Use of joint supervised machine learning algorithms in assessing the geotechnical peculiarities of erodible tropical soils from southeastern Nigeria

Abstract

ABSTRACT Multiple machine learning algorithms were integrated in this study to assess the geotechnical peculiarities of tropical soils from erosion sites in Nigeria. Laboratory analyses of the soils, which followed standard methods, revealed that they are erodible in nature. Results of correlation, principal component and factor analyses revealed the relationships between geotechnical variables, which were later used for artificial neural network (ANN) modelling. Soil particle distribution was predicted and analyzed using ANN1 (with sigmoid output activation) and ANN2 (with identity output activation). However, ANN2 gave more reliable prediction than ANN1, with R2 averaging 0.913 and 0.522, respectively. Low ANN model errors were also reported. Furthermore, soil erodibility potential, with emphasis on the grainsize distribution, was predicted using logistic regression analysis (LRA). The LRA results showed that the model accurately classified soil erosion events by 90%, and further revealed that sand content is the priority influencer of soil erodibility, more than gravel and fines contents. Thus, the likelihood of high soil erosion events in the area increases with sand %. The logistic regression model was tested for reliability based on Cox & Snell and Nagelkerke R-squares – R2 = 0.593 and R2 = 0.791, respectively – indicating that the model is acceptable and reliable.