Timber Harvest Planning Using Reinforcement Learning: A Feasibility Study

Abstract

This study developed a forest management plan model using reinforcement learning (Q-learning) to optimize both the economic and ecological functions of forests. Management objectives for national forests were established, and forest conditions were analyzed using GIS spatial data and administrative records. A 60-year forest management plan was formulated to predict timber production and management performance across different regions and time periods. Our analysis revealed that Scenario 3 (Carbon Storage Priority) demonstrated the highest economic value, starting at approximately KRW 576.2 billion in the initial period and escalating to KRW 775.7 billion over six 10-year periods, totaling 60 years. In addition to its economic performance, Scenario 3 effectively improved forest age class structure and ensured a stable timber supply, making it the most balanced approach for sustainable forest management. By focusing on carbon storage as a key management goal, this approach highlights the potential for achieving both economic and environmental benefits concurrently. These results suggest that reinforcement learning is a powerful tool for developing long-term forest management strategies that address multiple objectives, including economic viability, ecological sustainability, and resource optimization.

1. Introduction

In 2015, the Paris Agreement, adopted under the United Nations Framework Convention on Climate Change, acknowledged the urgent need for global efforts to achieve carbon neutrality by 2050. To this end, member countries have submitted their Nationally Determined Contributions, outlining strategies to reduce greenhouse gas emissions [1]. Articles 4 and 5 of the Paris Agreement emphasize the critical importance of securing greenhouse gas sinks and preserving carbon reservoirs, with forests playing a pivotal role in sequestering and storing carbon dioxide [2]. In response to these objectives, the Reducing Emissions from Deforestation and Forest Degradation Plus (REDD+) program has emerged as a key initiative to support the Paris Agreement’s goals. REDD+ aims to prevent deforestation, mitigate forest degradation, and enhance carbon sinks through sustainable forest management (SFM) practices [3]. Notably, effective forest management is proposed as a crucial component in realizing REDD+ objectives, as it can significantly increase timber growth and enhance soil organic carbon storage [4].

As a result of intensive reforestation efforts during the 1970s and 1980s, South Korea now ranks fourth among OECD countries in forest resource volume, with approximately 63% of its national territory (roughly 6.3 million hectares) covered by forests [5]. Since the 1990s, efforts have shifted toward achieving various forestry objectives aligned with SFM principles [6]. In 2013, the Korea Forest Service enacted the “Act on the Management and Improvement of Carbon Sink” and introduced a forest carbon offset system, enabling companies and forest owners to voluntarily engage in maintaining and enhancing carbon sinks through forest management practices [7]. Furthermore, the 6th National Forest Master Plan (2018–2037) established a resource recycling policy based on forest management [8]. Notably, with approximately 72% of the national forestland now classified as forest age class IV or higher, where more than 50% of the forest stands are between 31 to 40 years old, securing new carbon sinks through traditional afforestation or reforestation has become challenging. Consequently, forest management has been proposed as an innovative approach to secure new carbon sinks [9,10].

Forest management plans are developed through a comprehensive analysis of a forest’s current state, enabling managers to understand its production capacity and limitations and subsequently select optimal strategies [11]. Historically, these plans primarily focused on timber production, adhering to the principles of “statutory forests”. Consequently, research during this era predominantly centered on calculating sustainable timber harvest volumes. The 1960s marked a significant shift with the introduction of forest management plan models based on linear programming in the forestry sector. Studies continued to emphasize timber production-centric linear programming models well into the 1990s [12,13]. However, as societal demands evolved to include non-economic factors, forest management objectives diversified. This shift necessitated the development of more complex plans that incorporate spatial elements and consider environmental functions [14]. In line with this trend, the 2000s saw a growing body of research on multiobjective linear programming, which evaluates the multifaceted values that forests provide to society, going beyond traditional timber production models [15].

