PCA & Imagecube before train test split cause data leak. about hybridsn HOT 2 CLOSED

Hey, thanks for the feedback and for reaching out.

Can you specify what kind of other images did you use?
Also, yes the authors are aware of the fact that the proposed model fails to converge without PCA and this is attributed to the fact that hyperspectral images tend to contain a lot of redundant data in the large numbers of spectral bands. Approaches other than PCA have been explored and demonstrated to work successfully in other works but in case of our proposed model which is fairly simple in terms of size and complexity, PCA is one the approaches that help the model deal with the aforementioned redundancy by transforming the hyperspectral volume and selecting the most informative top-k bands to be sent to the network. I hope this helps to clarify your query.

About the data leak, creating image cubes before or after train-test split wouldn't make a difference as the cubes will at the end be derived from the original image volume. The only way to alleviate this is to sample cubes from discrete pixels in the volume such that there is no data leak and there isn't a huge class imbalance. The authors followed the existing sampling methods that were being used in the previous works but you are absolutely right in pointing out that this sampling method is prone to data leak. Alternative sampling methods have been proposed recently for dealing with this issue.

from hybridsn.

Thank you for your reply.
I use the model on hyperspectral images of grass, with the camera scanning different type of grass from about 1-meter distance.
Can you advise what is the approaches other than PCA that could apply on the hyperspectral image?
I will try to use for example half side of image for training and the other part for testing the image cube.
I like the concept of this model, and I will look into other models for comparisons.
Thanks.

from hybridsn.