Because it might not show the answer to a human? In their methods they say that they employed a technique called multiple endmember spectral mixture analysis which looks to me to be a probabilistic model which performs well over a large data set, like the one they draw from. It's not just a single image they use, but a bunch, each with small spectra differences which indicate areas of cover.

EDIT: Their methods also say that `Bull kelp forests are readily identified by multiple existing high spatial resolution satellite and airborne platforms because their floating surface canopies have strong reflectance in the near-infrared, similar to terrestrial vegetation`, so you wouldn't be able to see the difference anyways.

Are you looking at the paper or the image in the article?