In a nutshell:
In 2024 the Weather4cast competition at NeurIPS aims to improve rain forecasts world-wide on an expansive data set with over a magnitude more hi-res rain radar data, allowing a move towards Foundation Models through multi-modality, multi-scale, multi-task challenges. Besides keeping data fusion and super-resolution tasks,
this year, on the road to flexible foundation models, we move from basic precipitation prediction to testing generalization performance and emergent capabilities of probabilistic models on a set of downstream tasks, such as cumulative rainfall and unusual event prediction.
Competition tasks
To assess how well models develop emerging skills, in this year we introduce downstream tasks on which we will test the generalization performance of the models. In particular, we provide two application scenarios:
- Cumulative rainfall prediction (total rainfall in a given time span and small area, like a city / urban region)
- Prediction of rare extreme weather event (event location and extent in time and space, event severity)
…where you will be provided with 1 hour of before the prediction timeframe (either satellite only or satellite+radar input will be provided).
For the model pre-training task we share with you a dataset including satellite data and rain rates from OPERA high resolution radar. The satellite data consist of 11-band spectral satellite images. These 11 channels show slightly noisy satellite radiances covering so-called visible (VIS), water vapor (WV), and infrared (IR) bands. Each satellite image covers a 15 minute period and its pixels correspond to a spatial area a few kilometer across (4 km at the equator, 12 km in the North). In addition, this year, for the first time, we also release the 2 km resolution rain rates for the whole context regions , and these high resolution rain data can also be used as an input in the model training phase. You can pre-train your model entirely as you like, including but not limited to
- translation, e.g., satellite input to radar output (same time)
- future prediction, e.g.,
- future satellite from past satellite data
- future radar from past radar input data
- future radar from past satellite + radar data
- future radar from past satellite data only
- spatial gap filling
- super-resolution
You can train at different spatial and temporal resolutions.
You can also incorporate any additional public data or model output, as long as you declare that and these are available to other scientists free of charge for academic / non-commercial use.
Downstream tasks are defined on the high-resolution 2 km grid.
To start playing with the data and train your networks, you can use our Starting Kit from the last year’s competition. Please note that it predicts OPERA rain rates for small region of interest form the satellite context data.
The competition leaderboards are coming soon! Please note that this year we will test the performance of your models on different downstream tasks, such as cumulative rainfall prediction and extreme precipitation event prediction. We will not assess whole frame rain rate prediction directly as in past years.
We will jointly publish a NeurIPS Databased & Benchmarks Proceedings article this year. We trust that, like last year, the NeurIPS organizers will provide free registrations for leading submissions.