Abstract
Evidence of the rapid losses of coastal marshes calls for the application of remote sensing data. Nonetheless, many features indicative of incipient marsh loss, such as widening of tidal creeks and formation of small ponds, are often not readily detectable at the nominal 30 m×30 m resolution of Thematic Mapper (TM) imagery, the general source for conventional satellite sensor-based data on wetlands. Spectral mixture modelling, where the proportional representation of land cover types can be estimated within pixels, offers a potential solution to the problem of assessing initial indications of loss in coastal marshes. Nevertheless, the simple linear mixture models most commonly employed can be subject to significant errors when applied to marshes due to the considerable variety of soil/sediment types in these environments. A new method is presented here which not only successfully reduces the spectral variability of soils, but also in the other principal components of vegetation and water, to a three-endmember model for estimating their fractional representation in TM data. Fractional representations for these variables in this method yield more reliable results than those obtained from unmixing corrected reflectance data. Though unmixing individual scenes may still be best achieved by extensive ground-referencing and image analysis, this technique is a robust approach for large-scale, semi-automated processing of many scenes in investigations of marsh surface condition.
Acknowledgments
This research was funded by a NASA Regional Application Center grant no. NAGW3758 under the auspices of the Mission To Planet Earth Program, Mr Alex Tuyahov, programme manager. The authors also wish to thank Drs Shun-lin Liang and Satya Kalluri for critical discussions of various aspects of the study.