Three-dimensional hyperspectral indices: Application of proximal NIR and SWIR spectroscopy for robust stratified soil...

Accurate estimation of the vertical distribution of soil moisture content (SMC) at the field scale remains inherently challenging, particularly under vegetated conditions where optical observations are only indirectly related to subsurface moisture dynamics. This challenge is further exacerbated by the depth-dependent attenuation of soil moisture information as it propagates from the soil profile to the canopy, raising fundamental questions about how canopy spectra encode moisture signals originating from different soil layers. In this context, this study investigates whether integrating multi-band information through three-dimensional spectral representations can improve the characterization of stratified soil moisture and provides an explicit assessment of how spectral-moisture coupling varies with soil depth. The analysis was conducted during the tasseling stage of spring maize and focused on three soil layers (0-20, 20-40, and 40-60 cm). Proximal hyperspectral measurements spanning the 350-1830 nm range were used to construct a comprehensive set of optimal spectral indices (OSIs), including three-dimensional formulations designed to jointly encode moisture-related spectral responses across multiple wavelength regions, as well as conventional two-dimensional indices. These indices were derived in both reflectance and first-order derivative domains, and their performance in estimating depth-specific SMC was evaluated using partial least squares regression, support vector machine, and XGBoost models. All constructed spectral indices exhibited statistically significant correlations with measured SMC (P < 0.05), while the strength of these relationships showed a clear and systematic decline with increasing soil depth. The maximum correlation coefficients decreased from 0.562 in the 0-20 cm layer to 0.514 in the 20-40 cm layer and further to 0.464 in the 40-60 cm layer, indicating a progressive attenuation of soil moisture signals at greater depths. Across all soil layers, first-order derivative-based indices consistently outperformed reflectance-based indices, and three-dimensional OSIs demonstrated superior and more stable performance relative to their two-dimensional counterparts. Among the evaluated regression approaches, XGBoost achieved the highest estimation accuracy, with the optimal model for the 0-20 cm layer yielding an R² of 0.786, an RMSE of 0.790, and a mean relative error (MRE) of 4.129%. Overall, the results indicate that canopy hyperspectral sensitivity to soil moisture is inherently depth dependent and constrained by vegetation-mediated information transfer processes; within this framework, the superior performance of three-dimensional optimal spectral indices highlights their conceptual value for more effectively encoding distributed, multi-band moisture signals, while simultaneously underscoring the fundamental limitations of optical hyperspectral sensing for retrieving deeper soil moisture under vegetated conditions.