Enhancing spectral estimation accuracy of soil organic carbon by using geographic region and clay content as covariates.

Accurate monitoring of soil organic carbon (SOC) content, a core indicator of soil quality and ecosystem health, is important for sustainable agricultural development. Visible-near infrared (Vis-NIR) spectroscopy has become an important monitoring method owing to its fast, non-destructive, and cost-effective advantages. Regional soil data is difficult to cover soil attribute variations in complex environments. Soil spectral library (SSL) significantly enhances the comprehensiveness of data by integrating spectral and attribute data from multiple areas and types of soils. However, SSL has strong heterogeneity, and its soil samples may differ significantly in their attributes from the soil samples of the target area, which may lower the model accuracy. In this study, 210 soil samples were collected from three irrigated agricultural areas in Xinjiang, China, to construct a regional soil spectral library, and spectral preprocessing was combined with regression algorithms for modeling. Stratification modeling and covariate modeling strategies based on geographic region, soil type, clay content, and spectral type were proposed and evaluated. The results showed that the modeling based on first-order differential combined with random forest (FD-RF) showed optimal performance in all modeling strategies. Stratification modeling strategies and covariate modeling strategies significantly increased the SOC estimation accuracy compared with the entire dataset-based modeling. The geographic region-based stratification model (R = 0.788) had a higher estimation accuracy than soil type-, clay content-, and spectral type-based stratification models (R = 0.521-0.717), and the model synergizing spectral data with geographic region and clay content had optimal SOC estimation performance, with an R of 0.871, which was 44.21 % higher than the entire dataset-based model. This study demonstrates that using geographic region and clay content as covariates for modeling can reduce the influence of soil spatial heterogeneity on SOC estimation accuracy, and provides a new strategy for spectral estimation of soil properties.