Inversion of citrus SPAD value and leaf water content by combining feature selection and ensemble learning algorithm using UAV remote sensing images

Soil and Plant Analyzer Development (SPAD) value and leaf water content (LWC) are critical physiological parameters for agricultural irrigation and growth monitoring in late-maturing citrus. Accurate monitoring of citrus SPAD value and LWC is of great significance for guiding precision irrigation, improving water use efficiency, and enhancing yield. To rapidly and efficiently obtain the SPAD value and LWC of citrus orchards, this study extracted vegetation index (VI) and texture feature (TF) of late-maturing citrus at different growth stages based on UAV multi-spectral images. Feature variable selection methods (decision tree (DT) and least absolute shrinkage and selection operator (Lasso)) were combined with Support vector machine regression (SVR), AdaBoost (Ada), SVR-AdaBoost (SVR-Ada) and WOA-SVR-Ada. Models for estimating SPAD value and LWC in citrus orchards were constructed using VI, TF, and VI+TF as inputs. The results showed that the DT algorithm demonstrated superior capability in identifying feature variables compared to the Lasso. The integration of VI and TF can enhance the inversion accuracy of citrus SPAD value and LWC models. Compared to the SVR, Ada and SVR-Ada, the WOA-SVR-Ada model, constructed by combining the DT algorithm with VI+TF as inputs (WOA-SVR-Ada_D3), exhibited the highest estimation accuracy for both SPAD value and LWC. Therefore, combining feature variable selection methods with ensemble learning algorithms, along with the fusion of multi-feature information from UAV multispectral, holds promise for providing precise and robust estimations of SPAD value and LWC for late-maturing citrus in the seasonal drought regions of Southwest China.