Suppose I have a temporal stack of images of shape $m \times n \times k$ where shape of each image is $m \times n$ and $k$ represents the temporal dimension. In this context, I am trying to detect and classify temporal patterns.
So far my idea is to first train an autoencoder where the encoded data will be of shape $m \times n \times 1$. When the decoder is reliably able to reproduce the same input of shape, the clustering will be performed on the encoded $m \times n \times 1$ data.
Graphical representation of the idea is as follows:
However I am not sure if my idea is correct or not.
Assuming it is the right approach, what would be best choice of the autoencoder in this case?
It is obvious that any kind of pooling can't be performed here which will modify the shape so what would be the correct approach of convolution? I should clarify that, the spatial patterns also matters here.
How the clustering approach will differ from normal image clustering / segmentation / classification?
If this approach is not correct or if there are better alternatives kindly mention the alternatives with a bit of background and details.