Hello, Joe here. With the carto-freak David, we make up the GIS side of the research group’s “Spatial Command”. We don’t like to get our fingers dirty and so are instead lucky enough to spend hours on end in a windowless computer lab making minute adjustments to processes that we know will inexplicably fail anyway. Not that I’m bitter about the work I do or anything…
Anyway, Spatial Command is currently looking at a series of watersheds along the Mekong river, tying isotope data (specifically 210Pb and 137Cs) from soil samples to other parameters, like relief, rainfall, and land use. The isotopes get counted in Harbin, our reliable germanium detector. Data like relief, we can calculate from elevation models, and in fact, we used the wonderful GDEM dataset from NASA to that effect. Rainfall was a bit harder to do in-house, so we turned to the quite useful APHRODITE dataset from the Research Institute for Humanity and Nature (RIHN) and the Meteorological Research Institute of Japan Meteorological Agency (MRI/JMA). That leaves us with land use data, the source of some exciting new developments for Spatial Command, and the subject of this blog post.
Harbin, showing enormous patience with one the bozos from the lab, this is something Harbin has to put up with on a daily basis, as everyone who works in this lab is quite a joker.
Previously, we had been using the GlobCover dataset from the European Space Agency for our larger scale analysis and an in-house supervised classification of Landsat 8 data trained by yours truly for a smaller scale study of three basins. Spatial Command was all set to use GlobCover for our larger project, when word came in that a newer, more precise dataset existed, GlobeLand30, a global dataset with a spatial resolution of 30 meters. Needless to say, our cartographic natures were excited by this possibility. But we were still cautious. Where had this dataset come from? Sure it was more precise, but was it more accurate? We decided to take one step beyond guessing, and put it to the test!
GlobeLand30 is a product from the National Geomatics Center of China, and was released in 2014. The image data used to create GlobeLand30 were primarily 30 meter Landsat TM and ETM+ images from within a year of 2010. They undertook an extensive accuracy assessment, and reported an accuracy of 83.51%, but we decided it would be prudent to assess the accuracy of our study area specifically.
When comparing the GLOBCOVER and GlobeLand30 against accuracy assessment regions defined by me from Landsat 8 data, certain issues presented themselves. Namely, that each dataset was divided into different land use classes. The GlobCover dataset was divided into 23 different classes, GlobeLand30 had 10, and because we were primarily interested in quantifying agriculture, we had only 4 classes. Both the accuracy assessment and eventual application necessitated some amount of class combination.
Eventually, all data was combined into four classes: Agriculture, Forest (comprising all natural vegetation), artificial surfaces, and water. Some classes that were irrelevant to the areas used for accuracy assessment, like “tundra”, were ignored. Then it was smooth sailing to the goods; we plugged the data into ENVI, generated some confusion matrices, and then we were Good To Go!
The overall accuracy of GlobCover for our area was 72.22%, and the overall accuracy of GlobeLand30 was 76.76%. The accuracy for each individual class was also good, although classes that were not widely represented in the data used for accuracy assessment tended to have lower calculated accuracy.
As you can see from the maps comparing the GlobCover and GlobeLand30 datasets, the precision afforded by the GlobeLand30 means that the boundaries between classes are much more accurate. However there are also still areas of disagreement between GlobeLand30 and the hand-classified dataset used for accuracy assessment. The classification I did by hand was primarily from Landsat 8 satellite images, between that and my unfamiliarity with the region, it is certainly possible that I have misclassified things that the GlobeLand30 classification system classified correctly because they used a wider range of data sources that may have helped distinguish between things that were visually similar in the satellite data, like cropland and grassland.
Still, it seems clear that GlobeLand30 is the better choice, and we at Spatial Command are excited to get a chance to work with it. The next time you hear about it from us will be when we have some preliminary results about how land-use correlates with our isotope counts.