colm-openbench 3.0.0b3

Open Source Land Surface Model Benchmarking System

OpenBench

Open Source Land Surface Model Benchmarking System

OpenBench is a fully automated, cross-platform framework for benchmarking land surface models (LSMs) against curated reference datasets. From a single YAML config it preprocesses model output and references onto a common grid/time window, computes a consistent metric and score suite, renders publication-ready comparison figures, and assembles an HTML/PDF report — for any number of variables, models, and references at once.

pip install colm-openbench
openbench smoke-test   # verify the installed package with bundled sample data
openbench init           # generate openbench.yaml interactively
openbench check openbench.yaml   # validate config + data availability
openbench run   openbench.yaml   # evaluate, plot, report

---

Table of Contents

• Why OpenBench
• Key Features
• Installation
• Quick Start
• How It Works
• Configuration
• CLI Reference
• Metrics & Scores
• Comparison & Visualization
• Classification Group-By (IGBP / PFT / Köppen)
• Reference Data Management
• Model Profiles
• Output Structure
• Performance Tuning
• Migrating from Earlier Versions
• Requirements
• Citation
• Links
• License

---

Why OpenBench

Evaluating an LSM normally means hand-wiring regridding, time alignment, masking, metric code, and plotting for every model/reference pair — and redoing it when a file changes. OpenBench turns that into one declarative config:

• One config, many comparisons. Variables × models × references are evaluated as a full cartesian product, each pair kept in its own output so nothing clobbers anything else.
• Consistent science. The same regridding, masking, weighting, and metric definitions apply to every pair, so cross-model comparisons are fair by construction (cumulative unified masking enforces identical spatial coverage across models).
• Grid and station references are both first-class.
• Incremental & resumable. Results are cached by a content fingerprint of the config, code, and input files; an interrupted run safely resumes without recomputing finished work or producing partial summaries.
• Reproducible. Resolved configs, content fingerprints, and resumable caches make runs auditable and repeatable.

Key Features

• Unified configuration — a single openbench.yaml replaces the 4–6 file setup of earlier versions, with _defaults merging and !include splitting.
• Reference data registry — 101 built-in reference datasets (grid and station) you can list, inspect, scan, and register.
• Model profiles — 22 built-in profiles (CoLM2024, CLM5, NoahMP5, JULES7, VIC5, ERA5-Land, GLDAS2, MATSIRO, CaMa-Flood, WRF, …) supplying variable name mappings and unit conversions automatically.
• Rich science suite — 25+ point-wise metrics, 8 normalized skill scores, and 17 statistical analysis modules.
• Comparison & visualization — Taylor, Target, Portrait, heat-map, parallel coordinates, KDE, ridgeline, and more — generated automatically.
• Classification group-by — per-class aggregation by IGBP land cover, PFT, or Köppen climate zone; the masks ship with the package, so it works out of the box.
• CLI-first — run, check, init, ref, sim, model, migrate, cache, smoke-test, gui.
• GUI wizard — interactive configuration and run management (optional [gui]).
• Remote execution — SSH-based job submission to HPC clusters (optional [remote]).
• Migration tool — convert existing JSON / Fortran NML configs to YAML.

Installation

Recommended clean environment:

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
python -m pip install --upgrade pip

Install from PyPI:

# Default install — fully featured (CLI, statistics, plotting, HTML reports,
# legacy migration). Best for HPC/servers.
pip install colm-openbench

# Optional: graphical configuration wizard (PySide6/Qt)
pip install "colm-openbench[gui]"

# Optional: remote SSH execution (paramiko)
pip install "colm-openbench[remote]"

# Optional: PDF reports (xhtml2pdf; builds pycairo, needs system cairo)
pip install "colm-openbench[report]"

# Optional: all of the above
pip install "colm-openbench[all]"

Using uv:

uv pip install "colm-openbench[all]"

conda / conda-forge

OpenBench ships a conda-forge recipe (conda/meta.yaml, noarch: sourced from PyPI, with Cartopy/NetCDF/PROJ resolved by conda). Once published to conda-forge, a single command installs everything (no local compilation):

conda install -c conda-forge colm-openbench
mamba install -c conda-forge colm-openbench   # or the faster mamba

Note: the current release is a Beta pre-release (3.0.0b1) and is not on conda-forge yet. Until it is, use one of the two paths below.

