Skip to main content
SLU publication database (SLUpub)

Conference paper2005

Impact of climate change uncertainty on optimal forest management policies at stand level

Backeus Sofia, Eriksson Ljusk Ola, Garcia Frédérick

Abstract

Forest management is classically modeled as a deterministic planning problem where decisions are determined in advance unconditionally to future events. This approach has led to the development of efficient optimization methods based on linear programming which can solve large forest management problems as encountered in Scandinavian countries. Recent research results have nevertheless established that it could be essential from an economical point of view to consider stochastic phenomena in the definition of long term forest management problems such as natural hazards and market uncertainties. Climate change will have important effects on forest ecosystems in the long run. It is a new fundamental reason to go beyond deterministic planning approaches in forest management. According to the last scenarios published by IPCC, the average temperature increase in the world might be in the range of 1.4 to 5.8° by year 2100. At regional level, a model developed by SWECLIM for Sweden predicts an increase between 2.5°C to 4.5°C by 2100, but there is also a considerable uncertainty regarding the future climate trajectory. In the study presented in this paper, we investigated the impact of such climate change uncertainty on the solutions of forest management problems for typical Swedish stands with several species. Our main objective was to determine whether taking into account uncertainty can improve the best deterministic solutions obtained by considering average temperature change scenarios. Such stochastic planning problems are theoretically solved by dynamic programming or related methods. However, these approaches can rarely be employed when complex growth and yield models are used. The second methodological objective of this study was thus to assess the effectiveness of new alternative stochastic simulation methods like simulationbased optimization or reinforcement learning for solving forest management planning problems with stochastic features. In the work described in this paper, we analyzed approximate optimal policies obtained for forest management problems under climate variability and change. These problems were defined on the basis of a forest management model called GAYA and a climate change model developed for the purpose of this study. Approximate optimal policies were obtained with Linear-Q-learning(), a reinforcement learning algorithm. They were compared with approximate optimal plans obtained by considering average climate change scenarios. We studied fictive stands with one, two and three species. The stands were located in southern and northern Sweden and contained the species pine, spruce and birch. The stands were simulated for 20 five-year long periods years. Reinforcement learning converged toward approximate optimal policies after few simulations, even for complex stands with several species, either for problems with deterministic or stochastic scenarios. Ours results showed that there was a gain in considering stochastic models, but also that the magnitude of this gain was small. Considering climate change uncertainty improved the value of approximate optimal programs, but deterministic plans was sometime still optimal. This important conclusion can be explained by the relatively small estimated value of the growth effect, that lead to a quasi-deterministic dynamics of the stand state, and by the symmetry of the temperature change distribution around the average scenario. Our analysis lies entirely on the stochastic climate change model we developed, based on the SWECLIM regional climate modeling for northern Europe. Its main limitation is its stationary assumption on probabilities of the future climate change scenarios. Modeling today the probable fall of uncertainty in the future may be a complicated task. Note however that such a decreasing uncertainty model should strengthen our present conclusions

Published in


Publisher: Modelling and Simulation Society of Australia and New Zealand Inc

Conference

MODSIM05