In the context of the Fourth Industrial Revolution, artificial intelligence (AI) technologies, such as machine learning, have begun to be applied to forest management [16]. In particular, machine learning is effectively used for data analysis and predictive model construction, making it a valuable tool in complex decision-making processes like forest management. Machine learning is broadly categorized into supervised, unsupervised, and reinforcement learning techniques. Among these, reinforcement learning involves interacting with dynamic environments to learn actions that maximize rewards [17]. Reinforcement learning is increasingly recognized as an effective decision-making tool in forest management, particularly when dealing with diverse environmental variables. Recent studies have shown that reinforcement learning can be a useful tool for developing forest management plans that address irregular environmental factors such as landslides, pests, and wildfires [18,19]. For example, Bone and Dragićević [20] developed a multipurpose forest management model using reinforcement learning, which prioritizes both timber harvest and ecological landscape preservation, while Malo et al. [21] applied reinforcement learning to establish forest management plans by considering natural disaster occurrence rates and forest growth rates.

Traditionally, linear programming and other conventional methods have been used to predict timber harvest volumes or focus on individual environmental factors. These methods are suited to simpler forest management goals that account for a limited set of variables. However, modern forest management requires consideration of multiple objectives simultaneously, as well as the handling of complex environmental factors and large datasets. Consequently, traditional methods struggle to incorporate all these elements effectively. To address this limitation, forest management planning increasingly requires quantitative decision-making methodologies like reinforcement learning. These approaches can process large amounts of complex data, enabling more accurate and consistent decision making.

This study aims to develop a forest management plan model that leverages reinforcement learning to achieve both economic value and public benefits. By comparing and evaluating this model against traditional forest management methods, this research seeks to assess its applicability and effectiveness in modern forest management planning.

2. Materials and Methods

2.1. Study Site

For this study, we selected the Bongpyeong pilot forest management zone, situated in Pyeongchang County, Gangwon Province, South Korea (Figure 1). Since 2013, pilot forest management zones have been established nationwide to develop and disseminate intensive forest management techniques for private forests, capitalizing on the stable management conditions of state-owned forests [22]. The study site encompasses an area of 8606 ha, located between 128°15′24″ E and 128°26′39″ E (west to east) and 37°34′46″ N and 37°42′02″ N (south to north). This area represents 12.5% of the state-owned forest under Pyeongchang County’s jurisdiction (68,809 ha). Approximately 81.7% of the total area consists of highly mountainous terrain exceeding 800 m in elevation. The site is divided into 71 forest compartments and 353 sub-compartments, with an average sub-compartment size of 24.4 ha.

The initial forest survey and management plan for the study site were conducted in 1964, with the current forest management plan covering the period from 2023 to 2032. Regarding forest composition, deciduous forests dominate, covering approximately 6158 ha (71.6%), followed by coniferous forests at approximately 2106 ha (24.5%) and mixed forests spanning about 311 ha (3.6%). Areas classified as age class VII or higher account for approximately 4451 ha (51.7%), indicating that most of the area has reached the timber-harvesting stage. Trees in forest age class VI and above constitute approximately 82.0% of the total age class area.

In South Korea, forest age classification (Yeonggeup) is divided into 10 categories based on tree age. For example, class I includes trees aged 0 to 10 years, while class VII and above indicates trees over 60 years old, which are typically ready for harvesting. This classification is essential for understanding the structure of the forest and planning management activities effectively. The classification used in this study reflects the current conditions of the forest area and plays a key role in forest management strategies.

The map in Figure 1 includes the coordinate system, reference datum, and data sources to ensure the study’s accuracy and reproducibility. Furthermore, the climate of the study site is influenced by its proximity to the East Sea and the Taebaek Mountains. The average annual temperature is approximately 13.5 °C, ranging from −8 °C in winter to 28 °C in summer, with an annual rainfall of 1444.9 mm. This temperate climate, combined with mountainous terrain, creates diverse microclimates that significantly affect forest growth and ecological patterns.

2.2. Research Resources

The research resources utilized in this study encompass various geographic information system (GIS) data, including digital topographic, forest soil, forest function classification, regulatory land use, forest parcel, and forest road network maps. Additionally, administrative data from the Phase 2 Bongpyeong Forest Management Plan ledger, provided by the Pyeongchang National Forest Management Office, were incorporated. This comprehensive ledger documents 28 distinct items, including compartment and sub-compartment numbers, GPS coordinates (X, Y), addresses, tree species, and area measurements. For the purpose of analysis and to facilitate the establishment of an effective management plan, essential information was extracted from these resources. These crucial data include compartment and sub-compartment numbers, forest stand characteristics, age classes, area measurements, and timber volumes (

Table 1. Research resources utilized in this study.

