Indirect Estimation of Subsoil Aggregate Indices Based on Hyperspectral Vis-NIR Data in the Loess Hilly-gully Region

[Purpose] Soil aggregate indices, crucial indicators of soil structure quality, exhibit spatial and temporal variations influenced by soil conditions. Traditional methods for determining these indices, such as dry-sieving or wet-sieving, are resource-intensive. Previous research has proposed the use of hyperspectral visible near-infrared (Vis-NIR) data for topsoil aggregate index (TAI) estimation in croplands. However, subsoil aggregate index (SAI) spectra are challenging to obtain directly. Regions with severe erosion typically comprise grassland or forestland with steeper slopes rather than cropland. The study analyzes the variation of soil aggregate indices under different land use types of cropland, grassland, and forestland. The potential for indirectly predicting SAI from hyperspectral Vis-NIR is explored. Topsoil and subsoil macro-aggregate values and aggregate stability metrics are observed to be the highest in forestland with a greater slope, gradually increasing with prolonged forest duration. [Methods] A binary particle swarm optimization combined with an artificial neural network proves effective for TAI prediction under diverse land use types. [Results] Secondary soil properties (mean weight diameter, geometric mean diameter, percentage of aggregates destruction, and fractal dimension) outperform direct soil aggregate fractions (macro-aggregate, micro-aggregate, and organo-mineral aggregate) in predicting accuracy. Significant correlations are noted among TAI, among SAI, and between TAI and SAI. Leveraging the strong correlation between TAI and SAI, SAI can be directly predicted from measured TAI or indirectly from predicted TAI based on hyperspectral Vis-NIR. [Conclusions] The study underscores the critical role of spectra in TAI and SAI prediction, particularly in soils prone to erosion under different land use types.