Soil organic carbon (SOC) is a key indicator of soil fertility and a major component of the terrestrial carbon cycle, and rapid hyperspectral prediction of SOC has become increasingly important for precision agriculture and regional carbon monitoring. The accuracy of such hyperspectral predictions, however, may depend on soil depth because the spectral signal of SOC varies systematically with depth-related differences in carbon composition, soil texture, and water content. Here we collected 388 soil samples from four depth layers (0–20, 20–40, 40–60, and 60–80 cm) at 97 sampling sites in a Loess Plateau cropland in Changzhi, Shanxi Province, China. Reflectance spectra (350–2500 nm) were measured with an ASD FieldSpec 4 spectrometer under controlled laboratory conditions. Raw spectra were denoised using a Savitzky–Golay filter and resampled at 10 nm intervals. The Kennard–Stone algorithm partitioned the data into calibration and validation sets at a 2:1 ratio, and bands with statistically significant Pearson correlations with SOC (p < 0.05) were retained for modeling. We compared three modeling approaches—partial least squares regression (PLSR), grid search support vector regression (GRID-SVR), and particle swarm optimization-based support vector regression (PSO-SVR)—under two strategies: a depth-specific (single-layer) strategy and a fully integrated strategy that pooled all depths. None of the single-layer models reached a residual predictive deviation (RPD) above 1.4, mainly because the within-layer coefficients of variation were small (19.5%–27.3%). The integrated PSO-SVR model achieved validation R²=0.760, RMSE=1.496 g kg⁻¹, and RPD=2.051, satisfying the criterion for excellent prediction. When the validation predictions of the four single-layer PSO-SVR models were merged, the overall accuracy (R²=0.830, RMSE=1.312 g kg⁻¹, RPD=2.434) exceeded that of the integrated model. The findings imply that, although individual single-depth models may underperform in narrow-range data, depth-stratified modeling followed by prediction merging is a useful strategy for improving overall accuracy of hyperspectral SOC inversion. The framework is potentially applicable to other soil attributes that show strong vertical heterogeneity.
Yan et al. (Thu,) studied this question.