NEARC Fall 2016

AUTHORS: Sean MacFaden, Jarlath O'Neil-Dunne - University of Vermont Spatial Analysis Laboratory

ABSTRACT: Automated feature extraction of small, irregularly-distributed landscape features has long been a challenge for remote sensing-based land-cover mapping. Natural features such as vernal pools and wetlands may be obscured by vegetation or easily confused with other landscape elements, complicating feature discrimination and reducing classification accuracy. Anthropogenic elements such as previously-unmapped logging roads are similarly difficult to identify and then isolate from other linear features. The growing availability of high-resolution imagery and LiDAR is changing this dynamic, however, especially when these datasets are combined with object-based image analysis (OBIA) techniques. In a series of projects in Vermont and other northeastern states, we developed OBIA modeling routines that improved capture of these difficult-to-map features, using LiDAR-derived digital elevation models (DEMs) to first identify candidate features and then evaluating them with a combination of object characteristics and contextual criteria. The high-resolution DEMs were essential to mapping workflows, permitting characterization of fine-scale landscape transitions with compound topographic indices (wetlands), slope-derived depressions (vernal pools), and geomorphometric indices such as landform and dissection (logging roads). Overall, modeling focused on over-prediction (errors of commission are much harder to diagnose) and was efficiently performed on large geographic extents (county-sized) using the enterprise processing capabilities of OBIA software.