Feature Selection and Spectral Indices for Identifying Maize Stress Types.

This study aims to identify different types of stress on maize leaves using feature selection and spectral index methods. Spectral data were collected from leaves under heavy metal, water, fertilizer stress, as well as under normal healthy conditions. Preprocessing steps such as continuum removal (CR), standard normal variable (SNV) transformation, multiple scattering correction (MSC), detrend correction (DT), and first-order derivative (FOD) were applied to the raw spectra. Various feature selection methods including ReliefF, chi-square test, recursive feature elimination (FRE), mutual information (MI), random forest (RF), and gradient boosting tree (GBT) were employed to determine the importance scores of different spectral bands, thus identifying sensitive spectral features capable of distinguishing various stress types. Spectral indices for stress type differentiation were constructed using label correlation method. Classification models were built using support vector machine (SVM), K-nearest neighbors (KNN), Gaussian naive Bayes (GNB), extreme gradient boosting (XGBoost), RF, and adaptive boosting (AdaBoost) algorithms. Results indicate that the characteristic spectral bands for differentiating stress types are primarily distributed around the red edge (near 700-800 nm) and water absorption valley (near 1900 nm). Spectral indices constructed using combinations of spectral bands around the near-infrared plateau absorption valley (near 1185 nm) and water absorption valley (near 1460 nm) effectively differentiate maize stress types. Among the modeling classification algorithms, RF and AdaBoost algorithms exhibited optimal performance, demonstrating high classification accuracy on both training and validation sets. These findings hold promise for providing new technical support for maize stress monitoring and diagnosis in agricultural production.