If your platform does not have wheels for geospatial dependencies such as Cartopy, install the scientific stack from conda-forge first, then install OpenBench with pip:

mamba create -n openbench -c conda-forge python=3.12 cartopy netcdf4
mamba activate openbench
pip install colm-openbench

Or build a local conda package straight from the bundled recipe:

conda install -n base conda-build
conda build conda/
conda install -c local colm-openbench

Install from a GitHub checkout for development or local testing:

git clone https://github.com/CoLM-SYSU/OpenBench.git
cd OpenBench
python -m pip install -e ".[dev]"

The IGBP / PFT / Köppen classification masks needed for group-by analysis are bundled in the package (compressed, ~1.1 MB total) — no extra download.

Verify an installation before preparing your own data:

openbench smoke-test

This unpacks a bundled Initial_test fixture and runs openbench check against the original grid and station cases: three variables, two simulation cases, and their paired reference sources. Use openbench smoke-test --run for a full evaluation smoke run; the full run may let Cartopy download Natural Earth map assets on a clean machine.

Quick Start

# 1. Generate a config interactively (picks variables, references, models)
openbench init

# 2. Validate the config and check reference data availability
openbench check openbench.yaml

# 3. Run the evaluation (metrics + scores + comparison figures + report)
openbench run openbench.yaml

# Re-run after editing the config — only what changed is recomputed
openbench run openbench.yaml

# Force a full re-run, ignoring the incremental cache
openbench run openbench.yaml --force

Results land in output/<project_name>/; start with reports/report.html.

How It Works

config.yaml
   │
   ▼
┌─────────────┐   ┌──────────────┐   ┌───────────────┐   ┌────────────┐
│ Preprocess  │ → │  Evaluate    │ → │  Compare      │ → │  Report    │
│ regrid /    │   │  metrics &   │   │  cross-model  │   │  HTML/PDF  │
│ time-align /│   │  scores per  │   │  figures &    │   │            │
│ unified mask│   │  (var,sim,ref)│   │  statistics   │   │            │
└─────────────┘   └──────────────┘   └───────────────┘   └────────────┘
        every stage is incrementally cached and resumable

1. Preprocess — each model and reference is regridded to a common resolution, clipped to the shared time window, and (optionally) masked so all models share identical valid coverage.
2. Evaluate — for every (variable, simulation, reference) triple, OpenBench computes the configured metrics and scores (grid → per-cell NetCDF, station → per-station CSV).
3. Compare — when comparison.enabled, results across all pairs feed shared figures (Taylor, portrait, heat-map, …) and optional statistics.
4. Report — everything is assembled into an HTML report (PDF with the [report] extra).

Configuration

A minimal config:

# openbench.yaml
project:
  name: my_run
  output_dir: ./output
  years: [2010, 2014]
  lat_range: [-90, 90]
  lon_range: [-180, 180]

evaluation:
  variables:
    - Evapotranspiration
    - Latent_Heat

reference:
  Evapotranspiration: ET_Xu_etal_2025_LowRes
  Latent_Heat: FLUXCOM

simulation:
  CoLM2024_test:
    model: CoLM2024            # matches a built-in profile → auto var mapping
    root_dir: /data/simulations/colm2024

A fuller config showing per-simulation overrides, metric/score subsets, and comparison/statistics toggles:

project:
  name: multi-model
  output_dir: ./output
  years: [2000, 2020]
  grid_res: 0.5
  num_cores: 16
  time_alignment: per_pair      # intersection | per_pair | strict
  unified_mask: true            # identical valid coverage across models
  weight: area                  # area (cos-lat) | mass | none

evaluation:
  variables: [Evapotranspiration, Gross_Primary_Productivity, Latent_Heat]

reference:
  Evapotranspiration: ET_Xu_etal_2025_LowRes
  Gross_Primary_Productivity: FLUXCOM-X-BASE_LowRes
  Latent_Heat: FLUXCOM

simulation:
  CoLM2024:
    model: CoLM2024
    root_dir: /data/CoLM2024
    variables:
      Gross_Primary_Productivity:
        varname: GPP            # override the profile's file/variable name
  CLM5:
    model: CLM5
    root_dir: /data/CLM5
    variables:
      Gross_Primary_Productivity:
        varname: FPSN

metrics: [bias, RMSE, correlation, KGE]
scores:  [nBiasScore, nRMSEScore, Overall_Score]

comparison:
  enabled: true

statistics:
  enabled: true
  items: [Z_Score, ANOVA]

