Maria Uriarte and Ben Braunheim. Columbia University
Silviculture in North America has seen a dramatic shift away from even-aged management and clear-cutting to partial harvesting and all-aged management. Landowners and forest managers must strive for what are still perceived as conflicting objectives: removal of forest products while maintaining the structural complexity and diversity of natural forests. Devising harvesting recommendations that maximize multiple goals requires the use of data-based simulation models of forest dynamics, which have the ability to capture the spatial patchiness of harvested stands. However, it is often difficult to make recommendations for management by evaluating a limited number of interventions because of the large number of parameters that need to be considered and the highly non-linear interactions between individual trees and processes. We present a new approach to meet these challenges that couples a spatially-explicit forest simulator with machine learning algorithms. The inputs to the simulator were periodicity of harvest, radius of harvest area, number of harvest sites, and the percent of trees to be harvested for each of the nine species included in the model. We calculated structural complexity by comparing the probability density function for a stand that had been harvested to the PDF for a stand that had not been harvested, over the period of time of interest to management. Using this approach, we could quickly screen over 100,000 random harvesting regimes. We present recommendations resulting from this approach that simultaneously maximize multiple goals of forest management, namely timber yield, biodiversity, and structural complexity, in a 9-species mixed northeastern US forest. We can accurately classify harvesting regimes that represent a 75-80% improvement in timber yields while achieving a 50% improvement in stand structural complexity (dbh distributions relative to unharvested stands for the 9 species) relative to random harvest regimes.