Modern weather forecasting relies on a vast data collection network. While adding high resolution remote sensors provides the basis for more accurate and precise weather prediction, it is also brings the challenge of how to most effectively process, understand and maximize all that data.
Traditionally, weather forecasting focuses on developing complex dynamic numerical models aiming at more accurate predictions. But due to the uncertainty of the weather and some model drawbacks such as compatibility within coordinates, this method can fail to meet the requirements of different use cases. AI and data-driven methods have come into play to bridge the gap.
AI is not new to meteorology, it has found applications in weather forecasting since the 80s when neural networks were first introduced. With AI models gaining power and momentum across a number of industries in recent years, meteorological researchers are now applying the tech in satellite data processing, nowcasting, typhoon and extreme weather forecasting and other business and environmental analytics areas.
The journal Earth and Space Science identifies AI technologies as key for reducing human forecasters’ workloads while delivering more accurate and timely predictions. The US National Oceanic and Atmospheric Administration (NOAA) also says incorporating AI and machine learning significantly increases the prediction ability of extreme weather such as thunderstorms and hurricanes.
Google is an industry pioneer. The tech giant presented new research into the development of deep learning models for precipitation forecasting in December, 2019. The team treated forecasting as an image-to-image translation problem and leveraged the power of the ubiquitous UNET convolutional neural network. U-Net has a network architecture where layers iteratively decrease the resolution of the images passing through them in an encoding phase, and the low-dimensional representations of the image created by the encoding phase are expanded back to higher resolutions in the following decoding phase. In tests, the proposed system outperformed three commonly used models: optical flow, persistence, and NOAA’s numerical one-hour HRRR nowcasting prediction.
Large companies are also partnering to ride the trend. IBM acquired the Weather Company in 2015. The two companies’ combination of technology and expertise in weather data engendered the Atmospheric Forecasting System, which delivers personalized, actionable insights to customers across the globe. The system provides a wide range of prediction services including weather-related power outages predictions 72 hours in advance using machine learning models. The system is said to be the first-ever operational global weather model to run on GPU-accelerated servers in order to handle the increased resolution and frequent updates.
The meteorological bureau of Shenzhen, China, has been exploring ways to enhance weather forecasting in the challenging coastal area of Guangdong, where severe convection weather is frequent. The bureau worked with tech giant Huawei to build a meteorological cloud platform with 5G and AI ability to expedite the development, training and deployment of prediction models from 1-2 weeks to 3 days or less.
Startups are also positioning themselves as game-changers in the industry. ClimaCell’s patented MicroWeather engine applies machine learning to historical gridded weather data to improve accuracy in weather forecasting. The company recently launched a historical weather data archive for AI-model training derived from a global network of wireless signals, connected cars, airplanes, street cameras, drones and other internet of things (IoT) devices.
Although AI will continue to play an important role in weather forecasting, it hasn’t been easy to attract AI talents to the field. Meteorological bureaus can’t compete with the salaries offered by tech companies focused on snazzy sectors such as self-driving cars and computer vision. What we see happening instead is tech giants partnering with local meteorological organizations or taking over the work themselves. Globally, we may expect more such arrangements in the future.
Author: Jingya Xu | Editor: Michael Sarazen