Dynamic stand structure optimization of secondary Pinus yunnanensis forests integrating two-period structural prediction and harvesting-replanting coordination

In the sustainable management of secondary forests, existing stand structure optimization methods often treat harvesting and replanting as sequentially independent operations rather than spatially coupled decisions. Conventional approaches, including genetic algorithms and simulated annealing, typically optimize harvesting first based on current stand state and then add replanting using simple gap-filling rules. This sequential decomposition ignores two critical interdependencies: replanting suitability depends on the spatial pattern of retained trees after harvesting, and the long-term structural benefits of replanting depend on how future competition with existing trees is anticipated. These limitations arise because traditional single-stage optimization models are designed for a single decision variable set and cannot simultaneously solve the combinatorial problem of which trees to remove and where to add new trees within a unified framework. To address these limitations, this study focuses on secondary Pinus yunnanensis forests on Cangshan Mountain in Dali and proposes a coordinated dynamic optimization framework based on MAPPO-BO for harvesting and replanting management. In terms of replanting optimization, a slope-weighted replanting feasibility index (SRFI) was developed by incorporating slope weighting, and replanting locations were dynamically determined in combination with stand density constraints for different tree species. For collaborative decision-making, a multi-agent deep reinforcement learning algorithm integrating Multi-Agent Proximal Policy Optimization with Bayesian hyperparameter optimization (MAPPO-BO) was employed to achieve coordinated optimization through parallel decision-making and collaborative learning between harvesting and replanting agents. Furthermore, a stand structure prediction model based on two-period inventory data was introduced to establish a closed-loop feedback mechanism of optimization-prediction-reoptimization, enabling the dynamic adjustment of stand structure over time. Experimental results showed that the MAPPO-BO framework improved the objective function by an average of 49.29% across five plots, outperforming MADQN (47.17%). The SRFI effectively guided replanting activities toward areas with relatively gentle topography. After dynamic optimization, the improvement in the objective function across three plots ranged from 44.43% to 64.69%, while multiple stand spatial structure indicators continuously improved, leading the stands to gradually approach the characteristics of natural forests. These findings indicate that the proposed framework can effectively enhance the efficiency of coordinated harvesting and replanting decisions, thereby providing a new theoretical framework and methodological support for the precise management of secondary Pinus yunnanensis forests.