Advancing Seagrass semantic segmentation with SAM2 models

Seagrass ecosystems in shallow waters are important for assessing environmental factors such as pollution and global warming, because they are highly sensitive to changes in water quality and temperature. This paper investigates advanced deep learning techniques for semantic segmentation of Seagrass images collected from shallow water areas, focusing on fine tuning the Segment Anything Model 2 (SAM2) and SAM2-UNet architectures. We conducted extensive experiments on five SAM2 models to optimize segmentation performance on two subsets of data obtained at different depths of the ‘Looking for Seagrass’ dataset. Additionally, we performed an ablation study using the best SAM2 model with a lesser number of annotated training set images. Furthermore, we analyzed the inference time of the different models to provide insights for further model selection. The experiment results demonstrated the effectiveness of the SAM2 models, with the highest Intersection over Union (IoU) of 0.923 and 0.935, and Pixel Accuracy (PA) of 0.967 and 0.972 in two sub-datasets; obtained using the SAM2-UNet model; outperforming the method provided in the original dataset paper. The ablation study using a lesser number of annotated images in the training dataset provided similar accuracy as the second-best performing model. The total processing time was the least for the SAM2-UNet model. This work also highlights the potential of real-time Seagrass monitoring using drone and underwater camera footage.