I have two Landsat 5 NDVI images (Surface Reflectance products) from the same period of the year (only a few days apart), but one is from the 1980s and one from the 2000s. I am trying to do a change detection procedure based on NDVI differencing, and the expected outcome (based on aerial photographs) is that NDVI should have increased due to a widespread reforestation process, so the result of the differencing process (2007NDVI - 1987NDVI) should highlight mainly a) unchanged areas and b) areas of ndvi increase.

I tried to reclassify the difference image using a threshold calculated as μ ± n·σ; (μ represents the mean, and σ the standard deviation). I used n=1. (reference paper: http://www.sisef.it/iforest/contents/?id=ifor0909-007 ) The difference image classified this way, however, shows mainly unchanged pixels, so is it possible that the difference image shows no increase simply because the data are not "normalized"? Since this is a SR product, it has already been atmospherically corrected, so it may be that due to climatic differences during those two years in one case the vegetative season was more advanced than the other.

These are the image histograms: 1989 image 2007 image

The first is from 1989 and the second from 2007.

Would in this case a histogram matching procedure be of any help (or any other procedure you might suggest)? And if so, how can I perform that using QGIS? I couldn't find any dedicated tool or plugin for it.

  • I don't have enough reputation to comment, so I write here: If I were you I would use a maximum value composite (Holben, 1986) over several/all NDVI images taken by Landsat during the vegetation period. Compare the two resulting images.
    – countryman
    Commented Oct 1, 2015 at 9:41
  • 1
    When considering histogram matching, you need to keep in mind what it is you are working on. You can match the data before calculating NDVI, or after. Furthermore, histogram matching can take multiple forms. One option is to simply calculate the mean value over a large area for both images, and then shift the mean of one image to match the other. Another way is to fit a function to the frequency distribution and use this to shift and reweigh all points. Commented Oct 1, 2015 at 13:08
  • @countryman I think that would work in the case of having several NDVI images, but unfortunately due to cloud cover I have one or no usable images for most years.
    – Simona
    Commented Oct 2, 2015 at 10:16
  • @MikkelLydholmRasmussen do you have any idea if/how that can be done in QGIS?
    – Simona
    Commented Oct 2, 2015 at 10:19
  • 1
    @Simona, you can use the answer to this question: gis.stackexchange.com/questions/23203/… and then get the mean from both rasters. Then find the difference between the means and then add that to one of them using a raster calculator. That'll give you an image with a shifted histogram where the means match. Commented Oct 5, 2015 at 6:11


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.