Information regarding the spatial extent and inundation state in the internationally important Wetlands as designated by Ramsar Convention is important to a series of research questions including wetland ecosystem functioning and services, water management and habitat suitability assessment. This study develops an expedient digital mapping technique using optical remotely sensed imagery of the Landsat Thematic Mapper (TM), ENVISAT ASAR active radar C-band imagery, and topographical indices derived from topographic maps. All data inputs were resampled to a common 30 m resolution grid. An ensemble classifiers based on trees (random forest) procedure was employed to produce a final map of per-grid cell wetland probability map. This study also provides a general approach to delineate the extent of flooding builds upon documented relationships between fields measured inundation state and SAR data response on each vegetation types. The current study indicated that multi-source data (i.e. optical, radar and topography) are useful in the characterization of freshwater marshes and their inundation state. This analysis constitutes a necessary step towards improved herbaceous wetland monitoring and provides ecologists and managers with vital information that is related to ecology and hydrology in a wetland area.
Knowledge of the spatial extent of wetlands is important to a series of research questions and applications such as wetland ecosystem functioning, water management, and habitat suitability assessment. This study develops a practical digital mapping technique using an optical image of a Landsat thematic mapper (TM), Envisat advanced synthetic aperture radar (SAR) image, and topographical indices derived from topographic maps. An ensemble classifier based on classification tree procedure [random forests (RFs)] is applied to three different com- binations of predictors: (1) TM imagery alone (TM-only model); (2) TM imagery plus ancillary topographical data [TM + digital terrain model (DTM)]; and (3) TM imagery, ancillary topographical data and radar imagery (TM + DTM + SAR model). Accuracy assessment results indicate that the radar and topographical variables reduce classification error of marsh. The kappa coefficients for the land cover classification increases significantly as radar imagery and ancillary topographical data are added. The per-grid cell probabilities of each land-cover types are estimated based on the RFs model making use of all available predictors. A final land-cover map is generated by defining pixels as the land-cover type with the highest probabilities. Compared with a single classification and regression tree and a conventional maximum likelihood classifier, RFs produce the highest overall accuracy (72%) with a kappa coefficient of 0.6474, and marsh wetland accuracies ranging from 81.2% to 83.33%. The current study indicates that multisource data (i.e., optical, radar, and topography) are useful in the characterization of freshwater marshes and their adjacent land-cover types. The approach developed in the current study is automated, relatively easy to implement, and could be applicable in other settings over large extents.