A few weeks ago, Google’s artificial intelligence (AI) used a machine learning model to improve screening for breast cancer,media reported. Now, the company has used convolutional neural networks (CNN) in instant forecasts of precipitation. Google AI researchers mentioned its use of CNN in short-term precipitation forecasts in an article called Machine Learning for Precipitation Now Fromcasting Radar Images.
The results look promising, and according to Google itself, the results are better than the traditional method: this precipitation forecast focuses on 0-6 hours of forecasts, which produce a resolution of 1 km and a total delay of only 5-10 minutes (including data collection delays). Even in the early stages of development, it outperforms traditional models.
It is understood that traditional methods need to combine prior knowledge of how the atmosphere works, while the “physical freedom” method used by this group of researchers has turned the problem of weather forecasting into a separate image-to-image conversion problem. So the team-trained CNN — U-Net — only needs to estimate atmospheric physics of the training examples provided to it.
To train U-Net, the team used multispectral satellite images. Data collected from the continental United States from 2017 to 2019 were used for initial training. More specifically, the data is divided into four-week blocks, the last week of which is used as an evaluation data set and the rest is used as a training dataset.
Unlike the high-resolution rapid refresh (HRRR) numerical forecasting, light flow (OF) algorithms, and persistence models used by traditional instant forecasting, the U-Net model uses precision and recall charts to perform better forecast quality.
In addition, the model provides instantaneous predictions. This is an additional advantage because traditional methods such as HRRR calculate latency of 1-3 hours. This enables machine learning models to work with new data. However, the numerical model used in HRRR can make better long-term predictions, in part because it uses a complete 3D physical model — it is difficult to observe cloud formation from 2D images, making it more difficult for machine learning to learn the convection process.
Google believes that combining HRRR and machine learning models could be better because they enable accurate, fast short-term and long-term predictions. The company says it is considering applying machine learning technology directly to 3D observations in the future.