A complete result in one day.

Before kickoff: transfer and environment checks

Download, unpack, verify checksums, read technical instructions and check accounts, packages and compute with unrelated synthetic examples. You may use the neutral practice CSV for delivery checks. Challenge-data exploration, derived features, training, tuning and visual insights start at kickoff, including work performed by AI agents.

Focus on the main challenge first

Task 1 is the priority: build a model, submit an accepted final prediction, document your development result and method, and prepare your presentation. Tasks 2 and 3 are optional extensions for teams that finish early. Your model comparisons, findings and limitations can provide the insights and visualizations for judging.

Choose a route you can finish

Start with the supplied CPU model and a working keyed export after kickoff. Keep a short log of hypothesis, features, split, seed, runtime and result. Change one important choice at a time. Optional five-fold CV helps compare promising models; it is not required and does not earn points on its own.

Use the early test-inference window

1. Run the common German development comparison and rehearse practice delivery early.
2. Develop one or two useful model comparisons and visualize the results so you can explain your choices.
3. Copy required encrypted test inputs onto local runtime storage before release. Keep a tested table model as a fallback.
4. At 13:30 German time, unlock and verify the final inputs, infer all 63,590 requests and validate the complete sample key set.
5. Open your team workspace, select the CSV and use Encrypt & submit. Wait for a numbered receipt and an accepted result. Each newer accepted final file becomes the starred choice. To choose another accepted receipt, click its star and confirm before the deadline. The leaderboard follows that choice. Share method/code references and upload the presentation through the linked Google form before the deadline.

Explain your main result

Present what you learned from building and evaluating your model: which choices helped, where it struggles and what the data can support. Use clear visualizations of your experiments or model errors and explain uncertainty and reporting biases.

Use five minutes for the question, approach, common-development result, two or three findings and limitations. The jury then has two minutes for questions. Open the Google presentation form from your team workspace. Upload a PDF up to 10 MB and, optionally, an editable source up to 100 MB. Google sign-in is required. Retain the confirmation page; organizers use the latest complete valid response before the deadline.

Upload your presentation →

Finished early? Choose an optional extension

Once your submission, method references and presentation are ready, Task 2 offers additional exploration of the German bird observations, such as coverage, source composition or seasonality. Task 3 tests whether other species help your model. Both are optional. For any cross-species claim, compare with and without helper information on identical focal labels, supports and splits. A well-controlled negative result can be valuable.

Questions teams often ask

Why train on Germany instead of all Europe?

The learning and exploration release has a clear German boundary. Final evaluation tests geographic transfer to nearby Central Europe while foreign bird observations and exact evaluation locations stay with the organizers. Every team receives the same permitted data.

Will a city park disappear in a coarse average?

It can. The native vector and 10 m local patch retain finer environmental detail, while context summaries and wider patches describe the landscape. Compare supports rather than assuming one resolution is best. Fine pixels do not remove uncertainty in the reported bird position.

Are 100 m and 1 km features single coarse pixels?

No. They summarize circular neighborhoods of those radii. Native vectors, context summaries, 10 m patches and 40 m patches are different environmental supports; the data schemas explain each representation.

Can I use an ordinary image model?

You can develop spatial architectures, but inputs have 64 environmental channels and validity masks, not RGB colors. Apply documented decoding and normalization. Generic brightness augmentation or ImageNet color preprocessing has no automatic meaning here.

Must I train five models for CV?

No. Everyone reports the common German development result; five supplied spatial folds are optional. Learned preprocessing and helper stages must still obey the partition and exclusion rules.

Can I refit on training plus development?

Yes, after model selection and with a declaration. Retain the earlier honest held-out development score and use the same final inference interface.

Can I download additional bird observations?

No. Additional bird labels, range maps and pretrained bird-distribution predictions are prohibited. Additional environmental sources require equivalent organizer-provided training and test inputs for all teams.

Does background zero mean the bird was absent?

No. It is a matched recording opportunity sampled for this benchmark. Unreported species remain unknown. Scores rank reported seasonal presence against that fixed background rather than measuring calibrated occupancy.

How can other species help at final locations?

Learn transferable helper signals inside the permitted training partitions and predict them from the supplied final inputs. Observed helper labels at final locations are not supplied. Use out-of-fold or another leakage-safe construction and a matched control.

What if my scoring receipt arrives after the deadline?

The server receipt timestamp controls eligibility. A timely valid submission can finish evaluation later. Keep the numbered receipt from your team workspace. The upload must be verified and its receipt confirmed before the deadline; uploading a file alone is not enough.

Is the private score selected from my best attempt?

No. Both the provisional leaderboard and final performance use your starred accepted prediction. Each newer accepted final CSV becomes starred automatically; you may star another accepted receipt and confirm until the deadline. Five accepted final predictions apply across the whole event, not per day.

Do slide uploads need dataset editing permission?

No. Use the Google presentation form linked from the team workspace and sign in with Google. Its response files are private to the organizers. Challenge data remains a separate download-only folder. Submit your presentation before the work deadline.

Reproducibility handoff

Complete the method report with feature choices, seeds, runtime, common-development result, optional CV, AI/API and compute use, permitted-data and preparation declarations, any final refit and the actual CSV SHA256. Provide the exact code/checkpoint reference. Keep source notices with derived work and retain the limitations of annual embeddings, coordinate uncertainty and reported observations.

Read the full rules → · Check and encrypt predictions → · Data sizes, species and licenses →

Registration, submissions and receipts

1. Sign in with GitHub and create your team by entering its name. You join automatically as team lead; share the invitation link so teammates can sign in with GitHub and join. Teams have two or three people.
2. Wait for organizer approval, then rehearse using the practice CSV.
3. Select your prediction CSV in the team workspace. The app encrypts and uploads it privately.
4. Keep the numbered receipt. Queued means evaluation is pending; accepted means the file passed; rejected includes a reason.
5. Upload the PDF through the linked Google form and check your starred final submission before the deadline.

Receipt codes

FORMAT: check header, exact row IDs and finite scores in [0,1]. ENVELOPE: encrypt the original CSV again. DUPLICATE: identical predictions were already accepted, even if reordered. WINDOW / CLOSED: received outside the open window. QUOTA: accepted allowance used. RATE: invalid-request limit reached; contact an organizer. TEAM: team-member authorization failed.

Pending evaluation preserves the server receipt time. If upload completed but confirmation failed, retry receipt confirmation; the same ticket cannot create a second receipt.