I ordered and downloaded several (Surface Reflectance) Landsat 5-8 images from USGS Earth Explorer with the purpose of choosing a few of them to calculate some Vegetation Indices for each (NDVI, SAVI, EVI) and compare the evolution of the vegetation cover from the 1980s to the present.

I need to select the same month for each year I'm going to consider, but I'm not sure which is the best month to ensure the maximum separability between each vegetation type (I need to discriminate between grassland, closed forest, open forest/shrubs and bare soil in a mountainous environment in the Italian Alps).

Among the Landsat images I downloaded, I have a series of images from the same year (5 images for 2005, all falling within the vegetative season), which I want to use to determine the best month for classification.

Does it make sense to perform a semi-automatic classification on each 2005 image separately, and then compare these classifications to a high-res reference image of the same period to see which month gives the most accurate result?

Or is it better if I calculate NDVI for each image first, and perform the classification based on that?

My worry is that, since I will calculate other indices, not just ndvi, basing my choice of month only on NDVI might not be appropriate.

In addition, do you have different suggestions about how to do this?


Considering only one image per year is a pit-fall that you should try to avoid. One of the keys to correctly classifying different types of vegetation is the phenology, and to tap into this resource of information you should employ a multitemporal classification approach.
Basically, this means that instead of using the multispectral bands and a few indicies from a single image, you stack use multiple images (and the indices based on them) from the same year and perform classification on that stack. This means that you suddenly have a lot more information to work with in your classification scheme, which in turn can provide much better results.

This approach obviously also adds some "noise" in the sense that intra-annual changes to land use / vegetation type will behave strangely in the statistics, but by only considering a single year, this problem is usually minimal.

  • Thank you for your answer. What about performing separate classifications on each available image within one year and compare them? I'm not sure how to implement the kind of classification you are suggesting, so if you have any practical advice I would be grateful (I'm using Qgis). Another important point is that I need to be able to perform a change detection between images from different years, and in most cases I only have one image available for each year (generally from late summer). – Simona Aug 5 '15 at 21:42
  • 1
    You can do separate classifications on each available image, but it might be difficult to tell one type of forest from another. It also takes a lot longer to do that many classifications. As for doing multitemporal classification - you use the 'Merge'-tool from GDAL to combine all the images from one year into one image with many bands, on which you then perform classification. – Mikkel Lydholm Rasmussen Aug 6 '15 at 6:05

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.