- Dalcimar Casanova (Master's advisor - UTFPR Pato Branco)
- Heitor S. Lopes (Teacher - UTFPR Curitiba)
- Thiago H. Silva (Teacher - UTFPR Curitiba)
Accurate soil classification is crucial in several areas, for instance, agriculture and geology. However, traditional soil classification approaches, which involve considerable labor of trained professionals in field and laboratory research, are costly and do not scale easily, especially in places with diversified terrain. Thus, this study proposes a new soil classification method that uses modern computer vision and machine learning methods to classify soil images. The proposed approach, jointly with a domain expert, could speed up soil classification, reduce costs, and leverage new applications, especially in the developing world, where resources can be scarce. The proposed intelligent model is designed for two types of soil: oxisols (U.S. Soil Taxonomy) and inceptisols (U.S. Soil Taxonomy), which are very common in Brazil, especially in the South region. The proposed approach achieves an impressive balanced accuracy in the test set of 100.0%. This approach, together with the knowledge of a domain expert, could be useful to speed up soil classification helping to reduce costs and leverage new applications.
Keywords: Oxisol, Inceptisol, Transfer-Learning, Resnet50