1

Just want to ask one question regarding the classification using the hyperspectral or Multi-spectral data, If we have hyperspectral or Multi-spectral data, and we only use the data of R G and B frequencies and we got the good classification result. Does it make sense? Or we have to use all the possible the frequencies which have the information.

  • If it works, and you're content with the results, that's great.. it depends on what you're classifying, if you just want bare surfaces then RGB could be enough, if you want to identify predominant tree species then expect to use some infrared bands. – Michael Stimson Nov 24 '17 at 0:53
  • The Dataset I have is 4 different kind of weeds which is not so difficult. So I first tried to use RGB bands to see the performance using CNN and it is working fine. I just want to classify each category of the weed. I have taken out 27 RGB images and divide the image into 64 patches. Considering each patch as an image for that class. Then I apply the data augmentation method i.e. rotation to increase the dataset Any suggestions where I can make scenario difficult for the RGB bands? – Aadnan Farooq A Nov 24 '17 at 1:07
1

Assuming your accuracy assessment is valid and you are satisfied with the classification accuracy, by all means use only the visible spectrum bands. Many studies limit the spectral bands to the visible spectrum and near infrared for processing purposes or to maintain native spectral resolution. This is completely valid depending upon your research question and the desired results.

However, incorporating additional spectral information into your land cover classification will almost always increase your accuracy and decrease your commission and omission error rate. Therefore, unless you have some very particular reason not to incorporate the additional spectral information, I would opt to use it.

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.