Source-linked AI summary
Machine Learning for Precipitation Nowcasting from Radar Images
Shreya Agrawal, Luke Barrington, Carla Bromberg, John Burge, Cenk Gazen, Jason Hickey
TL;DR
High-resolution precipitation nowcasting needs timely, fine-scale forecasts for extreme weather, but optical-flow and numerical methods have important limitations. This paper uses a U-Net image-to-image model to predict one-hour-ahead precipitation from MRMS radar images and reports better short-term performance than persistence, optical flow, and HRRR, while HRRR is stronger at longer horizons.
Problem
High-resolution precipitation nowcasting requires actionable, low-latency forecasts, while optical flow misses storm initiation and decay and numerical models typically take hours to infer.
Method
The paper treats one-hour precipitation nowcasting as image-to-image translation from sequences of MRMS radar images using a U-Net convolutional neural network.
Results
The U-Net performs better than persistence, optical flow, and HRRR for short-term nowcasting, while HRRR consistently outperforms it at approximately 5 hours.
Takeaways & Limitations
Straightforward U-Net data-driven forecasting can outperform traditional numerical methods such as HRRR when the prediction window is on the order of a few hours.
Takeaways & Limitations
The approach predicts independently on geographical tiles, creating border effects when rain near a tile boundary originates outside the tile.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper frames high-resolution precipitation nowcasting as a low-latency forecasting problem where machine learning may address limitations of optical-flow and numerical approaches. It predicts one-hour-ahead precipitation from Doppler radar using a U-Net image-to-image model at 1 km resolution.
- High-resolution nowcasting forecasts near-future precipitation at high spatial resolution while requiring large-scale, low-latency data processing.
- Optical-flow models track image motion but cannot represent storm initiation or decay, while numerical models simulate atmospheric physics but typically require hours for inference.
- Climate-change-driven changes in weather patterns and more frequent extreme events increase the need for actionable, high-resolution predictions for planning and crisis management.
- The study predicts whether precipitation exceeds trace, light, and moderate-rain thresholds one hour ahead from Doppler radar over the continental United States at 1 km resolution.
- Forecasting is formulated as image-to-image translation from a sequence of radar images to a future radar image using a U-Net convolutional neural network.
2 Data setup
The study uses MRMS precipitation maps updated every two minutes at 1 km resolution, transforms them into threshold-based labels and independent geographic tiles, and oversamples rainy tiles. The example data juxtaposes inputs, predictions, and one-hour ground truth.
- MRMS combines radar, surface observations, and numerical weather prediction to provide precipitation rates every two minutes on a 1 km × 1 km grid.
- The MRMS dataset removes non-meteorological artifacts and projects observations from overlapping radar coverage onto a rectangular grid.
- Precipitation labels use four ranges defined by thresholds at 0.1, 1.0, and 2.5 millimeters of rain per hour.
- The United States is partitioned into 256km × 256km tiles, with rainy tiles oversampled so 80% contain at least one rainy pixel.
- Figure 1 presents the input image, quantized one-hour prediction, and quantized one-hour ground truth from left to right.
3 Evaluation and Results
The model is evaluated using pixel-level precision and recall across three rain thresholds against persistence, optical flow, and NOAA’s HRRR forecast. It outperforms all three baselines for short horizons, while HRRR becomes stronger at approximately five hours.
- Evaluation measures binary classification precision and recall for each output pixel across three precipitation thresholds.
- The comparison includes MRMS persistence, optical flow, and NOAA’s one-hour HRRR numerical forecast.
- MRMS persistence predicts that each location will rain at its current rate during the next hour.
- The model performs better than persistence, optical flow, and the one-hour HRRR forecast.
- HRRR consistently outperforms the model once the prediction window increases to approximately 5 hours.
4 Future Work
The paper identifies additional input modalities and model refinements as future directions, while noting that independently predicted tiles can create border effects. Combining radar with satellite data could broaden geographic coverage.
- Future work includes incorporating ground or satellite measurements as additional input modalities.
- Determining how to combine additional data effectively in a deep-learning model remains an active research problem.
