Skip to main content

NASA NRT Global Flood Product Update

Abstract:
NASA’s Global Flood Product has provided near real-time (NRT) daily inundation maps since 2012 using the Moderate Resolution Imaging Spectro radiometer (MODIS). With MODIS data acquisition set to end in 2026, NASA launched a new NRT flood product in April 2025 using the Visible Infrared Imaging Radiometer Suite (VIIRS). Building on this MODIS–VIIRS transition and motivated by the need to provide public users with accurate flood information informed by proven methodological advances, we developed the VIIRS Machine Learning Fractional Water Detection (ML-FWD); a new deep-learning-based global inundation model trained on moderate-resolution Sentinel-2 observations. Using six bands of VIIRS surface reflectance imagery and a custom U-Net Convolutional Neural Network, ML-FWD estimates the fraction of area inundated in each 375 m pixel. We trained the model using an extensive global dataset of more than 95,000 “weak labels” of surface water extracted from Dynamic World, a high-frequency, high-resolution land use land-cover classification derived from Sentinel-2. On a hold-out of over 9,700 unseen surface water instances, ML-FWD achieves an R2 of 0.83 with minimal bias (Mean-Error of-0.6% inundated area) and largely consistent performance across 14 global biomes. In a comprehensive validation encompassing 6.8 million km2 across 100 flood events, ML-FWD outperforms NASA and NOAA’s existing VIIRS water detection models by 0.27 and 0.12 in absolute R2 (a 141% and 35% relative improvement), respectively. In a subset of nine global flood events with coincident Sentinel-2 and Sentinel-1
coverage, ML-FWD even comes close to the accuracy of inundation maps from Sentinel-1 (R2 of 0.79 compared to 0.83 for Sentinel-1 and 0.61 for the next best VIIRS product). Importantly, ML-FWD greatly reduces the likelihood of false-positive water detections with VIIRS, especially in cases of cloud shadows and volcanic material. Public deployment via NASA’s Land Atmosphere Near-real-time Capability for EOS is planned for late 2026, making the upcoming version of NASA’s NRT Global Flood Product (i) the first machine learning based operational flood model with global daily coverage and (ii) the first public flood monitoring service at NASA using machine learning. Our study demonstrates that deep learning and moderate resolution observations greatly improve global surface water extent estimates from coarser-resolution sensors such as VIIRS. This update to NASA’s flood mapping services ensures the public have access to accurate information with daily revisit and short latency, critical for flood monitoring, response, and relief across a broad range of sectors.
 
Acknowledgements
Many thanks to Jonathan Giezendanner (Massachusetts Institute of Technology, previously University of Arizona) who initiated the project and developed the global deep learning framework including the model training data generation. Thanks also to PIs Beth Tellman (University of Wisconsin-Madison, previously University of Arizona) and Fritz Policelli (NASA Goddard Space Flight Center).
 
This research includes contributions from Ph.D. Student Alex Saunders. Learn more about his role and related work on his website: https://alex-saunders00.github.io/global-inundation-viirs/
 

Research Personnel: