Implementation of NIRS Spectroscopy in Predictive Models for Organic Matter Fractionation in Volcanic Soils

Abstract

The fractionation of soil organic matter (SOM) and labile carbon (LC) plays a crucial role in soil health, agricultural productivity, and climate change mitigation. Nevertheless, most research has focused primarily on the quantification of total organic carbon (TOC), without adequately addressing its specific fractions, which limits a comprehensive understanding of carbon storage, stabilization, and decomposition processes in volcanic-ash-derived soils. This study aimed to evaluate the fractionation of SOM and LC using multivariate statistical models based on near-infrared (NIR) spectral information in volcanic-ash-derived soils under coffee production in Poás, Costa Rica. Soil samples were collected under different preparation and preservation conditions to determine the sensitivity of spectral signatures to physical changes and to assess their influence on model predictive capacity. Particulate organic carbon (POC), LC, and TOC were quantified, and Partial Least Squares (PLS) and Random Forest (RF) models were developed to predict these fractions. PLS models did not reach satisfactory predictive performance, whereas RF models exhibited consistently high accuracy under all study conditions. Dry and sieved samples produced the most accurate model for TOC (R2 = 0.936), fresh samples yielded the best model for LC, and dry samples were most suitable for POC, indicating that the optimal preparation depends on the fraction of interest. No significant differences were detected between fresh and stored-fresh samples, nor between dry and sieved-dry samples, supporting the applicability of the models under multiple preservation and processing conditions while simplifying laboratory workflows. Overall, the combination of NIR spectroscopy with RF modeling represents a robust and accurate approach for monitoring soil carbon fractions in tropical volcanic-ash-derived soils under coffee, reducing analytical time and complexity, and offering strong potential for implementation in sustainable soil management, agricultural extension programs, and specialized laboratory settings. Because these soils are dominated by short-range-order minerals in the andic sectors and models were validated using random k-fold rather than spatially independent partitions, these locally calibrated models should be validated before transfer to non-volcanic soils.