I have a 5.8 m resolution satellite image (LISS IV aboard IRS-P6, in the wavelength (micrometers): 0.52 - 0.59 (band 2),0.62 - 0.68 (band 3) and 0.77 - 0.86 (band4). The swath covers a densely forested region in the Eastern Himalayas. My objective is to predict regions of canopy cover. Regions of canopy vs. non-canopy.

I intend on using several predictor variables in light of: Incorporating terrain data to predict canopy cover using randomForest in R and The stability of randomForest models after increasing predictor variables.

The area covered in my image ranges from 700 m to 3500 m and comprises at-least 3 forest types excluding high altitude grasslands and other landcover (water, rock etc.). As I want to map canopy cover over this entire region, what would be the ideal set of training data to use?

Should it cover all spectral classes of vegetation? As my objective is just canopy I suspect the training points should constitute a wide range of tree crown types.

Edit: Wish to map canopy vs. non-canopy within classes of canopy coverage (canopy cover percentage per pixel).

Definitions: by canopy I specifically mean the crown area of a primary stand. Secondary stands with a crown are not the target.

  • 1
    Could you please define "spectral classes of vegetation"? Are you looking to digitize training data on-screen or generate training data in the field? Are you only interested in two classes: canopy cover and everything else?
    – Aaron
    May 29, 2015 at 16:49
  • 1
    By spectral classes I mean, different vegetation types would have a range of spectral values. Deciduous forests then form one spectral class of vegetation. Hence, should my training data span all the vegetation classes? I am looking to digitize data on-screen as well as use some of the field data on forest type and canopy shadow. Interested only in canopy cover vs. everything else at the moment.
    – csheth
    May 29, 2015 at 16:55
  • 1
    You need to have a very clear idea of what your intent is. Do you want areas that "could" have canopy cover or those that currently do? By incorporating abiotic "terrain" variables you could end up with potential suitable areas in the estimate because optimal conditions are being incorporated into the multivariate space of "canopy" class. This is where one must be mindful of specifying a niche or current classification model. May 29, 2015 at 17:19
  • @Jeffrey, current intent is to use on-screen training data of tree crowns and predict a map of canopy cover. Can you also elaborate on "potential suitable areas". Do you mean that even if there isn't a tree crown the model may predict one based on terrain similarity?
    – csheth
    May 29, 2015 at 17:48


Your Answer

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