Task List #2

[Diacod-I]

• [x] 1) Combine the RGB bands together and the infrared bands together using image registration
• [ ] 2) Perform image stitching + tiling on the raster images as an alternative to mosaicking and combine the masks in a similiar fashion
• [ ] 3) Implement the previous architecture and review the difference in performance
• [ ] 4) Implement Custom U-Net as a new architecture and test its efficiency in terms of segmentation in comparison with the other models.

[Diacod-I]

1) Combined RGB bands together and Infrared Bands via Dimension Stacking, hence removing the need for Image registration (a much more computationally expensive approach)

A new issue found was the massive class imbalance in the masks.. Upon tiling and splitting the tiles into two groups, one with more whites over blacks and the other with more blacks than whites, we find the overall magnitude of difference between the number of black pixels over white pixels.