Detection of invasive submersed aquatic vegetation using a machine learning-based object detection program

Invasive submersed aquatic vegetation (SAV) is an ecological threat to all waterbodies. SAV are species that are introduced to an area outside of their natural ecosystems and will thrive and outcompete native species. This can lead to the devastation of an ecosystem by eliminating the natural species resulting in a consequent upheaval in food supply for animal life and, potentially, a hazard to human health and activities such as boating and fishing. Point intercept surveys are a standard tool to detect SAV. However, they are prone to missing patches of low-abundance species. To solve this issue, we propose a method of identifying SAV in real time by utilizing an object detection program. Unfortunately, there is little data of the necessary type to facilitate training of a machine learning (ML) algorithm already existent. Therefore, we collaborated with researchers knowledgeable of both field work and the subject matter to obtain the necessary data. An underwater remote operated vehicle (ROV) equipped with a camera was utilized in several different water bodies for video acquisition. After collection of a substantial dataset, frames were extracted from the gathered videos and annotated by a group of individuals for detection of the SAV of interest (dioecious Hydrilla verticillata). The frames and annotations were then collated into a singular dataset to use as the training dataset for a ML model. A popular object detection model (EfficientDet) was adapted to detect hydrilla in both video and images. Our model detected hydrilla with an intersection of union (IoU) of 0.5 at an accuracy of 81.2% and had a peak mean average precision (mAP) of 58.2% based on the COCO evaluation metric. Future plans include adding more species for detection to the model and integrating the model in an existing ROV platform for live detection of SAV.

PRESENTER

Jeong, Han Saim
han.s.jeong@erdc.dren.mil
214-604-5720

ERDC CEERD-ITL

CO-AUTHORS

Cheng, Jing-Ru
Ruth.C.Cheng@erdc.dren.mil

Schad, Aaron
Aaron.N.Schad@usace.army.mil

Rycroft, Taylor
taylor.e.rycroft@usace.army.mil

CATEGORY

Artificial Intelligence / Machine Learning usage for HPC Applications

SYSTEMS USED

Nautilus

SECRET

No