Thank you University of Maryland: AVHRR Land Cover Data

Turns out that The University of Maryland Department of Geography has already done a land cover classification project, which highlights urban areas very well! You’d think that would make me feel quite bad, after spending much of last week trying to do this BUT unfortunately their project used AVHRR (Advanced Very High Resolution Radiometer) data which has a larger resolution than LandSat and the data is older. But there is still an awful lot of analysis I can do with 20 year old 1km resolution data, if nothing else I can try to use some of their methods with the more up to date LandSat data I was using before.

So this is how to download, import into Qgis and use the AVHRR Land Cover Data

Firstly go to http://glcf.umiacs.umd.edu/data/landcover/ for some info on the whole project, click on the Download via Search and Preview Tool (ESDI) if you want to browse the World and select the data you want OR head straight to the FTP server. Choose your area, projection and resolution (I’m using North America, LatLong and 1km), eventually you will be presented with a bunch of files from which we download the file ending asc.gz as Qgis can handle this not the bsq.gz. Extract the file after downloading and boot up Qgis. Now hit ‘Add a Raster Layer’ and select the .asc file you downloaded and extracted, the file will be pretty big (NA is 1.39 GB) so expect a wait. Now you will have a greyscale image of your region, the brightness depends on the land cover classification calculated by UMD. Here are the code values which can be found on their website.

Code Values for 1km and 8km data

Value	Label	RGB Red	RGB Green	RGB Blue
0	Water	068	079	137
1	Evergreen Needleleaf Forest	001	100	000
2	Evergreen Broadleaf Forest	001	130	000
3	Deciduous Needleleaf Forest	151	191	071
4	Deciduous Broadleaf Forest	002	220	000
5	Mixed Forest	000	255	000
6	Woodland	146	174	047
7	Wooded Grassland	220	206	000
8	Closed Shrubland	255	173	000
9	Open Shrubland	255	251	195
10	Grassland	140	072	009
11	Cropland	247	165	255
12	Bare Ground	255	199	174
13	Urban and Built	000	255	255

To make the map in Qgis look at all interesting we need to apply a colormap algorithm. You can select your own by right clicking the file and selecting properties and changing the settings or go to properties select ‘Load Style’ and load the style I created which can be downloaded here (right click and ‘Save target as’). Now you will have a wonderful map showing land use in your region, the area highlighted in white is what I am particuarly interested in and worked on for much of last week. The next step is to see if I can work out how to generate the code values easily and apply it to LandSat data, then compare to OSM.

Here are some image Qgis outputs for North America and San Francisco.

Refining urban area highlighting: the previous method didn’t work around the World!

So the previous method with the lovely flowchart and everything proved to be great for London but when I tried to use exactly the same method in New York it took far too much urban area away leaving us with a rather dull image when it should be very bright to indicate the NYC metropolis. Take a look at the difference between New York and London! So I essentially went back to the drawing board and thought, “okay what are the main features to a LandSat image?” answer: urban areas, rural areas and water! So that means I need to distinguish between urban and rural and then distinguish between urban and water, then somehow combine the two. Thats exactly what I did.

The method is very similar to before, except I’m now using both hue and saturation from the 742 false colour image instead of simply hue. The pretty flowchart for the process is shown below and for those you you who thought the last one was a mess wait ’til you see this one!

The method essentially works due to the different properties of hue and saturation. There is a great difference between the hue value for urban and rural areas, therefore hue is used to distinguish between rural and urban. Here is the hue component of the 742 image with a colour threshold to highlight urban areas (notice that that water is also heavily highlighted).Now we apply a colour threshold to the saturation layer and take advantage of the great difference in value for land and sea here. So now we have two layers one highlighting urban and sea and one highlighting urban and rural, we need to pick out the urban. This is done by using a clever layer mode algorithm in GIMP called ‘darken only’. This compares the pixel value in the two layers (hue and saturation) and displays the lower value pixel, in this way water and rural areas are removed from the image. Result!

Here are the resulting finished images for London and New York.

You can compare the previous method and the new one here:

Old London
Old New York
New London
New New York

By the way if anyone is really interested in this stuff and would like more guidance or some high resolution images just get in touch.

Highlighting urban areas in LandSat

The hypothesis I’m working on at the moment is that there should be more OpenStreetMap data (i.e. more nodes, more ways etc) in urban areas. To find out where these urban areas are I’m using freely available LandSat 7 data downloaded from http://www.landcover.org/data/landsat/. The data comes in the form of 8 different greyscale images corresponding to 8 different spectral bands ranging from visible blue light (0.45-0.52 µm) to thermal IR (10.40-12.5 µm).  Using bands 1,2,3 corresponding to blue, green and red a ‘true’ colour image of an area can be built, however this is not the best combination to use to highlight urban areas it turns out that the best combination is 7,4,2 for red, green and blue. Head over to my page on using LandSat imagery in GIMP to find a tutorial on all this.

What I’ve been working on over the past few days is how to use the false colour image I’ve produced using bands 7,4 and 2 to highlight the urban areas.Below is the 742 false colour image, and we can see that urban areas appear quite brown and purple, the aim is to extract that information and make everything else invisible. The problem is that some of the sea to the East is a similar sort of colour to the cities.

I decided the best way to extract the urban areas from  the image above would be to decompose the image into hue, saturation and value. The hue is the most interesting component, giving a good contrast between rural and urban areas, but as we can see there is little difference in colour still between some sea areas and some cities which could prove a problem in extracting just urban areas.

The next step is to apply a colour threshold on this image to try to pick out only urban areas and black out everything else. After a great deal of playing around with filters the urban threshold here appears to be within light levels 196-212, after applying this filter the image below is obtained.

As we can see the urban areas are nicely highlighted in white and everything else is black. Now the aim is to compare the brightness of every pixel in this image with OSM data. We know the coordinates of the image above and we kno that each pixel is (30m x 30m) so this should be easy enough using a simple bounding box query of the OSM database. Lets hope that there is a relationship now between OSM data and the brightness of the pixels above. I’ll be back with any results when I have them and then hopefully be rolling this out across the world.

Click on the image below for a nice flowchart of the whole process.

New project: Lets use LandSat

Not sure about the licensing requirements for my beautiful Qgis model (which shows roughly which areas of the UK are complete and which incomplete) so I don’t think I can publish the full results from that yet. Essentially though I have found that there is a strong correlation between the length of road in a particular area and the population in that area, its good enough to accurately predict the length of road there should be in an area and compare that to OSM road length. But enough about that project for now, hopefully soon I will publish all my findings.

So instead its onwards and upwards to a new project which was inspired by the view of the Osmarender layer on OpenStreetMap, shown below. It is clear to see that there are vast areas in Asia and South America, where there is no OpenStreetMap data, the question is whether there is actually nothing there or OSM is just missing cities, roads etc. I plan to find out using open source aerial imagery.

The plan is to use LandSat and other forms of freey available imagery to work out where there should be cities and roads and where there shouldn’t be. Then take this information and compare to OSM. Easier said than done, I’m sure but it wouldn’t be a project if it wasn’t challenging. So below is a LandSat image of London and to the North, which I have applied a Yellow Contrast Gradient Map to, using GIMP. As you can see it emphases cities and rural areas quite well and I’m sure this is the starting point to predicting accurately where there should be OSM data.