Monitoring flooding and vegetation on seasonally inundated floodplains with multifrequency polarimetric synthetic aperture radar

The ability of synthetic aperture radar to detect flooding and vegetation structure was evaluated for three seasonally inundated floodplain sites supporting a broad variety of wetland and upland vegetation types: two reaches of the Solimoes floodplain in the central Amazon, and the Magela Creek floodplain in Northern Territory, Australia. For each site, C- and L-band polarimetric Shuttle Imaging Radar-C (SIR-C) data was obtained at both high- and low-water stages. Inundation status and vegetation structure were documented simultaneous with the SIR-C acquisitions using low-altitude videography and ground measurements. SIR-C images were classified into cover states defined by vegetation physiognomy and presence of standing water, using a decision-tree model with backscattering coefficients at HH, VV, and HV polarizations as input variables. Classification accuracy was assessed using user's accuracy, producer's accuracy, and kappa coefficient for a test population of pixels.; At all sites, both C- and L-band were necessary to accurately classify cover types with two dates. HH polarization was most. useful for distinguishing flooded from non-flooded vegetation (C-HH for macrophyte versus pasture, L-HH for flooded versus non-flooded forest), and cross-polarized L-band data provided the best separation between woody and non-woody vegetation. Increases in L-HH backscattering due to flooding were on the order of 3--4 dB for closed-canopy varzea and igapo forest, and 4--7 dB, for open Melaleuca woodland.; The broad range of physiognomies and stand structures found in both herbaceous and woody wetland communities, combined with the variation in the amount of emergent canopy caused by water level fluctuations and phenologic changes, resulted in a large range in backscattering characteristics of wetland communities both within and between sites. High accuracies cannot be achieved for these communities using single-date, single-band, single-polarization data, particularly in the case of distinguishing flooded macrophyte from non-flooded forest vegetation. However, the large changes in backscattering caused by flooding make it possible to achieve good accuracies (>85%) using multi-temporal data. Where river stage records are available, SAR-based maps of inundation status on a series of dates can be linked to long-term stage data to define wetland habitat types based on flooding regime and low-water vegetation cover.