Category	Resources	Year Obtained	Providers
GIS Data	Digital Topographic Map	2022	National Geographic Information Institute
	Forest Soil Map	2021	Korea Forest Service
	Forest Function Classification Map	2020
	Regulatory Land Use Map	2021	Ministry of Environment
	Forest Parcel Map	2023	Pyeongchang National Forest Management Office
	Forest Road Network Map	2023
Admin. Data	Forest Management Plan Ledger	2023

).

2.3. Methods

This study aimed to develop a forest management plan model utilizing reinforcement learning algorithms and apply it to a real forest management site to compare and evaluate the expected management performance. First, legal conditions and areas feasible for logging were identified by reviewing relevant laws and studies specific to the study site. For spatial analysis, ArcPro (version 3.3.1, ESRI, Redlands, CA, USA) was utilized to process GIS data, enabling the identification of current forest conditions and management areas. GIS spatial and administrative data were collected and analyzed, allowing us to assess the current forest conditions and management environment. This process led to the identification of manageable areas and the construction of a database. In collaboration with national forest management officers, we reviewed the forest function classification map to systematically establish management objectives. The values of each sub-compartment were converted into economic metrics based on these objectives, and weights were assigned accordingly.

The reinforcement learning model was constructed and analyzed using Python. Specifically, the Q-learning algorithm was employed to enable the agent to learn optimal actions by adjusting reward rates according to scenario-specific weights. This approach allowed the model to simulate decision-making processes under various management alternatives. A long-term forest management plan was established to cover 60 years, divided into six 10-year periods. The model predicted timber production and management performance across compartments and sub-compartments over these periods, providing a comprehensive comparison of different management strategies.

To evaluate the model’s applicability, the results derived from the reinforcement learning approach were compared against an existing forest management plan to assess its potential for real-world implementation (Figure 2).

The initial step in developing a forest management plan model is establishing clear management objectives and constraints. These objectives provide the fundamental direction for forest management and serve as the basis for selecting the objective function in mathematical models [23]. The study site, a pilot national forest management zone, necessitated the formulation of management objectives that balance economic and public interests. To inform decision making in this process, in-depth interviews were conducted with the manager of the Bongpyeong pilot forest management zone and other national forest management officials.

The primary economic function was centered around timber production, while the public interest functions revolved around identifying primary and composite functions using a forest function classification map and maintaining existing roles of national forests. The criteria for selecting public interest functions included data accessibility and quantifiability. The model was used to maximize the benefits from the economic and public interest functions of national forests, leading to the selection of “timber production” as the economic function and “carbon storage” and “water source recharge” as public interest functions.

Timber production, the first objective, is a cornerstone of forest management aimed at consistently producing and supplying high-quality timber. Given South Korea’s aging forests and low timber self-sufficiency, there is a pressing need to increase timber production. Consequently, ensuring sustainable resource supply through effective timber harvesting and securing long-term resource availability by improving the forest age class structure are essential [24].

The second objective, carbon storage, is vital in addressing climate change. The Paris Agreement underscores the importance of forests as carbon sinks, requiring all signatories to enhance and maintain this function. Forests are indispensable for reducing greenhouse gas emissions and adapting to climate change, and SFM is key to maintaining a healthy forest ecosystem. Thus, plans must be devised to produce a consistent annual timber yield while simultaneously increasing forest carbon storage.

Water source recharge, the third objective, is considered a critical function, considering natural environment conservation and forest disaster prevention functions [25]. Given South Korea’s geographical and climatic characteristics, managing water resources is challenging, emphasizing the need for forest management strategies that secure water resources [26].

2.3.1. Estimating Forest Management Areas

In forest management, blind spots exist in forestry operations due to legal, topographical, and technical constraints. Considering these constraints carefully, identifying and selecting areas where logging is both feasible and productive are crucial. Restricted areas can be categorized into legally protected no-logging areas and topographically and technically challenging areas [27].

