Classificação e segmentação de imagens de tomografia de solos via planos de entropia

Abstract

Although it is one of the first concerns in the study of the characterization of images in the most diverse contexts, the segmental process, that is, of separating objects of interest from the rest of the image is not always an obvious or trivial procedure. One of the materials in geosciences that undergoes more investigation, including in the area of segmentation of images, are the soils. They play diverse and critical roles in the biosphere, regulating biogeochemical cycles, providing a habitat for microorganisms and a medium for plant roots, storing nutrients and minerals and conducting fluids (such as water and gases). The identification of the pore structure of a given soil segment is important as it provides information on soil mechanics and how it can perform its function. In this sense, this thesis proposes a method for adequate segmentation, correctly identifying the pure phases of voids and solids, as well as mixed voxels, that is, those that do not immediately fit into voids or solids, this separation being one of the challenges in the segmentation of soil tomography images, as well as many other digital images. The construction of reflective entropic planes by the penetration of images in the direction of gravity using sample entropy and Shannon information allowed from the correlation of these planes to construct a multisegmentation curve that identifies pure phases from a valley and a peak, suggesting the method Entropic Plans Correlations (EPC). This curve was compared with the Pure Voxels Extraction (PVE) histogram, which also performs pure region searches and is then calibrated to correctly allocate the mixed voxels to the empty or non-void phases through edge detection theory. The proposed method of EPC, when compared with PVE, served the purposes of good image segmentation, since it obtained a high spatial correspondence and strongly diminished the subjectivity found in some stages of the latter. The entropic plans also provided a possibility of a consistent classification of these images, evidencing low density regions, that is, pores and organic matter and regions of high density, that is, harder rocks and minerals, as two relevant regions in the separation of soils of native forest and of the cultivation of sugar cane. These results showed that comparing entropic plans obtained from soil images, either through correlation or multivariate analysis, can bring relevant findings within the segmentation and classification of these images.