One problem, stated in the FAQ, is "The Terrapattern project uses tile-based search, not pixel-based search. In other words, the tool finds places that are similar to the map tile you clicked in—not the exact object you clicked on. As a result, the system may be serving up tiles that match your selection in ways that are different than the ones you're expecting."
So, it's accuracy would not be as high as visually searching. Terrapattern may be more like what Tn is doing to teach the computer to recognize shapes, which is why we humans assist with marking the computer's polys for accuracy. Just guessing, I suppose it could take billions of the humans' ratings before computers learn to detect shapes and objects versus shadows. However, Tn's computer algorithms are often fairly sharp, such as when it spotted just corners of cars parked under trees in a recent campaign! Of course, it still misread a tree's leaf edges as "corners"-- no car, just a tree. Or, when it picks out painted markings on streets, and marks them as vehicles... I always shake my head and say, "No, Server Johnny, not a car!" LOL (Just how do you teach a computer these nuances?!)