Legally prohibited no-logging areas comprise public interest forests under the Mountainous Districts Management Act and prohibited timber-harvesting areas under the Guidelines for Sustainable Forest Management. Public interest forests encompass natural recreation forests, wildlife protection areas, water source protection zones, wetland conservation areas, and Baekdudaegan protection regions. Although logging is permissible in some public interest forests through a permit/report system, this provision does not apply to areas where logging is inherently prohibited. Prohibited timber-harvesting areas include regions where regeneration is challenging, such as rocky terrain and areas at risk of degradation, as well as zones within 30 m of a water body’s full water level. To apply legal constraints, we extracted 24 land use areas and districts constituting public interest forests from the forest function classification map. Additionally, 30 m buffer zones were established using river network maps for national and provincial rivers and streams. Forest edge areas were also selected based on the forest direction indicated on the forest-type map.

Due to topographical limitations, areas with slopes exceeding 40° were excluded from consideration. This criterion is based on the maximum operational slope for the HAM300 tower yarder, developed by the National Forestry Cooperative Federation (NFCF), Seoul, South Korea, and commonly used forestry tractors. To address technical constraints, areas more than 300 m from forest and general roads were also excluded, as this represents the maximum logging distance for the state-of-the-art HAM300 cable logging machine. A 300 m buffer zone was applied to the spatial extent of forest roads and road networks, utilizing only roads situated within forested areas for the network analysis.

Management-eligible areas were determined by considering legal, topographical, and technical constraints. To establish forest management objectives and operational strategies for the study site, expert surveys were conducted and management plan documents were thoroughly reviewed. Expert interviews were conducted with the manager of the Bongpyeong pilot forest management zone and national forest management officials. Additionally, management plan documents were examined to assess harvest volumes, logging areas, and management plan objectives (

Table 2. Timber production restriction areas.

Constraints	Description
Legal	Public interest forests under the Mountainous Districts Management Act
	Within 30 m of a water body’s full water level
	Forest edges under the Guidelines for Sustainable Forest Management
Topographical	Areas with slopes exceeding 40°
Technical	Areas more than 300 m from forest roads and general roads

). Regarding legal constraints, these consist of three distinct categories: public interest forests as defined by the Mountainous Districts Management Act, areas within 30 m of a water body’s full water level, and forest edges as outlined in the Guidelines for Sustainable Forest Management. Each of these represents a separate restriction that must be considered when determining timber production eligibility.

2.3.2. Calculation of Economic Value for Each Management Objective

Timber Production

Timber harvest volume and residual stand growth were predicted using the yield table developed by Kang et al. [28] based on data from the 5th and 6th National Forest Inventory. This comprehensive table estimates growth and yield for 14 tree species at the stand level, providing detailed information such as average diameter, basal area, average height, tree count, volume, periodic annual increment, periodic annual growth rate, and mean annual increment according to age and site index. Stand-specific information for each sub-compartment was derived from the forest-type map and measured volume data extracted from management plan documents. Annual growth rates were calculated based on a site index of 16. The average annual growth rate was then applied to predict future harvest volumes and yields for both coniferous and deciduous forests. The economic value of timber production was determined by calculating the net profit, obtained by subtracting timber production costs from the revenue generated through timber sales. The sales revenue for primary timber was calculated using timber sale prices from the Pyeongchang National Forest Management Office. Forestry operation costs were based on the management plan of the Bongpyeong pilot forest management zone.

Carbon Storage

According to Article 2 of the Act on the Maintenance and Improvement of Carbon Sinks, forest carbon sinks encompass standing trees, bamboo, dead organic matter, soil, wood products, and forest biomass that absorb and sequester carbon. This study focused exclusively on carbon absorption by standing biomass affected by forest management activities. Standing biomass comprises aboveground biomass (stems, branches, and leaves) and belowground biomass (roots). The quantity of carbon stored (tCO₂) was calculated by applying wood density, biomass expansion factor, root-to-shoot ratio, carbon conversion factor (0.5), and carbon dioxide conversion factor to the tree volume [29,30]. The economic value of carbon storage was determined by multiplying the amount of stored carbon (tCO₂) by the carbon credit trading price. This price was based on greenhouse gas reduction performance through forests, utilizing the average price of KRW 28,879 per tCO₂ from 2018 to 2022 [31].

