Desenvolvimento de banco de dados para aplicação de algoritmos de machine learning na classificação de solos tropicais

Abstract

Correctly classifying soils used in engineering works is essential to contribute to structural safety and performance. However, traditional classification methods require detailed laboratory tests, which can be time-consuming and costly. Therefore, this study proposes an alternative approach, with the objective of structuring a database for future application of machine learning algorithms in the identification and classification of tropical soils in a faster and more accessible way. The soils that will feed the algorithm were selected based on the Tropical Compacted Miniature (TCM), which best applies to the reality of Brazilian soils, distinguishing lateritic and non-lateritic soils. For the database, Scanning Electron Microscopy (SEM) images collected from academic works in the areas of Geotechnics and Civil Engineering 4 including TCCs, dissertations and theses 4 were used. The data were treated with cleaning, normalization and correction of missing values. Although the model has not yet been fully implemented, it was possible to structure the database organization process and perform preliminary experiments with the convolutional neural network (CNN) architecture. The initial accuracy achieved was not satisfactory, which can be attributed to the model's serial approach, which did not consider the zoom level of the SEM images in the classification, in addition to the lack of refinement in the data used. As a proposal for improvement, it is recommended for future studies to develop a model with parallel input, which receives both the SEM image and the corresponding zoom level. In addition, techniques such as Data Augmentation can be used to improve data processing, increasing the diversity of examples and enhancing the algorithm's performance.