Proximal policy optimization integrated with Bayesian optimization for stand structure optimization in Pinus...

Area of Science:

• Forestry Science
• Computational Ecology
• Artificial Intelligence in Ecology

Background:

• Secondary forests often exhibit suboptimal stand structures and spatial arrangements, hindering their ecological functions.
• Current deep learning methods for forest stand optimization face challenges in efficiency, accuracy (Q-value overestimation), and site-specific felling constraints.

Purpose of the Study:

• To develop an advanced stand structure optimization model for secondary *Pinus yunnanensis* forests.
• To enhance computational efficiency and accuracy in forest management decision-making.
• To integrate site-specific factors like slope into felling constraints for ecological suitability.

Main Methods:

• Developed a stand structure optimization model using individual trees as decision units, based on field survey data.
• Implemented the Proximal Policy Optimization (PPO) algorithm, integrated with Bayesian Optimization (BO) for hyperparameter tuning (PPO-BO framework).
• Incorporated slope constraints (excluding trees on slopes >45°) and evaluated six optimization schemes.

Main Results:

• The PPO-BO algorithm yielded higher maximum objective function values (up to 19% improvement) compared to the Deep Q-Network (DQN) algorithm.
• The PPO-BO approach reduced overall training costs and improved computational efficiency.
• Selective felling based on the model significantly enhanced spatial structure indices, making the stand structure more akin to natural forests.

Conclusions:

• The PPO-BO framework offers a more efficient and accurate method for optimizing secondary forest stand structures.
• This study contributes novel theoretical and methodological insights for sustainable forest management practices.
• The integration of AI algorithms with ecological constraints advances the field of computational forestry.