Sunday, September 20, 2015

Ground Truth: Error Matrix & Overall Accuracy

This week I modified my map from last week, Land Use Land Classification. The challenge last week was to not use any tool that would help you verify the type of land I was classifying. I could only use the aerial image (.tif) of Pascagoula, MS. and make my best guess at what type of land I was classifying. This week, the challenge was to take an unbiased sampling of the land I classified and compute the overall accuracy. Under normal circumstances, you might travel to the location and conduct fieldwork to determine the correct land use/classification by conducting a Random, Stratified Random, or Systematic Sampling. Since I could not actually travel to Pascagoula, the tool of choice was Google Maps - Street View. This is an excellent proxy for actually being there.

Another tool I created and used was an Error Matrix. I conducted a virtual Random Sampling of my map starting at the North West border and working my way South to cover the entire study area and I marked as "True" or "False" the areas I had previously classified. I used Google Maps to confirm what I classified was correct - or not. I found that I miss-identified a few land types such as:

- A Commercial / Industrial area (Category 15, USGS Andersonian framework for national land use and land cover) - What looked like an Industrial area to me turned out to be a shopping center.
- A Residential area (category 11, USGS Andersonian framwork) - I correctly identified most of the residential areas, but one area I classified as residential turned out to be a cemetery!

My Overall Accuracy was 76%. Here's the map I created.


No comments:

Post a Comment