This master thesis describes using convolutional neural network models for blue carbon estimation. The dataset is gained through drone imagery and ground truth surveys. The estimation is of a seagrass type called eelgrass (Zostera marina). Different types of datasets have been used to train and validate the models. The first model was given sections of stitched orthomosaics as input and classified as to whether the sections contained seagrass. The classified seagrass sections were used as input for the next model, which classified the density of the sections. Furthermore, the density classified sections were multiplied with reference values found in the literature for carbon content in seagrass biomass. The obtained ground truth was used to train the models, not for calibrating the carbon estimation due to delays in the laboratory work. The models did provide a quantitative way for carbon budgeting with density predictions. Although the results were not as good as hoped for, the method might improve with more training data.