Objective To address the challenge that single height statistical features struggle to accurately invert plant height due to the drastic changes in rape canopy morphology across growth stages, this study investigated a rape plant height inversion method based on unmanned aerial vehicle light detection and ranging (UAV LiDAR).

Method The study took 81 rape experimental plots as the research objects. Based on UAV LiDAR point cloud data, 17 height statistical features were extracted, and multi-dimensional geometric structural features such as convex hull volume, canopy surface area, and coverage rate were introduced. Plant height inversion models were constructed under four scenarios: Three individual growth stages (seedling stage, bolting stage and flowering stage) and one combined scenario incorporating the above three growth stages. A nested cross-validation strategy was adopted in combination with the minimum redundancy maximum relevance (mRMR) feature selection algorithm to compare and analyze the inversion accuracy of three machine learning models, including partial least squares regression (PLSR), Lasso regression, and support vector regression (SVR). The SHapley Additive exPlanations (SHAP) method was used for interpretability analysis.

Result After fusing multi-dimensional geometric structural features, plant height inversion accuracy improved at all individual growth stages, with the greatest improvement observed at the bolting stage, where the optimal model, Lasso, showed an increase in R² from 0.69 to 0.82, and a decrease in RMSE from 7.0 cm to 5.3 cm. Under the combined multiple-growth-stage scenario, the Lasso model achieved the highest inversion accuracy, with an R² of 0.97, RMSE of 8.3 cm, and rRMSE of 10.9%. SHAP interpretability analysis revealed that geometric structural features such as convex hull volume and canopy surface area played important roles across all growth stages. Particularly at the seedling and bolting stages, the arithmetic mean of absolute SHAP values of these two features ranked the top two, and both exhibited positive contributions.

Conclusion Multi-dimensional geometric structural features can effectively compensate for the insufficient representation of single height statistical features under complex canopy conditions, significantly improve the accuracy of early-stage rape plant height inversion, and provide a more adaptable technical pathway for remote sensing monitoring of plant height in crops with variable canopy morphology.