Flood Mapping from Satellite Imagery

September 18 2014

We’ve recently begun a new collaboration, the Crisis Mapping project, with the Google Crisis Response team. Our goal is to develop tools to aid victims in a crisis through the use of map imagery processed by Google Earth Engine.

Google Earth Engine is a new tool for parallel computation on satellite imagery, enabling data processing on enormous scales using Google’s cloud infrastructure. Earth Engine takes care of tiling, georeferencing, and varying image resolutions automatically, letting researchers focus on the core aspects of their algorithms. Many popular georeferenced datasets are already loaded into Earth Engine for quick and easy use, such as 40 years of Landsat data, making historical data easily accessible. Earth Engine also provides an interactive sandbox for quickly prototyping image processing algorithms, and lazily computes results only for the parts of the image which are visible.

As a first step in the Crisis Mapping project, we decided to investigate flood mapping from satellite imagery. With flood maps, responders can know the extent of a flood and plan a more effective response.


We used data from MODIS (upper image, bands 1, 2 and 6 as RGB), which has the advantage of capturing an image of nearly every point on the Earth’s surface daily, making it suitable for fast responses to flooding. However, MODIS is blocked by clouds, which are often present in flooded areas. In the future we plan to investigate radar-based approaches to address this shortcoming. The MODIS data is coupled with images from Landsat (lower image, same region as the MODIS image), from which we derive ground truth data both for training and to evaluate the results of the algorithms on MODIS data. Landsat data is available infrequently and is thus less suitable for flood mapping.

We compared a number of existing flood mapping approaches in a trade study:

  • Thresholding Approaches: Threshold four combinations of MODIS bands and indices, including the difference of bands 1 and 2, a ratio of bands 1 and 2 (used by the Dartmouth Flood Observatory), and two other combinations of MODIS indices [1] [2].
  • Earth Engine Classifiers: CART, SVM and Random Forest classifiers, built into Earth Engine, are learned from training data with the features proposed in [3].
  • Dynamic Nearest Neighbor Search (DNNS) [4]: Find a “water fraction” of how much each pixel is flooded based on nearby land and water pixels.
  • DNNS with Digital Elevation Map (DEM) [5]: Given water fractions from DNNS, estimate the water height and flood cells in the higher resolution DEM.
  • Historical Difference Thresholding: Threshold the difference of two bands, but learn the appropriate threshold from historical data.


We evaluated all the algorithms on five test regions: two floods along the Mississippi from different months, a flood in Pakistan (shown above), New Orleans after Katrina, and the San Francisco Bay Area as an unflooded test case.

The trade study showed that none of the approaches was always the “best”, but all had tradeoffs. The two thresholding approaches from the literature missed many flooded areas but had few mislabelings, while the DNNS with DEM approach missed the fewest flooded pixels but had more mislabelings. Overall, the difference threshold, ratio threshold, CART, and historical approaches seemed to have the best balance between both high precision and high recall.


Above: Detected flooded pixels in New Orleans after Katrina according to the Dartmouth Flood Observatory method, displayed atop a Landsat image.

In terms of a more qualitative analysis, the thresholding approaches had the advantage of not requiring any training, while the Earth Engine classifiers required manual annotation of nearby training regions. The DEM approach had the advantage of being able to form higher resolution maps based on the DEM. DNNS with DEM fared particularly poorly near levees since it did not consider their effect on water flow.

As a next step, we are looking into using radar data to map floods since it is still effective in the presence of cloud cover. After further research, we intend to publish the full results and methodology of the study in a peer-reviewed publication.


[1] Islam, Bala, and Haque. “Flood Inundation Map of Bangladesh Using Modis Surface Reflectance Data”, International Conference on Water and Flood Management, 2009.
[2] Xiao, Boles, Frolking, et. al. “Mapping Paddy Rice Agriculture in South and Southeast Asia Using Multi-Temporal MODIS Images”, Remote Sensing of Environment, 2006.
[3] Sun, Yu, and Goldberg. “Deriving Water Fraction and Flood Maps from MODIS Images Using a Decision Tree Approach”, Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2011.
[4] Li, Sun, Yu, et. al. “A New Short-Wave Infrared (SWIR) Method for Quantitative Water Fraction Derivation and Evaluation with EOS/MODIS and Landsat/TM Data.” IEEE Transactions on Geoscience and Remote Sensing, 2013.
[5] Li, Sun, Goldberg, and Stefanidis. “Derivation of 30-m-Resolution Water Maps from TERRA/MODIS and SRTM.” Remote Sensing of Environment, 2013.

This entry was posted on Thursday, September 18th, 2014 at 6:31 pm and is filed under Mapmakers. You can follow any responses to this entry through the RSS 2.0 feed. Post a comment or leave a trackback: Trackback URI.

Post a Comment

Your email is never published nor shared. Required fields are marked *



  • Measuring Lake Tahoe from Space

    This fall, the crisis mapping team extended their work on flood mapping to measure the water level of Lake Tahoe over time ... more

  • Flood Mapping with Adaboost

    The Crisis Mapping team recently wrapped up our research on flood mapping. Previously, we evaluated existing MODIS flood mapping algorithms across a diverse range ... more

  • Crisis Mapping Toolkit Released

    We are excited to announce the open source release of the Crisis Mapping Toolkit under the Apache 2.0 license! The ... more