Inspiration

Atmospheric models need to transfer to facilities they have never seen. Aether studies that challenge with reproducible site-held-out evaluation rather than random row splits.

What it does

Predicts the five official OBSCURA z-score targets from the supplied environmental variables and Sentinel-2 chips. It is a research pipeline, with saved folds, model configurations, out-of-fold predictions and validated Kaggle prediction files.

How it was built

We compared regularized linear models, CatBoost, LightGBM, clear-pixel spectral summaries, SWIR tail statistics, a frozen MIT-licensed SeCo-Eco multispectral encoder, fire-specific regressors and a compact image-plus-tabular neural network. The selected candidate combines a target-specific nested ensemble (75%) with the compact neural model (25%), averaged over two five-fold site assignments. Every learned transform and calibration weight is fitted using training sites only. Missing labels are masked separately for each target.

Results and lessons

The fixed hybrid achieved masked mean column-wise RMSE of 0.88263 and 0.88653 on the two grouped assignments. Averaging their site-held-out predictions gives 0.88244; this is validation, not a private test score. The previous image/tabular Kaggle submission scored 1.08275 publicly (submission 58554926). The validation/public gap remains substantial. No private score or final ranking is known.

Challenges

Aggressively standardized pretrained features extrapolated badly to held-out sites and were rejected. Unit-normalized embeddings and restrained calibration were more stable. Spectral fire summaries added complementary information, while methane remains difficult. Site grouping reduces leakage but cannot fully reproduce the undisclosed future-time shift.

Reproduction and attribution

Source and exact commands: https://github.com/AnshOP2008/obscura-aether-2026 . The uploaded research report records experiments, provenance, limitations and submission evidence. Kaggle participant: anshop2008; solo Devpost participant: anshurgupta1work.

The official OBSCURA starter informed baseline comparisons. Frozen SeCo-Eco weights are from TorchGeo, pinned to revision 3ede08c40f70534eb1132057ee338b99e54319f0 (MIT). Contains modified Copernicus Sentinel data and information from NASA FIRMS, ECMWF ERA5-Land and Copernicus DEM supplied by the organizers. Restricted competition data and predictions are excluded from the public repository. AI coding assistance was used for implementation and documentation; all reported scores come from executed experiments.

Built With

Share this project:

Updates

Submission history