Infilling of Missing Rainfall Radar Data with a Memory-Assisted Deep Learning Approach

Abstract. Incomplete spatio-temporal meteorological observations can result in misinterpretations of the current climate state, uncertainties in early warning systems, or inaccuracies in nowcasting models and can thereby pose signficant challanges in hydrology research or similar applications. Traditional statistical methods for infilling missing precipitation data demand substantial computational resources and fail over large areas with sparse data – like temporary outages of weather radars. Although recent machine learning advancements have shown promise in addressing missing meteorological or satellite observations, they typically focus on spatial aspects, overlooking the complex spatio-temporal variability characteristic of precipitation, especially during extreme events. We propose a deep convolutional neural network enhanced with a temporal memory component to better account for temporal changes in precipitation fields. This approach can analyse arbitrary sequences from before and/or after the incomplete observation of interest. Our model is trained and evaluated on the hourly RADKLIM dataset, which features 1-km resolution precipitation derived from combined radar and weather station data across Germany. By infilling both synthetic and actual data gaps of RADKLIM, the study demonstrates the model's effectiveness, providing detailed insights into its capabilities during significant rainfall events, such as those in May 2012 and July 2021, including those responsible for the Ahrtal flood. This novel approach represents a step forward in hydrological applications, potentially improving the way we predict and manage water-related events by increasing the accuracy and reliability of precipitation data analysis.