DeepScale

Modular downscaling, calibration, and verification for seasonal climate forecasts.

DeepScale turns coarse global-model (GCM) forecasts and fine-resolution observations into calibrated, high-resolution forecast products, and scores them with cross-validated skill metrics. It operates on xarray arrays and is agnostic to where the data came from, so it pairs naturally with a data layer like Rosetta but does not require it.

Downscaling methods, skill metrics, and ensemble strategies are looked up by name from a registry, so you select them with plain strings and can add new ones without changing the orchestration code.

Installation

pip install accord-deepscale

The distribution is published as accord-deepscale; the import name is deepscale:

import deepscale

The shapefile and region-clipping helpers additionally require Rosetta:

pip install accord-rosetta

DeepScale requires Python 3.10 or newer.

Core API

import deepscale

# Bias-correct and downscale one model against observations.
result = deepscale.downscale(gcm, obs, method="bcsd")

# Turn a predictor into below/normal/above tercile probabilities.
probs = deepscale.calibrate(predictor, obs, method="ereg")

# Try several methods and keep the most skillful.
best = deepscale.optimize(gcm, obs, methods=["bcsd", "cca"])

# Combine multiple models into one forecast.
mme = deepscale.ensemble([model_a, model_b], obs, strategy="uniform")

# Score a forecast against observations.
report = deepscale.skill(forecast, obs, metrics=["rpss", "roc"])

Inputs are xarray arrays with CF-style coordinates. A GCM hindcast has dimensions (year, member, lat, lon) and observations have (year, lat, lon). Outputs are continuous or tercile forecast products plus skill summaries and maps. Terciles are ordered [0, 1, 2] for below-normal, normal, and above-normal.

What is included

Everything below is selected by name.

Downscaling and bias-correction methods, passed as method= to downscale() and optimize():

Calibration methods, passed as method= to calibrate():

Ensemble strategies, passed as strategy= to ensemble(): uniform, skill_weighted, bma, drop_worst.

Skill metrics, passed as metrics= to skill(): rpss, roc, roc_area_below_normal, roc_area_above_normal, generalized_roc, pearson_r, spearman, 2afc, root_mean_squared_error, mean_square_skill_score (msss), continuous_ranked_probability_skill_score (crpss), heidke_skill_score, reliability, spread_error_ratio, spread_error_correlation.

Cross-validation schemes: loyo (leave-one-year-out), lko (leave-k-out), blocked, expanding.

Daily aggregations

Rainy-season timing and dry-spell statistics, for products that need a date rather than a seasonal total. These take continuous daily rainfall rather than the seasonal arrays above, and are module-qualified rather than selected by name:

from deepscale import aggregations as agg

onset = agg.onset(daily, season="MAM")                       # default criterion
cessation = agg.cessation(daily, season="MAM", after=onset)
length = agg.season_length(onset, cessation)
spells = agg.dry_spell(daily, season="MAM")

Onset defaults to 20 mm across 3 consecutive days, rejected as a false start if a 7-day dry spell falls within the following 21 days. Every threshold is a keyword argument, so other regional definitions are one call away, and the values used are recorded on the output for provenance.

Timing results carry a three-state occurred field distinguishing a season that failed from a cell with no data, which a NaN date alone cannot express. Output dims are (year, lat, lon), the same shape the rest of the library takes for observations. Full detail, including how much daily data each function needs past the season end, is in skills/deepscale/references/aggregations.md.

Example workflow

The repository ships a runnable end-to-end demo:

python examples/demo_forecast.py

It uses Rosetta to fetch ERA5 temperature observations (obs/era5) and ECMWF seasonal hindcasts (c3s/ecmwf-monthly), reshapes them into DeepScale inputs, then runs optimize, tercile conversion, and skill scoring. Rosetta handles the remote retrieval and normalization; DeepScale starts from the prepared xarray datasets.

The demo needs CDS credentials in ~/.cdsapirc with the relevant dataset licences accepted (see the Rosetta README for setup). examples/README.md lists all demos and their prerequisites.

Calibration

deepscale.calibrate() produces tercile probabilities with dims (tercile, lat, lon) directly. Use it when the predictor is already on the target grid, or when a scalar index drives the forecast.

Ensemble regression (method="ereg")

eReg fits each model independently with per-grid-cell ordinary least squares: the ensemble-mean hindcast predicts the observed field, and the chosen forecast year is converted to parametric tercile probabilities. Multiple models are averaged after each produces its own probability map.

probs = deepscale.calibrate(
    {
        "ecmwf": (ecmwf_hindcast_on_obs_grid, ecmwf_forecast_on_obs_grid),
        "ukmo": (ukmo_hindcast_on_obs_grid, ukmo_forecast_on_obs_grid),
    },
    obs,
    method="ereg",
    forecast_year=2026,
)

Each hindcast needs a year dimension, an optional member dimension, and spatial dimensions named lat/lon, latitude/longitude, Y/X, or y/x. eReg calibrates, it does not regrid, so put model fields on the observation grid first. If every forecast contains exactly one year, forecast_year is inferred.

