karpos-downscaling-baronnies

karpos-downscaling — Baronnies v1

Reference weights for the residual U-Net with FiLM-conditioned terrain modulation described in:

• karpos-downscaling: A meteorological super-resolution framework bridging numerical weather prediction and machine learning (JOSS, in submission)
• FiLM-conditioned residual U-Net downscaling of frost-risk temperature fields (GMD, in submission)

Code repository: https://github.com/maurinl26/downscaling

Model description

Four-level U-Net (≈ 19 M parameters; exactly 19,053,030 across 246 tensors) with FiLM modulation x ← γ(DEM) · x + β(DEM) applied at every encoder and decoder level. A small pyramidal DEM encoder produces spatially-varying (γ_ℓ, β_ℓ) maps shared between encoder and decoder. The output is residual: ŷ = t2m_input + δ(x_met, DEM). Training combines a reanalysis target (stage A) with tail-weighted in-situ supervision against Sencrop minima (stage C).

Training data

• ERA5-Land 9 km (Copernicus CDS, open access) — meteorological inputs
• CERRA 5.5 km (Copernicus CDS, open access) — stage-A target
• COP-DEM 30 m → 1 km (ESA, open access) — terrain conditioning
• Sencrop 48 stations, Baronnies Provençales (proprietary usage agreement) — stage-C ground truth

Training period: 2015–2021 (TRAIN). DEV held-out: 2022–2025. PURE TEST 2026 reserved for a single end-of-cycle evaluation.

Intended use

• Reproducing the JOSS and GMD figures and ablations
• Benchmarking new super-resolution architectures on the same temporal splits
• Research on FiLM-conditioned image-to-image regression in meteorology
• Building blocks for AIWP fine-tuning pipelines that need locally-supervised target fields

Out-of-scope use

• Direct operational frost detection for parametric insurance — for production use, see the Karpos commercial service. The released weights are MVP-grade reference checkpoints, not production-calibrated.
• Extrapolation outside the Baronnies relief class without retraining. Out-of-domain behaviour is documented as a known limitation (cf. paper §7.2).
• Real-time operational decisions without a documented downstream calibration and quality-monitoring pipeline.

Performance (DEV held-out 2022–2025)

All figures averaged over 5 seeds with bootstrap confidence intervals. See the GMD paper §6 for the complete table, ablations, and pairwise McNemar tests against the QDM statistical baseline.

How to use

from downscaling.hub import load_baronnies_v1

model = load_baronnies_v1(device="cuda")
# model: torch.nn.Module ready for inference
# expected input shape: (B, C_meteo + C_dem, H, W) with H × W = 168 × 120

For the full evaluation harness, baselines, and Hydra configurations, see the GitHub repository.

Limitations and known biases

• Discriminative ceiling at AUC ≈ 0.96, imposed by the CERRA reanalysis target. Adding meteorological predictors (dew point, wind) does not raise the ceiling on this dataset.
• MSE-based in-situ calibration degrades detection (mean-regression pathology). Tail-weighted supervision is required for the cold tail; see the GMD paper §5.
• Trained on the Baronnies pre-alpine basins. Behaviour on hillside vineyards (Côtes du Rhône, Bourgogne, …) is the subject of ongoing follow-up work.
• No surface-state conditioning yet. SURFEX integration (soil moisture, canopy temperature, vegetation type) is planned for v2; see paper §7.3.

Citation

Software release:

@software{karpos_downscaling_baronnies_v1,
  author       = {Maurin, Loïc},
  title        = {karpos-downscaling — Baronnies v1},
  year         = 2026,
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.20783563},
  url          = {https://huggingface.co/karpos26/karpos-downscaling-baronnies}
}

Companion papers — see CITATION.cff in the GitHub repository.

License

Weights released under Creative Commons Attribution 4.0 (CC-BY 4.0). Code in the companion GitHub repository released under Apache 2.0. The Sencrop sensor data used during training is proprietary and is not redistributed with these weights; retraining requires an independent Sencrop usage agreement.

Authors and acknowledgements

Loïc Maurin (External Lecturer in Stochastic Filtering, École Nationale de la Météorologie, Météo-France, Toulouse).

This work was conducted in the context of the Karpos project. The author thanks Sencrop for sensor-network access, Copernicus for CERRA and ERA5-Land, and the SURFEX/AROME development teams at Météo-France/CNRM.

Release strategy

The Baronnies weights are released as a public reference MVP: they support reproducibility of the JOSS and GMD papers and academic follow-up work, without exposing any commercially calibrated configuration. Region-specific production weights (Côtes du Rhône and beyond) will be released publicly after the commercial moat is built in each region (typically T+18 to T+24 months after the first signed insurance contract, one validated frost season, and a renewed customer cycle).