Evidence of chronic flooding from a global 10-meter flood‑occurrence dataset 

Rohit Mukherjee

Postdoctoral Researcher

December 16^th, 2025

PNNL is operated by Battelle for the U.S. Department of Energy

Global flood datasets spatially and temporally incomplete

• Need for global, continuous coverage across space and time
• + urban resolving

​

• Adverse impacts of chronic or nuisance flooding

​

• Google’s Dynamic World land cover product offers water class at 10-meters

​

• Google Earth Engine allows processing of millions of scenes over 10 years (2015 to now)

2

Global Flood Database (Tellman and Sullivan et al. 2021)

Increasing urban exposure to flooding (Cao et al. 2022)

Urban resolving global flood dataset (2015 to 2025)

3

Hahnville, west of New Orleans

Dynamic World water class

For each Sentinel-2 scene,

• DW produces a map if cloud cover is at max 35%
• Produces a water band from 0 to 1
• A separate band where each pixel is labeled as the likeliest class across all ten classes

​

We look at each land pixel since 2015, and ask, can this pixel flood and can it be observed by a satellite sensor? If yes, how many times did this pixel flood in the last 10 years?

Isolating flooded pixels from inundation

4

Hahnville, west of New Orleans

Dynamic World label – remove water occurring more than 50%

Dynamic World water class probability over 0.5

ESA WorldCover

permanent water, wetlands, mangroves

Potential flooded pixels

we start with

remove

remove

Applying physical constraints

5

Building footprints

Slope higher than 5 degrees

Flooding rate for every pixel since 2015

Hahnville, west of New Orleans

remove

remove

What chronic flooding looks like over 10 years

6

Flooding tendency

• 359 valid Dynamic World observations (2015–2025)
• 6% of pixels flooded at least once
• 2.1% of pixels flooded two or more times

​

Land-cover classes experiencing flooding

• Flooded at least once:

Bare/sparse vegetation (22%), built-up (13%)

• Flooded at least twice:

Bare/sparse (17%), grassland (4.4%), built-up (4.1%)

​

Prior land cover of flooded built-up pixels

• 70% of flooded built-up pixels were already urban in 1985
• Remaining 30% were formerly forest, salt marsh, water body, or cropland

Carolina beach, NC

Global evidence of chronic urban flooding (2015–2025)

7

Higher local flooding (urban > rural) is more common on coasts (22.1% vs 14.2%) and in the Global South (20.9% vs 12.3%).

urban

rural

How accurate is Dynamic World water class?

8

Low intersection over union

but high precision

Craig, Missouri

March 2019 Missouri River floods

*

*

Urban Flood Observations (Mukherjee et al., 2025)

Sentinel-1 extends flood coverage

Sentinel-1 can capture signal despite cloud cover, with a 5-day revisit at 10-meters

​

Extending our dataset with Sentinel-1

• Trained a deep-learning model to generate Dynamic World-like water predictions from Sentinel-1

​

Increased temporal sampling

• Valid observations increased from ~400 to ~900 at this location
• Effective revisit time improved from ~5.6 days to ~2.4 days

​

9

Camp Mystic

Our flood dataset

+ Sentinel-1

No single perfect flood product, but complementary solutions

• Optical: more accurate but cloud cover
• Floods often occur under cloudy conditions

​

• Radar: cloud free, but less accurate
• Sentinel-1 struggles in arid and dense urban areas

​

• Commercial imagery: high accuracy, lower accessibility
• PlanetScope sensor inconsistency reduces robustness

10

Did this location flood in the last 10 years? (Example: coastal Panama)

11

Rohit Mukherjee

rohit.mukherjee@pnnl.gov

rohitmukherjee.space

Pacific Northwest National Laboratory

TC Chakraborty

Earth Scientist

tirthankar.chakraborty@pnnl.gov

​

FloodPlanet (Zhijie Zhang, Jonathan Giezendanner and Rohit Mukherjee, Beth Tellman, Alexander Melancon, Matt Purri, Iksha Gurung, Upmanu Lall, Kobus Barnard, and Andrew Molthan, 2025)

Urban Flood Observations (Rohit Mukherjee, Hannah K. Friedrich, Beth Tellman, Ariful Islam, Zhijie

Zhang, Jonathan Giezendanner, Upmanu Lall, and Venkataraman Lakshmi)

19 global flood events

12 global urban flood events

Our dataset provides the first global, urban-resolving assessment of chronic flooding based on a decade of satellite observations

Thank you