Empowering Soil Spectral Modeling with Interpretable Feature Wavelength Transfer Techniques in Deep Learning

With the widespread use of deep learning in soil spectral modeling, interpretable models have been introduced to explain the impact of characteristic wavelengths. Notably, Shapley Additive exPlanations values from game theory have been applied successfully to interpret deep learning soil spectral models. This study explores the use of computational results from interpretable models for spectral feature wavelength migration and data preprocessing. Using the LUCAS spectral database, we selected nine soil features and their spectra to construct convolutional neural networks (CNN) models. Four interpretable methods were employed to determine the contribution of different wavelengths to model training, with the best-performing method chosen for each soil property. Bands with lower contribution scores were gradually eliminated, and the remaining spectra were passed into the screening model (PLSR)l for fitting, determining the best spectral bands and data dimensions based on the results. Our findings indicate that feature wavelengths learned by the CNN model are applicable to the PLSR model, enabling feature downscaling and wavelength screening. Using high- and medium-accuracy CNNs, we achieved better performance than the original spectrum by retaining 38% of the original wavelength features. Additionally, extracting significant wavelengths from low-precision CNNs improved the PLSR fitting effect, allowing wavelength features to be compressed to 8% of the original without significant impact. The study also discusses the basis for feature wavelength migration from data characteristics and physical properties perspectives and verifies the feasibility of migrating feature wavelengths from high-parameter to low-parameter models. Combining interpretable methods and screening models offers a versatile data processing method, advancing deep learning applications in soil hyperspectral modeling.