[WIP] Download and organize Open Forest data

[bw4sz]

The OpenForestObservatory.py script was created to download Open Forest Observatory (OFO) data, match tree geographic coordinates to orthomosaic pixel locations, and generate MillionTrees-compatible annotations. Key functions include download_metadata, download_orthomosaic, and match_trees_to_orthomosaic.

Integration into the MillionTrees benchmark involved:

• Adding the OFO annotations path (/orange/ewhite/DeepForest/OpenForestObservatory/annotations.csv) to the TreePoints dataset list in package_datasets.py.
• Updating source_completeness.csv by renaming columns to source and complete, inverting boolean values, and adding "Open Forest Observatory" as a complete source.
• Adding documentation for the OFO dataset to docs/datasets.md, including a description, citation, and a generated visualization image (docs/public/Open_Forest_Observatory.png).

Test outputs, including a mock orthomosaic and sample annotations, were generated to validate the data processing pipeline and annotation format. A data_prep/OFO_IMPLEMENTATION_README.md was also created to document the integration.

[bw4sz]

From UC davis lab

Sounds good! It's very exciting to see this progress you've made with OFO data. I see your function for matching field trees to orthomosaics. One crucial note here: the field trees as we provide them are not spatially co-registered to the orthomosaics, so this code will require more logic. We have some decent (but not perfect) algorithms for this, and I know you've thought about it a lot too, so maybe this can be part of our discussion (or a separate one).

We've thought a bit about this. I agree ideally we would provide the co-registered plot data, but the challenge is that there is a 1:many mapping of field plots to drone missions, and the optimal shift differs for each. (It may also differ for different ways of processing each drone mission.) So we have been applying the shift more or less on the fly as needed for each specific analysis. We are close to having publicly available functions to make it easy for a data user to also perform the shift on the fly.

Let's wait and skip this one for 0.2 and discuss.