Estimating mineral-associated organic carbon saturation and sequestration potential using MIR spectral based local quantile regression

Abstract

Associating with mineral surfaces, mineral-associated organic carbon (MAOC) is able to persist against fast decomposition via chemical bonding or physical occlusion, considered as key to soil organic carbon (SOC) stabilisation. In this study, the feasibility and capability of using mid-infrared (MIR) spectral models to predict MAOC and optimising the estimation of theoretical MAOC saturation limits was tested. Based on measured MAOC from physical carbon fractionation, the spectral MAOC model (R2 = 0.86, RMSE = 4.41 g C kg−1) predicted MAOC values from a large regional scale spectral library. Based on measured MAOC from physical carbon fractionation, the model with a medium RMSE (R2 = 0.86, RMSE = 4.41 g C kg−1) among 41 randomizations was identified as the most generalized and was selected to predict MAOC values from a large regional-scale spectral library. As SOC increased, the rate of MAOC accumulation diminished, indicating the presence of a theoretical saturation limit. Hence, quantile regression at 95th was performed on the whole dataset based on the relationship between MAOC and silt + clay to estimate theoretical MAOC saturation limits. Using this approach, estimated theoretical MAOC saturation limits was 67.5 ± 2 g C kg−1 with a 95 % confidence interval ranging from 64.0 to 71.4 g C kg−1. To advance this, a new data-driven approach combining quantile regression and MIR spectral library was proposed using a spectral neighbourhood framework, called ‘local quantile regression’, to improve the estimation of theoretical MAOC saturation limits in quantile regression. By defining neighbourhoods around each soil sample based on spectral dissimilarity, quantile regression was conducted within these neighbourhoods, and inverse distance weight averaging was applied to improve the robustness of the estimates. MAOC theoretical saturation limits estimated in local quantile regression varied from 44 g C kg−1 to 82 g C kg−1. In contrast to the constant theoretical upper limit in global quantile regression, local quantile regression using MIR data captures chemical information, specifically, clay minerals related to carbon storage that offers potentially more realistic assessment of MAOC saturation. Moreover, based on correlation analysis and variable importance used in random forest model, soil mineralogy related properties, such as CEC and different cations, followed by land management related covariates, like available phosphorus and climatology, were identified as primary and secondary driving factors behind this variation of MAOC saturation limit. Hence, local quantile regression provided a conservative but more feasible MAOC sequestration target, overcoming limitations in global quantile regression and offering a better framework for regional-scale carbon sequestration estimation. Based on local quantile regression estimated MAOC saturation, the MAOC sequestration potential was calculated, reaching 53.04 Mt C could be sequestered in total at 5–20 cm as MAOC on mineral soil in grassland in the northern half of Ireland, revealing a huge potential for C sequestration.