Decadal Change Characterization in Northern Wetlands Based on Analysis of L-band SAR Satellite Data

Northern wetlands are believed to have hitherto served as carbon sinks, sequestering about one third of the total global pool of soil carbon. The warmer, drier conditions occurring throughout the Arctic as a consequence of global warming may now be causing them to evolve from carbon sinks into major sources of greenhouse gases. The ability to characterize long-term changes in the condition of northern wetlands is therefore essential to the development of accurate global carbon budgets. L-band synthetic aperture radar (SAR) offers a unique tool for monitoring changes in the characteristics of vegetated wetlands. It is sensitive to vegetation structure, biomass, and moisture content, and can penetrate vegetation canopies to detect standing water underneath. We have been using L-band SAR imagery from two different spaceborne sensors separated in time by approximately one decade, JERS and PALSAR, to produce a thematic map of change in the type and extent of wetlands in Alaska. Summer and winter JERS imagery characterizes the wetlands status for the 1997 time frame while dual-polarized PALSAR imagery captures the wetlands status for the summer of 2007 time frame. To produce each classified wetlands map, the SAR imagery is supplemented with ancillary information derived from the SAR and other sources. The classification algorithm applied to each set of imagery is based upon the Random Forests technique, by using a multitude of decision trees. To ensure that a sufficiently wide spectrum of ground reference points are included in each map segment to be able to develop a representative set of decision trees, we classify over large geographic regions. Since our PALSAR imagery is provided at resolutions 3-8 times that of our JERS imagery, classifying it requires much more computer memory than does classifying JERS for the same size region. We have investigated resampling strategies to address computational limitations, and have found it necessary to average the PALSAR imagery to the same resolution as JERS. We have also investigated 1) adding the ability to apply weights to the different data layers, and 2) adding the ability to base classifications on a combination of ground reference pixels available locally and a saved decision tree forest from a more remote location. The latter feature would be used when local ground reference data is sparse but not entirely absent. We are also exploring options for segmenting the data prior to classification. The accuracy of the resulting thematic change map is verified using ground reference data. The results are expected to demonstrate the utility of multi-platform satellite L-band SAR observations for characterizing transitions in the extent and type of vegetated wetlands as a result of climate change. This work is part of the Global Inundated Wetlands Mapping MEASURES project and is being carried out at The University of Michigan, and at the Jet Propulsion Laboratory, California Institute of Technology, under contract to National Aeronautics and Space Administration. It has been undertaken in part within the framework of the JAXA ALOS Kyoto & Carbon Initiative. PALSAR data have been provided by JAXA EORC.