Factsheet 15: Bias correction
Climate models provide critical information for us to understand, prepare for and adapt to the future climate. However, climate models are simplifications of the real climate and do not simulate the climate perfectly. In climate science, the mismatch between an observation and data from a climate model is called 'bias'. Biases in climate models are managed by a rigorous process of removing errors from the climate model outputs, commonly referred to as 'bias correction'.
The high-resolution climate change projections available on Queensland Future Climate have undergone an extensive evaluation (see Factsheet #13 - Model Evaluation, for more details). Model evaluation outcomes show that our models represent Australia's historical climate accurately, and improve upon Global Climate Models. The precision of high-resolution climate projections available on Queensland Future Climate is further enhanced by correcting biases for their usage in climate risk assessments (e.g., climate extremes, hydrological hazards, multi-sectoral impact assessments). The data on the Queensland Future Climate Dashboard have all been bias-corrected except for the variables in the Mean Climate Theme. This factsheet explains the bias correction process.
While we bias-correct the data for release on the Dashboard, so it can be used in hazard and risk assessments, we don't bias-correct the data in our scientific publications when our aim is to understand model behaviour.
What is bias correction?
Bias correction is a process of removing biases, or errors, in climate model outputs. This is achieved by 'calibrating' the raw climate model outputs against reliable observational climate data. This process can be as simple as correcting the mean value of a variable, but it can also mean correcting other measures such as variability of a variable and correlation with other variables. Bias correction can be applied to one variable at a time (univariate), or it can be applied to multiple variables at a time (multivariate). Multivariate bias correction methods aim to maintain physical relationships between the different variables and are generally recommended when looking at impacts that depend on multiple variables.
There are many bias correction methods available, and there is no one 'best' method. Different methods may perform better in different locations, seasons, for different climate models, and for different purposes. For this reason, it is important to evaluate different methods before applying them.
Which bias correction method did we use?
Bias correction requires robust, long-term observations for calibrating the modelled climate data. Where possible, we used the Australian Gridded Climate Dataset (AGCD; Jones et al., 2009) developed by the Bureau of Meteorology (BoM) as the source of historical observational climate data used in our bias correction methods. When data were not available from this product, other observational products were applied (further details below). The correction functions applied to the calibration period (1981-2020) are also applied to the future climate of each climate model on an individual grid cell basis. Separate correction functions are derived for every grid cell of the model.
We evaluated several bias correction methods, which were selected based on a review of the literature. Bias correction methods were evaluated for their impact on average bias, daily bias, extremes and the climate change signal. For Queensland Future Climate, we chose the multivariate method, Cannon's N-Method (MBCn; Cannon, 2018), applied on a monthly basis. This method corrects multiple properties of the model output and preserves trends in the climate change signal. We found applying the bias correction method on a monthly basis was better suited for correcting seasonal cycles than applying the method on all months grouped together. Prior to applying MBCn, we also corrected the frequency of wet and dry days, as we found this to be important across Queensland's arid areas.
To maximise the performance of the MBCn under limited computational resources, we tailored the corrections to the climate indices used by the Queensland Future Climate Science Program as follows:
- For extreme temperatures, we corrected maximum and minimum temperature.
- For humid heatwaves, we corrected average temperature and relative humidity.
- For droughts (SPI and SPEI), we corrected precipitation and FAO reference evaporation, which was obtained from the BoM's Australian Water Outlook. This corrected precipitation was also used to calculate extreme precipitation indices.
- For forest fire hazard based on the Forest Fire Danger Index (FFDI), we used the Multi-Source Weather (MSWX; Beck et al., 2022) dataset for the correction of wind and relative humidity, which were bias corrected separately from precipitation and maximum temperature. This was done as wind and relative humidity are taken at 3pm, while daily data was adopted for precipitation and maximum temperature.
An alternate approach was adopted for parts of the Torres Strait region where there was a lack of AGCD observational data. For regions where data was available, we first fit a function comparing the climate simulations to historic observations, which is calculated separately at every individual grid cell. This function is then applied to correct the historic and future climate simulations. For the Torres Strait region without data available, the bias correction 'function' from the nearest spatial grid cell with observational data available was applied to the spatial grid missing observational data. In this way, we were able to correct the climate model outputs for the whole of Queensland, even in regions where observations were lacking.
References
Beck, H. E., van Dijk, A. I. J. M., Larraondo, P. R., McVicar, T. R., Pan, M., Dutra, E., & Miralles, D. G. (2022). MSWX: Global 3-Hourly 0.1° Bias-Corrected Meteorological Data Including Near-Real-Time Updates and Forecast Ensembles. Bulletin of the American Meteorological Society, 103(3), E710-E732. https://doi.org/10.1175/BAMS-D-21-0145.1
Cannon, A. J. (2018). Multivariate quantile mapping bias correction: An N-dimensional probability density function transform for climate model simulations of multiple variables. Climate Dynamics, 50(1-2), 31-49. https://doi.org/10.1007/s00382-017-3580-6
Jones, D. A., Wang, W., & Fawcett, R. (2009). High-quality spatial climate data-sets for Australia. Australian Meteorlogical and Oceangraphic Journal, 58, 233-248.
