Skillful precipitation nowcasting using deep generative models of radar šØļø [Paper Summary]
An overview of DeepMind's paper Skilful precipitation nowcasting using deep generative models of radar.
Skilful precipitation nowcasting using deep generative models of radar
What
DeepMind trained Deep Generative Models on radar observations in the hopes of creating a model that can accurately model & predict precipitation events within the immediate future (1-2 hours).
Why
Nowcasting (the act of predicting the weather in the next hour or so) is an extremely important task for a variety of fields.(Sports, transportation, news) Accurate weather predictions can improve safety, increase quality of life, and optimize logistical decisions. Current weather prediction systems have a long compute times & struggle with accurate predictions within 2-hours from a given point in time. Machine Learning methods can fill in this gap and provide quick weather forecasts for the immediate future.
How
- The GAN approach is used for their Deep Generative Models.
- The generator context is fed four consectuve radar observations
- Two loss functions are used.
- The first loss function/spatial discriminator is a CNN that attempts to distinguish real radar observations from generated ones.
- The second loss/temporal discriminator is a 3D CNN that attempts to distinguish real & generated radar observations, but enforces a temporal consistency & penalizes predictions that skip or jump forward in time.
- The precipitation events are 256 x 256 crops extracted from the radar stream.
- The stream length is 110 minutes which results in 22 frames.
- Trained on radar observations from 2016-2018
- Radar observations from 2019 are used as a test set.
Key Findings
- DeepMindās Deep Generative Models achieved high ratings from meterologists surveyed. Over 89% ranked it as their preferred choice compared to existing methods.
- More work is needed to improve accuracy on rare & intense weather events
- The model can generate predictions within seconds on just a NIVIDIA V100 GPU.
- DeepMindās work shows you do not need to encode physical assumptions about the weather to achive promising results. This is not news to ML practioners but more proof of a networks ability to model complex phenomena.
Open Questions
- Would a models performance improve if some assumptions/meterological knowledge was encoded?
- If trained on radar observations from a variety of regions, would a models ability to predict improve?
My perspective
This is my first time attempting to summarize a paper in this field. So Iāll save you my opinion until its informed by more papers & model building experience. A lot of the information I gleaned from this paper is all new to me. You can load the model & make predictions using this notebook provided by DeepMind. If you find the methodology interesting go ahead & give the paper a read for details I did not include!