Progress on SAR-Based Mapping and Change Detection for Boreal Wetlands of North America

In the face of ongoing anthropogenic global warming, extensive areas of boreal wetland may be evolving into major sources of atmospheric carbon. High-resolution data on the locations, types, and extents of northern wetlands thus constitute an essential prerequisite to the development of accurate global carbon budgets. As a step towards creating these data, we have been constructing high-resolution thematic maps of North American boreal wetlands. Space-based L-Band synthetic aperture radar (SAR) offers efficient, high-resolution visibility into vast, often-inaccessible wetland regions. This capability, along with its inherent sensitivity to vegetation structure, moisture content, and biomass, plus its ability to detect underlying standing water, makes space-based L-band SAR an excellent tool for mapping boreal wetlands. Previously, we used two-season L-band SAR imagery collected from the Japanese Earth Resources Satellite (JERS) in the 1997-1998 time frame to produce a 100 m resolution wetlands map of Alaska. We now present an updated version of the map incorporating improved slope and texture data inputs. Additionally, we have been using dual-polarized L-band SAR imagery collected from the Phased Array L-Band Synthetic Aperture Radar (PALSAR) in the 2007 time frame to form a second 100 m wetlands map of Alaska. We present our results of that mapping effort as well, along with a thematic map of differences in the two maps that reveals changes occurring during the 1997-2007 decade. Finally, we present JERS- and PALSAR-based wetlands maps for certain regions in Canada. We develop our maps using an algorithm based on the "Random Forests" ensemble decision tree classifier. Inputs to the classification include the mentioned SAR imagery plus ancillary information such as texture, image collection dates, slope, and latitude. Our classification approach also requires ground reference data for both training and validation; these are taken from national wetlands and land cover databases as well as from in situ data. We calculate producer and user error statistics for each wetlands map. Averaging results across all regions completed to date, the aggregate classification accuracy for our JERS-based 1997 wetlands map is 88.0% and the aggregate classification accuracy for our PALSAR-based 2007 wetlands map is 88.7%. Comparing the two maps, the most prevalent changes observed so far are emergent wetlands changing into scrub/shrub wetlands; there are, however, also extensive areas in which scrub/shrub wetlands have transitioned to emergent wetlands or scrub/shrub wetlands have changed into forested wetlands. This work was done in part within the ALOS Kyoto & Carbon Initiative, with portions carried out at the University of Michigan and at the Jet Propulsion Laboratory under contract to National Aeronautics and Space Administration. PALSAR data were provided by JAXA/EORC and the Alaska Satellite Facility.