We know that the threshold value varies by geography. For the reason identifying this value is challenging and there is a considerable risk of identifying the incorrect value. That is why my primary goal is to precisely determine threshold value so that I can separate vegetations for a specific area from the rest of the landscape.
NDVI values vary depending on the volume and condition of the vegetation cover. In the same region there may be crops with soil gaps, or meadows (with high/medium/low green (fresh) or dry yellowed grass, and forests (coniferous or deciduous)/ Each of these sites/vegetation states may correspond to different NDVI values. For this reason, each of these landscapes will have a different range of NDVI values. That's why your question misses the main point - from which landscape do you want to separate your particular site? What index values will be found in your study area?
Dear Elena A. Baldina
You are absolutely right,
I have chosen two separate sites: agriculture fields and a forest, which is basically a mangrove forest. For such places, I used NDVI and got values ranging from -0.35 to 0.81 and -0.21 to 0.93, respectively. In this range, my sites have vegetation (large and canopy), bear soil, water bodies, rivers and so on.
I this instance, I intend to use a machine learning-based method to segregate with greater precision and capture the desired ndvi value for my specific landscapes in my sites. In order to that, I can recommend ndvi value for my specific landscape that can distinguish vegetation from other landcover.
Could you kindly advise me on what I should do or how I should proceed?
I assume you are considering these two types of landscape separately. And I also think you know that if NDVI takes negative values, it's highly likely that it's water or snow/ice. On the other hand, if you have an NDVI value close to +1, there's a high possibility that it's dense green vegetation. But when NDVI is close to zero, there aren't green leaves and it could even be a bare soil, rocks, or an urbanized area.
If you're writing about machine learning techniques, then you must first teach the computer program to distinguish between the classes of interest, that is, NDVI values corresponding to each type of vegetation cover.
In our university's Images Interpretation course, students do hands-on work to learn how to map NDVI values to different types of objects on study different methods of geographic interpretation. It's supposed to get (calculate) NDVI values for a mountainous region with a pronounced altitude zonation and compare the whole range of numeric NDVI values to different altitude zones, which should be known from literature, maps, interpretation of high-resolution imagery etc. (in my example is a map of terrain zones). We don't use machine learning here, because the goal is to learn to see objects and corresponding them numerical NDVI values. You can use machine learning methods successfully only if you have a good understanding of how objects differ in the feature you are interested in and can teach the algorithm on a sufficient number of concrete examples.
(I've attached an example of a student’s result. Note that some types of plant communities (e.g., mountain tundra (where vegetation is sparse) and mountain-steppe (where vegetation is mostly dry) cannot be distinguished by NDVI. In this case, you need other bases, in particular the location of the objects (in my example, up in the mountains or on the southern slopes below)).
Elena A. Baldina
I Appreciate for your insightful response, and I am working on it.
Is it possible to get Images Interpretation course materials?
One more thing,
I have plan to apply unsupervised machine learning so that my algorithm will analysis and try to find the pattern of value. These types of algorithms capable to identify several sorts of clusters and I assume that among the clusters, one cluster must contain or distinguish ndvi value form other landscape. And I will repeat this for each landscape.
I have developed a flowchart for you explains my strategy. I'd appreciate it if you could have a look at it and give me some input.