Everyday Dashcam Photos Could Help Cities Find Flooded Blocks That Nobody Reported

When a big storm overwhelms a city’s drains, the people hit hardest are often the ones least likely to be counted. During New York’s flash flood of September 29, 2023, the researchers write, basement apartments housing lower-income residents were among the places hit hardest. If no sensor sits on the block and nobody files a complaint, the flooding can go unrecorded, and so can the case for fixing it.

A new method called BayFlood looks for that missing flooding in an unexpected place: photos from dashboard cameras in cars driving around the city. It comes from researchers at Cornell Tech and the University of California, Berkeley, and was published on July 29, 2026 in Nature Communications. Applied to that one storm, it flagged Census tracts home to about 1.05 million people, 12 percent of the city, as high flood risk. About 100,150 of those people live where none of the city’s usual signals (flood sensors, 311 complaints or stormwater maps) showed any flood risk at all.

The storm, and the cameras that happened to be there

The paper describes the September 29 storm as a state of emergency: subway suspensions, flooded highways, homes and businesses inundated, and dozens of rescues. The team analyzed 926,212 dashcam images of New York from that day, provided for research by Nexar, a dashcam maker that aggregates image data.

Nobody could label that many photos by hand. So the team asked an open-source vision-language model, Cambrian-1, one plain question about each picture: “Does the street in this image show more than a foot of standing water?” The model needed no flood training data, which matters, because labeled flood photos are typically unavailable. It flagged 0.2 percent of images as flooded.

Checking the machine’s homework

The researchers did not take the model’s word for it. They hand-checked 500 images it called flooded and 500 it called dry. About 65.8 percent of the “flooded” calls were right, and only 0.6 percent of the “dry” calls were actually flooded. Imperfect, but a strong signal: an image the model flags is about 110 times more likely to show a flood.

A statistical (Bayesian) model then combined those calls with the hand checks and other data, smoothed across nearby areas and estimated how flooded each Census tract was. The team also tested the approach on three other storms, two in New York and one in the San Francisco area, for 1,440,184 images in all, and report that the classifier gave a strong signal for floods across cities and time periods.

What it showed about who gets heard

New York’s flood sensor network, FloodNet, had around 250 sensors when the paper was written, expanding to more than 500 at a cost of $7.2 million, in a city with roughly 40,000 intersections. Of the people in high-risk tracts, 832,929 lived where there was no FloodNet sensor at all.

The dashcam estimates also exposed a pattern in the city’s 311 complaint line. After controlling for estimated flooding, the authors found that tracts with higher fractions of White residents, lower fractions of Hispanic residents, higher average household incomes and higher fractions of children were statistically significantly likelier to have a 311 flood report. They note this fits past evidence of biases in 311 reporting.

The team used its map to suggest 25 new sensor locations and submitted them to FloodNet, which, the authors write, expressed interest in the data as one source of signal alongside its placement process, which is largely driven by community engagement.

Where this could go

For a family in a basement apartment on a block no sensor watches, the hope is simple: that the next storm’s flooding gets noticed, and that the next sensor or drain repair lands on their street. The authors say real-time monitoring would likely need smaller models running on the dashcams themselves, which is the honest next hurdle.