Exploring the predictability of using crowdsourced observations in statistical postprocessing of NWP based precipitation nowcasts by machine learning

Crowdsourced observations are incorporated as one of the predictors for statistically postprocessing precipitation nowcasts by a numerical weather prediction (NWP) model using machine learning (ML) in Norway. Specifically, we use neural networks and random forests to predict precipitation with a one-hour lead time. Our study indicates that ML based statistical postprocessing combined with crowdsourced observations can improve the NWP precipitation nowcasts in terms of both reducing errors and decreasing uncertainties of the nowcasting. The most important predictor in our study is the accumulated precipitation one hour before the forecasting time by crowdsourced observations from a dense network of Netatmo personal weather stations (PWS). The quality of the crowdsourced observations of precipitation can influence the predictive skills of the postprocessing. In addition, the skewness of precipitation data due to frequent zeros is a limiting factor of the predictive skills. Overall, our results support that one single ML model trained on sample data from a large geographic region can be generalized to other areas in the region, provided that the areas are covered by crowdsourced observations of high quality.