A comparison of techniques for generating forest ownership spatial products

To fully understand forest resources, it is imperative to understand the social context in which the forests exist. A pivotal part of that context is the forest ownership. It is the owners, operating within biophysical and social constraints, who ultimately decide if the land will remain forested, how the resources will be used, and by whom. Forest ownership patterns vary substantially across the United States. These distributions are traditionally represented with tabular statistics that fail to capture the spatial patterns of ownership. Existing spatial products are not sufficient for many strategic-level planning needs because they are not electronically available for large areas (e.g., parcels maps) or do not provide detailed ownership categories (e.g., only depict private versus public ownership). Thiessen polygon, multinomial logit, and classification tree methods were tested for producing a forest ownership spatial dataset across four states with divergent ownership patterns: Alabama, Arizona, Michigan, and Oregon. Over 17,000 sample points with classified forest ownership, collected as part of the USDA Forest Service, Forest Inventory and Analysis (FIA) program, were divided into two datasets, one used as the dependent variable across all of the models and 10 percent of the points were retained for validation across the models. Additional model inputs included a polygon coverage of public lands from the Conservation Biology Institute’s Protected Areas Database (PAD) and data representing human population pressures, road densities, forest characteristics, land cover, and other attributes. The Thiessen polygon approach predicted ownership patterns based on proximity to the sample points in the model dataset and subsequent combining with the PAD ownership data layer. The multinomial logit and classification tree approaches predicted the ownership at the validation points based on the PAD ownership information and data representing human population, road, forest, land cover, and other attributes. The percentage of validation points across the four states correctly predicted ranged from 76.3 to 78.9 among the methods with corresponding weighted kappa values ranging from 0.73 to 0.76. Different methods performed slightly, but statistically significantly, better in different states Overall, the Thiessen polygon method was deemed preferable because: it has a lower bias towards dominant ownership categories; requires fewer inputs; and is simpler to implement.