Summary of content : According to the forecast by the Meteorological Department, on August 12, Beijing will usher in the strongest rainfall since the flood season this year, which has aroused the attention of all walks of life, and also made people pay attention to the weather forecast and the scientific prediction methods behind it. Nowadays, in meteorological observation, artificial intelligence has also been added.
Keywords : weather AI, image recognition, neural network
According to the forecast of the Meteorological Department, on August 12, Beijing will usher in the strongest rainfall since the flood season this year, and various departments have issued warning notices.
After a long day and night waiting, the intermittent rainstorms provided a lot of inspiration for netizens and created a lot of widely spread stories. It also makes everyone wonder how to forecast the weather? How to make the weather forecast more accurate and timely?
Who cares about the weather forecast?
In the past, weather forecasting was based on various weather observation instruments and multiple weather stations, which measured temperature, humidity, and air pressure and other indicators, and compiled the observation results on a map.
This picture shows the changes of the atmosphere at different heights and levels, so as to predict the possible weather.
Meteorological data is very complex, often from dozens of sources and types
Different weather will use different facilities for detection. For example, the ground station directly measures wind and precipitation, etc., and can also perform temperature and pressure wet wind observations and lightning observations.
Radar observations, such as Doppler radar, can detect precipitation in real time, as well as automatic remote sensing observations.
The Fengyun satellite we are familiar with is a meteorological satellite, which provides multi-spectral imaging, such as day and night visible light, infrared cloud images, ice and snow cover, vegetation, ocean water color, sea surface temperature, etc.;
In September 2017, WeChat changed the opening picture to
A panoramic view of the motherland taken from space by the geostationary meteorological satellite Fengyun-4A
Nowadays, more objective methods such as numerical forecasting models and algorithm forecasts have been added to weather forecasting, as well as more complete forecasting systems and observational data.
Meteorology, complex enough to reflect national strength
Meteorological research is not just about wind and rain, but from the ocean to the sky, covering the five major circles of the atmosphere, water springs, lithosphere, biosphere, and cryosphere.
The chief engineer of the Film and Television Center of the China Meteorological Administration mentioned in an interview with "I am a scientist iScientist": "In meteorological research, we need to use physics to explain the movement of the atmosphere and the ocean, and we need to use chemistry to understand the changes in matter, Mathematics is required for statistics and calculations. Behind the few numbers of weather forecasting lies the accumulation of knowledge in a large number of comprehensive disciplines, which is the strongest computing power and space exploration ability of a country."
my country’s weather observation network has formed three-dimensional observations. According to the news from the China Meteorological Administration in May this year, my country’s meteorological department has more than 70,000 ground weather observation stations, with a coverage rate of 99.6% in towns and villages across the country, and the data transmission time limit has been increased from 1 hour. To 1 minute.
The National Meteorological Science Data Center provides various public data
A new-generation weather radar network composed of 216 radars has successfully launched 17 Fengyun series weather satellites and 7 in orbit, providing services to more than 2500 domestic users in more than 100 countries and regions around the world.
At present, there are more than 1,000 weather satellites around the world in space, which can provide a large amount of weather data such as wind, rain and temperature. There are hundreds of thousands of national and enterprise-level weather stations on the earth, all of which are constantly collecting real-time data.
National-level weather stations provide convenience for people’s lives, while enterprise-level weather stations provide commercial services, such as providing finer-grained weather data for large farms, sports events, and aviation.
Unpredictable weather, AI can't predict
According to recent data from China Industry Information Network: China's meteorological service industry revenue is expected to reach 300 billion yuan in the next five years.
Many large companies such as GE, IBM, Google, Panasonic, etc. have expanded and provided meteorological data services.
AI Measure: Neural Network
In the paper "Machine Learning for Precipitation Nowcasting from Radar Images" released by Google earlier this year, Google AI researchers proposed a new method for "development of machine learning models for precipitation prediction" research method.
The new method in the thesis is to use data-driven, no atmospheric physical model to establish a short-term precipitation forecast model. Only the neural network is used to learn to fit atmospheric physics through training data sets, instead of using a priori basic knowledge of atmospheric physics.
In this method, precipitation forecasting is regarded as a conversion problem from picture to picture, and a U-net structure convolutional neural network is used to achieve the purpose of forecasting.
AI survey: image recognition
In weather forecasting, radar data is converted into images. By extracting features such as hue, saturation, and brightness of the image, image recognition methods are used to distinguish different weather phenomena, such as rain, snow, hail, dew, frost, and fog. (haze).
The first three images in the upper row show the radar images 60 minutes ago, 30 minutes ago, and 0 minutes ago. The rightmost image shows the radar image 60 minutes later, which is the ground truth of nowcasting.
The left image in the bottom row is for comparison, the vector field generated by advection modeling the data from the top three panels by applying the optical flow (OF) algorithm.
Optical flow OF is a computational vision method developed in the 1940s and is often used to predict short-term weather evolution.
The right picture in the lower row is an example forecast made by OF. It predicts the amount of precipitation, but fails to explain the attenuation intensity of the storm.
AI Survey: High Performance Computing
IBM runs the world's highest resolution global weather forecast model-Global High Resolution Atmospheric Forecast (GRAF). It is the first global weather model that is updated every hour and can predict small-scale weather systems like thunderstorms almost anywhere on the planet.
IBM's luxurious data center for GRAF
In order to support the operation of a large system such as GRAF, IBM supports 84 AC922 nodes, each with 4 Nvidia V100 GPUs and 3.5 PB of IBM Spectrum Scale Storage, which can process up to 10 TB of weather data per day.
AI survey: AI doesn’t count
Although it now appears that artificial intelligence has provided scientific research acceleration in many aspects for weather forecasting and weather forecasting. But after interviewing professional scholars in the industry, we learned that in weather forecasting, there are thousands of factors that affect weather changes. Whether it is sunlight, sea water, ocean currents, every variable is constantly changing, and will also affect climate change.
The more variables involved, the higher the requirements for artificial intelligence training data and computing capabilities. For example, the heavy rainfall in Beijing was predicted and warned one day in advance during strong convective weather. There will be certain errors, whether it is a comprehensive study. It is still AI, and there is still a long way to go in terms of weather data forecasting.
But being able to take advantage of the sudden rainfall in Beijing to let more people understand the scientific knowledge and investment in scientific research behind the weather forecast is also a rewarding and timely rain.
Reference materials:
-I’m a scientist iScientist: "Why is the weather forecast inaccurate? We talked with the meteorologist"
- Google：《Using Machine Learning to “Nowcast” Precipitation in High Resolution》
-The Heart of the Machine "Strengthen Data Analysis, Accurate Weather Forecast, Artificial Intelligence Empower Meteorological Research"
- IBM https://www.ibm.com/weather/industries/cross-industry/graf