Key project options

Option	Values / default	Meaning
years	[start, end]	Evaluation time window
lat_range / lon_range	[min, max]	Spatial domain
grid_res	float (degrees)	Common grid the data is regridded to
num_cores	int / null	Parallel workers (null/0 → all CPUs)
time_alignment	intersection (default) / per_pair / strict	How model & reference time windows are matched
regrid_backend	openbench_conservative (default) / cdo_remapcon / xesmf_conservative / basic_interpolation	Regridding engine
unified_mask	true (default)	Cumulative NaN mask across sims for cross-model fairness
weight	area (default when omitted) / mass / none	Spatial aggregation weighting
IGBP_groupby / PFT_groupby / climate_zone_groupby	false	Per-class aggregation (see below)
generate_report	true	Emit HTML/PDF report

Advanced config helpers

• _defaults block — merge shared settings into every simulation entry.
• !include path/to/partial.yaml — split large configs across files.
• Environment variables override YAML at runtime (handy for quick experiments) — see Performance Tuning.

CLI Reference

Run openbench <command> --help for full options on any command.

Top-level commands

Command	Description
openbench run CONFIG	Run evaluation, comparison, and report from a YAML config
openbench check CONFIG	Validate config and check reference/data availability
openbench smoke-test	Validate the installed package with bundled Initial_test sample data
openbench init	Interactively generate an openbench.yaml
openbench ref ...	Manage the reference dataset registry (see below)
openbench model ...	Manage model profiles (see below)
openbench sim scan DIR	Scan directories for simulation cases → config fragments
openbench cache ...	Inspect/clear runtime caches (see below)
openbench migrate OLD	Convert a JSON or Fortran NML config to YAML
openbench gui	Launch the GUI wizard (requires [gui] extra)
openbench version / --version	Print version

Useful run flags: --force (ignore cache), --comparison-only (re-run only comparisons on existing outputs), --dry-run, --cores N, --variable VAR (repeatable), --output-dir DIR, --dump-config.

openbench ref — reference registry

Subcommand	Description
ref list	List all available reference datasets
ref show NAME	Show details of a dataset
ref scan DIR	Scan a directory and auto-register datasets you already have
ref register NAME	Register or update a reference dataset
ref register-profile	Register a reference profile (variable mappings)
ref path NAME	Print the local path for a dataset
ref optimize	Convert a dataset to Zarr for faster reads
ref generate-station-list	Auto-generate a station-list CSV from NetCDF files
ref delete NAME	Delete a user reference entry/overlay
ref convert-old	Convert an old-format reference YAML to v3

openbench model — model profiles

Subcommand	Description
model list	List all available model profiles
model show NAME	Show variable mappings for a profile
model status	Show model registry status
model register / rename / delete	Manage user profiles
model export / import	Move a profile in/out as standalone YAML
model alias	List or create user model aliases
model remove-var	Remove a variable from a profile
model validate	Check a profile for completeness issues
model path NAME	Show whether a profile is bundled or user-defined

openbench cache — runtime caches

Subcommand	Description
cache status	Show cache status (e.g. --regrid for the regrid-weight cache)
cache clear	Clear selected cache files

Metrics & Scores

OpenBench computes 25+ point-wise metrics per (variable, simulation, reference) pair. By default all applicable metrics are produced; restrict with the top-level metrics: list.

