Neural Estimation of Localization and Classification of Soils for Use in Low-Traffic-Volume Roads

This research proposes the development of a method of localization and prediction of AASHTO soil classification that can contribute to the decision-making process in projects for low-traffic-volume roads in the metropolitan area of Fortaleza, Ceará, Brazil. Geoprocessing and artificial neural networks were used as modeling techniques, and biophysical and spatial variables were used to explain the modeled phenomenon. The characteristics investigated (pedology, geology, geomorphology, vegetation, altimetry, and position) were correlated with the AASHTO classification. The soil AASHTO classification data were taken from projects and preexisting studies and totaled 876 points. In the development of this geotechnical estimates generation model, artificial neural network topologies were calibrated, validated, and tested so as to find a model that could best fit the set of tests. The model had an accuracy rate of 92.6% for the generation of estimates of AASHTO soil classification in the Fortaleza region, on the basis of the biophysical variables studied. The tested model was used to construct neural geotechnical maps that can be used to predict the AASHTO classification of soils not yet characterized in the lab and to predict the type of subgrade of a region not yet explored. Results show that the artificial neural networks technique is promising for geotechnical studies of low-traffic-volume roads.