Try running the epitomes or clustering models

[lzachmann]

Epitomes: https://github.com/anthonymlortiz/epitomes_lsr Clustering: https://github.com/malkin1729/cluster-lsr

[jessjaco]

Some background docs I found: https://openreview.net/pdf?id=rkxwShA9Ym https://arxiv.org/pdf/2004.11498v1.pdf https://openaccess.thecvf.com/content_CVPR_2019/papers/Robinson_Large_Scale_High-Resolution_Land_Cover_Mapping_With_Multi-Resolution_Data_CVPR_2019_paper.pdf

[lzachmann]

This is great, thanks Jesse. Regarding the first link (Malkin et al. 2019), this is promising:

our proposed method is able to utilize both data with low-resolution labels and any available high-resolution labels, which we show improves performance compared to a network trained only with the same amount of high-resolution data

I do wish they had elaborated a bit more on the change detection problem. They say (in Appendix D)

We conclude that using the raw NAIP satellite images as signals yields poor results, whereas using land cover predictions by our model as signals yields reasonable land cover change detection results.

I suppose this can be taken as more evidence that change detection is best applied not as its own model (in which change is the response), but rather by inspecting a land cover map (in which cover is the response) for change over time.

On a separate note http://landcovermap.eastus.cloudapp.azure.com:4040/ (Appendix E) appears to be broken, but maybe we can ask those guys if there is a tag (and the requisite model files, etc.) in the land cover mapping repo that we can use to check this out ourselves?

One thing I don't entirely understand yet is the role of any sort of DL architecture here. I don't yet understand how / where the U-net model mentioned in a couple of places comes in, and does it apply to the high-res model only or is it also used in their super-res model? Maybe we can make this part of a convo with @tonychangmsu.