基于无人机多维特征的水稻估产模型对比与归因解析OA

Accurate quantitative prediction of crop yield is a crucial foundation for ensuring national food security and formulating smart agricultural decisions.During the rice growth cycle,the full heading stage is the most critical period for biomass accumulation,and its canopy spectral characteristics contain rich information on biochemical parameters and yield potential.This study explores the application advantages of combining multi-dimensional UAV remote sensing features with advanced machine learning algorithms for rice yield prediction at the full heading stage.Using a multispectral UAV to acquire images,this research systematically constructed six categories of feature sets,including red-green-blue(RGB),multispectral(MS),multispectral to hyperspectral(MSTPP),spatial texture(Texture),vegetation index(Ⅵ),and multi-dimensional fusion(Fusion).At the algorithmic level,eight machine learning algorithms were comprehensively evaluated,with a focus on comparing kernel methods represented by support vector regression(SVR)and various nonlinear ensemble models.The experimental results indicate that MS data exhibit outstanding prediction potential among single-dimensional features.The strategy of fusing spectral and texture information achieved excellent prediction stability,wherein the SVR model combined with multi-dimensional fusion features achieved an ideal fitting effect(coefficient of determination R2=0.828),calibrating the spectral signals impaired by canopy closure during the full heading stage,and effectively correcting the systematic estimation bias of single-dimensional models in extreme yield intervals.Furthermore,this experiment verified through the introduction of the Shapley additive explanations(SHAP)framework that the green normalized difference vegetation index(GNDVI)played a core driving role in yield prediction,objectively revealing the safeguard mechanism of population growth balance on the final yield.The approach of"multi-dimensional fusion combined with nonlinear algorithms and attribution analysis"explored in this study provides a feasible approach for constructing quantitative yield estimation models with both high robustness and agronomic interpretability..