I applied supervised maximum likelihood classification to a yearly stack of Landsat images (resolution 30m). The land cover classes are related to coniferous forest, i.e., forest, clear-cut, fire, bark beetle. I had only few years of available reference aerial photos (resolution 0.5m). Aerial photos do not fully cover the extent of Landsat images (AOI). To assess the accuracy of the classification for single year scene, I extracted a overlapping part between the extent of Landsat and aerial photos:

enter image description here

I calculated the measures of accuracy for the extracted part of fully classified Landsat by set of stratified sampling points per mapped class.

I followed the methods published here:

  • Olofsson, P., Foody, G.M., Herold, M., Stehman, S. V., Woodcock, C.E., Wulder, M.A., 2014. Good practices for estimating area and assessing accuracy of land change. Remote Sens. Environ. 148, 42–57. doi:10.1016/j.rse.2014.02.015 and
  • Olofsson, P., Foody, G.M., Stehman, S. V., Woodcock, C.E., 2013. Making better use of accuracy data in land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimation. Remote Sens. Environ. 129, 122–131. doi:10.1016/j.rse.2012.10.031

to assess the accuracy of the classification.

Great R code summarizing this can be found here: https://github.com/openforis/accuracy-assessment/blob/master/error_matrix_analysis.R

The error matrix, by this method, is not based on the simple sample counts, but on the estimated proportions of the area. Including the proportion of the mapped classes in accuracy assessment, it is possible to quantify the uncertainty attributable to variability in the sampling. Simply, to adjust the mapped area to map classification error. The error-adjusted area estimates should be accompagnied by confidence interval.

My problem:

I classified 7 Landsat images. I calculated accuracy for extracted parts of them for two years, because of no availability of another reference data sets.

My question:

  • Can I apply the accuracy measures established from the partially covered area (intersection between extent of Landsat and reference aerial photography) to the whole classified scene Landsat?


If your area of overlap is large enough and the land cover in the scene is homogeneous, you can make the hypothesis that your area is representative. Under that hypothesis, you can use a stratified sample to estimate the marginal error of each class, and weight them based on their proportions in the classified scene in order to predict the overall accuracy.This would of course not be as precise as a random sample over the whole area, but if you can show that the scene is homogeneous then you could justify your extrapolation for map user. If you aim at publishing a paper, then I would recommend to draw your conclusions only for the overlapping area.


The problem, in effect, is that your training data is a spatially biased. Its a matter of whether you can argue that this bias does not affect the validity of the result. But the issue is not dissimilar to whether the distribution of training areas selected is sufficient within an image.

I suspect it would come down to a case by case evaluation given the classes applied and the spectral homogeneity of the land cover each class applies to.

If the spectral signature of all land covers in the extended data is very close to that of the training set with reasonably distinct class boundaries AND you can make a case for there being no confounding land covers likely (e.g. different cover with similar spectral signature) perhaps by reference to other landcover datasets.. then it is not unreasonable to suggest the classification is likely to be comparable. But thats quite afew "If"s.

You could support the argument with some descriptive stats .. e.g. by comparing the per class spectra for random sub regions of the Landsat image to demonstrate consistency. Ultimately it is still just an argument, you can't prove the same classification works everywhere without the necessary data.


You are probably going to have a hard time convincing the reviewer that this is justifiable. One workaround is to see if you can use Google Earth imagery as the reference data. Google Earth may have very high resolution imagery in your study area. If this is indeed the case, you may consider using them since that is better than doing nothing at all.

  • No, googleEarth does not have the needed reference data. Only for year 2014, the whole area is covered. – maycca Oct 9 '16 at 9:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.