV multispectral monitoring of canopy LAI and WOFOST data assimilation

Abstract

The estimation accuracy of crop model is influenced by model parameters, model inputs, and model structure. Data assimilation was frequently employed to enhance model performance. To evaluate the feasibility of data assimilation by UAV remote sensing and WOFOST in improving maize yield prediction, a set of field experiments was conducted in Mangshi, Yunnan Province, from 2023 to 2024. Based on the canopy remote sensing data collected by UAV, five inversion models of leaf area index (LAI) were developed using machine learning methods, i.e. Random Forest (RF), Partial Least Squares (PLS), Ridge Regression (RR), k-Nearest Neighbors (KNN), and Extreme Gradient Boosting (XGBoost), the best-performing inversion model was selected for data assimilation. Field trials data were used to calibrate the WOFOST model, and the ensemble Kalman filter (ENKF) was applied to assimilate inverted LAI. The results showed that the RF-based inversion model provided the highest accuracy in estimating LAI, with R² of 0.82 and NRMSE of 18 %. For the calibrated model, the NRMSE of yield and LAI were 12 % and 34 %, respectively. After assimilation, the NRMSE for yield and LAI decreased to 4 % and 15 %, respectively, and the average yield error was reduced by 808 kg/ha. Multiple rounds of assimilation reduced both the error range and bias caused by parameters uncertainty. This study demonstrates that assimilating UAV-inverted LAI with RF into the WOFOST model effectively enhances its ability to simulate dynamic crop growth and reduces uncertainty. The effect of data assimilation on the interaction of various uncertainties in the model needs further research. This research offers valuable insights into applying UAV remote sensing and data assimilation technologies for precision maize management.