Post-processing
Post-processing involves several steps and methods, such as bias adjustment and downscaling, that are used to transform the raw data outputs of climate prediction systems into more actionable and reliable climate information that users can apply to decision making.
What is post-processing?
Climate prediction systems rely on Earth system models, which simulate the complex processes taking place in the Earth system and provide predictions on the expected climate conditions in the future weeks, months or years, such as temperature and precipitation. However, the raw data that come out of climate prediction systems have biases and other limitations, such as low spatial resolution, since models might not be able to reproduce some atmospheric/ocean processes.
Post-processing involves a number of steps applied to the raw data outputs, which can help correct or improve systematic errors and provide more useful (e.g. higher spatial resolution) and usable (e.g. information about the quality of the predictions) climate information to users.
How do we do this?
Our team performs post-processing of data by applying several methods, selected according to their relevance and expected outputs. Some of these methods are briefly summarised below:
- Bias adjustment: The predictions of past conditions are compared against the past observations to identify systematic errors or inconsistencies. Then, corrections can be made to remove any biases and bring the predictions closer to the observations. Different methodologies can be used for adjusting these biases, depending on the aspect we are looking to improve (e.g. mean values, other statistics of the distribution, ensemble calibration).
- Forecast quality assessment: In order to assess the quality of the predictions (i.e. how much skill there is in the predictions), we conduct systematic comparison between the past observations and past predictions (hindcast). We can thus provide the users with skill information that can help determine if it is best to use these predictions instead of past observation to forecast future conditions.
- Downscaling: Climate information is typically provided at coarse resolutions that are not enough for many applications. For example, users might be interested in receiving information at a parcel or regional level that the climate forecast system cannot provide. To overcome this, we apply different downscaling techniques to obtain information at a finer spatial scale. As downscaling often requires the use of higher resolution observations, it can also be understood as a bias adjustment, since the predictions are improved according to these observations.
- Indicators and impact modelling: The post-processing steps can also be applied to climate information beyond the essential climate variables (such as temperature, precipitation, wind speed), which can be useful for decision making. Some of this information provided by our team include specific climate indicators, risk indices, weather regimes and teleconnections. Our aim is to apply a seamless approach to improve predictions and to combine data at different timescales, ensuring that they are tailored to the user needs.