Extreme weather events, whether scorching temperatures that ruin crops or killer storms that drown coastal towns, are likely to be more frequent and more powerful with climate change.
Quantifying the increase in these extreme events (and their economic and public health costs) requires combing through thousands of gigabytes of data that climate models generate every day.
Scientists can’t just look at the results of their climate models and count hurricanes or droughts. Instead, they are turning to to find such extreme weather events in their models’ data.
For decades, modelers have relied on heuristics—mathematical definitions of an object of interest—to pinpoint extreme weather events. But heuristics cannot capture the complexity and variability of weather, which are very difficult to condense into a set of values and threshold conditions. Although many algorithms based on heuristics can detect atmospheric rivers (air currents that transport water from the tropics to the poles), for instance, their results tend to be unreliable, with large discrepancies between algorithms. Researchers hope to replace them with a new generation of () algorithms.
To build these better algorithms, climate scientists have turned to a particular subset of techniques known as . Deep learning systems do not require a set of human-defined rules and values to guide their output. Instead, researchers “train” a system with hundreds or thousands of solved examples that the system then analyzes to create its own relevant rules.
Deep learning was first developed in the 1970s in the field of computer vision, with the goal of training computers to “see” or recognize certain objects in digital images.
“In the 80s, 90s, and 2000s, people kept coming up with heuristics for defining what makes a pedestrian, what makes a car, what makes a face, and so on and so forth,” said Prabhat, a computer scientist who leads big-data initiatives at the Lawrence Berkeley National Laboratory in California. “In the last 10 years it has been conclusively proved that and, in particular, techniques are truly well suited for solving [the computer vision] problem. We felt that we could apply the same idea to finding extreme weather patterns.”
To test this approach, Prabhat and his colleagues attempted to train a network to recognize and draw labels around two types of extreme weather in high-resolution simulation data: tropical cyclones and atmospheric rivers, both of which are associated with heavy rainfall. After being trained with over 500 labeled examples, their system can detect most atmospheric rivers and tropical cyclones in simulation data it has never seen before. Their work is described in a preprint article in the open-access journal Geoscientific Model Development.
Experienced climate scientists have been “very convinced that the network is identifying the right patterns,” says Karthik Kashinath, a scientist leading the project at the National Energy Research Scientific Computing Center. That’s not to say the project is perfect—the system still yields a number of false positives and false negatives—but the researchers hope it can be improved with more training data.
Generating Training Data
Obtaining reliable training data is one of the main challenges for applications. For extreme weather patterns, the labeling has to be done by experts in the field. “We’re going directly to the climate experts, the meteorologists, who have been looking at these patterns for many years, so they have a good sense of what they look like,” Kashinath said. But labeling these images is a tedious and time-consuming process.