Presentation of the three training workflows of the segmentation models presented in this paper. Model A: two-class segmentation model, Model B1: three-class segmentation model with cascade training method and model B2: three-class segmentation model with direct training methodIn this study, DeepLab model (Chen et al., 2018b) with a ResNet-101 backbone (Wu et al., 2019) trained on COCO 2017 dataset (Lin et al., 2014) of 21 classes is used to train the three models. The final layer of the pretrained model was modified to predict either 2 or 3 categories, depending on the task. Transferring the weights of a semantic segmentation model trained on vast dataset to retrain the models permits to fine-tune the models for a new task with a minimum number of images, which also accelerates convergence during training.