- Refining the neural network’s topology and hyperparameters, including possible GAN-based approaches, is another proposed direction.
- Independent geographical tiles can create border effects because the CNN lacks information about rain arriving from outside a tile.
- Satellite inputs could enable predictions virtually anywhere on the planet rather than only where radar data are available.
5 Conclusion
The paper frames precipitation nowcasting as a data-driven image-to-image translation problem and finds that straightforward U-Net models can outperform traditional numerical methods for short-term forecasts.
- 5 Conclusion: The approach maps a short history of MRMS rain images to the rain state one hour later, avoiding explicit modeling of atmospheric physics.The input is a sequence of MRMS images, and the output is the state of rain one hour afterward.
- 5 Conclusion: U-Nets, a convolutional neural network architecture used for image translation, produce better short-term nowcasts than traditional numerical methods such as HRRR.The paper characterizes this as a straightforward use of U-Nets for prediction windows on the order of a few hours.
A Related Deep Learning Work
Prior precipitation-nowcasting work broadly uses recurrent networks that model time or CNNs that transform images. In their optical-flow comparison, the authors find worse performance, likely because optical flow assumes rain amount remains unchanged.
- A Related Deep Learning Work: Prior deep-learning approaches divide into recurrent models that explicitly model time and CNNs that transform input images into output images.This categorization distinguishes temporal architectures such as RNNs from image-translation approaches based on CNNs.
- A Related Deep Learning Work: Examples include ConvLSTM, Trajectory GRU, PredNet, U-Net, baseline CNN, perceptron-based, and multi-task CNN approaches.The cited methods differ in how they represent spatial relationships, temporal structure, or the nowcasting stage.
- A Related Deep Learning Work: Optical flow performed worse in this study than in the cited Lebedev et al. work.The authors suggest this may reflect optical flow assumptions that are violated in precipitation, including constant rain amount over time.
B Problem Formulation
The paper formulates nowcasting as thresholded probability prediction rather than regression of expected rain rate. This classification approach represents distinct probabilities for different precipitation intensities while avoiding the difficulty of full per-pixel distributions.
- B Problem Formulation: The ideal target is a calibrated per-pixel probability distribution for precipitation quantities, but Bayesian approaches are difficult with large data volumes.The formulation conditions precipitation rate on recent MRMS images.
- B Problem Formulation: Regression of expected rain rate can be ambiguous because different atmospheric events may share the same expected value.A certain light shower and a less probable heavier thunderstorm can both have an expected rate of 1mm per hour.
- B Problem Formulation: The paper instead predicts probabilities that precipitation exceeds several rain-rate thresholds.This middle ground can distinguish, for example, a 100% chance of 1mm/hr rain from a 10% chance of 10mm/hr rain.
C Modeling
The model uses a U-Net for image-to-image precipitation forecasting, with convolutional encoder-decoder blocks and skip-connections. Training uses pixelwise cross-entropy on stacked MRMS inputs enriched with temporal and geographic features.
- C Modeling: The CNN maps an MRMS precipitation image to the MRMS image collected one hour later, following an image-to-image translation formulation.In this task, input pixels are mapped to pixels in a target image representing future precipitation.
- C Modeling: The U-Net consists of an encoder with downsampling blocks and a decoder with upsampling blocks.The architecture uses basic, downsample, and upsample convolutional blocks.
- C Modeling: The basic, downsample, and upsample blocks combine convolutions, batch normalization, LeakyReLU activations, pooling, and nearest-neighbor resizing.The upsample operation resizes feature maps using nearest-neighbor interpolation.
- C Modeling: Long and short skip-connections connect encoder and decoder blocks and help gradients update more efficiently during training.Long skips follow the standard U-Net design, while short skips resemble connections used in ResNets and some U-Nets.
- C Modeling: Training uses pixelwise cross-entropy and ADADELTA, with seven downsampling and upsampling blocks and 3 × 3 convolutions.Downsampling uses 2 × 2 max pooling.
- C Modeling: Seven MRMS tiles collected ten minutes apart over an hour are concatenated as channels, with added time-of-day, latitude, and longitude channels.The target is collected one hour after the last input image.