Hi everyone,

Thank you for your patience, and thank you to those who raised concerns about the original challenge data. We have reviewed the issue carefully with the competition organisers and have made updates to the challenge files to protect the integrity of the competition. We apologise for the delay in communicating this.

We have replaced the challenge files with a revised dataset and challenge setup. The original hidden test set will no longer be used for evaluation and has been incorporated into the public training data with labels instead. The new setup is:

• Train: original train data + original test data with labels
• Test: new data supplied by the organisers
• Leaderboard: reset using the new public/private references
• Submission limits updated: now 5 per day 100 total

Longitude and latitude have also been removed from the public files. Important change to the test data: The new test file is designed to better reflect a real-world partial-observation setting. Each test row still represents one sample, and you will still submit one prediction per test ID. However, each test sample now only includes a consecutive block of 4, 5, or 6 months of satellite observations. The remaining months are masked as -9999. This means that models should be robust to incomplete temporal observations and should not rely too heavily on a single month or feature. The challenge is intentionally designed this way to prevent models from relying solely on summer spectral signatures - robust pond detection requires using whatever consecutive window is available, regardless of season.

What do participants need to do?

• Please download the updated files and submit using the new sample submission format.
• Previous submissions were made against the old test set and are no longer valid. The leaderboard will be reset so that all participants are evaluated on the updated data.

Scoring format reminder The submission file contains two prediction columns:

• TargetF1: binary prediction, 0 or 1
• TargetRAUC: probability between 0.0 and 1.0

Please make sure both columns are included in your submission.

Note on class distribution: The training set contains approximately 40% positive (pond) examples. The test set may have a higher proportion of positives. This is an inherent property of the dataset split and cannot be changed. We recommend participants account for this.