Team Pixel Prophets (3 people) at the StatsGeeks Society Data Science Hackathon, September 2026. Won the Best F1 Score category out of six teams.
Building-age transfer learning · Madrid → Amsterdam. We classify 30 m Landsat pixels into four construction-era classes using Madrid source knowledge and a small labelled Amsterdam support set.
Coordinate-RF — macro-F1: 0.749299 ± 0.005716.
Locked at 200 labels/class, across 200 paired, fixed-city random-pixel episodes. The ± value is population SD, not a confidence interval. This is a transductive, full-pool result—not an organizer-held-out or unseen-city score.
Judges: follow this order; no execution is needed.
- FINAL_SUBMISSION/ — the complete current submission.
- Presentation: PDF · PowerPoint.
- Written justification — approved revised explanation.
- Notebook — inspect saved outputs without running cells.
- Model — frozen state, configuration and inference interface.
- Results — locked tables and report.
- Repository documentation — methods, limitations, reproducibility and presentation guide.
| Folder | Purpose |
|---|---|
FINAL_SUBMISSION/ |
Authoritative current presentation, PDF, justification, logo and unchanged scientific payload |
docs/ |
Judge navigation, methodology, reproducibility and one active presentation guide |
presentation/ |
Navigation to the final deck; older decks/tooling are historical, not upload targets |
FINAL_SUBMISSION/results_summary/ |
Submitted results; there is no root results/ folder |
FINAL_SUBMISSION/model/ |
Submitted model; there is no root model/ folder |
candidate/ |
Canonical scientific reports, models, predictions and experiment history |
reproducibility/ |
Integrity records and historical release receipts |
figures/ |
Existing figures; canonical scientific figures also remain under candidate/ |
scripts/ |
Existing maintenance and validation utilities; older release checks target older layouts |
data/, original/, working/ |
Supplied data, original notebooks and preserved research work |
archive/ |
Superseded presentation materials and historical documentation; not current submission |
| Deliverable | Exact location |
|---|---|
| Presentation | FINAL_SUBMISSION/presentation.pptx |
| FINAL_SUBMISSION/presentation.pdf | |
| Written justification | FINAL_SUBMISSION/written_explanation/written_justification.txt |
| Notebook | FINAL_SUBMISSION/solution_notebook.ipynb |
| Model | FINAL_SUBMISSION/model/pool_state.pkl + configuration |
| Submission README | FINAL_SUBMISSION/README.md |
| Requirements | FINAL_SUBMISSION/requirements.txt |
Scientific artifacts are frozen: models, predictions, notebooks, evidence, reports, CSVs, methodology and scientific hash files are unchanged by final synchronization. Saved outputs, locked configurations and existing replay evidence document the submitted work.
Safe byte check: from FINAL_SUBMISSION/, run sha256sum -c checksums.sha256. Only documentation/presentation entries in the distribution manifest were updated. No training, inference or notebook execution was performed. Runtime reproduction is documented in the submission README; it fits support forests and is not a release-audit step. Large ignored binaries may be missing from a Git clone—verify the actual upload.
The artifact requires the original ordered Amsterdam pool. The additional 800-label development bank, historical audit exposure, weaker spatial results and unresolved organizer eligibility are disclosed in the justification. See the inventory and remaining checks.
Pixel Prophets
Presentation guide · License (MIT, team code only) · License and redistribution status
MIT, see LICENSE. Organiser-supplied competition data is not covered; see license status.