ARVA: Agricultural Robustness across Varied Areas

This dataset accompanies the ARVA paper and provides the frozen representations, split assignments, and result tables used in its empirical evaluation. ARVA studies geographic transfer for Earth-observation foundation models while varying target-region label access and holding the target test examples fixed.

Repository structure

• results/consolidated-final-grid-8seed-8b647d1/: per-run metrics, predictions, and provenance for the eight-seed evaluation.
• results/final-analysis-8seed-8b647d1/: analysis tables for regional performance, learned-versus-raw comparisons, calibration, split sensitivity, target-label allocation, regional screening, and selective prediction.
• splits/ and splits.json: frozen split assignments and their central manifest.
• cache/benchmark/: assembled caches for CropHarvest, EuroCropsML, BreizhCrops, and PASTIS-R.
• cache/embeddings/: frozen model and raw-input representations.
• RELEASE_SCOPE.json and FILE_INVENTORY.tsv: machine-readable release metadata and a complete file inventory.

Reproduction

The corresponding code and usage documentation are available at https://github.com/akshithio/arva. The reported evaluation uses scientific revision 8b647d1ea9747bb80a4da44ebedea76cd0628e9d.

The result directories contain completion manifests and artifact hashes for checking provenance before reproducing or extending an analysis.

Licensing

Benchmark data, model outputs, and derived artifacts remain subject to their respective upstream licenses and terms.