https://arxiv.org/pdf/1912.12132.pdf

In this paper, we focus on the sub problem of predicting the
instantaneous rate of precipitation one hour into the future from Doppler radar.
Specifically, we provide three binary classifications that indicate whether the
rate exceeds thresholds that roughly correspond to  trace rain, light rain,
and moderate rain. Our forecasts are at 1km spatial resolution, are within the
continental United States and are based on data from NEXRAD.

We treat forecasting as an image-to-image translation problem where we are given
a sequence of (n) input radar images that start at some point of time, t(in1),
and end at t(inn). Our task is to generate the radar image at some point in the
future, t(out). At the time scales we are working with, horizontal atmospheric
advection is the primary driver for changes in the radar images, which represent
the dynamics we are capturing in our neural network model.

Conclusion:

We explore the efficacy of treating precipitation nowcasting (very short term
forecasting) as an image-to-image translation problem. Instead of modeling the
complex physics involved in atmospheric evolution of precipitation, a time
consuming and computational intensive practice, we treat this as a data-driven
input/output problem.

The input is a sequence of MRMS (multi-radar
multi-sensor) images providing a short history of rain in a given region
and the output is the state of rain one hour afterwards. We leverage the power
of U-Nets, a type of Convolutional Neural Network commonly used in image
translation problems, and demonstrate that straight-forward uses can make better
predictions than traditional numerical methods, such as HRRR, for short-term
nowcasting predictions presuming the window for the prediction is on the order
of a few hours.

An open question remains as to whether pure Machine Learning data-driven
approaches can outperform the traditional numerical methods, or perhaps
ultimately, the best predictions will need to come from a combination of both
approaches.

Thanks to Gary Osoba.