Salt stress is a significant abiotic factor affecting the growth and development of alfalfa. Malondialdehyde (MDA) serves as a critical biomarker for assessing alfalfa’s salt tolerance. Traditional methods for measuring MDA are often time-consuming and labor-intensive. Recent advances in remote sensing technology have made non-destructive estimation of metabolites feasible, positioning the accurate estimation of MDA content in alfalfa as a key focus in intelligent breeding. To address the challenge of detecting subtle changes in MDA content, this study developed a partial least squares regression (PLSR) model specifically for Medicago truncatula. This study utilized leaf reflectance hyperspectral data across the visible near-infrared–shortwave infrared (VIR-NIR-SWIR) spectrum, applying multi-order spectral transformation methods, including continuous wavelet transform (CWT), fractional differential (FD), and multi-granularity spectral segmentation (MGSS). Feature selection techniques, such as sequential forward selection (SFS), Least-Squares Boosting (LSBoost), and feature selection using neighborhood component analysis for regression (FSRNCA), were employed to enhance the efficiency of the MDA estimation. The findings revealed that the optimal PLSR model for MDA estimation was achieved by integrating CWT features across orders 1–30 with the SFS method. This model demonstrated robust estimation capabilities under varying salt stress conditions, significantly outperforming the original spectral data (R2 = 0.654, RMSE = 22.567 vs. R2 = 0.242, RMSE = 33.411). A comparative analysis of feature selection methods confirmed that SFS was the most effective for estimating MDA content in alfalfa. These results provide valuable insights and methodologies for MDA estimation and evaluating salt tolerance in alfalfa.
Read full abstract