Segmentation of particulate organic matter in X-ray Computed Tomography images of soil aggregates with deep convolutional networks

Description

Abstract (English)

Identifying the spatial distribution of particulate organic matter (POM) in soil aggregates is essential to unravelling the mechanisms that favor the storage and stability of carbon in soils, as part of a strategy to mitigate the effects of global climate change and improve ecosystem functions. The location of carbon in soil aggregates affects whether it remains protected or is consumed by soil-dwelling microorganisms, thus indicating the need to decipher the role of soil structure in the physical protection of carbon. The use of X-Ray Computed Tomography (XCT) has enabled progress to perform in situ investigation of the processes involved in the decomposition and physical protection of POM in soils at the micron scale. However, one challenge associated with the use of this technique is the identification (or segmentation) of POM in XCT images, especially when chemical staining techniques are not considered. The present study aims to propose a method for direct segmentation of POM without the need for chemical staining. We built a database of organic fragments from 3D XCT images of soil aggregates, by segmenting them manually and by interactive machine learning strategies, which then fed the training of a neural network for POM segmentation by deep learning using the Annotat3D software. The 3D images were acquired on the XCT IMX beamline at the Brazilian Synchrotron Light Laboratory, with a voxel size of 0.82 μm. The convolution neural networks (V-net (3D) architecture) were trained using an artificial data creation resource (augmentation) and by mixing two different sets of images to diversify the input data. Lastly, we applied the proposed method to segment POM from 17 soil aggregates and found fair separation between the massive regions of POM and its internal pores. As the trained neural networks can be customized and optimized to be applied to various image sets, the approach proposed herein has the potential to be used in various studies aimed at providing a deeper understanding of the role of soil structure in controlling carbon processes.