OOS 39-3
Integrating image segmentation, LIDAR, and modeling for high resolution, detailed and accurate vegetation maps of Sonoma County and western national parks
Sonoma County requires a fine scale vegetation and habitat map for conservation planning, management for watershed protection, fire and fuels management, and for assessing climate mitigation and adaptation strategies. Existing vegetation map products derived from Landsat TM data are too coarse for parcel-scale management. The National Vegetation Classification (NVC) provides a hierarchical classification system that is well suited for fine scale vegetation and habitat mapping. Because of the large size of Sonoma County, a semi-automated mapping approach was chosen over a wholly manual one to improve map consistency, add detail, and reduce map cost. Using very high resolution imagery and high-density LiDAR, semi-automated methods will be used to map Sonoma County in two phases: first to general, broad “lifefeform” classes, and then to the more detailed alliance level of the NVC. Alliance level plot data is critical to the approach, both for training machine learning classifiers and for accuracy assessment.
Results/Conclusions
The first phase of the mapping, lifeform mapping, relies on automated image segmentation/classification in object oriented image analysis software to create polygons and label the polygons with lifeform labels. Rule sets are developed heuristically and used to create image objects and label them. Important datasets for segmentation and rule set development are 6 inch resolution, 4- band imagery, LiDAR derived ground and canopy height models, LiDAR derived canopy density profiles, thematic layers, and various hyperspectral indices (AVIRIS). Accuracy of the lifeform map is evaluated by spatially comparing plot data with the lifeform map. The second phase of mapping relies on machine learning techniques (CART, Random Forests), to refine lifeform to NVC alliance. For Sonoma County, alliance level mapping will take place in 2015, so we will discuss results from vegetation mapping of the Grand Canyon and the National Parks of Hawaii and American Samoa. Our discussion will focus on practical lessons learned: the role of NVC field data in the mapping workflow, the independent variables that we’ve found consistently to be most important in machine learning, the limits of machine learning, the importance of manual editing, and the benefits and challenges of mapping to the NVC.