To understand how processes in ecosystems work and how they are connected the analysis of soil systems is essential. Since traditional computer vision methods for analysing soil cores reach their limits the next step is to integrate deep learning methods. Therefore a sufficient amount of labeled ground truth data is needed. Since labeling this large data is a tedious and time consuming procedure an efficient annotation strategy is developed and following a semantic segmentation model is trained. Thus, it is sufficient to label only 0.5% of the data to obtain high quality predictions. The trained model can then be used to label the missing 99.5% of the training data or to predict completely new images. The obtained information can be used to analyse the soil structure or the flow of water as well as to measure the biological activities.