SMAC

Please report issues to the Forum in CodaBench during the competition.

Dataset

The dataset comprises Sentinel-1 SAR imagery with two multispectral bands (VV/VH). We provide two classification labels (unaffected / affected by earthquake) and a real value representing the earthquake magnitudes for each sample.

Each sample is composed of:

• image with four channels. It contains VV and VH channels for two images at times t0 and t1 (where t0 < t1)
• label contains a binary value (0 for unaffected and 1 for affected area)
• magnitude contains a real value in the range 0-10, representing the magnitude in mb

NOTE: The dataset is implemented in TorchGeo (see main.py in the starter-kit)

Submission

The file submission.csv contains a sample submission with the following columns:

• key: unique identifier
• magnitude: predicted magnitude (should be in the range 0-10)
• affected: binary label (0-1)
• flops: resource consumption expressed in FLOPs by PAPI

You can use submission_creator.py with arguments --predictions {prediction_csv_file} and --flops {estimated_flops} to "compile" a submission.csv with FLOPs and your predictions. This file also does some basic checks on your submission.

Starter Kit

In the starter-kit folder, you can find the code to run the baseline using main.py.

requirements.txt contains the libraries required to run the code.

You can run inference thanks to inference.py simply passing your saved checkpoint with --checkpoint {checkpoint} to the command line.

Private Set

You can find the private test set on HuggingFace. It is compatible with QuakeSet class of TorchGeo and you can load it in this way: ds = QuakeSet(root="private_set", split="test").

You can iterate through the dataset and the IDs in the following way:

predictions = []
for metadata, sample in zip(ds.data, ds):
  out = model(sample)
  predictions.append({"key": metadata["key"], ...})

NOTE: the labels and magnitudes are only placeholders in this case. They are not the ground truth.