Water Source Recharge

Kim et al. [32] calculated the water recharge capacity of forests by multiplying the average soil depth by the macroporosity of soil layers A and B. We obtained the average soil depth from forest site maps and derived macroporosity using a calculation formula provided by Kim et al. [32]. To estimate the economic value of water source recharge, we employed the replacement cost method. This approach assesses the economic value of environmental resources by calculating the cost of preventing or mitigating environmental damage or substituting environmental functions through alternative means [33]. In this study, the economic value was determined based on the cost of storing an equivalent amount of water in a dam, applying a social discount rate of 5%. Kim et al. [32] estimated the dam construction cost at KRW 710.34 per ton and the dam volume maintenance cost at KRW 7.1 per ton. Consequently, the economic value of water source recharge was set at KRW 717.44 per ton.

2.3.3. Scenario-Based Forest Management Using Weights

In developing various management alternatives, we assigned differential weights to management objectives based on their priorities, employing a weighted methodology to maintain equilibrium among these objectives. Priorities were determined by analyzing the primary and secondary functions of each parcel, utilizing the forest function classification map.

At our study site, the Bongpyeong pilot forest management zone, timber production was identified as the most critical function and thus accorded the highest priority. Carbon sequestration received the second-highest weight, recognizing the potential of high-age-class forests as new carbon sinks. Finally, the function covering the largest area in the forest function classification map was designated as the third priority among value-added functions, with weights calculated using the ranking method (Equation (1)).

This study established four management scenarios centered on three primary forest management objectives: timber production, carbon storage, and water source recharge. Each scenario applied distinct weights to these objectives. In the first scenario (S1), timber production was assigned the highest weight of 0.5, followed by carbon storage at 0.3 and water source recharge at 0.2. The second scenario (S2) prioritized timber harvesting exclusively, allocating the maximum weight of 1 to timber production. The third (S3) and fourth (S4) scenarios assigned maximum weights to carbon storage and water source recharge, respectively (

Table 3. Weighting by management scenario.

Management Objectives	Management Scenarios
	S1 (Weighting)	S2 (TP Priority)	S3 (CS Priority)	S4 (WSR Priority)
Timber Production	0.5	1	0	0
Carbon Storage	0.3	0	1	0
Water Source Recharge	0.2	0	0	1

(TP: Timber Production; CS: Carbon Storage; WSR: Water Source Recharge).

).

$$W_j={\displaystyle \frac{n-r_j+1}{\sum \left(n-r_j+1\right)}}$$

(1)

2.3.4. Reinforcement Learning Model Construction

Reinforcement learning is a probabilistic technique that derives optimal solutions through simulations. It involves learning an optimal policy that maximizes cumulative rewards through interactions between an agent and its environment [34]. In this study, we framed the forest management problem as a Markov decision process (MDP) to construct a reinforcement learning model. An MDP comprises states (S), actions (A), a transition function (T), and rewards (R). The agent selects optimal actions within a given environment to maximize rewards [35]. The environment in reinforcement learning is defined by states, actions, and rewards, and this construction of the environment plays a crucial role in model optimization. The environment is categorized into continuous and discrete spaces. For this study, we simplified the complex continuous space by discretizing it. The forest management model incorporated sub-compartment data, including forest type, age class, area, and volume and timber production, carbon storage, and water source recharge values. Forest type and age class were expressed as distinct conditions within the discrete space.

To further validate the model, we conducted a retrospective analysis by comparing model predictions with actual historical timber harvest data from the Bongpyeong forest management site. This comparison allowed us to assess how well the model’s outputs aligned with real-world results, thereby verifying its effectiveness in predicting timber harvest volumes and management outcomes over time.

The agent learns the optimal policy through multiple attempts and iterative trial-and-error processes. In this study, we employed a “goal-based agent” to select the most effective actions based on predefined objectives. The agent selects actions from the current state space, receives feedback in the form of rewards from the environment, and progressively learns toward the optimal policy. The primary goal of the agent is to maximize economic value by determining whether to harvest each sub-compartment while adhering to the constraint of a maximum harvest volume of 150,000 m³. The action space encompasses all possible choices available to the agent. In this study, actions were defined as either post-harvest reforestation or maintenance of the existing stand. By evaluating options within this action space, the agent selects the optimal course of action to achieve the management objectives.

