Método baseado em machine learning e deep learning para segmentação da matéria orgânica particulada em imagens de Microtomografia de Raios X de agregados de solo
Data
2023-03-01
Autores
Oliveira, Aline Barbosa de
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A identificação da distribuição espacial da fração da matéria orgânica particulada (MOP) em agregados de solo é de fundamental importância para decifrar os mecanismos que determinam o aumento da estocagem e estabilidade do carbono no solo. A localização do carbono no agregado de solo pode afetar se ele permanecerá protegido ou será consumido por microrganismos que habitam o solo, por isso a necessidade de buscar compreender o papel da estrutura do solo na proteção física do carbono. O uso da técnica de Microtomografia Computadorizada de Raios X (μTCX) tem proporcionado avanços na investigação in situ de processos relacionados à decomposição e proteção física da fração MOP no solo. No entanto, uma dificuldade que limita o progresso do uso dessa técnica está na identificação (ou segmentação) da MOP em imagens de μTCX, especialmente quando não são realizados procedimentos de marcação química dessa fração. Por este motivo, o presente estudo tem como objetivo propor um método de segmentação da fração MOP. Buscou-se construir uma base de dados de fragmentos orgânicos, contidos em imagens de μTCX, segmentados manualmente, utilizando o software Avizo, ou por meio de estratégias de machine learning, no software Annotat3D, para treinar uma rede neural de segmentação da MOP via deep learning. As imagens 3D consideradas para desenvolvimento do método foram adquiridas na linha de luz de μTCX do Laboratório Nacional de Luz Síncrotron, com tamanho de pixel de 0,82 μm. Para o treinamento das redes, considerou-se: o uso de um recurso de criação de dados artificiais (augmentation) e a mistura de dois conjuntos de imagens diferentes para diversificação dos dados de entrada. As duas melhores redes neurais treinadas obtiveram identificação parcial ou total dos fragmentos orgânicos nas imagens de μTCX com uma estimativa de acurácia de 57% e 69%. Os resultados mostraram que o método foi bem sucedido para a segmentação da fração MOP em 17 agregados de solo, sem a utilização de marcação química. O recurso augmentation proporcionou melhor continuidade dos fragmentos orgânicos, e a mistura de imagens de diferentes conjuntos resultou em maior número de fragmentos identificados. Portanto, o método apresenta potencial para ser empregado em diferentes propostas experimentais com objetivos semelhantes, uma vez que a rede neural treinada pode ser adaptada e otimizada para a sua aplicação em diferentes conjuntos de imagens.
Identifying the spatial distribution of particulate organic matter (POM) fraction in soil aggregates is fundamental to deciphering the mechanisms that determine the increased of carbon storage and stability in soils. The location of carbon in the soil aggregate can affect whether it remains protected or is consumed by soil inhabiting microorganisms, hence the need to seek an understanding of the role of soil structure in the physical protection of carbon. The use of X-ray Computed Microtomography (μCT) has provided advances in the in situ investigation of processes related to the decomposition and physical protection of POM in soil. However, one difficulty limiting progress in the use of this technique is the identification (or segmentation) of POM in μCT images, especially in the absence of chemical staining procedures. For this reason, the present study aims to propose a method for direct segmentation of POM fraction, without chemical staining. We sought to build a database of organic fragments, contained in μCT images, segmented manually using the Avizo software, or by machine learning strategies, in the Annotat3D software, in order to train a neural network for POM segmentation through deep learning. The 3D images considered for method development were acquired on the μCT IMX beamline of the Brazilian Synchrotron Light Laboratory, with a pixel size of 0.82 μm. For the training of the networks, we considered: the use of an artificial data creation feature (augmentation) and the mixture of two different image sets for diversification of the input data. The two best trained networks obtained partial or total identification of organic fragments in the μCT images, with an estimated accuracy of 57% and 69%. The results showed that the method was successful in segmenting the POM fraction in 17 soil aggregates without the use of chemical staining procedure. The augmentation feature provided better continuity of organic fragments, and the mixing of images from different sets resulted in a greater number of identified fragments. Therefore, the method has the potential to be employed in different experimental proposals with similar objectives, since the trained neural network can be adapted and optimized for its application on different sets of images.
Identifying the spatial distribution of particulate organic matter (POM) fraction in soil aggregates is fundamental to deciphering the mechanisms that determine the increased of carbon storage and stability in soils. The location of carbon in the soil aggregate can affect whether it remains protected or is consumed by soil inhabiting microorganisms, hence the need to seek an understanding of the role of soil structure in the physical protection of carbon. The use of X-ray Computed Microtomography (μCT) has provided advances in the in situ investigation of processes related to the decomposition and physical protection of POM in soil. However, one difficulty limiting progress in the use of this technique is the identification (or segmentation) of POM in μCT images, especially in the absence of chemical staining procedures. For this reason, the present study aims to propose a method for direct segmentation of POM fraction, without chemical staining. We sought to build a database of organic fragments, contained in μCT images, segmented manually using the Avizo software, or by machine learning strategies, in the Annotat3D software, in order to train a neural network for POM segmentation through deep learning. The 3D images considered for method development were acquired on the μCT IMX beamline of the Brazilian Synchrotron Light Laboratory, with a pixel size of 0.82 μm. For the training of the networks, we considered: the use of an artificial data creation feature (augmentation) and the mixture of two different image sets for diversification of the input data. The two best trained networks obtained partial or total identification of organic fragments in the μCT images, with an estimated accuracy of 57% and 69%. The results showed that the method was successful in segmenting the POM fraction in 17 soil aggregates without the use of chemical staining procedure. The augmentation feature provided better continuity of organic fragments, and the mixing of images from different sets resulted in a greater number of identified fragments. Therefore, the method has the potential to be employed in different experimental proposals with similar objectives, since the trained neural network can be adapted and optimized for its application on different sets of images.
Descrição
Palavras-chave
Agronomia, Estrutura do solo, Matéria orgânica, Processamento de imagens