Technical Description, DFO-GSFC Surface Water Mapping Algorithm
Updated August 20, 2012
G. Robert Brakenridge
Director, Dartmouth Flood Observatory
Senior Research Scientist
CSDMS, INSTAAR, University of Colorado
Campus Box 450
Boulder, CO 80309-0450 USA
DFO water mapping using NASA's two MODIS sensors has evolved since Terra MODIS "first light" in late 1999, and our earlier flood mapping using this sensor. The results are, however, very similar. Most changes concern how clouds and cloud shadows are removed. Currently we use the NASA LANCE-MODIS processor. Earlier, we used the MODIS Rapid Response Subset data, and before this, swath data while performing our own "bow tie correction" and geocoding. Use of the LANCE-MODIS 10 deg x 10 deg mosaic scenes, already in Geotiff format, improves the ability to generate new maps very quickly. In late March, 2011, a new format for the Surface Water Record has also been initiated, with all map sheets covering 10 degree lat. by 10 degree long. land areas and in an unprojected (simple latitude and longitude, Plate Carree) coordinate system.
A cooperative research project funded by the NASA Earth Science Applications Program in 2010 incorporated DFO MODIS flood mapping methods into an automated processor. Fritz Policelli at NASA Goddard Spaceflight Center (GSFC) was project Principal Investigator. Project personnel included J. Sun, D. Slayback, D. Ouzounov, Science Systems and Applications, Inc. at GSFC, and, at DFO, also A. Kettner. Jeff Schmalz of the MODIS Rapid Response team was instrumental early-on in facilitating the utilization of the Rapid Response Subset data for flood mapping. NASA support of the LANCE-MODIS project now provides even more recent MODIS data (latency < 3 hours), and the source MODIS band 1 and 2 data are digitized to 12 bit rather than 8 bit formats. Cooperative work between D. Slayback and A. Kettner, in particular, has now provided a near real time MODIS data transformation from satellite acquisition to GIS (.shp file) surface water polygons.
Three products are now or will soon be available with global coverage:
1) "NASA NRT Experimental Flood Maps". These are near real-time (NRT) maps of current flood water, produced within a few hours of the Aqua satellite's overpass (at approximately 13:30, depending on latitude). One portal to these raster image map products is: http://floodobservatory.colorado.edu/. Click on the global map there. A NASA portal is also available.
2) The GIS data behind the NRT maps are also transferred to an archive at the Flood Observatory as they are produced. These data may be accessed here: http://csdms.colorado.edu/pub/flood_observatory/MODISlance/ The data are organized into 10 deg long x 10 deg lat folders and by Julian date.
3) DFO's Surface Water Record, Version 2, is available from clickable map tiles at the DFO home page. Click on "Surface Water Record" on each map page. also makes use of these NRT MODIS surface water files. As compared to the Experimental Flood Maps, the Record map sheets are larger and show much more detail, and include the DFO historical flood mapping data. Unlike the Flood Maps, there is no cloud mask. Instead, the Record displays accumulate NRT GIS output over 10 days in order to provide maximum geographic coverage and removal of cloud obscuration. Commencing in 2012, DFO also began providing access to Geotif and kmz format versions of these large-format displays.
Water Classification Procedures
Water classification is accomplished by a ratio approach (band2 + A)/(band1+ B), with constants A and B determined empirically. Below a numerical threshold, all pixels are classified as water and all above as land. Threshold values have ranged, as experience was gained with this algorithm, from 0.6 to 0.9. Numerical thresholds in band 1 and 7 data can also be used to remove clouds and partially shadowed clouds.
Cloud shadows on land and terrain shadows may exhibit spectral signatures in the MODIS bands used that are nearly identical to that of clear water. To remove cloud shadows, our approach evolved with experience: from a 4 scene set including two days and two scenes per day (Terra, morning, and Aqua, afternoon), to also employing a 3 day, 6 scene set in regions where such noise is a problem. In this approach, cloud shadows are removed by requiring 3 "water" classifications for each pixel to be accepted as water in the final product. Cloud shadows are largely removed, because their positions change over a three day interval. Only three, in any order, of the 6 images need show water: intermittent cloud cover is partially removed also by this approach, which fills coverage between clouds, and while at the same time removing noise caused by cloud shadows. Terrain shadows also move from morning to afternoon and may be partially removed, but some deeply shadowed land areas in mountainous terrain remain misclassified as water.
Data Product Notes
The NRT GIS vectors, fit around each final water classification product, show in their file names single dates, but currently represent a result from four scenes, over two days. Work is underway to reduce cloud shadows.
For data older than 10 days, the color of the "flood" layer changes from red to light red. Flood extents from all previous years are shown in light blue. This method allows: 1) the most recent data obtained to be illustrated and also made available in GIS formats, and 2) the extents of flooding for each year to be accumulated and integrated into the long term archive of surface water variability.
A drought product is being prepared for each map sheet. Instead of showing current and past surface water extension, it will show current and past surface water contractions. We are working to provide a map display of maximum surface water contraction (maximum observed drought conditions) and to allow a similar comparison of current condtions to the long term, so-far-observed range of variability.
Flood Observatory Homepage