The state space portrays the forest’s condition at a specific point in time, with each sub-compartment’s state determined by key forest attributes such as forest type, age class, area, and volume. This comprehensive state space incorporates data crucial for optimizing multiple forest management objectives, including timber production, carbon sequestration, and water source replenishment. We employed the Q-learning algorithm to enable the agent to learn the optimal policy from the environment. Q-learning is a prominent model-free reinforcement learning technique that derives optimal control strategies based on rewards (R), learning rate (α), and discount factor (γ) [36]. The algorithm initializes all Q-values to zero and progressively refines them through iterative updates of states and actions across multiple episodes, ultimately converging on the optimal Q-values. The learning rate and discount factor were set at 0.1 and 0.95, respectively, after several iterations to balance short-term versus long-term rewards and ensure efficient learning.

Data input into the model was carefully processed using GIS tools to ensure accurate integration of spatial and administrative data. Key inputs included forest attributes such as type, age class, area, and volume, which were discretized for use in the Q-learning algorithm. These inputs helped refine the model’s state space and ensured that the agent’s decisions were based on the most relevant forest characteristics.

The reward system is a crucial component that quantitatively assesses the outcomes of the agent’s actions. In our reinforcement learning model, rewards were calibrated to reflect changes in economic value, guiding the agent’s behavior through positive and negative reward points. A reward of 200 points was awarded if the economic value was maximized while adhering to the constraint of a maximum harvest volume of 150,000 m³. Conversely, a negative reward of −100 points was imposed if these conditions were not met. If no harvesting occurred, a penalty of −50 points was applied.

Finally, the results derived from the reinforcement learning model were compared to an existing forest management plan to evaluate its applicability in real-world settings. This comparison helped validate the model’s potential for practical use and provided insights into how the reinforcement learning approach can be applied to improve long-term forest management strategies (Figure 2).

3. Results and Discussion

3.1. Estimating Timber Harvest Volumes Using the Reinforcement Learning Model

In this study, we analyzed timber harvest volumes across six consecutive 10-year periods using four distinct forest management scenarios. We estimated the timber harvest volumes for each scenario across the six periods (Figure 3). Our analysis revealed that S3 (Carbon Storage Priority) yielded the highest total harvest volume at 957,351 m³, followed by S4 (Water Source Replenishment Priority), S2 (Timber Production Priority), and S1 (Weighting Method). The maximum allowable harvest volume per period was constrained to 150,000 m³, resulting in a total permissible cumulative harvest volume of 900,000 m³ over the six periods. S2 came closest to this constraint, with an estimated total harvest volume of 905,994 m³, exceeding the limit by a mere 0.7%. S1, on the other hand, fell short of the constraint with a total harvest volume of approximately 851,296 m³, meeting the per-period limit in all but the sixth period.

To verify the accuracy of the reinforcement learning model’s predictions, we conducted a retrospective analysis. This involved comparing the model’s predictions against actual timber harvest results from the Bongpyeong pilot forest management zone. By evaluating historical harvest data provided by the Bongpyeong Forest Management Plan, we assessed how closely the model’s predictions aligned with real-world outcomes. This comparison helped validate the model’s effectiveness in predicting harvest volumes over time and demonstrated its practical applicability in forest management.

A detailed analysis of each scenario’s outcomes by period revealed that S1 adhered to the constraint (harvest volume ≤ 150,000 m³) in all periods except the sixth, maintaining an average 10-year harvest of 132,615 m³. In the sixth period, however, the estimated harvest volume reached 188,221 m³, surpassing the constraint by 38,221 m³. S2 approached the constraint threshold in the first two periods but exceeded it in the third, with an estimated harvest volume of 192,984 m³—approximately 1.3-fold higher than the limit. Excluding this outlier, S2 had an average 10-year harvest of 142,602 m³. S3, with a total harvest volume of 957,351 m³, exceeded the total constraint of 900,000 m³ by approximately 6.4%, making it the scenario with the highest total harvest volume. Notably, this scenario achieved peak harvest volumes in the fourth and fifth periods, totaling 423,156 m³ and accounting for approximately 44% of its total harvest volume. Compared with the other scenarios, S4 demonstrated overestimation in the second and third periods, with harvest volumes of 166,132 m³ (1.1-fold the constraint) and 168,592 m³ (1.2-fold the constraint), respectively. Excluding these two periods, S4 maintained an average 10-year harvest of 145,420 m³.

