Integrating soil spectral libraries with laboratory hyperspectral imaging for profile organic carbon prediction in paddy soils

Abstract

To promote soil organic carbon (SOC) storage and to meet growing food demand with limited land, it is essential to understand the spatial characteristics of SOC across the entire profile of paddy soils. The establishment of soil spectral libraries (SSLs) at various geographical scales has made near-infrared (NIR: 700–1100 nm) and shortwave-infrared (SWIR: 1100–2500 nm) hyperspectral imaging (HSI) more feasible for the rapid and cost-effective estimation of SOC. This study aimed to integrate SSLs with HSI for fine-scale mapping of elemental concentrations with a high spatial resolution of 1 mm per pixel (image size: 980 × 160 pixels) with a spectral range of 900–1700 nm. We apply SOC-reflectance calibrations from Global Soil Spectral Library (GSSL) to an independent local field site for the entire profile in paddy soils to a depth of 1-m, combined with spectral similarity with continuum removal (SS-CR) analysis, three spectral matching methods (e.g., Euclidean distances [ED], Mahalanobis distances [MD], and Spectral angle mapper [SAM]) and two modeling algorithms (e.g., random forest [RF] and Cubist). Additionally, we compared the performance of different Global and Local models in characterizing the distribution of SOC across the entire profile. Results indicated that although the Cubist-Local model provided good prediction accuracy (R² ≥ 0.77, RMSE ≤ 0.77 %, RPIQ ≥ 1.90), its ability to fine-scale mapping of the profile SOC was limited. In contrast, the RF-Local model based on the ED spectral matching method (ED-RF-Local) not only achieved the best performance (R² = 0.80, RMSE = 0.78 %, RPIQ = 1.86), but also successfully mapped SOC across the entire soil profile. This model used two-fifths of the samples compared to the Global model. The findings of this study provide a valuable reference for SOC prediction and mapping at the field scale and across the entire soil profile using HSI techniques and GSSL, emphasizing the potential for predictions in paddy soils with high vertical resolution.