Machine Learning for Precipitation Nowcasting from Radar Images

TL;DR

This work treats high-resolution precipitation nowcasting as an image-to-image translation task, predicting the 1-hour-ahead radar image from a sequence of MRMS radar frames using a U-Net. It demonstrates superior performance to persistence, optical flow, and HRRR baselines for short lead times, highlighting fast inference suitable for real-time nowcasting. The data setup uses 1 km MRMS grids tiled into 256 km regions with rainy-tile oversampling and four rainfall thresholds, and the model is trained on 2018 data with 2017/2019 held-out tests. The findings suggest strong potential for data-driven approaches in operational nowcasting, with future work exploring additional modalities and hybrid ML-physics integrations.

Abstract

High-resolution nowcasting is an essential tool needed for effective adaptation to climate change, particularly for extreme weather. As Deep Learning (DL) techniques have shown dramatic promise in many domains, including the geosciences, we present an application of DL to the problem of precipitation nowcasting, i.e., high-resolution (1 km x 1 km) short-term (1 hour) predictions of precipitation. We treat forecasting as an image-to-image translation problem and leverage the power of the ubiquitous UNET convolutional neural network. We find this performs favorably when compared to three commonly used models: optical flow, persistence and NOAA's numerical one-hour HRRR nowcasting prediction.

Figures (2)

• Figure 1: Sample MRMS Image and Predicted Precipitation
• Figure 2: Precision-Recall Curves For Rain Prediction