Imputing Potential Productivity of Pacific Northwest Forests over Space and Time
Temesgen Hailemariam, Greg Latta, Tara Barrett, Darius Adams
Forest Engineering, Resources and Management, Oregon State University, Corvallis, OR, United States; Forest Engineering, Resources and Management, Oregon State University, Corvallis, OR, United States; Forestry, Pacific North West Research Station, Anchorage, AK, United States; Forest Engineering, Resources and Management, Oregon State University, Corvallis, OR, United States

Site productivity is a critical variable in forest planning and assessments. Yet, forest analyst and planners are faced with missing forest productivity values over space and time. Using climate and productivity data measured on a network of 3356 sample plots, we developed a simultaneous autoregressive (SAR) model to evaluate the impacts of climatic parameters on potential productivity of Pacific Northwest (PNW) forests of the United States, and evaluated the performance of the model against three widely used methods – nearest neighbor (NN), linear regression (MLR) and thin-plate splines (TPS).

Based on Monte Carlo simulation results, the SAR model outperformed the NN, MLR and TPS models in terms of precision, bias, and other fit statistics. By utilizing information regarding spatial error, the SAR model substantially improved the accuracy and precision of potential productivity estimation over the conventional methods that assume the observations are independent. The SAR model also had no evidence of spatial autocorrelation which was significant and visible in the error maps of the other three methods.

Site productivity is explained by the interaction of annual temperature, precipitation, and climate moisture index. Increasing the number of unit responses from 20% to 50% resulted in increased accuracy, with no noticeable improvement with a further increase to 80%. No large gain was noted in using more than one substitute stands. We also discuss the suitability and predictive abilities of SAR and NN methods to impute forest productivity under different error structures.

Keywords: Nearest neighbor imputation; Non-sampled polygons; Mapping climate change; Simultaneous autoregressive model

Biography: Temesgen Hailemariam is Associate Professor in Forest Biometrics and Measurements at Oregon State University (OSU). He holds a Ph.D. degree in Forest Biometrics from University of British Columbia (UBC), M.Sc. degree from Lakehead University, and B.S. degree in Plant Sciences from Alemaya University of Agriculture. His main research areas include developing efficient imputation and sampling methods to assess, monitor, and analyze forest resources. He teaches undergraduate and graduate courses in Forest Biometrics and Measurements at OSU.