(C) Crown Copyright, Met Office. All rights reserved. See LICENCE.txt in the root of the repository for full licensing details.

EMBCCA‑UNSEEN: multivariate bias correction for UNSEEN compound extremes

This repository contains the analysis code used in the manuscript:

A new fast multivariate bias correction technique: a case study for compound events in Hunan Province, China, using the UNSEEN approach

The code applies the EMBCCA‑UNSEEN bias correction method to DePreSys4 initialised hindcasts, using ERA5‑Land as the observational reference, and evaluates fidelity using:

• Multivariate statistical feature consistency (SFC) testing
• Support Vector Machine (SVM)-based separability testing

---

Repository structure

• Multi-DePreSys4-Paper-area_avg_final_multiscatter.py
  → Hunan Province (area-mean) analysis
  → Reproduces main manuscript figures
• Multi-DePreSys4-Paper-area_full_final_China.py
  → China-wide spatial analysis
  → Produces correlation maps and comparison of time taken for the different multivariate methods when applied China-wide
• fidelity_test_cube.py
  → Helper module for UNSEEN-style fidelity testing

---

Installation (conda)

conda env create -f environment.yml
conda activate embcca-unseen

Tested with:

• Python 3.12.10
• SBCK 1.4.2 (https://github.com/yrobink/SBCK-python)

---

How to run

This code is designed to be run by editing constants at the top of each script.

Each script contains:

# =============================================================================
# USER SETTINGS (edit these to run the workflow)
# =============================================================================

You must edit:

• DATA_DIR
• OUTDIR

---

1. Hunan case study (main results)

python Multi-DePreSys4-Paper-area_avg_final_multiscatter.py

Produces:

• Line plots (not shown in manuscript)
• Scatter plots comparing joint distributions across methods
• Statistical fidelity plots
• SVM ROC curves
• Extreme event probabilities

Outputs saved in:

OUTDIR/

subfolders:

• Line_plots
• Scatter_plots
• Statistical_Comparison
• SVM_Comparison
• Exceedance_Comparison
• Fidelity_Testing/

---

2. China spatial analysis

python Multi-DePreSys4-Paper-area_full_final_China.py

Produces:

• Spatial correlation maps
• Correlation anomaly maps
• Calculation of time taken for each multivariate bias adjustment method to be applied across China

Outputs saved in:

OUTDIR/China/

---

Figure mapping (script outputs → manuscript figures)

This table helps reproduce key figures from the paper.

Main figures

Manuscript Figure	Description	Script output
Figure 3	Univariate fidelity (temperature mean shift)	Fidelity_Testing/Original model data_temperature.png and Fidelity_Testing/Univariate mean shift_temperature.png
Figure 4	Six-panel temperature–precip scatter	Scatter_plots/Scatter_sixpanel.png
Figure 5a	China correlation maps	China/maps/correlation'.png
Figure 5b	Correlation anomaly maps	China/maps/correlation_diff'.png
Figure 6	Correlation fidelity distributions	Statistical_Comparison/Correlation.png
Figure 7	SVM ROC curves	SVM_Comparison/SVM_ROC_sixpanel.png
Figure 8 (left)	Dry / hot probabilities	Exceedance_Comparison/exceedance_comparison_bar.png
Figure 8 (right)	Joint probability	Exceedance_Comparison/joint_exceedance_comparison_bar.png
Figure 9	Threshold sensitivity plots
→ precip decrement		Exceedance_Comparison/joint_exceedance_by_precipitation_decrement.png
→ temperature increment		Exceedance_Comparison/joint_exceedance_by_temperature_increment.png

Appendix B figures (SFC testing)

Appendix Figure	Description	Script output
Figures 10–11	Mean distributions	Statistical_Comparison/*_Mean.png where * = temperature or precipitation
Figures 12–13	Standard deviation	Statistical_Comparison/*_Standard_Deviation.png where * = temperature or precipitation
Figure 14	Skewness	Statistical_Comparison/Skewness.png
Figure 15	Kurtosis	Statistical_Comparison/Kurtosis.png

---

Data requirements

This repository does not include data. Data can be provided on request in .nc format.

You need:

Hunan (area-mean)

• DePreSys4 JJA temperature + precipitation
• ERA5-Land JJA temperature + precipitation

China (gridded)

• Regridded ERA5-Land
• DePreSys4 gridded output

Expected variable names

Variable	Name
Model temperature	mean_jja_temperature
Model precipitation	total_jja_precipitation
Obs temperature	t2m
Obs precipitation	tp

---

Reproducibility

• Seeds are used for bootstrapping during SFC testing, during SVM resampling, and when splitting the data into training and testing, so that results are reproducible.

Analysis period:

1992–2021 (30 years)

---

Plotting / HPC usage

The scripts use:

MPL_BACKEND = "Agg"

Switch to "Qt5Agg" only for interactive use.

---

Code and data availability

• Data: available on reasonable request
• Code: provided in this repository

---

Citation

Please cite:

• The associated manuscript

---

Notes

• SBCK methods included: dOTC, MRec, R2D2
• Unlike these methods, EMBCCA‑UNSEEN preserves variance (critical for UNSEEN applications)