Maps have been with us for a very long time, even as their uses — and the technology required to create and interpret them — has changed greatly. Today, individuals and companies use maps to plan routes, demarcate political territories and understand trends as time and demographic changes take hold.
Artificial intelligence is rapidly improving the quality and usefulness of our maps. In some cases, it’s even unlocking brand-new business models.
Funny-looking cars with cameras strapped to their roofs — like Google’s “street view” mapping cars — may not be long for this world. Major technology companies and startups alike build products and revenue streams around this type of “manual” mapping process. The trouble is, it’s extremely time-consuming and usually expensive.
Artificial intelligence appears nearly ready to make this process a thing of the past. A paper written jointly by researchers at MIT and Qatar Computing Research Institute describes how mapping companies can instead use AI to tag roadway and topographical features in satellite images, both accurately and without human intervention.
The research so far is promising. It has implications for ride-sharing drivers and delivery logistics companies that sometimes face delays due to out-of-date maps in quickly developing areas. The developers of the “RoadTagger” system say the system already predicts the number of lanes on roads with 77% accuracy and identifies road types with 93% accuracy.
AI Delivers More Accurate and Timely Public Health Alerts
There are two main challenges for AI developers working in the mapping field. These are the quantity and diversity of data used in “training” neural networks. The City of Los Angeles provides a case study in how to overcome these challenges as well as a look at how useful AI-generated maps can be.
The American Lung Association places the greater Los Angeles-Long Beach area in the top spot among the “most polluted cities” in the United States. Improving public health in areas stricken with high level of air pollution begins with understanding the underlying social and geographical causes and predicting when conditions will be at their worst.
The WHO estimates that 7 million people die globally each year from fine particulates in compromised air. Alerting the most vulnerable members of the population quickly, including caregivers for children and the elderly, is essential.
To build a complete picture of the factors at work, researchers began pooling data from the Pediatric Research Integrated Sensor Monitoring System (PRISMS) and data from the OpenStreetMap database. The result was an AI model that predicts where and when fine particulate levels will be at their highest.
Researchers were able to draw meaningful correlations between land use, traffic patterns, the weather and road design (including proximity to residential areas) to create predictive models for high-risk air quality conditions. Best of all, the longer projects like these run, the better the AI model becomes at generating meaningful predictions and improving urban design.
AI Improves Competitiveness and Location Intelligence for Businesses
Growing businesses have a considerable number of variables to study and interpret as they plan the way forward in competitive markets.
AI is proving instrumental in collecting unstructured data from various sources. For instance, it’s able to collect customer data, supply chain data, data from competitors and the larger market, weather patterns, and even geopolitical events, and then draws meaningful conclusions from it.
One area with significant growth potential includes using AI to map customer sentiment and shopping habits in real-time. One company, called Jetpac, used public data from Instagram to analyze photos about local businesses and determine traffic volume and hours of peak demand.
Google bought Jetpac in 2014 and soon began incorporating the company’s tech into its own. The lesson for AI startups and their customers is clear: It’s time to find uses for the 60 to 73% of data that companies accumulate but never put to analytical use. This data is a goldmine for training neural networks to deliver actionable data — but it has to begin with a clear mission or a problem to solve.
AI Creates Detailed 3D City Maps and Improves Disaster Response
There are many ways in which major metro areas can put detailed, always-up-to-date 3D city maps to use. 3D models help us engage in better urban planning and appraise the potential environmental impact of new construction projects. Other applications involve identifying areas of high risk during natural disasters and mounting a timely response after severe weather strikes.
The question is how to turn 2D aerial photography into 3D models that are useful to engineers and disaster response agencies. Naturally, the answer lies in artificial intelligence.
In a recent article, data scientist Dmitry Kudinov described a method to train a neural network to extrapolate detailed and accurate 3D models of entire cities using 2D aerial photography. One of Kudinov’s colleagues, Shairoz Sohail, explained the significance of this technology in a follow-up post called “AI for Good — Disaster Response.”
When the team combined this neural network with satellite imagery, the system was able to turn 2D representations of buildings and structures into highly accurate 3D models, down to building height and the shape of the gables.
It also identified with impressive accuracy which buildings had been damaged and required disaster response and which roads and pathways had become impassable during natural disasters. It’s not hard to see the applications, from better tools for first responders to web or mobile apps for residents as they attempt to move about the area during the aftermath.
Training AI to Build Better Maps
Projects like these drive home the importance of rich datasets. For AI engineers, AI companies and their prospective customers, it’s also a reminder that for all its promise, artificial intelligence is still in its infancy.
Any claims made today about AI adding value to a company or boosting its predictive capabilities comes with a caveat: An AI’s “intelligence” is only as strong as its “training” — and the results it delivers correlate directly with the quantity and quality of data it uses to make them.
Jenna Tsui is a technology journalist with writing experience in future & disruptive technologies, AI, medtech, and scientific development. To see more of Jenna's work, visit The Byte Beat, follow her on Twitter or check her out on LinkedIn.