Abstract: Citrus (Citrus reticulata) was selected as the research object, and UAV multispectral data and citrus leaf area index (LAI) data were collected. After band selection and combination of the multispectral data, three feature processing approaches (Boruta algorithm, RFECV method, and no feature selection) were employed, each combined with three machine learning regression models (support vector regression (SVR), random forest regression (RFR), and backpropagation neural network regression (BPNNR)) to construct nine combined models for LAI estimation. The model parameters were optimized using the GridSearchCV method, the accuracy and stability of each model were compared, the optimal LAI prediction model was selected, and a spatial distribution image of citrus LAI was generated. The results showed that the Boruta algorithm could effectively select feature variables and reduce model overfitting. Among the nine combined models, the Boruta_BPNNR model performed best in citrus LAI estimation, exhibiting low data dispersion and a high degree of fit between the regression curve and the diagonal line. The LAI retrieval results indicated that the spatial distribution of LAI in the study area showed a distinct north-south gradient difference, with LAI generally higher in the northern region than in the southern region. This was basically consistent with the spatial pattern observed in the field survey, where citrus growth was lush in the north region and relatively sparse in the south region.

Key words: citrus (Citrus reticulata), UAV multispectral, leaf area index (LAI), Boruta_BPNNR model, retrieval