Underwater AI: CDS Capstone Team Builds Fish Detection System for Sustainable Caribbean Tourism
By 10 meters underwater, the red channel of visible light has essentially disappeared, leaving behind the blue-green tint familiar from any reef video. That physical fact — along with turbidity, glare, and the sheer scarcity of clean training footage — turned out to be the central technical obstacle for a team of CDS MS students who built a computer vision system to identify Caribbean reef fish in raw video shot by an underwater robot. The project came together through the CDS Capstone Project, a course in the final year of the MS program that pairs students with industry partners and research labs to apply data science to real-world problems.
Working with the UNDP Accelerator Lab for Barbados and the Eastern Caribbean, CDS MS students Amaan Mansuri, Vishwa Raval, and Shravan Khunti developed a real-time pipeline that detects and classifies three reef species — surgeonfish, parrotfish, and grunts — from footage gathered by BlueBOT, an autonomous underwater rover built by Bajan Digital Creations Inc. The work was selected by the Japan Cabinet Office as part of the 2025 Japan SDGs (Sustainable Development Goals) Challenge and is now featured on the UNDP Barbados blog, with a video walkthrough demonstrating the prototype. The team’s poster, “Detecting and Classifying Species Biodiversity in the Waters of Barbados and the Eastern Caribbean,” was selected as a Best Capstone Poster of 2025.
The application is what the UNDP calls a Sustainable, Virtual AI Marine Prototype: a digital diving experience where tourists can explore reefs through labeled, species-aware video instead of physically entering the water. Reef tourism is central to the Caribbean economy, but every diver also stresses the ecosystem. A virtual alternative widens access for everyone — non-swimmers, people with disabilities, and the elderly — while reducing pressure on coral.
“Instead of the tourists going for the underwater dive each and every time, we’d just create a model which will give them a virtual dive within their edge device, like their mobile phone,” Raval said.
The same pipeline can also automate biodiversity monitoring, replacing the hours a marine biologist would otherwise spend hand-labeling ROV footage.
The team started with 93 raw videos shot in real reef conditions and built on YOLOv11, a real-time object detection model. As Khunti explained, YOLO — short for “you only look once” — handles both detection and segmentation in a single pass, which made it well suited for the small dataset and the under-70-megabyte size constraint required for the rover.
What they found surprised them. Increasing model size didn’t help — a 130-million-parameter model actually performed worse than their 9.4-million-parameter version. After more than sixty training runs hit an accuracy ceiling regardless of architecture, the team ensembled their five best models, pushing accuracy past their 70 percent target while staying within the size constraint.
“The problem with the accuracy percentage is not the model architecture, but the data quality,” Mansuri said. “If we increase the number of samples we have from 90 videos to maybe 200 videos, we might achieve more accuracy than what we’re getting right now.”
That finding — that dataset quality, not model capacity, was the binding constraint — is the kind of unfashionable result that rarely makes it into headline AI research but matters enormously for real-world deployment.
Each of the three brought prior computer vision experience to the project. Mansuri previously published research on multilingual handwritten digit recognition in IEEE. Khunti’s earlier work at NYU Langone Health applied computer vision, SLAM, and LiDAR to improve indoor navigation in New York City subway stations for blind and physically disabled commuters. Raval spent the summer of 2025 at Mount Sinai’s Icahn School of Medicine, where she optimized a Polygenic Risk Score generation pipeline used by clinicians and MDs on the research team. All three are from Gujarat, India, and were classmates in the CDS MS program before forming the team.
The project was mentored by Stevenson Antonio Hollingsworth of Bajan Digital Creations alongside Jordanna Straker and Veronica Millington of the UNDP Accelerator Lab, and grew out of a Blue Economy challenge the Lab first hosted in 2019. The team is in conversation with UNDP about extending the work, with a possible patent on the table.
“It’s a dream to go there once,” Mansuri said of the Australian reef. “I just want this thing to be impactful in such a way that I’m able to do the virtual dive.”
By Stephen Thomas
