Enhanced implementation of structural equation modeling for bulk density, carbon and carbon-to-nitrogen ratio within a boreal study region in Northern Ontario, Canada

Structural equation modeling (SEM) was explored to discern linkages between bulk density (BD), total carbon (C) concentration and carbon-to-nitrogen ratio (C:N) at the 5–15 cm depth of mineral soil within a boreal study region experiencing landcover conversion. A framework for facilitating SEM was devised by encompassing a multiapproach feature selection of environmental covariates relating to soil formation factors, with mitigation for multicollinearity. SEM yielded enhanced interpretability and structured representation of interrelated soil properties. Supportive goodness of fit metrics were attained with comparative fit index (CFI) of 0.969 > 0.95 and standardized root mean square residual (SRMR) of 0.037 < 0.08. Validation coefficient of determination (R2) for C with SEM increased to around 0.4 from 0.2 with a baseline random forest (RF) trained on the full predictor set. Gains in R2 for RFs trained on features retained for SEM versus baseline RFs improved by 0.15 and 0.24 for BD and C, respectively. Slope attributes in conjunction with canopy height model were important for BD, corresponding to land use differentiation between agricultural and forested locations. BD, digital elevation model and plan curvature were significant for C, relating to forest cover, with C and ground wetness covariates important for C:N.