Contact us through the forums in case of questions or issues.
We are proud to present the Weather4cast 2024 Competitions.
In this Competition Challenge, you are asked to pre-train a generic model, which is then finetuned to individual Downstream Tasks.
This page here is for Downstream Task #1 – the prediction of cumulative rainfall in a given small area like an urban region or around an airport at a particular time.
Leaderboard and Submission links are below.
Cumulative rainfall should be averaged over a 32×32 pixel area of hi-res radar rain rates. It should be aggregated over 4h into the future (16 time slots à 15 mins). As rain rates are per hour and there are 4 slots à 15 mins per hour, that means averaging per hour and summing the 4 hours, i.e., averaging the 16 slots and multiplying by 4 (or summing the 16 slots and diving by 4).
Predictions can use 1h of input of whole-region (252×252 pixel) 11-channel multi-spectral weather-satellite data (4 time slots). One satellite pixel covers 6×6 hi-res rain radar pixels.
You can finetune your generic model with cumulative rainfall targets that you can extract from the 6 training regions and arbitrary 32×32 areas and sequences of 16 time slots at arbitrary time-points. The Leaderboard will start with a test of 40 area/time-point combinations in one unseen new region and expand to 3 such unseen new regions subsequently.
Input EUMETSAT satellite band data for the context region for which the predicion is made including satellite data for 1h ouf input are now available via sftp link in a following location:
2019/ 2019/HRIT/ 2019/HRIT/roxi_0008.cum1test19.reflbt0.ns.h5 2019/HRIT/roxi_0009.cum1test19.reflbt0.ns.h5 2019/HRIT/roxi_0010.cum1test19.reflbt0.ns.h5 2020/ 2020/HRIT/ 2020/HRIT/roxi_0008.cum1test20.reflbt0.ns.h5 2020/HRIT/roxi_0009.cum1test20.reflbt0.ns.h5 2020/HRIT/roxi_0010.cum1test20.reflbt0.ns.h5
This file is accompanied with the additional dictionary files:
mapping the indices of h5 dataset to the cases as well as the information about pixel OPERA radar pixel coordinates of test areas. Please note that the indices here are given in a python format. Below we provide the exemplary line and explanation of its conent:
Case-id,year,timeslot-start,timeslot-end,x-top-left,x-bottom-right,y-top-left,y-bottom-right TOC2,2019,20,24,437,469,972,1004
- TOC2 – indicates case ID
- 2019 – year indicating in which input dataset the timeslot exists (roxi_0010.cum1test19.reflbt0.ns.h5 or roxi_0010.cum1test20.reflbt0.ns.h5)
- timeslot-start – the stating index of the input timeslot
- timeslot-end – the ending index of the input timeslot (please note there are four input timeslots for each case so the indexes are given in a python indexing scheme)
- x-top-left – x top left corner pixel of the test area for which prediction needs to made
- x-bottom-right – x bottom right corner pixel of the test area for which prediction needs to made (please note the are size is 32×32 pixels and the pixel indexes are given in a python indexing scheme)
- y-top-left – y top left corner pixel of the test area for which prediction needs to made
- y-bottom-right – y bottom right corner pixel of the test area for which prediction needs to made
Predictions are evaluated using a Continuous Ranked Probability Score (CRPS) that support (but do not require) probabilistic forecasts. Submissions need to submit empirical cumulative distribution functions (eCDF) for each case. Tables must have 3 columns giving the ID of the forecast case, the cumulative rain rate for 4h over the 32×32 area, and the cumulative probability for that rain rate. One table is required collecting all cases to be forecast for each region and year (e.g., ‘roxi_0010’ & ‘2019’). So for 3 regions and 2 years your submission will be a zip archive with 6 tables.
You are encouraged to start each eCDF with a value where you forecast the cumulative probability to be zero and end it with a value where you forecast the cumulative probability to be one (otherwise the scoring function will insert such values).
For non-probabilistic forecasts, the prediction table would simply hold the case ID, the forecast value, and the cumulative probability of 1.0.
Submission zip archives must have the following structure, with each file being in CSV format without headers:
2019/ 2019/roxi_0008.test.cum4h.csv 2019/roxi_0009.test.cum4h.csv 2019/roxi_0010.test.cum4h.csv 2020/ 2020/roxi_0008.test.cum4h.csv 2020/roxi_0009.test.cum4h.csv 2020/roxi_0010.test.cum4h.csv
Below we provide several exemplary submission files:
- non-probabilistic forecasts
- probabilistic forecasts for 17 bins representing probabilities for cumulative rain values: 0, 32, 64, 96, …. , 416, 448, 480, 512 [mm/h]
- probabilistic forecasts for 129 bins representing probabilities for cumulative rain values: 0, 4, 8, 12, …. , 500, 504, 508, 512 [mm/h]