Using visible-near infrared spectroscopy to estimate whole-profile soil organic carbon and its fractions

Abstract

Soil organic carbon (SOC) is crucial for soil health and quality, and its sequestration has been suggested as a natural solution to climate change. Accurate and cost-efficient determination of SOC and its functional fractions is essential for effective SOC management. Visible near-infrared spectroscopy (vis-NIR) has emerged as a cost-efficient approach. However, its ability to predict whole-profile SOC content and its fractions has rarely been assessed. Here, we measured SOC and its two functional fractions, particulate (POC) and mineral-associated organic carbon (MAOC), down to a depth of 200 ​cm in seven sequential layers across 183 dryland cropping fields in northwest, southwest, and south China. Then, vis-NIR spectra of the soil samples were collected to train a machine learning model (partial least squares regression) to predict SOC, POC, MAOC, and the ratio of MAOC to SOC (MAOC/SOC – an index of carbon vulnerability). We found that the accuracy of the model indicated by the determination coefficient of validation (Rval2) is 0.39, 0.30, 0.49, and 0.48 for SOC, POC, MAOC, and MAOC/SOC, respectively. Incorporating mean annual temperature improved model performance, and Rval2 was increased to 0.64, 0.31, 0.63, and 0.51 for the four carbon variables, respectively. Further incorporating SOC into the model increased Rval2 to 0.82, 0.64, and 0.59, respectively. These results suggest that combining vis-NIR spectroscopy with readily-available climate data and total SOC measurements enables fast and accurate estimation of whole-profile POC and MAOC across diverse environmental conditions, facilitating reliable prediction of whole-profile SOC dynamics over large spatial extents.

Keywords

• Mineral-associated organic carbon
• Particulate organic carbon
• Carbon sequestration
• Carbon vulnerability
• Machine learning
• Dry croplands