Forests are precious multi-use resources with high economic, ecologic, and societal value. Forests not only produce wood - a renewable resource that is increasingly replacing fossil-based materials - but also preserve biodiversity and sequestrate CO2 from the atmosphere. Optimal forest management is therefore crucial for combating climate change and for reaching several of the United Nations' (UN) Sustainability Development Goals. However, determining optimal harvest timings and intensities is one of the oldest - and still unsolved - problems in forestry. Optimizing forest management operations presents a complex, dynamic, discrete-time control problem. Complications arise from discontinuities, nonconvexities, a large number of decision variables, a hybrid action space, and a long planning horizon. Conflicting stakeholder interests and uncertainty - for example in forest growth dynamics, timber prices, currency exchange rates, or natural disasters - further complicate the problem. Existing forestry optimization methods need to either simplify the problem to remain feasible or they require days or even weeks to find an approximate solution. This leads to sub-optimal forest management decisions, which in turn lead to economic losses and unnecessary environmental destruction. Against this backdrop, this doctoral dissertation contributes novel methods and insights on optimal and sustainable forest management by applying AI-based optimization techniques that have not been previously used in economic forest research. In Article I, we use multi-objective evolutionary algorithms to compute and evaluate multi-objective forestry strategies, without the need for policy makers to assign preferences a priori. Article II marks a methodological shift and explores the necessary preconditions for successful real-world application of reinforcement learning. In Article III, we then use reinforcement learning to solve a high-dimensional optimal harvesting problem that correctly includes stochasticity in forest growth and in the occurrence of natural disasters. Our method is the first to simultaneously consider both clear-cutting and continuous cover forest management, and to calculate near-optimal harvesting schedules purely based on the long-term goals of forest owners. We find that multi-species continuous cover forestry is often more profitable and sustainable than current single-species clear-cut practices; especially when including the risk of natural disasters. Moreover, our work helps to navigate conflicting goals (economic profit vs. carbon storage vs. biodiversity). Finally, we establishes forest management as a multi-disciplinary research area by bridging economic forest research with AI research. In summary, this thesis contributes novel methods and practical insights on optimal and sustainable forest management, with far-reaching implications for forest owners, policy makers, asset managers, ESG investors, and for reaching several UN Sustainability Development Goals.
|Translated title of the contribution||AI for Optimal and Sustainable Forest Management|
|Publication status||Published - 2022|
|MoE publication type||G5 Doctoral dissertation (article)|
- forest management
- evolutionary computation
- reinforcement learning