Leveraging Machine Learning for Autonomous Ultrasonic Mitigation of Harmful Algal Blooms Along the Pacific Coast

Harmful algal blooms (HABs) pose significant threats to aquatic ecosystems, public health, and local economies by degrading water quality, causing fish kills, and producing harmful toxins. Traditional methods of detecting and mitigating HABs are often labor-intensive and expensive, targeting only smaller water bodies. Additionally, many detection systems currently work on a microbial scale or focus only on detection rather than classification. Existing research indicates that ultrasonic waves at specific frequencies can effectively destroy algal cells, but upscaling is necessary before implementation in the field becomes practical. We address this by leveraging machine learning to locate and classify multiple types of oceanic HABs and then outputting ultrasonic emissions at varied frequencies corresponding to each targeted HAB. The system analyzes satellite images with a convolutional neural network (CNN) to identify and categorize different types of harmful algae, allowing for rapid, large-scale monitoring of water bodies and providing timely data for environmental management. To train the CNN, we also construct a novel dataset containing 4998 satellite images of algal blooms along the west coast of North America. After the visual processing, we identify the optimal ultrasonic wave frequency for each identified HAB. Our model successfully identifies eight different algal genera with an accuracy of 98.2%. To tackle practical application, we propose a prototype to implement our machine learning model for algal classification with an ultrasonic irradiation device. It consists of two parts: a drone for autonomous scanning across the ocean, and an underwater device for treatment. Thus, we demonstrate the potential of machine learning to advance ultrasonic mitigation towards large-scale implementation.