Last week I got started on some (very preliminary) work for my actual project. We are looking for a good, reliable way to find/estimate tree mortality from forest fires, based on some kind of remote-sensed data. The easiest would be Landsat, because it's ubiquitous and free and there is some information already out there on that. The other thing we'll investigate will be LiDAR, which is more complex and expensive and I haven't started playing with it but I'm sure I'll talk about it more later.
But for now, I'm playing with known indicators of fire severity that can be derived from Landsat data. I'm doing the processing in GIS, which is the thing that I came back to school to learn, basically, so I'm excited about it (I'm sure that will wear off, soon enough). Landsat data (Landsat 7 if you want to get particular) comes in 8 bands - the first 3 are portions of the visible spectrum and roughly correspond to blue-green, green, and red, the next four are slices of infra-red, and the last is some silly thing that no-one uses.
There are four commonly used measures of fire severity: NBR (normalized burn ratio), NDVI (normalized differenced vegetation index), dNBR, and dNDVI (these are just the postfire thing minus the prefire ). One is calculated as (band 4 - band 7) / (band 4 + band 7) and the other is the same formula, different bands.
The data (photo, basically) has a 30 meter pixel size - what do you want, from a satellite? - so what you end up with is a 30 meter grid of some arbitrary values that you can arbitrarily assign color values to, which looks pretty much like any other false-color aerial photo if you view it at a pretty large scale. Unfortunately, I didn't screencap that when I was at school, so you'll have to take my word for that. Anyway, what we want to do, to test the accuracy, is examine the values ONLY for places where we know what died and what survived.
I suppose this would be a good time to back up and say that the area I'm studying is the school forest up around Davenport, which caught on fire in 2009. We have really, really good data going back to the early 90s on all the trees in a series of 1/5 acre circular plots, laid out (nominally) on a 500 foot grid.
So, then, what you see below is a small area of the forest, with NDVI symbolized in grayscale, and the same NDVI within the plots symbolized in color:
I found this was a nice way to visualize what was going on - the plot data jumps out because of the color, but you can see the context from the background. At this scale, the 30m pixel looks incredibly blocky, and yet, keep in mind that this is an area roughly a mile square. Landsat is really only good for broad-brush interpretation, which is one of the things we'll be trying to improve on, later on.
Here is the dNDVI for comparison:
But for now, I'm playing with known indicators of fire severity that can be derived from Landsat data. I'm doing the processing in GIS, which is the thing that I came back to school to learn, basically, so I'm excited about it (I'm sure that will wear off, soon enough). Landsat data (Landsat 7 if you want to get particular) comes in 8 bands - the first 3 are portions of the visible spectrum and roughly correspond to blue-green, green, and red, the next four are slices of infra-red, and the last is some silly thing that no-one uses.
There are four commonly used measures of fire severity: NBR (normalized burn ratio), NDVI (normalized differenced vegetation index), dNBR, and dNDVI (these are just the postfire thing minus the prefire ). One is calculated as (band 4 - band 7) / (band 4 + band 7) and the other is the same formula, different bands.
The data (photo, basically) has a 30 meter pixel size - what do you want, from a satellite? - so what you end up with is a 30 meter grid of some arbitrary values that you can arbitrarily assign color values to, which looks pretty much like any other false-color aerial photo if you view it at a pretty large scale. Unfortunately, I didn't screencap that when I was at school, so you'll have to take my word for that. Anyway, what we want to do, to test the accuracy, is examine the values ONLY for places where we know what died and what survived.
I suppose this would be a good time to back up and say that the area I'm studying is the school forest up around Davenport, which caught on fire in 2009. We have really, really good data going back to the early 90s on all the trees in a series of 1/5 acre circular plots, laid out (nominally) on a 500 foot grid.
So, then, what you see below is a small area of the forest, with NDVI symbolized in grayscale, and the same NDVI within the plots symbolized in color:
I found this was a nice way to visualize what was going on - the plot data jumps out because of the color, but you can see the context from the background. At this scale, the 30m pixel looks incredibly blocky, and yet, keep in mind that this is an area roughly a mile square. Landsat is really only good for broad-brush interpretation, which is one of the things we'll be trying to improve on, later on.
Here is the dNDVI for comparison:
The idea behind this is that, by subtracting the calculated values from before the fire, you're removing any variation not caused by the actual fire. One thing I'll be looking into is whether this is actually any better. One thing that you might have noticed is that our plots are very close to the same size as the pixels (6.9 feet broader, but who's counting) - pure coincidence. But it makes calculating an average value for each plot, which I have to do, interesting, especially given the slightly random way the two grids interact. You can see from the picture that the crew that established the plots didn't get them on a precise grid - hard to accomplish in the forest, although they missed pretty bad in a few places. Someone went out and GPS'd the actual plots much later, and it's not like you can move them once they've been sampled and resampled for 10 or 20 years, so that's what we've got to work with.
Um, so, yeah. That's what I've been doing, in a tiny windowless room, for the last week or so. Actually it was only probably 8 hours of work or so, just spread out over several days and with a few false starts.
Um, so, yeah. That's what I've been doing, in a tiny windowless room, for the last week or so. Actually it was only probably 8 hours of work or so, just spread out over several days and with a few false starts.
