2020 saw the worst wildfire season on the West Coast of the U.S. since records have been kept. By the end of the year, roughly 4% of California’s acreage had been burned, an area roughly the size of the state of Rhode Island. The fires also broke the weather forecasts.

When there are active wildfires, intuitively, there is a lot of smoke. Where that smoke lingers, it can block out sunlight, making it cooler on the ground than it otherwise would be. Most atmospheric simulation models used operationally to forecast the weather do not anticipate wildfire smoke or incorporate its effects, nor do the AI methods trained to correct and blend the model forecasts. So when AI tried to forecast the California weather after being trained on historical data that didn’t include instances of record-setting wildfires, it didn’t anticipate these temperature anomalies. The results were forecasts that skewed far too warm due to an omission that would have been obvious to anyone on the ground.

This episode highlights one of the main weaknesses of AI when it makes predictions on its own, without the additional insight of a human being. AI almost always makes the forecast better. But when it’s trained with historical information, AI is not so great at making sense of events that haven’t happened before, or have occurred only rarely.

When it comes to the weather, that’s an important limitation. Not only are extreme weather events getting increasingly common, but they’re having a growing impact. The year 2020 saw 22 weather disasters that caused more than $1B in losses each.[1] The solution is to marry the speed and skill of AI with the wisdom of experienced human forecasters “over the loop.” In the case of last year’s wildfires, meteorologists at IBM’s The Weather Company recognized that our AI forecasts were too warm in smoky areas and jumped in to supply corrective nudges. The AI’s machine learning, which is constantly updating, began to correct for the new conditions as it learned from each new data point, then adjusted back again when the smoke cleared.

Incorporating the human perspective is a crucial part of both keeping automated forecasts in line with reality and tailoring them to the needs of individuals or businesses. When forecasts incorporate a human perspective, they can go beyond what the weather will be to what it will mean for people and their activities: for example, predicting flu transmissibility, pollen counts, the best time to go for a run or a safety threat that requires action. And humans are still unmatched in translating the forecast and communicating what people and businesses need to know, particularly when faced with an unprecedented weather event.

Harnessing AI’s strengths

One reason that AI is such a powerful tool is that AI and human beings actually have surprisingly complementary strengths.

While humans are better at subjective observations or perceiving meaningful anomalies, AI’s strengths in lightning-fast data processing, curation, and fusion are essential for taking the fire hose of weather information and making it comprehensible to humans in a timely way. Weather data comes from a wide variety of sources, from satellites and ground sensors to radars, cell phones and even connected vehicles, producing billions of data points every second from all around the world. Without AI, humans would find all that information overwhelming. For example, IBM’s The Weather Company handles more than 400 terabytes of data daily. AI helps clean up data sets so that they are high quality and accessible for people to study and understand. It also combines disparate data sources and blends them together to make hyper-local weather predictions anywhere in the world in the blink of an eye.

Because it’s so good with numbers, AI is also an incredibly powerful tool for determining the range of possible weather outcomes and how likely different scenarios are. These weather probability predictions can have huge business implications. Take the example of a grower deciding whether to deploy blowers or mister spray to mitigate crop damage in the face of a possible freeze. Knowing the likelihood of freezing temperatures helps them weigh the cost of an intervention against the potential loss if no action is taken. AI can also help model the surviving crop value, predicting the expected cost and reward for each outcome and allowing the farmer to make a more optimal decision. Situations in which quantifying weather uncertainty helps people make better decisions abound in business, government and our personal lives.

Anticipating human needs

One of the main reasons people look to weather forecasts is to enhance their ability to plan, from the type of clothing they need to be comfortable on their morning commute to what time the weekend picnic should start. And here, as well, human perspective, along with AI, is important. AI may not be ready to help you plan a weekend by itself, but a meteorologist who understands what the weather means for activities in your location and season can interpret the forecast and provide just the guidance you need. Combining the human perspective with an AI-driven forecast of the weather, its uncertainty and its impact can automate this process, creating a personalized “decision service.”

At IBM, we believe there is even greater potential for AI and humans to work together to create more accurate and relevant forecasts and use them to guide decisions. For instance, AI can combine the latest observations and atmospheric simulations to create precise forecasts for precipitation, down to the minute that the rain is likely to start in your neighborhood, for those who have opted to share that information. With human input, we can figure out who needs that kind of granularity, or what it may mean in terms of flooding hazard or evacuation routes. And with feedback on what people value, AI can learn to guide us to superior choices. By working together, humans and AI can enhance how we adapt to changes in the weather, a crucial step toward helping individuals, organizations and enterprises better prepare for the future.

By John K. Williams, Ph.D., | Senior Scientist of Forecasting and Machine Learning at IBM’s The Weather Company
Source IBM