I try to extract woodland areas from a greyscale orthophoto. I used the "Haralick Texture Extraction" from the Orfeo Tool Box to calculate texture properties of the greyscale orthophoto. As output I get a raster with 10+ bands. Each band is representing a texture property like Contrast, Homogeneity, Dissimilarity and so on. Calculating texture properties is parameterised. You need to state a specific radius.

For the differentiation of woodland and non-woodland areas, my idea was to create a stacked raster with the texture features for 5 different radiuses, this raster would have 50+ bands, followed by a supervised classification. I already created a training dataset for woodland and non-woodland areas.

My question is: How do I identify bands of this stacked image that are relevant for the differentiation of woodland and non-woodland areas prior to the classification?

2 Answers 2


If you want to figure out which inputs contribute largely to a classification and which offer very little, you are usually looking for a PCA ( = prinicipal component analysis).
PCA is the scientific approach to "running many possible input combinations and seeing if the results worsen/improve significantly", which you also could do manually (which is very tedious and vague).

If you search for "PCA hyperspectral", you'll find a number of approaches that can probably be applied to your use case.

If you need more specific recommendations, please provide more detail in your question (programs/code language used, data source details etc.) .


I would recommend using a Random Forest classifier for this task for its ability to handle non-parametric data and its robustness against overfitting and outliers.

@JeffreyEvans created a utilities package for random forests called rfUtilities which may be of interest to you. The multi.collinear (p.4 documentation) command allows you to test for multicollinearity in data using qr-matrix decomposition.

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.