- SEVIR ( 5 sensor inputs easy for ml format )
- Numerical Weather Model Output + Potentially with overlapping data sources
- Synthetic Ocean Model Proxies + Potentially scrap data sources
- Satellite imagery (planet data, landsat, modis, sentinel, chespeake, hyperspectral)
- Interpretability and Visualization
How can I intuitively represent predictive data for climate and weather?
How can use other datasets or self supervision tasks to learn features which can be broadly applied to other climate and weather problems
- Challenge 1 idea : Provide a feature embedding that can be fine tuned on a hidden task and evaluated
- Challenge 1 idea : Perform a task well even with noise
- Challenge 2 idea : Enforce that a predictive model follows physical constraints of a system
- Challenge 3 idea : Break a model for nwp or classification under some input pertubation constraint
- Challenge 4 idea : anomaly detection in input
How can we develop algorithms which can be indirectly evaluated
- Challenge 1 idea : estimate flow field from satellite images and synthetic imagery
- Challenge 2 idea : Sensor fusion or prediction (given 4 sensors predict 5)