Improving seasonal forecasts for early warning systems in Germany

Climate forecasts at seasonal timescales are critical for various sectors, and play a key role in decision-making processes, helping to mitigate risks associated with climate variability and extreme events. However, model outputs are typically insufficient for many practical applications due to coarse resolution and systematic biases, requiring the employment of post-processing techniques to enhance their usability and target stakeholders’ interest such as early warning systems. Post-processing techniques such as downscaling and bias correction can translate model outputs into higher-resolution, bias-corrected forecasts that are more relevant and best appropriate for local applications. We present a physics-informed CNN-based framework for downscaling and bias correction of ECMWF SEAS5.1 seasonal temperature and precipitation forecasts over Europe from 1° to ~1.2km, which represents a downscaling factor of ~60. The approach considers several climate drivers of atmospheric surface variables from SEAS5.1 as input and takes European Meteorological Observations at 1.2 km as ground truth data. We use an analog-based approach to account for the mismatch between long-range model outputs and observations due to model drifting, which is a problem for supervised neural networks algorithms running on climate datasets. Finally, we present a detailed evaluation of the performance for the period 2017-2022, by comparing our results to the raw output. In most cases, the post-processed forecasts outperform the raw predictions in terms of bias reduction, spatial representation and capturing the extremes. This work has potential implications for reducing uncertainties, improving spatial representation, and addressing systematic biases present in raw ECMWF seasonal products.