Lab 04 - Ground Truthing and Accuracy Assessment

This week we looked at the various methods of accuracy assessment for Land Use, Land Cover Classification maps. There are various standards, with varying limitations, and no one method is right for every project. Using the LULC map we created last week, we would be doing some ex-situ "ground truthing" to conduct an accuracy assessment. As in-situ ground truthing wasn't realistic for the lab, we used higher resolution data, in this case Google Maps to analyze our sampling points.

The first step was to create 30 sample points. There are different ways to sample a study area, including random, stratified, and sequential. I chose stratified as there were various sizes of classes in the study area, for example a large residential area vs some smaller forest classes. Using stratified allowed me to distribute my random points to be heavier in the larger sampling areas. I created a PointCount field to specify the number of points I wanted within each class, and then I used the Create Random Points tool to let the computer randomly set the points. From there, it was a matter of visually locating each sample point in Google Maps and determining if the classification code I originally chose was true or false. After much discussion posting with other students and professor, I also decided to do a radius of 200 ft x 200 ft for some spatial context of my sampling unit as my original 6 ha MMu was coarse.

This was simple in the more homogeneous areas, like the bay or the non-forested wetlands, where pattern and color were obvious, and these areas had the highest accuracy at 100%. The urban areas were more challenging, where buildings and land use was more diverse and difficult to identify from a vertical photo. Here Google Map's StreetView was handy to look at a sign on a building, or get a new perspective on a building. My two mixed classes 12/13 (Commercial and Services/Industrial) and 12/14 (Commercial and Service/Transportation, Communications, and Utilities) were the least accurate, at 0%. For both of these, the 4 sampling points were determined to be only commercial and service, and further inspection identified the polygons as only commercial and service (12) as well. Overall accuracy ended up at 76%.

All in all it was interesting to learn the "art and science" of accuracy assessment on our own LULC map. Just like classification, accuracy assessment can be subjective if methods and standards are not determined before a project begins, and even after setting a few standards for myself, I still felt like there was a lot of ambiguity in my assessment.

Accuracy assessment of Lab 03's LULC map, using 30 stratified sample points and Google Maps as higher resolution reference data. Overall accuracy was 76%.