Precipitation estimation from the Arctic Weather Satellite: An initial study using probabilistic deep learning

Abstract

Accurate precipitation information is important for understanding the hydrological cycle, improving weather forecasts and supporting hydrological applications. How ever, precipitation is difficult to observe with sufficient spatial and temporal coverage. Rain gauges provide only local measurements and ground-based radars are limited to regions where they are available. Satellite observations can complement these sys tems by providing precipitation information over larger and sparsely instrumented domains. Passive microwave observations are widely used for precipitation estimation, and the newly launched Arctic Weather Satellite (AWS) offers a novel extension of this capability. AWS is a compact satellite carrying a 19-channel microwave radiometer and the first mission to include sub-millimetre channels around 325 GHz for this purpose. This thesis primarily investigates whether AWS observations can be used to retrieve surface precipitation rates using supervised machine learning. A secondary aim is to determine if these novel sub-millimetre channels have a positive effect on the retrieval performance. To achieve this, a dataset is constructed by matching AWS antenna temperatures with Multi-Radar Multi-Sensor (MRMS) radar precipitation estimates over the contiguous United States. Quantile regression neural networks are trained to predict conditional quantiles of the surface precipitation rate, providing both deterministic estimates and probabilistic estimates of retrieval uncertainty. Results demonstrate that AWS antenna temperatures contain valuable information for surface precipitation retrieval. The trained model captures broad precipitation structures and achieves near-zero overall bias on the independent test set. Although the predictions do not fully capture the intensity of the precipitation the spatial mapping is consistent with the radar reference. An additional model trained with out the sub-millimetre channels evaluates the specific impact of these sensors. Com parisons indicate that the sub-millimetre channels provide a modest but consistent improvement across the overall evaluation metrics. These findings serve as an initial assessment of AWS-based precipitation retrieval. This investigation is particularly relevant due to the upcoming launch of EPS-Sterna, a constellation of AWS satellites which could significantly enhance the temporal sampling of global precipitation ob servations. Further efforts could explore using other reference precipitation products ii to expand the geographical domain, as well as incorporating longer data records as newer AWS observations become available.