Is it plausible to use NDVI along with other regular bands for image classification related data processing?

Recently, I came across a comment that RED and NIR band might interfere with NDVI or vice versa during the processing and overfit the model.

I am working on a project where I want to classify healthy and stressed crop patches using satellite imagery, and I am using all four bands (RGB and NIR) along with NDVI to train the model and classify the satellite images.

For the classification, I am using Random Forest algorithm.


Including NDVI with your spectral bands is a very common practice in land cover classification and will almost certainly increase your classification accuracy. The Random Forests (RF) algorithm is fairly robust against overfitting. Leo Breiman and Adele Cutler, pioneers in Random Forests, claim that Random Forests does not overfit (reference). However, you can test for multi-collinearity using qr-matrix decomposition in the R rfUtilities package (multi.collinear). For your own reference, I would recommend running the classification with and without NDVI and see how the accuracy assessment differs.

Here is a link to a paper where we used qr-matrix decomposition to remove multi-collinear variables prior to using a Random Forest classification approach.

| improve this answer | |
  • @JeffreyEvans Are you aware of a reference I could use to back up my assertion that the Random Forests algorithm is fairly robust against overfitting? – Aaron Aug 26 '19 at 18:05
  • 1
    the original Breiman 2001 paper directly addresses the overfit issue as well as, for an ecological context, the Evans et al., (2011) book chapter. However, one assumption of ensemble learning is that the replicates are not overly correlated. If this is not the case, you will most certainly have an overfit model because of lack of independence across the bootstraps. This can often be observed in utilizing the entire population (all pixels) in remote sensed data to build a model. You can see this directly by taking multiple 33% random samples of imagery and looking at the correlations. – Jeffrey Evans Aug 26 '19 at 21:26
  • 1
    Please note that I added a permutation option to the rfUtilities::multi.collinear function to address the "hinge variable" problem. This is where one variable can overlay other variables, in scaled multivariate space, erroneously identifying them as collinear. In a permutated randomization you can use the selection frequency to mitigate this issue. – Jeffrey Evans Aug 26 '19 at 21:34

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.