home contents changes options help

last edited 4 years ago by cranmer
These are some images of from LandSat? of the Amazon being deforested. In order to asses the economic extent of the logging (i.e. where / how much / species type / etc... of what is being logged combined with how much it costs to harvest) the fellas at Imazon need to identify the logging roads from the LandSat? images. The roads are unpaved for the most part, thus they can be identified from "soil fraction" layers that are produced from per pixel stastical analysis across several different (near) optical bands. The difficulty is that many of the roads appear to be broken and some roads are sitting in pastures that also have high soil fractions. The way they do it now is basically making some threshold cut on on a grey scale soil fraction image which produces a binary (black and white) image. The binary image is then fed into some other algorithms. The problem is that a simple threshold cut is probably not the optimal way to do it. For instance, the entire pasture passes the threshold cut, so it looks like one gigantic road. What would be nice is to try to provide some algorithms that chew on the greyscale (pre processed?) image and identify roads. Seems like a neural net might do pretty well, maybe something simpler. It would be also nice if there was some sort of algorithm that would piece together the broken / obscured parts of the road - even if it just enhanced pixels that were probably part of a road (colinear with other "road" pixels) that would be a big improvement. Below are a couple of images that might help. Click to see something bigger. If you have any thougths, please hit the edit button at the bottom right and share them. Thanks
Kyle

Seed list (i.e. "edit me"):