I'm doing an extraction of spectral trajectories for disturbed forest pixels using multitemporal (single scene for 20 years) Landsat Surface Reflectance products.

Usually, forest disturbances are detected using automated algorithms (mainly Landtrend for Landsat data) and statistic validation of the results is done using confusion matrix (overall accuracy ecc.), Kappa, Pearson correlation and similar methods.

In my case, the disturbed pixels are identified by visual interpretation (so they can be classified as ground truth data) and no predicting algorithm is used. I'm only extracting the pixel values and plotting them along the temporal series.

My question is: after noise removal (outliers detection, data cleaning filters ecc.), how can I tell how accurate the trajectories obtained from the pixel values are?

  • Questions about using GIS for statistical analysis are on-topic here as well as on stats.stackexchange.com. If you don't get any good answers here (after waiting maybe a week or so) you might want to try the other site. Just a suggestion, since I'm having a hard time understanding your question. How can you calculate accuracy if you don't know what the "correct" value is? – csk Apr 5 '19 at 21:00
  • You understood it right, it was exactly my point. In short, I was asked to provide consistent trends to train an algorithm in another phase. Along with the trends, I was asked to provide some kind of accuracy method but, as you said, my trends are built on observations (truth), not predictions. I posted here just to be sure that I wasn't missing something...it seems that I am not the only one who's puzzled by this request. – D.K. Apr 5 '19 at 21:30

Your Answer

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

Browse other questions tagged or ask your own question.