Researchers Turn a Digital Eye on the Weather, Converting Cameras Into Rainfall Sensors
Novel machine learning model provides high-accuracy measurement and prediction using existing surveillance infrastructure.
Researchers from Tianjin University are looking to rapidly deploy city-scale rainfall monitoring and prediction systems, without the need for any new hardware — by repurposing existing camera systems.
"Our system leverages widely available surveillance infrastructure and advanced AI [Artificial Intelligence] to fill gaps left by traditional rainfall monitoring techniques," says senior author Mingna Wang of the team's work. "What's most exciting is that we can now provide highly accurate, real-time rainfall estimates using existing urban technology, even under challenging conditions like night-time or high-density rainfall. This opens the door to smarter flood management systems and more resilient cities in the face of climate change."
Traditionally, rainfall is monitored using physical sensors — often as simple and analog as a bucket that fills and tips, triggering a reading each time. While this is entirely functional, it means readings rely on the deployment of hardware and coverage, as a result, is incomplete. Cameras, by contrast, are everywhere, and it's these that the researchers' proposed alternative uses in order to monitor rainfall over a wide area with high granularity.
The team's software uses a feature extraction module (FeM) to analyze frames from existing camera systems and extract texture features through an image quality signature system — highlighting streaks of rain even in noisy and low-light conditions. A separate rainfall estimation module (RiM) uses a combination of depthwise separable convolution (FSC) and gated recurrent unit (GRU) machine learning approaches to turn those data into rain event patterns across both locations and time.
In testing, using real-world camera systems in the cities of Tianjin and Fuzhou, the model returned rainfall readings matching those of a traditional rainfall gauge while providing increased accuracy in predicting future rainfall events, the team says, even in cases of poor visibility.
The team's work has been published in the journal Environmental Science and Ecotechnology under open-access terms.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.