Dharumarajan S., Gomez Cécile, Kusuma C. G., Vasundhara R., Kalaiselvi B., Lalitha M., Hegde R. (2025). Prediction of soil organic carbon stock along layers and profiles using Vis-NIR laboratory spectroscopy. Catena, 257, p. 109150 [13 p.]. ISSN 0341-8162.
Titre du document
Prediction of soil organic carbon stock along layers and profiles using Vis-NIR laboratory spectroscopy
Dharumarajan S., Gomez Cécile, Kusuma C. G., Vasundhara R., Kalaiselvi B., Lalitha M., Hegde R.
Source
Catena, 2025,
257, p. 109150 [13 p.] ISSN 0341-8162
There is a growing need to estimate soil organic carbon (SOC) stocks at both local and global scales. This study explores the use of Visible-Near Infrared (Vis-NIR) laboratory spectroscopy as an alternative to traditional wet chemistry methods for SOC stock estimation. Two approaches were tested: an indirect method, which uses Partial Least Squares Regression (PLSR) models to predict SOC content and bulk density separately and then multiplies them by measured layer depth; and a direct method, where PLSR models predict SOC stock per layer directly. The estimates were then aggregated to calculate the total SOC stock per profile. We evaluated both approaches using 361 samples from 84 soil profiles collected across three villages in Kerala, India. Two calibration scenarios were tested: (i) non-clustering, where 75 % of the dataset was used for calibration and 25 % for validation, and (ii) clustering, where models were trained on samples from two villages and validated on the third. The results showed that the indirect approach consistently outperformed the direct approach, both at the layer and profile scale. The non-clustering calibration scenario provided variable accuracy, with R2val values ranging from 0.52 (direct approach) to 0.70 (indirect approach). The clustering scenario produced more variable results depending on the calibration set used. Overall, this study confirms that Vis-NIR spectroscopy is a promising, rapid, nondestructive, and cost-effective method for SOC stock estimation. However, scaling up its application across agricultural landscapes will require substantial data collection and further methodological refinement.
