Segmentation of soil constituents in X-ray CT imagery is critical for advancing our understanding of soil structure dynamics, but challenges arise due to the overlapping X-ray attenuation. 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) heterogeneous soil structure with numerous material classes sharing gray value ranges, (2) fine roots in noisy images, and (3) permafrost core 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. For Dataset 1, the segmentation accuracy surpassed that of ilastik for the particulate organic matter class. Dataset-specific challenges, such as annotating fine structures (Dataset 2) or poorly defined boundaries between material classes (Dataset 3), impacted the performance of nnUNet, compared to more specialized methods. The mean validation Dice scores were 0.71, 0.90, 0.92, for Dataset 1, 2 and 3, respectively, which suggested overall good model performance. 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.