Checklist IOAI 2025 At-Home Round · Task 2
Satellite Weather Forecasting
Segment rain and no-rain pixels in GOES-16 satellite image patches, and decide whether each patch contains any rain at all.
The task
Radar and ground sensors measure rain well but are expensive and often missing in remote regions. As an intern at a climate lab, you test whether rainfall can be detected from satellite imagery alone, helped by context such as sun angle, time and location.
The inputs are 16-channel GOES-16 ABI patches (128 × 128 or 256 × 256) from 2024, with per-pixel rain masks from NOAA's MRMS system and metadata such as patch location and capture time. A pretrained U-Net baseline is provided.
Training patches are biased towards rain (at least 3% rainy pixels), while evaluation data mirrors real conditions, with many dry patches and some deliberately missing or corrupted channels. You may fine-tune the U-Net, use the metadata and change post-processing, but not use outside data or other pretrained models.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- 16-channel GOES-16 patches (128×128 or 256×256) with binary MRMS rain masks and metadata (top-left latitude/longitude, start and end times, a sun-elevation helper); test inputs come as
X_test.npzwithmetadata_test.csv. The official repo doesn't include the training data; the task notebook points to a Google Colab notebook for downloading it. - You submit
submission.zipwithpred_a.npz(public test set) andpred_b.npz(private test set), each holding boolean arrays Y_pred_128 (51×128×128) and Y_pred_256 (183×256×256).- Scoring
- Final Score = (Mean Dice Score + Image-level Rain Accuracy) / 2, where Dice = 2 × |intersection| / (|prediction| + |ground truth|) and image-level accuracy checks whether any rain is present in the image.
- Rules
- No external datasets
- No external pretrained models other than the provided baseline U-Net
- Do not look up weather-forecasting articles on the internet
- Format
- At-home round (non-graded practice, 1-30 July 2025)