Comparative analysis of random forest, exploratory regression, and structural equation modeling for screening key environmental variables in evaluating rangeland above-ground biomass

Abstract

Above-Ground Biomass (AGB) is a key indicator of rangeland health and productivity, as well as ecosystem conservation and resource sustainability. Estimating rangeland AGB under complex environmental conditions requires selecting appropriate predictors and models. In this research, we compared three methods to select the most important predictors of AGB from sixteen environmental variables: soil (pH, Nitrogen (N), Phosphorus (P), Potassium (K), Organic Matter (OM), clay, silt, sand), climate (Mean Temperature (MT), Mean Precipitation (MP), Actual Evapotranspiration (AET), Land Surface Temperature (LST)), and topography (elevation, slope, aspect, hillshade). The methods were Random Forest (RF), Exploratory Regression (ER), and Structural Equation Modeling (SEM). Based on the selected predictors from each method, we developed AGB estimation models using multiple linear regression (MLR). We evaluated the precision and accuracy of the models using the Ideal Point Error (IPE). The results showed that the three methods differed significantly in their selection of key predictors. RF identified sand, MT, MP, slope, silt, and N as the most important predictors based on %IncMSE criteria. ER selected sand, MT, P, and hillshade as the optimal combination of explanatory variables based on adjusted R2, JB, VIF, and SA (0.50, 0.9, 1.22, and 0.5, respectively). SEM revealed that N, P, K, pH, silt, MP, MT, slope, and hillshade were the most effective predictors based on CMIN/DF = 1.33, CFI = 0.96, GFI = 0.9, RMSEA = 0.06. The validation results based on IPE indicated that SEM performed the best (IPE = 0.17) compared to ER (IPE = 0.18) and RF (IPE = 0.19). Although the three models - RF, ER, and SEM - do not have the capability to directly predict and map the response variable from selected important drivers, SEM provides a direct understanding of the interrelationships among predictor and response variables through data summarization. This makes it more reliable than RF and ER.