Logistic index calibration (method="logit")

logit fits a gridded logistic relationship between a scalar predictor index and observed tercile occurrence. Pass the hindcast index series as predictor and the forecast-year value as forecast.

index = deepscale.Index.named("wvg")
hindcast_index = index.reduce(sst_hindcast)
forecast_index = index.reduce(sst_forecast, climatology=sst_hindcast)

probs = deepscale.calibrate(
    hindcast_index, obs, method="logit", forecast=forecast_index,
)

For gridded SST predictors, LogitConfig reduces the fields through an Index before calibration:

probs = deepscale.calibrate(
    predictor_hindcast=sst_hindcast,
    predictor_forecast=sst_forecast,
    obs=obs,
    method=deepscale.LogitConfig(
        index=deepscale.Index.named("wvg"),
        detrend=True,
        significance=0.1,
    ),
)

Smoothed-coefficient regression (method="smoothed_regression")

Implements the postprocessing method of Kharin, Merryfield, Boer & Lee (2017, Mon. Wea. Rev. 145, 3545–3561). It rescales the ensemble-mean anomaly with a per-grid-cell regression coefficient, but where ereg fits each season independently, smoothed_regression smooths the coefficients across the seasonal cycle to suppress the sampling error that a ~30-year record leaves in each season's estimate. This recovers, and often improves, skill in weakly predictable regimes where naive per-season calibration degrades it.

It is season-aware: inputs carry a season dimension (up to 12 rolling seasons) that this method owns. temporal_sigma sets the smoothing: None per-season, a float for cyclic Gaussian smoothing across the calendar, or "constant" for a single year-round coefficient. The fit always comes from the hindcast (cross-validation is the caller's concern, as with ereg); the target is either a hindcast year (forecast_year=, retro-forecast/evaluation) or a separate out-of-sample forecast ensemble (forecast=, real-time). The two selectors are mutually exclusive.

Two output modes via output_type:

# Deterministic: rescaled ensemble-mean anomaly (scored with the `msss` metric).
adjusted = deepscale.calibrate(
    hindcast, obs,                       # (season, year, member, lat, lon) / (season, year, lat, lon)
    method="smoothed_regression",
    output_type="deterministic",
    temporal_sigma="constant",
    forecast_year=2024,
)

# Probabilistic: below/normal/above tercile probabilities (scored with `crpss`, `reliability`).
probs = deepscale.calibrate(
    hindcast, obs,
    method="smoothed_regression",
    output_type="tercile",
    distribution="gamma",                # "normal" for temperature, "gamma" for precipitation
    forecast_year=2024,
)

# Real-time: apply the hindcast fit to an out-of-sample forecast ensemble
# (e.g. OND 2026 against a 1993-2020 hindcast). The forecast members go through
# the hindcast-fitted coefficients, gamma parameters, and tercile boundaries.
probs_2026 = deepscale.calibrate(
    hindcast, obs,
    method="smoothed_regression",
    output_type="tercile",
    distribution="gamma",
    forecast=forecast_members,           # (season, member, lat, lon), not in obs years
)

The probabilistic mode additionally calibrates the forecast spread and, for precipitation, works through a gamma distribution so probabilities never fall on negative rainfall. deepscale.seasonal_coefficients(hindcast, obs, temporal_sigma=...) exposes the fitted, smoothed coefficient field for inspection or plotting.

Multi-model input uses the same {model: (hindcast, forecast)} shape as ereg, but where ereg calibrates each model separately and averages the tercile maps, smoothed_regression pools the members across models into one super-ensemble (with reindexed member ids) and calibrates that, matching the Kharin et al. experiment design. Hindcast years are intersected across models and with the obs before fitting.

Runnable examples: examples/demo_ensemble_regression.py (eReg) and examples/demo_logistic_wvg.py (logit).

Relationship to Rosetta

Rosetta handles data acquisition and normalization; DeepScale handles forecasting and verification. The interface between them is standardized xarray, so DeepScale stays source-agnostic and works with any data prepared the same way.

Development setup

git clone https://github.com/accord-research/deepscale.git
cd deepscale
uv sync

Some examples also use Rosetta for data acquisition:

cd ..
git clone https://github.com/accord-research/rosetta.git
cd rosetta
uv sync

The roadmap (PyCPT parity, additional methods, and machine-learning tiers) is tracked on GitHub Issues under the v1-roadmap label.