How wetland type and area differ through scale: A GEOBIA case study in Alberta's Boreal Plains

It is estimated that Canada comprises approximately 28% of the world's wetlands (~ 1.5 million km2) providing essential ecological services such as purifying water, nutrient cycling, reducing flooding, recharging ground water supplies, and protecting shorelines. In order to better understand how wetland type and area differ over a range of spatial and thematic scales, this paper introduces a multi-scale geographic object-based image analysis (GEOBIA) approach that incorporates new object-based texture measures (geotex) and a decision-tree classifier (See5), to assess wetland differences through five common spatial resolutions (5, 10, 15, 20 and 30 m) and two different thematic classification schemes. Themes are based on (i) a Ducks Unlimited (DU: 15 class) regional classification system for wetlands in the Boreal Plain Ecosystem and (ii) the Canadian Wetland Inventory (CWI: 5 classes). Results reveal that the highest overall accuracies (67.9% and 82.2%) were achieved at the 10 m spatial resolution for both the DU and CWI classification schemes respectively. It was also found that the DU wetland types experienced greater area differences through scale with the largest differences for both classification schemes occurring in classes with a large treed component. Results further show that the inclusion of geotex channels (generated from dynamically sized and shaped window that measures the spatial variability of the wetland components composing a scene, rather than of individual pixels within a fixed sized window) improved wetland classification.

► Overall accuracy for the CWI classification was comparable to other studies.
► Overall accuracy for the DU classification was only appropriate for the 10 m scale.
► Highest overall accuracy for both classification schemes was achieved at 10 m.
► Greater area difference for DU wetland types through scale.
► Wetlands with a large treed component experienced the largest area differences.
► Texture information was shown to be useful for aiding wetland classification.