Spring maize height estimation using machine learning and unmanned aerial vehicle multispectral monitoring

Monitoring crop height at different growth stages is essential for understanding crop growth conditions and optimizing field management. We employed five machine learning algorithms—partial least squares regression, elastic net regression, support vector regression, random forest, and gradient boosting regression tree (GBRT)—in conjunction with 13 multispectral unmanned aerial vehicle (UAV) vegetation indices (VIs) to estimate spring maize height at the field scale in Northeast China. The results revealed strong positive correlations between observed maize height and UAV VIs during the jointing, tasseling, silking, milk, and maturity stages, demonstrating the effectiveness of UAV VIs for estimating spring maize height. Among the models, GBRT consistently outperformed the others across all growth stages, with R2 values ranging from 0.79 to 0.99, RMSE values from 0.30 to 14.70 cm, and MAE values from 2.4 to 11.40 cm. In addition, using the GBRT model and Shapley Additive Explanations, the study identified the most influential VIs for height estimation at each growth stage. Specifically, MGRVI, RVI, EVI2, EVI, and MTCI were the key predictors during the trefoil, jointing, tasseling, silking, milk, and maturity stages, respectively. These findings provide valuable insights for precision agriculture management at the field scale and offer a reference for estimating crop height using satellite-based VIs at a regional scale.