Combining the vegetation indices and bionic optimization random forest for predicting SPAD values in maize leaves

Abstract

Leaf Soil and Plant Analyzer Development (SPAD) is an important physiological index reflecting the healthy growth and development of maize. Accurate estimation of leaf SPAD by vegetation indices is of great significance for regulating maize growth, optimizing nutrient management and improving yield formation. However, there is a lack of automated and transferable parameter adjustment processes for machine learning based on hyperspectral data characteristics at the field scale. A bionic optimization random forest (BO-RF) based on the spectral vegetation indices was proposed to predict the SPAD value of maize leaves. The SPAD and spectral reflectance of maize leaves at key growth stages were obtained by field experiments, and the pretreatment methods and predictors with the strongest correlation with SPAD were selected. Parrot optimization (PO), catch fishing optimization algorithm (CFOA), goose optimization (GOOSE) and RF predictor were combined to compare the SPAD inversion models of maize leaves. The results showed that the V12 and R1 stages were the characteristic growth stages of SPAD values of maize leaves. The GOOSE-RF under the driving of vegetation indices after SG-SNV pretreatment performed best, the test set R 2 and RMSE of which were 0.73 and 2.68 respectively. Compared with RF, PO-RF and CFOA-RF, the prediction accuracy of GOOSE-RF model were improved by 15.87%, 12.31% and 7.35% respectively. The PO, CFOA and GOOSE algorithms improved the RF model in the generalization ability of the model, suppressing overfitting and optimizing parameter configuration. In summary, the BO-RF method with sensitive vegetation indices as input could provide a theoretical basis for spectral monitoring of crop leaf nutrition information.