Machine learning-based prediction of soil aggregate stability for sustainable land management in the eastern Himalayan region of India

Abstract

Plot-scale rainfall-induced soil erosion data are scarce in the eastern Himalayas of India. Soil aggregate stability (SAS), commonly quantified by mean weight diameter (MWD), is a key indicator of erosion susceptibility. As laboratory determination of MWD is labor-intensive, machine learning (ML)-based pedotransfer functions (PTFs) using readily measurable soil properties offer an efficient alternative. In this study, 126 topsoil samples (0–15 cm depth) were collected from agricultural (n = 61) and forest (n = 65) lands in a sub-watershed of Sikkim, India, to develop and evaluate ML-based models for soil MWD prediction. Eleven soil physico-chemical parameters, i.e., ten ‘input’ (%Sand, %Silt, %Clay, %OC, BD, pH, Ca 2+ , Mg 2+ , CO 3 2− , and HCO 3 −), and one ‘target’ (MWD), estimated in the laboratory, were used in this study. Three supervised ML techniques, such as ANN, RF and SVM, along with empirical MLR, were used to predict soil MWD. The mean BD, %OC and MWD of agricultural land were 1.25 g/cm 3 , 3.37% and 1.09 mm, respectively, whereas those of forest land were 1.17 g/cm 3 , 4.23% and 1.11 mm, respectively. The sand fraction was found more in agricultural land (61.34%), while the clay proportion was found more in forest land (30.11%). All three ML models performed better than MLR in predicting soil MWD. The RF model performed best for the agricultural land (R 2 = 0.89, NSE = 0.89 for training; R 2 = 0.66, NSE = 0.66 for testing), while the ANN model was found to be superior for the forest land (R 2 = 0.92, NSE = 0.92 for training; R 2 = 0.91, NSE = 0.91 for testing). The SVM model predicted soil MWD more accurately for the pooled dataset (R 2 = 0.98, NSE = 0.98 for training; R 2 = 0.72, NSE = 0.72 for testing). No single technique is superior for all the land uses; however, the SVM model is either best or second best in all the cases. Therefore, the developed SVM-based SAS model can be employed to accurately predict soil MWD across the eastern Himalayan region and facilitate regional soil erosion investigations.