Probabilistic Weather Forecast Downscaler — Tunisia (28 km -> 5.5 km)

Open in Spaces

Live Forecast Player & Verification Explorer: Explore live ECMWF AIFS-ENS, IFS-ENS, and NOAA GEFS ensemble forecasts downscaled to 5.5 km over Tunisia — plus interactive 17-station WMO verification and spatial power spectra — in InfectV22/tunisia-downscaling-demo (fullscreen app).

An ensemble U-Net downscaler that sharpens coarse 0.25° (~28 km) 2 m temperature fields onto the Copernicus CERRA 5.5 km grid over Tunisia (32.0°N–38.0°N, 7.5°E–11.8°E).

Trained strictly on historical reanalysis pairs (ERA5 0.25° -> CERRA 5.5 km, 2014–2019) with Continuous Ranked Probability Score (CRPS) loss over land, the model never sees numerical or AI weather forecasts during training — allowing it to be applied zero-shot to any coarse forecast model (such as ECMWF IFS-ENS or AIFS-ENS) at inference time.

Property Value
Task Probabilistic downscaling of 2 m temperature (0.25° -> 5.5 km CERRA Lambert grid)
Region Tunisia (32.0°N–38.0°N, 7.5°E–11.8°E · 127 × 77 fine grid)
Architecture Multi-scale Probabilistic U-Net (~2.15M parameters · base_channels=32, depth=3)
Loss Fair CRPS over land (land_only=True, 4 train members / 8 eval members)
Data Split Train: 2014–2019 (8,764 6-hourly steps) · Held-out Validation: 2020 (1,464 steps)
Inputs 11 physical & spatiotemporal channels + 2 multi-scale noise channels (coarse_factor=[8, 16])

Forecast Downscaling Overview


1. Held-Out Validation Results (2020, 1,464 Steps vs. CERRA 5.5 km)

Evaluated across the full held-out year 2020 (00, 06, 12, 18 UTC) over Tunisian land grid cells (N = 9,290,544 pixel-time pairs):

Elevation / Subset Bilinear Interp RMSE (K) CRPS U-Net RMSE (K) Interp CRPS (K) CRPS U-Net CRPS (K) Spread-Skill Ratio (SSR)
All Land 1.383 0.932 (-32.6%) 1.046 0.475 (-54.6%) 0.860
Lowlands (0–200 m) 1.310 0.927 (-29.2%) 0.970 0.470 (-51.5%) 0.863
Hills (200–500 m) 1.377 0.941 (-31.7%) 1.057 0.482 (-54.4%) 0.847
Highlands (500–1000 m) 1.554 0.934 (-39.9%) 1.212 0.478 (-60.6%) 0.865
Dorsale Mountains (1000–1500 m) 1.731 0.933 (-46.1%) 1.403 0.477 (-66.0%) 0.899

Diurnal RMSE stays consistent across all synoptic hours (00 UTC: 0.920 K, 06 UTC: 0.946 K, 12 UTC: 0.910 K, 18 UTC: 0.952 K).

Validation Maps

Training Curves


2. Effective Spatial Resolution (227.5 km -> 13.8 km)

Unlike deterministic MSE regression (which blurs small scales toward the conditional mean), each stochastic CRPS member preserves fine-scale coastal and topographic variance across the Tunisian Dorsale and Mediterranean coastline:

  • Coarse input (0.25° ERA5): 227.5 km effective wavelength (at 50% target spectral power)
  • Downscaled 5.5 km member: 13.8 km effective wavelength (16.5× finer spatial resolution)

Spatial Power Spectrum


3. 10-Day Forecast Inference & Independent Tunisian Station Verification

Applied zero-shot to 10-day (+6h to +240h) ECMWF IFS-ENS 0.25° ensemble forecasts (10 coarse members × 4 stochastic realizations = 40 fine-grid ensemble members) and verified independently against both CERRA 5.5 km and 17 WMO / INM Tunisian synoptic weather stations retrieved from the NOAA Integrated Surface Database (ISD):

