Hyperspectral inversion method of garlic LAI based on multi-scale spectral decomposition and differentiated feature selection

As a critical component in hyperspectral modelling, feature selection directly affects the accuracy and stability of crop physiological parameter inversion. To address the challenges posed by the high dimensionality, strong redundancy, and significant frequency-domain heterogeneity of hyperspectral data, this study proposes a hyperspectral modelling method based on multiscale spectral decomposition and differentiated feature selection for garlic leaf area index (LAI) inversion under multi-growth phase conditions. First, Discrete Wavelet Transform (DWT) was employed to perform multiscale decomposition of the original spectra, extracting sub-band spectral features that reflect different physiological structural levels. Then, according to the frequency domain attributes of each sub-band, a differentiated feature selection method with frequency band adaptability is designed. Four Filter classes, four Wrapper classes and three Embedded classes are used to extract low-frequency, medium-frequency and high-frequency features respectively. Furthermore, a multiscale feature fusion method based on attention mechanism is constructed, and a lightweight CAM-CNN-Transformer structure is introduced to strengthen the weight distribution between features and spectral domain modelling capabilities. The experimental results show that the method performs well in LAI inversion. The CAM-CNN-Transformer model achieves the highest R2 = 0.918 and RPD = 3.497 on the test set. Stratified cross-validation further indicates its robust generalization capability across different phases and growth phases (mean RPD > 3.0). This study verifies the significant advantages of multi-scale spectral decomposition combined with differential selection method in improving the performance of hyperspectral modelling, and provides a new modelling path with generalization ability for remote sensing parameter inversion.