Soil Organic Carbon Fractionation Assessment in Areas with High Fire Activity Using Diffuse Spectroscopy and Tree-Based Machine Learning Algorithms

Abstract

Wildfires have a significant impact on Soil Organic Carbon (SOC) content and fractionation. Here we used Diffuse Reflectance Spectroscopy (DRS) and Machine Learning (ML) algorithms, particularly decision tree-based ones, to assess post-wildfire changes in labile, removable, and recalcitrant SOC fractions across six distinct areas with varying wildfire recurrence levels. Once spectral data had been acquired by DRS, the transformation of raw data through first and second derivatives enhanced the resolution of the measurements. In addition, Quantile Random Forest (QRF) emerged as the best algorithm to optimize unbiased models, with a notable goodness-of-fit. However, while QRF excelled in predicting recalcitrant C, it yielded slightly lower precision for the most labile C fraction (cold-water extracted C), with R2 and rRMSE (%) ranging from 0.62 to 0.83 and from 78.70 to 7.99, respectively, accompanied in both cases by acceptable RPD statistics. Moreover, the study underscored the importance of the NIR spectral range in accurately predicting SOC fractions. Moreover, our findings highlighted specific spectral regions related to clay content and organic C features, which are crucial for understanding post-wildfire SOC dynamics and useful for future determinations by remote sensing (drone, satellite).