Researchers at the California State University Monterey Bay, The Ocean Cleanup, and the University of California, San Diego, have developed and open-sourced an application that uses Deep Learning techniques to detect and quantify subsurface marine plastic using underwater photography.
Oceans all around the world have a pertinent issue. The increasing amount of sub-surface plastic debris has caused a severe threat to marine life and the environment. Plastics can cause damage to marine life via ingestion, suffocation, or restraint. Understanding marine plastic and removing it is critical to helping the oceans.
Current plastic quantification, monitoring, and removal, such as Manta Trawls, manual-collection are cost and labor-intensive. To solve this problem, researchers from the California state university at Monterey Bay, The Ocean Cleanup, and The University of California San Diego, in a joint effort, have developed a method that automates the process of detecting subsurface plastic. “DeepPlastic ” enables researchers and citizen scientists to use underwater photography (which can be attached to AUV’s and buoys) to detect marine plastic.
In the words of Gautam Tata- Lead Author and Researcher from CSU Monterey Bay, “Underwater photography mixed with Analytical Deep Learning Algorithms are a more reliable, faster and efficient way to understand, monitor and remove marine plastic.”
“There are several factors in the ocean, such as occlusion, turbidity, and noise, making it harder to detect marine plastic automatically. This is one of the main reasons why a handful of studies focused on automating the process. While automated sensing technologies are still at an early stage, we are confident the with future studies, our method could be applied to real-world scenarios”.
This method of automating the detection of plastic debris uses “Deep Convolutional Neural Networks,” a branch of computer science that mixes Artificial Intelligence with Computer Vision. The researchers went to the field to hand-collect and annotate more than 4000 images of marine plastic to train the model.
The algorithm was tested and validated on images the model had never seen before. The results were significant – 85% accuracy. This meant that the Neural Network could accurately detect subsurface plastic around 85% of the time. This was with a confidence score of 50! (i.e., The model has to be more than 50% confident that the object in focus is marine plastic — Usually, the higher the confidence score, the better, but in this situation, with multiple oceanic factors coming into play such as occlusion, the confidence score was more appropriate).
THE WAY AHEAD
While the team is in the process of getting their paper published into a journal, they are testing other methods to synthesize images without needing to go into the field. The team is also working with other researchers at the Monterey Bay Aquarium Research Institute to test the expansiveness of their model.
Watch their model in action: https://www.youtube.com/watch?v=QK2YZL3TCZ0