Using specialized software, such as various forms of image processing and artificial neural networks, precision agriculture solutions aim to generate accurate estimates of weeds and consequently, target automatic herbicide applications. One of the ways through which such estimations may be accomplished is the semantic segmentation of crop and weed regions. The paper proposes a two-stage segmentation neural network pipeline for tobacco and weed segmentation, trained and tested using a publicly available dataset of field images captured with a drone. The two stages are carried out by independent U-Net models with EfficientNet backbones. In the first stage, the model segments regions covered by vegetation, including tobacco plants and weeds. In the second stage, the model segments the tobacco plants. The final segmentation mask is created by combining the two intermediary masks. The model has good performance, with Dice scores exceeding 90% for crop and weed segmentation.