What is the best classifier for UAV images?


My question has two parts. I have multispectral (r, g, b, RE, NIR) UAV orthomosaic that I am trying to classify. I am doing the following steps:

  1. Large Scale mean shift
  2. Train Vector Classifier
  3. Vector Classification

My questions are:

  1. is there a general consensus as to what the best classifier in the Train Vector Classifier step is?
    So far, I have used the Random Forest classifier, which has done well, but I really just chose it at random.

  2. for my validation data, I have currently selected these myself but I am wondering if they are supposed to be random? Most of my image is a rocky substrate and if I choose them at random the majority of my checkpoints will be on the rocky class and miss out on my other classes.

I get good results using LibSVM classifier with SVM Kernal Type of Gaussian Radial Basis Function

For your approach:

  1. RF is good because less sensitive to noise in training data
  2. Its hard to say since it depends on what you want to measure. Ideally of course you should have enough elements to represent well the classes distributions.

Personally I prefer semantic segmentation (with convolutional neural networks) than OBIA based approach.

Can you perform semantic segmentation (with convolutional neural networks) using OTB?

Yes you can with the OTBTF remote module.

1 Like

Great thanks I’ll look into it.