Development of a Remotely Operated Underwater Drone for Crown-of-Thorns (COTS) Starfish Detection using Simple Convolutional Neural Network

The proposed system is an underwater ROV with a bottom-facing camera that records underwater footage of the coral reef environment. The design of the drone is based on the BlueROV2 model and was deployed in both controlled and open sea water for evaluation. The underwater drone is tethered to a topside computer with the developed Crown-of-Thorns Starfish (COTS) detection and counting software, wherein, the video footage from the drone is imported. The final number of the COTS population on the surveyed area will be tallied. Multiple machine learning models that focus on image detection were trained using a custom COTS dataset, the SSD MobileNet v1 with FPN and 640x640 resolution proved to be the most stable model fit, with a mAP score of 81.53%. A Two-Tailed Z-Hypothetical Testing was also conducted to validate if the observed mean of the detection and counting system than the actual object count. A map layout will also be generated postdeployment through panoramic image stitching from the recorded video footage, this will be laid out using QGIS3 software georeferenced from the surveyed locations. This system provides a wider coverage, a more accurate statistics, and a visual image of the underwater coral reef environment with lesser resources and risks per deployment compared to manual dive surveys.