Cyclones and The City: how deep learning can help us prepare, adapt and react to broad ranging challenges.
Earlier this month NASA released its Hurricane Intensity Estimator, an application that aims to accurately and automatically predict the direction and velocity of tropical cyclones using deep learning algorithms. This revolutionary technology has the potential to literally save lives.
Since the early 1970s, satellite imagery has been used to estimate the intensity of hurricanes using the Dvorak technique. The Dvorak technique, named after the American meteorologist Vernon Dvorak, relies on subjective pattern recognition and often produces broad-ranging and inconsistent predictions. Given the considerable loss of life and damage that is caused by these extreme weather events, the motivation to help scientists and forecasters more accurately predict a hurricane’s path and wind speed is clear.
Enabled by today’s unprecedented computational power and advancements in neural networks, NASA uses deep learning models to, first, identify hurricanes in satellite imagery and, second, actively monitor their occurrence and estimate their intensity. The output data can be used as a situation awareness tool and help with preparedness and disaster planning.
Machine learning is not magic. It relies on hundreds of thousands of data points. In this case, using satellite imagery from the US Naval Research Laboratory database, over 200,000 images were labelled with wind speed data from the National Hurricane Center to form the training data. The labelled images are propagated through convolutional layers of pattern spotting filters, which can detect the features or patterns that are indicators of a hurricane. These layers build the model.
Deployed to Amazon Web Services, the system actively monitors for hurricanes and a workflow is triggered when one is identified. Using the satellite imagery from GOES-16, the first of geostationary operational environmental satellites that was launched into space in 2016, the system estimates the wind speed in near real-time using the trained model. The predictions are recorded alongside the images so that the predicted outcomes can be compared to the actual observations and the model retrained and improved.
Deep learning algorithms such as these are inspired by the function of the brain and are used by FeedStock in our business analytics tools. We certainly can’t claim to be saving lives, but our software does help people harness data and adapt to new challenging situations and solve business problems. FeedStock mines new sources of business interaction data to give our clients insight into the way they are working, gain a deeper understanding of client preferences and deliver services on a hyper-individualized basis. Our artificial intelligence driven product will not steal jobs from humans, but empower you to work on interesting, more value-creating tasks by digitalising the mundane parts of your business.
This is just one example of how smart technologies, like those we harness, will change peoples’ lives for the better.