Using Deep Neural Networks to Identify Natural and Anthropogenic Disturbances in Seagrass Meadows Observed in Side-scan Sonar Images

Shallow-water coastal ecosystems suffer from deterioration and destruction of seagrass habitats, evident in both naturally formed bare patches and propeller scars left by motorized watercraft. Quantifying the severity and rate of seagrass degradation is therefore crucial for monitoring the health of these vital marine ecosystems. Compared to traditional remote-sensing techniques such as optically based methods, sonar mapping has been successful due to its high resolution, relatively low cost, and independence from meteorological conditions and water turbidity. However, automatically quantifying seagrass abundance and localizing features of seagrass degradation from sonar images has remained a challenge. Using side-scan sonar data collected from the Lower Laguna Madre in Texas, we trained a set of deep neural networks for identifying, localizing, and classifying disturbances in sonar images of seagrass meadows: one model for bare patches, and one for propeller scars. Our approach is therefore capable of discriminating between natural and anthropogenic disturbances. The bare patch model performs with a binary accuracy of 98.3%, and the propeller scar model has an accuracy of 99.5%. We will present the neural network approach, its strengths and weaknesses, and further information on our ongoing research for robust mapping of seagrass disturbance from side-scan sonar imagery. [Work supported by ONR and ARL:UT IR&D.]