Process_Description:
Object based image analysis (OBIA) using a rule-based expert system developed in eCognition 9.4 (Trimble). In a preliminary modeling step, developed features from the 2013 land-cover layer for Pennsylvania (Buildings, Roads, Impervious Surfaces) were burned directly into a draft map, eliminating these areas from further consideration. Water was also incorporated from the available land-cover data to ensure smooth transitions between wetlands and adjacent open-water features (e.g., lakes, ponds, rivers). Active agricultural fields were then estimated using vegetation height as derived from the available DEM and similarly set aside. All remaining areas were segmented into objects representing actual landscape features using a Compound Topographic Index (CTI) developed from the 2006-2008 LiDAR. This CTI layer incorporated topography and hydrological flow potential. In a last preliminary step, objects with topographic characteristics highly unlikely to support wetlands (e.g., steep slopes) were removed. reducing the number of objects to be evaluated in subsequent modeling.
In an initial classification step, candidate wetland objects were selected using CTI texture as calculated for each object using an edge extraction function in eCognition. CTI texture better captures transitions between wetlands and upland features compared to original CTI values. Additional classification steps included selection criteria with the spectral characteristics of leaf-off orthoimagery and 10-m statistical models summarizing climate, soils, and topography. Following segmentation at a finer scale, which created smaller objects, the classification was refined with error-checking routines that removed false positives (e.g., forested objects with bright spectral characteristics in leaf-off imagery, indicating an upland site). The shape and contiguity of identified wetland objects was then improved with gap-filling and smoothing routines.
For the final classification, the initial wetlands class was segmented by vegetation height and then categorized by primary wetland type: Emergent, Scrub\Scrub, and Forested. It was also generalized to a minumum mapping unit of 0.1 acres by consolidating small objects into adjacent larger ones, making the map comparable to existing classification protocols. The final map, including water as delineated by the 2013 statewide land-cover layer, was exported in vector format.