Spatial debiasing is a post-processing step applied to 0.25deg native forecasts and ERA5. It provides a better estimate for the weather for a specific ****lat/lon or a list of grid points in a location_file ****(down to a precision of 100m) by accounting for the vegetation, buildings, elevation, etc. Debiasing helps correct inherent biases present in spatial data like ERA5 Reanalysis, a source of  training data for Salient’s models. Salient uses two methodologies depending on the timescale via an API debiasing option(debias=True). In a validation context, debiasing should only be used when validating against daily station data.

<aside> 📌

Note: Currently, debiased forecasts are not available in Cloudflare.

</aside>

Debiasing Model - Daily Resolution

The debiasing model applies to Salient GEM and Salient Blend and ERA 5 Daily Downscale

The debiasing solution is a hybrid approach depending on the proximity of the location to a weather station. If the location is within 1 km of a good station, then the station-specific model uses a quantile-mapping approach and applies a different bias adjustment depending on the daily temperature (warm vs cold day). Otherwise, the debiasing uses the generalized model that uses the following training data to learn the complex relationship between ERA5 and local environment to calculate a bias correction factor.

Training data for the debiasing model includes:

• GHCNd (hourly) global station observations
  • Other station networks will be added
  • QC algorithm excludes stations with poor quality data
• Static high-resolution (<1km) remote sensing observations
  • DEM, built environment, distance to water and coast
• Training period: 1995-2021
  • ~100M samples

Validation

To ensure accuracy, the model validation consists of:

• Test period: 2021-2025
• Validation using station observations including out-of-sample stations

Endpoints and Data Fields

/forecast_timeseries for GEM at daily timescale

Data fields

• Vals_ens
• Vals

/downscale for Blend at daily timescale /data_timeseries for ERA5 at daily timescale

Supported Variables