Metric	Range	Meaning
bias, percent_bias	(−∞, ∞)	Mean (relative) difference; 0 = unbiased
RMSE	[0, ∞)	Root-mean-square error
ubRMSE / CRMSD	[0, ∞)	Bias-removed (centered) RMSE
mean_absolute_error	[0, ∞)	Mean absolute difference
correlation	[−1, 1]	Pearson correlation
correlation_R2	[0, 1]	Explained-variance fraction
NSE	(−∞, 1]	Nash–Sutcliffe efficiency
KGE, KGESS	(−∞, 1]	Kling–Gupta efficiency / skill score
index_agreement	[0, 1]	Willmott index of agreement
rv	[0, ∞)	Variance ratio σ_sim / σ_obs
kappa_coeff	[−1, 1]	Categorical agreement

8 normalized skill scores rescale metrics onto a comparable 0–1 axis (1 = best): nBiasScore, nRMSEScore, nPhaseScore, nIavScore, nSpatialScore, nSeasonalityScore, index_agreement, and the combined Overall_Score that drives portrait plots and rankings.

Beyond metrics/scores, the statistics block exposes 17 analysis modules (correlation, ANOVA, Mann–Kendall trend, three-cornered hat, Hellinger distance, partial least squares, false discovery rate, …).

Comparison & Visualization

When comparison.enabled: true, OpenBench renders cross-model figures automatically (PNG, with the underlying data saved alongside):

• Taylor diagram & Target diagram — correlation/variance/CRMSD at a glance
• Portrait plot (seasonal) — model × variable score grid
• Heat map & land-cover-based heat map
• Parallel coordinates, kernel density estimate, ridgeline, whisker plots
• Difference plots, relative score, radar map
• Single Model Performance Index (SMPI), Mann–Kendall trend, three-cornered hat, functional response, PLSR

Classification Group-By (IGBP / PFT / Köppen)

OpenBench can aggregate scores per land-cover or climate class. Enable any of:

project:
  IGBP_groupby: true          # MODIS/IGBP land cover (17 classes)
  PFT_groupby: true           # Plant Functional Type (16 classes)
  climate_zone_groupby: true  # Köppen climate zone (30 classes)

The classification masks (IGBP.nc, PFT.nc, Climate_zone.nc) are bundled with the package (compressed integer masks, ~1.1 MB total), so group-by works out of the box with no extra download. To use your own / higher-resolution masks, override the bundled versions via either:

export OPENBENCH_DATASET_DIR=/path/to/your/dataset   # highest priority
# or place files in ./dataset/<name>.nc relative to where you run openbench

Per-class results appear under output/<run>/comparisons/{IGBP,PFT,Climate_zone}_groupby/.

Reference Data Management

# Browse all 101 reference datasets
openbench ref list

# Inspect a single dataset (a base name resolves to its resolution family)
openbench ref show GLEAM_v4.2a

# Scan a directory and auto-register reference datasets you already have
openbench ref scan /Volumes/work/Reference

Example openbench ref list output:

Name                           Category     Type   Res      Years          Variables
────────────────────────────────────────────────────────────────────────────────────
AH4GUC_LowRes                  Urban        grid   0.5°     2010-2010      1
CERES_EBAF_Ed4.2_LowRes        Energy       grid   0.5°     ...            7
CLARA_3_LowRes                 Energy       grid   0.5°     1979-2024      8
CN05.1_MidRes                  Meteorology  grid   0.25°    1961-2021      2
Caravan_Daily                  Water        stn    stn      1950-2023      1
...                            ...          ...    ...      ...            ...
Total: 101 datasets

Many datasets come in multiple resolutions (_LowRes, _MidRes); a base name like GLEAM_v4.2a resolves to its family, or you can pin a specific one such as ET_Xu_etal_2025_LowRes.

Model Profiles

openbench model list            # list built-in profiles
openbench model show CoLM2024   # variable mappings for a profile

Example openbench model list output:

Name                 Type   Res      Variables Complete
────────────────────────────────────────────────────────────
CoLM2024             grid   0.5°     52        ✓
CLM5                 grid   1.25°    69        ⚠ 16
ERA5-Land            grid   0.1°     12        ✓
GLDAS                grid   0.25°    11        ✓
NoahMP5              grid   ...      70        ⚠ 34
...
Total: 22 model profiles

When simulation.<name>.model matches a known profile, variable file-name patterns and unit conversions are applied automatically. Fields in the simulation entry override the profile.