  • Against CERRA 5.5 km over Tunisia land:
    • Forecast RMSE reduced from 2.18 K -> 1.69 K (-22.5%)
    • Forecast CRPS reduced from 1.53 K -> 0.95 K (-37.9%)
  • Against 17 independent Tunisian weather stations (reports/summary.json):
    • Overall station CRPS reduced from 1.173 K -> 1.028 K (-12.4%) (reaching -38% at +24h and -32% at +48h)
    • Station RMSE reduced from 1.856 K -> 1.803 K (-2.8%), cold bias reduced from -0.868 K -> -0.693 K
    • Ensemble Spread-Skill Ratio (SSR) improved from 0.465 -> 0.644 (0.83 by Day 10)
    • 100% of inland/mountain stations (>200 m) improved (+12.9% CRPS skill at 200–500 m, +31.1% at 500–1000 m)
    • Extreme heat (>35 °C) detection hit rate rose from 68.2% -> 75.0% (Brier error reduced by -25%)

Coarse vs CERRA vs Downscaled

Error Reduction vs CERRA

10-Day Ensemble Plume at Tunisian Mountain Station

Evaluation Scorecard

Station Skill Improvement


4. Input Channels & Usage

All physical and static predictors are stored or computed on the 5.5 km grid and normalized using training-period (2014–2019) statistics bundled directly inside model.pt:

  1. Coarse 2 m temperature bilinearly interpolated onto the 5.5 km grid
  2. Fine-scale 5.5 km terrain elevation (orography) and elevation difference vs. the coarse surface
  3. Fine-scale land-sea mask and exponential coastal influence (25 km decay scale)
  4. Normalized latitude and longitude
  5. Time of day (sin, cos) and day of year (sin, cos)
  6. 2 multi-scale Gaussian noise channels (coarse_factor = [8, 16]) for ensemble generation

Python Inference

Using the weather-forecast-model-downscaler package:

from huggingface_hub import hf_hub_download
from downscale.downscaler import Downscaler

ckpt = hf_hub_download(
    repo_id="InfectV22/tunisia-temperature-downscaler",
    filename="model.pt",
)
model = Downscaler.load(ckpt)

# coarse_kelvin: (time, y, x) interpolated onto the model's 127x77 fine grid
# returns: (time, members, y, x) in Kelvin
fine_ensemble = model.predict(coarse_kelvin, valid_times, members=4)

model.pt carries the 127 × 77 Tunisia grid coordinates, static terrain, coastal distance field, and normalization statistics, so no external static files are required at inference time.


5. Limitations

  • Single-variable coarse input (t2m only): Without wind, humidity, or cloud cover inputs, the model learns terrain lapse rates, coastal gradients, and diurnal/seasonal cycles accurately, while flow-dependent mesoscale phenomena (such as sirocco intrusions or sea-breeze fronts) are captured only on average.
  • Reanalysis target inheritance: Because the model is trained against 5.5 km CERRA reanalysis over land, it inherits any local systematic biases present in CERRA; independent WMO station verification (reports/summary.json) provides the ground-truth benchmark.
  • Regional domain: Trained exclusively over Tunisia (32.0°N–38.0°N, 7.5°E–11.8°E); do not apply outside this domain without retraining.

6. Licence & Attribution

Pipeline architecture adapted from EmmaScharfmann/weather-forecast-model-downscaler (Do we really need bigger AI weather forecasting models to improve resolution?), extended to Tunisia with a NOAA ISD station verification provider.

Weights and evaluation artefacts derive from the following upstream datasets and retain their respective terms:

  • CERRA (5.5 km) & ERA5 (0.25°) Reanalysis: Copernicus Climate Change Service (C3S). Generated using Copernicus Climate Change Service information.
  • IFS-ENS (0.25°) Forecasts: European Centre for Medium-Range Weather Forecasts (ECMWF) via WeatherBench 2 (CC-BY-4.0).
  • Station Observations: Tunisian synoptic weather stations (WMO / Institut National de la Météorologie), accessed via the NOAA Integrated Surface Database (ISD) (ncei.noaa.gov).

Neither Copernicus, ECMWF, NOAA, nor INM endorses this work.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using InfectV22/tunisia-temperature-downscaler 1