Estimating wet soil aggregate stability from easily available properties in a highly mountainous watershed

Abstract

A comparison study was carried out with the purpose of verifying when the adaptive neuro-fuzzy inference system (ANFIS), artificial neural network (ANN), generalized linear model (GLM), and multiple linear regression (MLR) models are appropriate for prediction of soil wet aggregate stability (as quantified by the mean weight diameter, MWD) in a highly mountainous watershed (Bazoft watershed, southwestern Iran). Three different sets of easily available properties were used as inputs. The first set (denoted as SP) consisted of soil properties including clay content, calcium carbonate equivalent, and soil organic matter content. The second set (denoted as TVA) included topographic attributes (slope and aspect) and the normalized difference vegetation index (NDVI). The third set (denoted as STV) was a combination of soil properties, slope, and NDVI. The ANN and ANFIS models predicted MWD more accurately than the GLM and MLR models. Estimation of MWD using TVA data set resulted in the lowest model efficiency values. The observed model efficiency values for the developed MLR, GLM, ANN, and ANFIS models using the SP data set were 60.76, 62.98, 77.68 and 77.15, respectively. Adding slope and NDVI to soil data (i.e. STV data set) improved the predictions of all four methods. The obtained correlation coefficient values between the predicted and measured MWD for the developed MLR, GLM, ANN, and ANFIS models using STV data set were 0.24, 0.35, 0.84 and 0.73, respectively. In conclusion, the ANN and ANFIS models showed greater potential in predicting soil aggregate stability from soil and site characteristics, whereas linear regression methods did not perform well.