Is it possible, and if so how, to effectively classify vegetation types using false near infra red imagery?

Question

Background : I use near infra-red imagery to delineate wetlands and water bodies with good success. It's also obvious to distinguish between new healthy vegetation and older growth and land areas with no vegetation but i need to devide vegetation into further groups. To look at other examples, I downloaded the latest NRCAN (Natural Resources Canada) CANVEC (Canada Vector) vegetation data (polygons with attributes) and the data is incredibly detailed given the coverage (Canada) and scale (1:50000). It is my understanding the classification was done with using Landsat 7 & 8 multispectral imagery in the near infra-red spectrum. I believe this data is also continually updated as new imagery becomes available. There are at least 14 different vegetation types (screen capture with types below). How is this possible? It seems that to do my project manually would be possible but also an enormous undertaking at the detail / scale I need (approx 1:2000).

Questions:

How can I efficiently classify vegetation using custom false NIR high resolution imagery into similar categories?

Is there a fairly accurate automated process or is this done visually / manually?

What is a typical technique one may use to classify vegetation in this way?

CanVec Vegetation categories as classified from imagery:

Typical example of our custom imagery. I can delineate this area manually and perhaps work out some categories but this is only a fraction of our data and it would take VERY long time to finish the entire area.

Answer 1

In order to get at the classes you describe, you will need to incorporate a sophisticated classification algorithm and ancillary data derived from the imagery. I would recommend two approaches: 1) an object-oriented image segmentation (IS) approach using IS software such as eCognition or 2) a pixel-based non-metric, decision tree (Random Forest) approach using R. The new release of eCognition allows for Random Forests classification of segmented image objects.

With both approaches, I would incorporate several ancillary data layer including texture and vegetation indices such as NDVI and EVI. Both R and eCogition have the capacity to produce and incorporate these data layers in to the analysis. There is a study (presentation here) where the researchers where able to get 85% accuracy (Kappa 0.80) describing 5 classes of forest structure. I suspect you will need to take this approach to get at the dense, open and sparse characteristics described in CanVec.

Recommended Reading

Wood, E. M., Pidgeon, A. M., Radeloff, V. C., & Keuler, N. S. (2012). Image texture as a remotely sensed measure of vegetation structure. Remote Sensing of Environment, 121, 516-526.