Get an understanding from an EIS user group of their use cases to determine MVP features for the VEDA AP

[abarciauskas-bgse]

• [x] Meet with Elijah Orland
• [x] Review Eli's code
• [x] Establish next steps (1 or more tickets ready to be worked) to support Eli in the short term
• [x] Incorporate what we've learned into longer term plans to support EIS fire (high level tickets)
• [x] Incorporate what we've learned into longer term plans for VEDA

[abarciauskas-bgse]

I took these rough notes from our meeting Notes meeting with Eli Orland 08/25

[j08lue]

FWIW, here is a synthesis of Eli's workflow from a technical perspective.

High-level flow: Use MTBS as labels for generating training features from EVT data.

In order to run Eli's workflow in a cloud environment, we need to support the following steps:

1. Load annual MTBS raster data (third-party) [cloud]
2. Load annual EVT raster and vector data (third-party) [cloud]
3. Load fire boundaries vector data (user-provided - I assume) [cloud]
4. Subset MBTS and EVT to fire boundaries [local or cloud]
5. Reproject rasters to be pixel-aligned [local]
6. Combine all pixels and metadata into a tabular feature data store (Pandas dataframe, maybe persisted somehow) for subsequent model training and inference [local]
7. Train model on feature data. [local]

Not sure what exactly the model is supposed to predict / classify and whether the result should be published anywhere. But the key job of this workflow is to generate the training features by combining and selecting input data.

The input data sources are described here: https://docs.google.com/spreadsheets/d/1uCDLYUUkSBhKBAmkkaPIxmB65W2HjaDLHRStNSmQTbQ/edit?usp=sharing

[abarciauskas-bgse]

Nice thanks for doing this evaluation @j08lue

[j08lue]

Next step: A PoC JupyterLab environment where Eli's workflow can run, perhaps with some modifications to load data from the VEDA data store instead of from local disk.