Improving Hyperspectral Estimation of Fig Leaf Water Content Using Continuous Wavelet Transform and SHAP-Based Explainable Machine Learning: The Potential of Multiscale Wavelet Indices

Leaf water content (LWC) is an important indicator of plant water status, and its rapid estimation is essential for water diagnosis and cultivation management in fig production. Although hyperspectral sensing provides an effective means of estimating LWC, spectral redundancy and noise may hinder the extraction of water-sensitive information. Wavelet analysis can extract localized spectral information; however, single-scale wavelet features may not simultaneously preserve fine spectral details and suppress noise, and thus cannot fully characterize the complementary LWC-related responses across different scales. This study therefore developed multiscale double wavelet indices (MSDWIs) and multiscale triple wavelet indices (MSTWIs) to improve the hyperspectral estimation of fig LWC. Savitzky–Golay (SG) filtering and multiplicative scatter correction (MSC) were compared, and random forest (RF) and support vector regression (SVR) were used to evaluate the estimation performance of traditional vegetation indices (VIs), MSDWIs, MSTWIs, and their fused feature sets. The results showed that multiscale wavelet indices generally achieved higher estimation accuracy than traditional VIs, while multi-feature fusion further improved model performance. The SVR model based on the SG-preprocessed VIs+MSDWI+MSTWI feature set achieved the best validation performance (R2 = 0.760, RMSE = 0.0232, and MAE = 0.0152). SHAP analysis of the optimal RF and SVR models showed that MSTWI was the dominant feature category, accounting for 57.8% and 57.0% of the total SHAP importance, respectively. These findings demonstrate that integrating multiscale wavelet indices with traditional VIs can enhance the representation of LWC-related spectral information and provide an effective approach for estimating fig LWC.