I have undertaken supervised classification (maximum likelihood) and planning to survey the ground to validate the output.
How many validation points do I need to take?
The raster is a landsat tile and I have 8 categories for classification
Add a comment
|
2 Answers
According to Lillesand et al. (2004), you should obtain a minimum of 50 validation points per landcover class. Given 8 landcover classes, a minimum of 400 validation points would be required in total.
With a greater number of validation points, the overall accuracy of your classification can be determined more precisely. Also, the omission/commission errors can be characterised more precisely. In my opinion, having at least 100 or 200 validation points per landcover class is desirable, but sometimes impractical if you study area is large and inaccessible.
To avoid extensive ground surveys, you could use high-resolution imagery (e.g. Google Earth, aerial photography) to validate your classification. This would save considerable time and ensure evenly distributed validation points (as opposed to validation points clustered in one or two survey locations).
Lillesand, T., Kiefer, R. W., & Chipman, J. (2004). Remote sensing and image interpretation. John Wiley & Sons.
The number of points depends on the precision that you expect from your estimator and from the type of sampling design that you use. According to Cochran (1977), the equation to determine the number of point with a simple random sampling is
z^2*O(1-O)/d^2
where z is a percentile from the standard normal distribution (most of the time, the value of the 95% percentile (=1.96) is used), O is the overall accuracy (in proportion) that you target from your calssification and d is the size of the half width confidence interval (one or 2 percents in practice).
This equation is designed for the estimation of the overall accuracy. If you also want to provide information about the user and producer accuracy, you need more points because your dataset will be divided in each of the classes. You can estimate the number of points that a simple random sampling would give you based on the proportion of each class in the map. If some of your classes are underepresented (very low frequency), then a stratified random sampling is needed. Note that, in this case, you will need to weight each class by their area in order to accurately estimate the overall accuracy.
For more information, the paper from Olofsson et al (Good practices for estimating area and assessing accuracy of land change, Remote sensing of environment) is a very good summary IMHO.
