One of the chapters of my thesis (in review for publication) dealt with classifying different types of forest in a hyperspectral image. I wanted to track the expansion of tree plantations that were being subsidized by the Costa Rican government. And boy, Internet, was it a tough row to hoe. Working with hyperspectral imagery is computer-intensive, tricky, and full of pitfalls and challenges. But it is AMAZING what you can pull out of the data--there's just so much of it! Some of my image files were 50 GB in size, and with them I was able to classify tree plantations with sufficient (and satisfying) accuracy and make a great map.
I will give you that Tan et al.'s new algorithm is most likely incredibly computer-intensive, and it worked with a limited (and quite distinct) set of classes. Obviously I am not quite out of a job yet. But computer processing speeds are only increasing, and the cloud is doing more and more processing for scientists. It is entirely likely that one day, when I start a hyperspectral analysis, 95% of my work may already be done by the computer.
And I am perfectly happy with that. The sooner that we can relegate supervised classification (a time-intensive, iterative data collection process) to a minor component of the overall image analysis process, the better. If I could've skipped the hard work of delineating forest from nonforest using an algorithm and gone straight to delineating tree plantations, that would have been great. And if algorithms improve and I could skip the process entirely, just input my classes and get a tree plantation map, even better! Nothing is sweeter in science than asking a great question and getting the answer the very next day.
But wait, you may say. Doesn't that mean that you, an image analyst, would be out of a job? Well, there's a few things you should consider. First, I am interested in using satellite and aerial imagery for conservation and answering ecological questions. So automated techniques would only speed up my work and let me test more hypotheses. Second, there will always be technical challenges to overcome; working with images from space or a speeding plane or UAV is not straightforward, and you need experience and expertise to problem-solve. Third, there will always be new frontiers in remote sensing that require expertise and algorithm testing. Hyperspectral lidar, high-resolution thermal imagery, and hyper-temporal satellite imagery are frontiers that are just now opening.
I get excited just thinking of the new, innovative analyses that we will be able to do with new automated algorithms and new sensors. Remote sensing and GIS are in their early decades still, and the science and science applications are expanding rapidly. I have a friend who works with hyperspectral imagery of a different kind, of people. He is using it to do accurate, automated classification of skin blemishes, to improve cancer screening. And that, I think, is a good thing.