The Washington State Department of Natural Resources (DNR) manages over 800,000 hectares of forested state trust lands and 20,000 kilometers of forest roads in Washington State. Forest harvest and road reconstruction decisions greatly impact the agency’s cash flows and its ability to meet its fiduciary obligations. We introduce a mixed-integer programming model that integrates harvest and road scheduling decisions. We show how DNR embedded the new model in its workflows and applied it to the Upper Clearwater River Landscape in the Olympic Experimental State Forest. We find that the forest valuation of the Upper Clearwater increased by $0.5–$1 million (0.4–1.1 percent) because of the new method, which allowed the DNR to concentrate capital expenditures in support of harvest and road operations in both time and space. This led to a 14.5 percent reduction in the size of the active road network. DNR is now in the process of scaling the new approach to the entire forest estate.

Long time horizons, typical of forest management, make planning more difficult due to added exposure to climate uncertainty. Current methods for stochastic programming limit the incorporation of climate uncertainty in forest management planning. To account for climate uncertainty in forest harvest scheduling, we discretize the potential distribution of forest growth under different climate scenarios and solve the resulting stochastic mixed integer program. Increasing the number of scenarios allows for a better approximation of the entire probability space of future forest growth but at a computational expense. To address this shortcoming, we propose a new heuristic algorithm designed to work well with multistage stochastic harvest-scheduling problems. Starting from the root-node of the scenario tree that represents the discretized probability space, our progressive hedging algorithm sequentially fixes the values of decision variables associated with scenarios that share the same path up to a given node. Once all variables from a node are fixed, the problem can be decomposed into subproblems that can be solved independently. We tested the algorithm performance on six forests considering different numbers of scenarios. The results showed that our algorithm performed well when the number of scenarios was large.