Classification of expansive soils among problematic soils in the brazilian semiarid from artificial neural networks

Abstract

  The climatological and geomorphological conditions of the Brazilian semiarid region favor the formation of potentially expansive soils. The development of neural networks to identify and classify expansive soils in Pernambuco, and the need to export to soils across the semiarid region of Brazil is investigated in this study using a Multi-Layer Perceptron (backpropagation). The neural network was developed using 87 experimental data points, divided into three groups. The Training Group, consisting of 53 samples, using data inputs: sand percentage, clay percentage, plasticity index, activity index, and geological, pedological, and climatological classification. Selection Group, made up of 17 samples, was used to select the best network architecture, genetics algorithms formed of 7 inputs and 3 hidden nodes. The Test Group used 17 samples to evaluate the predictive accuracy of the expansion potential, obtaining an accuracy rate of 88,2%. This network was validated by applying it to 67 samples of problematic soils of the collapsible, expansive, and soft type from the Brazilian semiarid region, reaching an accuracy of 76,1%. Probabilistic Neural Networks were found to be efficient in evaluating the behavior of expansive soils, with the ability to deal with the absence of sample input data, demonstrating the ability to capture movement trends in the expansion of the soil surface, indicating the functions that introduce the effects of the composition potential on the expansion behavior, and determining the limit values of each of the input variables for the samples from the database used.