|
Hi, I'd like to validate a land cover change map. Only 11% op the map area is subject to change. I'm only interested in the estimates of the change classes. When I only take into account the change classes, I need about 3000 samples. Should I also include the area that is not subject to land cover change in my validation? Or can I omit them to limit the number of reference points? |
|
It depends on what you want to do... If you only include plots in the change area (according to the map you want to validate) you will only get an accuracy estimation of the change detected, but you will get no information on how the classifier performed on areas where it said there was no change, meaning that your accuracy assessment could says that you have a 100% accuracy on the change area but you would have missed that ( for instance) the classifier had failed to assess changes in an area that was even larger ( as you have not put any plots outside the area that was classified as change) EXAMPLEIf the areas in the map are 990.000 hectares of NO CHANGE and 10.000 hectares of CHANGE ( totalling 1.000.000 hectares) Then if you put 100 plots to validate the assessment of the CHANGE area, an 90 are correct (the plot had actually changed), then you can say that the accuracy on the CHANGE assessed was 90%....sounds good! But, say that now you put another 1.000 plots for validation in the area of NO CHANGE. Turns out that you have 10 plots within that area with CHANGE! This means that these plots, as they have been spread over the 990.000 hectares of NO CHANGE ( each plot represents thus 990.000/1.000 == 990 hectares), mean that you assess that there is 990 * 10 = 9.900 of actual CHANGE on the part that your map says it is NO CHANGE!!! So the area of change initially from the map after assessing the accuracy is corrected to around 9.000 hectares (since we said that 90% of the plots in the 10.000 hectares of CHANGE are correct) BUT your area of CHANGE in the part that the maps says there is NO CHANGE is 9.900 hectares. This means that there is more change OUTSIDE your CHANGE area in the map that inside! If you don't put plots on the NO CHANGE area you cannot give a really good assessment of the performance of your classifier and on the area data that you will use in your report as it would be biased, you are saying that according to the map the errors can only be in the CHANGE area and the NO CHANGE area is "perfect" |
|
Please refer to this guide on Map Accuracy Assessment and Area Estimation : http://www.fao.org/3/a-i5601e.pdf |