Unraveling the mystery of DeepMind’s rainfall nowcasting: a step-by-step tutorial for hydrologists

Since Hinton et al. introduced Deep Learning (DL) in 2006 [1], DL methods have led to breakthroughs in various scientific fields, such as speech recognition, medical, materials, and many more. Various early attempts to apply DL to short-term rainfall forecasting (nowcasting) were also reported. However, these early models did not lead to significant improvements as compared to non-AI nowcasting models such as STEPS [2]. It was TrajGRU model [3] which first demonstrated the potential gains that may be achieved with DL-based nowcasting models. Since then, a variety of DL models have been proposed or applied to tackle rainfall nowcasting, with the most iconic ones including U-Net, MetNet and DGMR [4-6]. Similarly to the Trajectory GRU (TrajGRU) model, the U-Net and MetNet models show clear improvements in predicting the occurrence of rainfall at high spatial and temporal resolutions and with a longer lead time, as compared to non-AI models. However, the predicted rainfall images from these three models (and their variants) become overly smooth rather quickly (at lead times of 15-20 minutes); this is a common ‘feature’ of many other DL models [7]. This means that significant amount of spatial rainfall details is lost, which is undesirable for certain hydrological applications, such urban flow and flood forecasting where small-scale rainfall variability -in particular localised peaks- may have tangible impacts [8]. In 2016, DeepMind [6] proposed a new type of DL-based nowcasting model called the Deep Generative Model of Radar (DGMR), which is based upon a Generative Adversarial Network (GAN) framework. The DGMR successfully improves the aforementioned smoothing drawback of other DL-models by incorporating noise into the rainfall forecast generator such that small-scale rainfall details can be preserved and, consequently, localised peak intensities can be better predicted. DGMR thus shows great potential for hydrological applications.

In spite of the success, the model structure of DGMR is complex and hard to digest by someone without proper training in DL. Therefore, even though the model structure has been published, it remains a mystery for most hydrologists, thus hindering its application.

In this work, we explore the success of DGMR with an in-depth analysis of its model structure. More specifically, through the process of re-constructing the DGMR model, we have developed a short tutorial on the different model components, in plain language and with example images and intermediate analyses. This will enable better understanding of the features and behaviour of the DGMR model and of the implications for hydrological applications. Additionally, a better understanding of the DGMR model components may instigate further improvements.

References:

[1] Hinton, G.E., et al., Neural Comput., 18 (7), 1527-1554, 2006.

[2] Bowler, N.E., et al., Q. J. Roy. Meteor. Soc., 132(620), 2127-2155, 2006.

[3] Shi, X., arXiv preprint arXiv:1706.03458, 2017.

[4] Agrawal, S., et al., arXiv preprint arXiv:1912.12132, 2019.

[5] Sønderby, C.K., et al., arXiv preprint arXiv:2003.12140, 2020.

[6] Ravuri, S., et al., Nature, 597, 672-677, 2021.

[7] Ayzel, G., Geosci. Model Dev., 13(6), 2631-2644, 2020.

[8] Ochoa-Rodriguez, S., et al., J. Hydrol., 531, 389-407, 2015.