A classification method to map wetlands in Quebec for the Canadian Wetland Inventory using a top-down approach with object-oriented segmentation
2005
Canadian Wetland Inventory (CWI) started in spring 2002 and was first aimed to develop a method to map 5 wetland classes stratification, based on the Canadian Wetland Classification System, with a minimal mapping unit of 1 hectare. The 5 wetlands classes for the CWI are : marsh, swamp, shallow water, fen and bog. Satellite image types used for CWI are RADARSAT-1 and Landsat-ETM. The Landsat-ETM images are acquired, when possible, in summer at a time when vegetation exhibit high level of activity and Radarsat images must be acquired during spring flood or during a period of high water in fall. The mapping method, developed by the Canadian Wildlife Service (CWS) Quebec region, applies spectral, spatial and texture classification of satellite images using a top-down approach (coarser to the finer spatial objects) with a segmentation algorithm. The classification is accomplish with the eCognition© software. The method was tested on 5 study sites : Radisson, lac Saint-Pierre, lac Saint-Francois (Appalachian), Grande-Plee-Bleue and Isle Verte. Theses sites provide a wide range of different wetland classes in varied environment which helped test the mapping method before applying it to large areas. Following the top-down approach requires that spatial objects in the images be identified starting at the coarser level, and gradually to finer levels, according to the size of the objects of interest : wetlands, water bodies, mineral surfaces, forest, etc. Generally 3 segmentation levels are required: The coarse levels allow toidentify large spatial ensembles. For example, large wetlands, water bodies, mineral surfaces, and « Potential Wetlands » are typically identified at that level. The medium level is the refinement of the coarse level, primarily on « Potential Wetlands » class. This level also allows to identify medium size wetlands. Lastly, the fine level is the refinement of the medium level, composed exclusively of « Potential Wetlands - Medium » class. The smallest wetlands are identified at this level, but always keeping in mind the spatial objects must be larger than 1 ha to be compatible with the CWI minimum mapping unit. Hierarchical classification optimizes wetland identification and reduces multi-class confusion. Sets of rules are assigned to each class based on membership functions using spatial and spectral attributes such as mean, ratio, texture, shape, neighbourhood, etc. Misclassified objects are reassigned manually after the automated classification which requires approximately 15-20 % of processing time. This time may be reduced with better contextual knowledge of landscape or with use of more specific membership functions. The spatial and temporal variability of wetland lead to a validation strategy using the best knowledge' approach instead of comparing the final classification with other map products or ground-based plots. 'Best knowledge' approach involves a qualitative confidence-building assessment where an expert interprets the original image enhanced, and when possible, with additional available datasets.
Keywords:
- Correction
- Source
- Cite
- Save
- Machine Reading By IdeaReader
0
References
0
Citations
NaN
KQI