Operationally Ready High-resolution Global Precipitation Nowcasting

Precipitation nowcasting is recognized as crucial information in numerous applications, ranging from individuals deciding whether to carry an umbrella during a stroll to meteorologists issuing timely warnings for rapidly developing heavy rainfall events. Nevertheless, conventional numerical weather prediction techniques frequently require several hours to generate a forecast, often resulting in the nowcast time range having passed by the time the forecast is disseminated. Predicting over tropical regions poses unique challenges due to elevated precipitation rates and the unpredictable and rapidly evolving nature of storms. In this work, we present a satellite-based Global Nowcasting model capable of predicting precipitation over the subsequent twelve hours at a spatial resolution of 5 kmX5 km and a temporal resolution of 15 minutes across the globe. We have expanded our machine learning-based nowcasting model, MetNet, on a global scale to encompass regions where radar data may be scarce or unavailable. Even in the absence of ground radar as an input, MetNet Global Nowcasting outperforms ECMWF’s high resolution forecast globally, particularly in data-sparse regions such as India, Brazil, and Africa. The HRES forecast is at a 9kmX9km spatial and hourly temporal resolution, making our model a considerable improvement over the benchmark in granularity.

The performance of our global precipitation nowcasting model is assessed in diverse regions and climatic zones worldwide. The Global Precipitation Measurement mission's dual-precipitation radar satellite serves as the ground truth in radar-sparse regions, complemented by ground radar data where accessible. Our evaluation metrics include the critical success index, frequency bias, and fractions skill score. We observe significant improvements across all metrics, encompassing low to high precipitation rates and lead times extending up to 12 hours. Notably, this performance enhancement is achieved under operational conditions, where model inputs incorporate delays associated with real-time data acquisition and inference. The already trained model can generate forecasts in under a minute using Google TPUs, facilitating easy deployment and utilization in various applications. MetNet Global Nowcasting represents a significant advancement in reducing disparities in global weather forecast quality. It offers the potential to assist forecasters in issuing early warnings in data-sparse regions, thereby enhancing preparedness for rainfall events.