Output Structure

output/<project_name>/
├── config.yaml          # the fully-resolved config actually used
├── run.log              # complete run log
├── data/                # preprocessed (regridded, clipped) sim/ref NetCDF
├── metrics/             # per-(var,sim,ref) metrics — grid: NetCDF, station: CSV
├── scores/              # per-(var,sim,ref) normalized scores
├── figures/             # per-pair visualizations (PNG)
├── comparisons/         # cross-model figures + group-by results (if enabled)
├── statistics/          # statistical analyses (if enabled)
├── reports/
│   ├── report.html
│   └── report.pdf       # only with the [report] extra
└── .openbench_cache.json # incremental cache index

Grid evaluations write one NetCDF per metric (<var>_ref_<R>_sim_<S>_<metric>.nc, dims (lat, lon) or (time, lat, lon)); station evaluations write an aggregated CSV (<var>_stn_<R>_<S>_evaluations.csv, one row per station). Read either directly with xarray / pandas for downstream analysis.

Performance Tuning

For high-resolution grids or many simulation files, tune IO and Dask in openbench.yaml:

project:
  num_cores: 16

  io:
    # true enables zlib for numeric NetCDF outputs; level 1 is the fast default.
    netcdf_compression: true
    netcdf_compression_level: 1

    # auto/null: use the resource planner; 0: disable; N: force N files per batch.
    mfdataset_batch_size: auto
    mfdataset_auto_batch_min_files: 200
    mfdataset_auto_batch_min_size_mb: 1024
    mfdataset_auto_batch_max_size: 100
    mfdataset_auto_batch_memory_fraction: 0.25

  dask:
    enabled: true
    n_workers: 4
    threads_per_worker: 1
    processes: true
    memory_limit: auto

Environment variables override YAML and take precedence, which is useful for temporary benchmarking:

OPENBENCH_NETCDF_COMPRESSION=1 OPENBENCH_NETCDF_COMP_LEVEL=1 \
OPENBENCH_MFDATASET_BATCH_SIZE=50 \
OPENBENCH_DASK=1 OPENBENCH_DASK_N_WORKERS=4 \
openbench run openbench.yaml --force

For repeatable measurements, run the same config from a clean output directory and keep the resolved config, run log, and environment metadata with the output:

openbench run openbench.yaml --force --dump-config

Migrating from Earlier Versions

# Convert an existing JSON config
openbench migrate old_config.json -o openbench.yaml

# Convert a Fortran NML config (requires the [migration] extra)
openbench migrate main.nml -o openbench.yaml

The v3.0 migration path is designed to preserve the intended evaluation setup while making defaults and generated YAML explicit. For publication comparisons, archive the resolved config.yaml and rerun a small overlap case before mixing old and new result sets.

Requirements

• Python 3.10 or newer
• Core dependencies: xarray, numpy, scipy, pandas, netCDF4, matplotlib, cartopy, dask, joblib, flox, PyYAML, Jinja2, click, tqdm, packaging
• Optional extras: gui (PySide6), remote (paramiko), report (xhtml2pdf), migration (f90nml)
• See pyproject.toml for the full pinned dependency list

Citation

If you use OpenBench in scientific work, please cite:

Wei, Z., Xu, Q., Bai, F., Xu, X., Wei, Z., Dong, W., Liang, H., Wei, N., Lu, X., Li, L., Zhang, S., Yuan, H., Liu, L., and Dai, Y.: OpenBench: a land model evaluation system, Geosci. Model Dev., 18, 6517–6540, https://doi.org/10.5194/gmd-18-6517-2025, 2025.

@article{wei2025openbench,
  author  = {Wei, Z. and Xu, Q. and Bai, F. and Xu, X. and Wei, Z. and Dong, W.
             and Liang, H. and Wei, N. and Lu, X. and Li, L. and Zhang, S.
             and Yuan, H. and Liu, L. and Dai, Y.},
  title   = {{OpenBench}: a land model evaluation system},
  journal = {Geoscientific Model Development},
  volume  = {18},
  pages   = {6517--6540},
  year    = {2025},
  doi     = {10.5194/gmd-18-6517-2025}
}

Links

• Source: https://github.com/CoLM-SYSU/OpenBench
• PyPI: https://pypi.org/project/colm-openbench/
• User's Guide (PDF): English · 中文
• Issues: https://github.com/CoLM-SYSU/OpenBench/issues

License

MIT License. See LICENSE for details.