3

I'd like to calculate the NDVI for multi-date satellite images of Landsat (7,8) to perform a change detection on the deforestation of the Amazon rainforest. I already acquired surface reflectance products of Landsat 8 and 7. After calculating the NDVI, I'd like to reclassify the NDVI raster with the help of a threshold into at least forest and non-forest classes. I tried to calculate the NDVI for two images (2000 and 2015) and found that I couldn't use the same threshold value to identify forest from non-forest in both images. So, my question is:

1) Does the threshold need to be the same for both images ? (I assume yes, but I'm a beginner.)

  • 1
    Welcome to GIS.SE. Please take the tour if you have not already. In general, you should only ask one question, and have the title closely reflect the question, to make it easier for other people to find in the future. – John Powell Jan 1 '16 at 12:16
  • 3
    Are the two images from the same month / growing-period / season? If not you can't possibly use the same thresholds since NDVI vastyl differs throughout the growing period - even with coniferous forest (if that is the type of forest you are looking for). – Kersten Jan 1 '16 at 19:32
  • 4
    Additionally: If you are looking at forest cover change from 2000 to 2014/15 with Landsat resolution why not use the dataset published by Hansen et al.? – Kersten Jan 1 '16 at 19:34
  • @Kersten, you bring up good and relevant points. Although the Amazon is an evergreen forest and had historically been thought to have consistent seasonal phenology, recent research has shown otherwise. I agree that it would be necessary to adjust the threshold between the time periods. – Aaron Jan 1 '16 at 19:54
  • @AAron totally overlooked the Amazon bit, thanks for pointing that out - but as you remarkedt the point still holds true for evergreen forests as well. The seasons might not be as distinct but still enough to throw a simple threshold classification of course. – Kersten Jan 1 '16 at 20:31
5

No, the NDVI threshold value will not be the same for the time series due to differences in phenology and unique conditions on the ground. As Kersten mentioned in the comments, you may want to consider using Global Forest Watch data, which is well respected in the environmental community.

You have uncovered one of the limitations of working with thresholding for image classification. This pixel-based method is easy to implement, although requires the user to guess at which threshold value is most suitable. For example, secondary growth forests 15 years post-clearcut will have very high NDVI values due to all the new growth. Some of the clearcuts will likely be converted to pasture, which will have different NDVI values.

Thresholding is valuable for many operations, although more sophisticated approaches often produce better results. For example, an image segmentation and classification approach that uses all of the Landsat spectral bands and derived products such as NDVI (and other vegetation indices) and texture may produce better results. Object oriented image processing methods, such as image segmentation, are often able to handle the temporal phenological differences better--allowing for a more automated approach across vast areas.

  • Many thanks for your information. Do you know if there is another way to determine the NDVI threshold apart from guessing it? I thought about comparing two methods (NDVI classification and supervised classification) but I might just stick to one (the supervised classifiation) – Maja Jan 2 '16 at 23:15
  • I'm afraid you have to eyeball it and use your best judgement when thresholding. You may want to consider using a supervised classification algorithm to classify forest/non-forest. Common approaches here include maximum liklihood or random forests. – Aaron Jan 2 '16 at 23:27
2

There is other possibility to do what you want. INPE used to MLME with good results. I have studied so much it and I am very pleasure with results which I reached with this methodology. Try to do this that I am sure that you will like.

  • 1
    Thanks for the contribution Diogo. It would be helpful to include a brief description of the acronyms you use and why you are pleased with the results. – Aaron Jan 4 '16 at 1:35

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.