[Paper Review] Machine Learning for Precipitation Nowcasting from Radar Images

Shreya Agrawal, Luke Barrington+4 more|arXiv (Cornell University)|Dec 11, 2019

Meteorological Phenomena and SimulationsEarth and Planetary Sciences14 references239 citations

TL;DR

The paper treats precipitation nowcasting as image-to-image translation using a U-Net to predict 1-hour ahead, 1 km resolution radar-based precipitation from MRMS radar sequences, outperforming persistence, optical flow, and HRRR for short-term forecasts. HRRR overtakes after about 5 hours.

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.

Motivation & Objective

• Motivate high-resolution nowcasting to support climate adaptation and disaster response.
• Develop a data-driven, low-latency nowcasting model using radar imagery.
• Compare CNN-based nowcasting to traditional methods (persistence, optical flow, HRRR).
• Operate on 1 km x 1 km tiles across the continental US and address data sparsity by oversampling rainy tiles.

Proposed method

• Treat nowcasting as image-to-image translation using a U-Net CNN.
• Input a sequence of seven MRMS images collected ten minutes apart over one hour; channels include time of day and pixel latitude/longitude.
• Output the MRMS image one hour after the last input image, quantized into precipitation thresholds.
• Label images are quantized into four ranges: [0,0.1),[0.1,1.0),[1.0,2.5),[2.5,∞).
• Use cross-entropy loss per pixel and ADADELTA optimization; incorporate long skip connections and short skip connections within blocks.

Experimental results

Research questions

• RQ1 Can a U-Net based CNN accurately forecast 1-hour ahead precipitation at 1 km resolution from a short radar history?
• RQ2 How does the CNN-based nowcasting model compare to MRMS persistence, optical flow, and HRRR one-hour forecasts?
• RQ3 What is the impact of tile-based training and oversampling rainy regions on predictive performance?
• RQ4 Do input modalities like time-of-day and location features improve nowcasting accuracy?

Key findings

• The CNN-based model outperforms persistence, optical flow, and the HRRR 1-hour forecast for short-term nowcasting.
• For 1-hour predictions, HRRR does not yet outperform the CNN approach; HRRR performance improves relative to the CNN only at longer forecast windows (around 5 hours).
• The approach uses 1 km x 1 km tiles across the continental US with oversampling so 80% of tiles have rain, addressing class imbalance.
• Predictions are framed as thresholds-based probabilities per pixel, enabling explicit probabilistic statements for rain rates.
• The model is trained on MRMS data from 2018 and tested on 2017 and 2019 half-years, indicating cross-year generalization within the study period.