Increasingly, meteorologists are turning to AI to improve weather modeling, from predicting short and long-term patterns to predicting life-threatening extreme weather events.
The monitoring of Earth, its environment, and its weather system have come on leaps and bounds in recent decades. From the early days of weather monitoring in the 19th century using telegraphs and telephones to the present day when multiple satellites monitor Earth from space.
The proliferation of Earth-observation satellites aren’t the only new sources of data for metrologists to consider. Air-pressure monitors in billions of mobile phones are constantly collecting atmospheric data.
That means whilst the technology used to monitor Earth’s environmental complex system has leveled up, so too has the volume of data it supplies. This has led to such a deluge of data that scientists simply can’t integrate it all into their models.
This has led to an increasing reliance on machine learning in Earth and environmental sciences — collectively known as Earth sciences, with significant progress being made in the field of numerical weather prediction (NWP). This application is even helping scientists predict extreme weather events, including tracking tropical storms as they evolve into violent and destructive hurricanes.
In April 2019, the National Oceanic and Atmospheric Administration (NOAA) held its first workshop concerning the use of AI in Earth sciences, bringing together hundreds of scientists, program managers and leaders in fields as diverse as academia and the private sector.
The meeting, which allowed experts to exchange ideas, experiences and suggestions with a wider community, has now been followed by a paper¹ published in the Bulletin of the American Meteorological Society. The paper explores concepts discussed at the workshop, the current status of AI in weather prediction and modeling, and how machine learning could provide a massive boost to these fields in the near future.
AI in Weather Prediction and Modeling
Without question, weather science is on the cusp of a major paradigm shift. Monitoring with satellites has delivered so much data that it is becoming vital to employ recent advances in machine learning to deal with it. This is coupled with the fact that we are becoming increasingly reliant on weather prediction as a society.
This has led to Earth scientists and meteorologists enviously eyeing advances in AI data-handling like graphical processing and other cost-effective hardware frameworks made in other fields such as finance, facial recognition and even in the development of self-driving cars.
This has led many researchers, institutes, and private sector companies to develop efficient and intelligent signal and image processing, pattern recognition, and prediction capabilities to combine data from diverse sources and build comprehensive weather models.
Such use of AI and machine learning in weather forecasting is expected to allow for the simulation of long and short-term weather patterns, enabling us to predict future climate trends and predict small weather events just hours or even seconds into the future.
One current application of AI in weather forecasting that proves these ambitions are achievable comes in the form of the Dynamic Integrated Forecast (DICast), a two-decade-long project initiated in 1998 that is at the heart of many applications.
Taking data from a range of sources, DICast applies automated forecasting techniques to create independent forecasts. These individual predictions are then combined using a machine learning program that employs ‘fuzzy logic’ to create completely automated, timely, accurate forecasts out to ten days at thousands of international locations.
AI Applications in Extreme Situations
The viability of AI in the Earth sciences has also been demonstrated more recently in the fight against wildfires. Increasingly firefighters are turning to satellites and remote drones to collect data about the spread of wildfires and even predict where such events could begin. And handling this data requires AI.
Twelve states in the US are currently testing a scheme in which satellites use microwaves and leaf color, amongst other factors, to determine how dry an area of forest is, the logic being that moist trees and leaves don’t burn. An AI then takes this data and collates it in a ‘forest dryness map’ which so far has demonstrated itself to be around 70% accurate.
Of course, wildfires aren’t the only way in which weather effects can threaten property and lives. Extreme weather like hurricanes and tornadoes are a significant and growing problem in many areas of the world.
This has inspired a $20 million USD investment by the National Science Foundation (NSF) into a program that uses image recognition algorithms to predict extreme weather, including tracking and predicting the path of tornadoes.
Will AI weather prediction replace traditional methods?
Despite the fact that AI weather predictions are becoming more and more precise and accurate, they are still yet to be used to create models completely independent of traditional methods of weather prediction. The reason for this may be because of how differently traditional methods and AI alternatives reach their conclusions.
Whilst standard methods use the laws of physics to model the world and environment, machine learning uses up to 100s of millions of fragments of data to build a model, employing past observations and pattern recognition to make predictions.
The fact is, traditional meteorological methods are still quite effective, preventing meteorologists from contemplating a completely AI-based alternative. This doesn’t mean, however, that machine learning won’t irrevocably change these traditional methods. Turning to AI will reduce that amount of computing power needed to create a comprehensive model of Earth’s weather, whilst simultaneously ‘baking in’ more and more data.
Sid Boukabara, principal scientist at NOAA’s Center for Satellite Applications and Research and lead author of the Bulletin of the American Meteorological Society paper predicts that this unification of methods could lead to significantly better weather modeling, “For certain components, it could be 10 times more efficient to 1,000 times more efficient.”
1. Boukabara. S., Krasnopolsky. V., Penny. S. G., et al, , ‘Outlook for Exploiting Artificial Intelligence in the Earth and Environmental Sciences,’ Bulletin of the American Meteorological Society, [DOI:https://doi.org/10.1175/BAMS-D-20-0031.1]