Short summary is that this isn't going to work.
I was involved with a company doing some development on whale detection AI. The vast majority of the time you just don't see enough of the animal to make reliable detections. And that is in a best case scenario, add in movement of the boat/camera, waves, sunlight reflections, and having to ignore a vast majority of potential things on/in the water that can look kind of like whales, but are not whales. Dolphins in particular.
This isn't a problem for AI/machine vision, IMO.