The Crisis Mapping team recently wrapped up our research on flood mapping.
Previously, we evaluated existing MODIS flood mapping algorithms across a diverse range of flood conditions. From this study, we realized that all of the existing algorithms will fail for some floods, but rarely do all of the algorithms fail for the same flood. Hence, we investigated how to combine these approaches to create a new algorithm that is more robust and reliable.
To do so, we turned to Adaboost, a common technique for combining multiple classifiers. Our algorithm takes in multiple “weak classifiers” generated from combinations of satellite bands, and combines the weak classifiers with a weighting scheme. The weights for the weak classifiers are learned from non-flood data and a permanent water mask. Information from a Digital Elevation Map (DEM) is then applied in a post-processing step to further improve the generated flood map.
In the above image, the top left segment shows three channels of the raw MODIS satellite data as an RGB image. The top center and top right segments show two different previously developed algorithms, which serve as weak classifiers for Adaboost. The bottom left shows the sum of all Adaboost’s weak classifiers with trained weights. Finally, the bottom center shows the thresholded output flood map from Adaboost, and the bottom right the results after post-processing with the DEM.
We evaluated Adaboost on six flood images and on over 750 non-flood images. Such an extensive evaluation was made possible by the power of Google Earth Engine. The results show that, although Adaboost may be outperformed by another algorithm on an individual flood image, in aggregate Adaboost fails much more rarely than any of the other algorithms and tends to achieve results close to the best algorithm for any given flood. Furthermore, Adaboost does not require any human input, and can easily combine information from multiple data sources, as we demonstrate by fusing MODIS and SAR data, as well as MODIS and Landsat data, to better map floods. The full results of our research are currently under submission to a journal. In the meantime, for those interested, our algorithm and all of the code used for this research is available open source.
What’s next for Crisis Mapping? Working with our partners at Google, we aim to develop a prototype fully automatic flood mapping tool based on the results of this research. Ultimately, such a tool could be used by Google to automatically deliver maps to victims in flooded regions.