| 21 Sep 2026
A new fast multivariate bias correction technique, EMBCCA-UNSEEN (v1.1.0): a case study of compound events in Hunan Province, China, using the UNSEEN approach
Abstract. The frequency of extreme climate events such as floods, droughts and heatwaves has been increasing due to anthropogenic climate change. However, the limited availability of historical data poses a significant challenge in accurately assessing these events and their likelihood. This study addresses this issue by employing the UNSEEN approach (UNprecedented Simulated Extremes using ENsembles) in combination with the UK Met Office's DePreSys4 initialized hindcast model to simulate a large ensemble of climate data. To enhance the reliability of the model data, we develop a multivariate bias and correlation correction method, Eigenvalue-based Multivariate Bias Correction with Correlation Alignment (EMBCCA-UNSEEN), which combines mean-shift correction and linear transformations to correct inter-variable correlations, ensuring consistency across multiple dimensions of the model data with the observational data. EMBCCA-UNSEEN v1.1.0 is distributed as an installable Python package (Yang et al., 2026). An evaluation of the corrected model data is conducted using two multivariate fidelity testing methods, including statistical feature assessment and a support vector machine-based testing approach. The test results demonstrate that applying the EMBCCA-UNSEEN correction method can significantly improve the consistency between the model data and the observational data, and is significantly more computationally efficient than alternative approaches.
Using a case study for Hunan Province, China, we demonstrate how the methodology improves the representation of extreme events and quantify the annual risk of extreme compound hot and dry events. The results indicate that the new bias correction method effectively captures inter-variable correlations while preserving variance, thereby enhancing the accuracy of simulations for compound extreme climate events.
This manuscript presents a method to apply multivariate bias correction to climate model data. The method is applied to a decadal hindcast model, and the bias corrected data is then used to assess risk of compound events within the Hunan Province of China.Â
The paper is well written and clear to follow. The bias correction does enable the model data to be transformed to closer agreement with the observational data available. However, I have concerns about the application of the bias correction to assess extreme events as applied in this study. Â
The UNSEEN method was developed to allow assessment of the risk of extreme events, by using large ensembles such as the decadal prediction system, DePreSys, to provide a much larger sample of 'observation-like' data. This enables the extreme events to be better sampled as multiple events at the 1-in-100 return level are available – unlike relatively short observational-based datasets such as ERA5. However, in situations where the model data shows bias in the extremes the UNSEEN method is harder to apply. If adjustments are made to shift the distribution of the model towards the observational distribution, then we lose the major benefit of the method. We are not sampling further into the plausible extremes as we are forcing the data to 'look like' the observations. Once more complex bias corrections are applied (e.g. this multivariate method, or adjusting both mean and stdev) then we may as well just assess the observational dataset alone, we lose any benefit of using the model data. Â
In section 3.4 it is commented that the raw DePreSys output overestimates the joint probability (8.4%) compared to the bias corrected data (2.5%). But what if the raw model is not 'incorrect', as the model does include our best understanding of the climate system? Â
A key element of UNSEEN is that it allows assessment of plausible, unprecedented events. As the multiple variables are being adjusted differently, the processes and feedbacks that are incorporated within the model system are no longer valid. Once statistical bias-corrections are applied the extreme events in the model may no longer be physically plausible.Â
There is likely a useful applications of the bias correction method presented, but I do not think UNSEEN is it. Â