Deep learning segmentation of soil constituents in 3D X-ray CT images

Accurate segmentation of soil constituents in X-ray CT imagery is critical for advancing our understanding of soil structure dynamics. However, difficulties arise because of the overlapping X-ray attenuation of soil constituents, which makes segmentation based on voxel intensity alone impossible. In this study, we explore the potential of nnUNet, a deep learning-based semantic segmentation model, for applications in soil science. We evaluated nnUNet on three challenging datasets: (1) complex soil structure with numerous material classes with overlapping gray value ranges, (2) fine roots in noisy images, and (3) permafrost with gradual gray value transitions between sediment types. The performance of nnUNet was compared to other reference methods, namely Ilastik, Rootine v.2, and manual thresholding. The Dice scores indicated overall good model performance across all three datasets. Compared to the reference methods, nnUNet outperformed Ilastik on Dataset 1. For Dataset 2, nnUNet produced segmentations with fewer false-positive roots than Rootine v2; however, this came at the expense of missing fine roots that were barely visible to the naked eye and, therefore, not annotated. In Dataset 3, we encountered challenges in annotating classes due to gradual transitions in voxel intensity. Our study underscores that deep learning models like nnUNet perform well for the segmentation of complex soil structures and could assist the development of a generalized segmentation model, thereby fostering standardization in soil structure analysis. To achieve this, increased access to diverse and well-curated annotations is still necessary.