L7 modality issue / timeseries

[ando-shah]

Hi,

I'm trying to use the L7 projector for a 6 band dataset from Landsat (m-forestnet). I get runtime errors of mismatching sizes when I try this with varying patch sizes [48, 320, 480..]. Whats the recommended patch size for an image of size, say 64px/120px? This is a single date image.

Additionally using single date images with any of the time series modalities (like s2) incurs high memory overheads. Currently I unsqueeze the time dimension in these cases.

Do you have recommendations on how to approach these cases?

Thanks!

[gastruc]

Hi,
The code has indeed be designed for square images. The model could completely process non squared images with the same weights but it will require some coding work. An easy way would be to perform 2 inference on 64x64 images and average the features on the overlap part.
For the single date images, the model perform a temporal attention even if the time dimension is 1. Once again, you could overpass temporal attention in these cases but I've not implemented this feature yet.
Best regards,