The study site’s characteristics, featuring a wide distribution of harvestable stands in age class VI and above, led to the selection of unharvested higher-age-class stands as the periods progressed. Consequently, all scenarios except S1 exceeded the target harvest volume in some periods following the initial period.

3.2. Estimating Harvest Areas Using the Reinforcement Learning Model

Consistent with the current management plan, the constraint on harvest area was established at 988 ha per period, with a maximum total harvest area of 5928 ha. Analysis of harvest areas for each scenario revealed that S3 yielded the largest total harvest area at 3983 ha, followed by S4, S2, and S1. Examination of the 10-year harvest areas for each scenario showed that S1 had the most substantial harvest area in the first period, spanning 813 ha or accounting for 23% of the total harvest area across the six 10-year periods. Although the harvest area gradually diminished from the second period onward, it rebounded to 569 ha in the sixth period, the largest sixth-period harvest area among all scenarios.

S2 exhibited the largest single-period harvest area of 888 ha in the first period, surpassing all other scenarios in any given period. This scenario averaged 764 ha of harvest area per period from the first to the third period, approximately 1.7-fold higher than the average harvest area of 455 ha per period from the fourth to the sixth period.

In S3, the harvest area initially decreased from the first to the third period but then surged to 859 ha in the fourth period, nearly matching the first-period harvest area of 868 ha, before declining again in the fifth and sixth periods. Notably, S3 demonstrated the largest harvest areas in the fourth and fifth periods compared with the corresponding periods of other scenarios.

S4 maintained larger average harvest areas of 815 ha from the first to the third period compared with the fourth to the sixth period. Particularly, the second period had an estimated harvest area of 834 ha, approximately 13% larger than the average second-period harvest area of 736 ha across all scenarios (Figure 4). While the harvest volumes in all scenarios were overestimated, the total and 10-year harvest areas adhered to the established constraints, resulting in compliant outcomes.

3.3. Analysis of Age Class Structure Changes According to Forest Management Plans

We analyzed changes in the age class structure according to the forest management objectives set by each management scenario (Figure 5). Currently, stands in age class VII or higher comprise 43% of the study site, while the proportion of young forests (age classes I to III) is as low as 9%, indicating a significant imbalance in the age class distribution. This current structure is markedly unbalanced, and without appropriate intervention, the sustainable supply of timber resources may become increasingly challenging. To address this imbalance in the long term, increasing timber harvesting is imperative, thereby creating successor stands and improving the circulation of resources accumulated in the forest. The management plans project that after six 10-year periods, the proportion of stands in age class VII or higher will improve most significantly under S3, reaching approximately 32%, followed by S4, S2, and S1. Furthermore, all the scenarios demonstrated an increase in the area of young forests over time, indicating a gradual improvement in the overall age class structure.

Our findings are consistent with those of Kim, Han, and Chung (2021), who observed similar trends in their analysis of optimal forest management. In their study, forest compartments with age class IV or higher initially dominated the landscape, but through planned harvesting and regeneration, the age class distribution gradually balanced out. Similarly, in our scenarios, after multiple periods of management interventions, the age class structure transitions toward a more stable distribution, ensuring a sustainable supply of timber. This parallel suggests that planned harvesting can effectively address age class imbalances and promote long-term forest sustainability under various management objectives [37].

In S1, the proportion of mature stands (age class VII or higher) experienced a modest decline, decreasing from the current 43% to 39% after six periods—a reduction of approximately 4%. This represents the smallest decrease among all scenarios evaluated. Conversely, the proportion of young forests (age classes I to III) increased to roughly 36% after three 10-year periods. In subsequent periods, all age classes except for the mature stands (VII or higher) maintained a relatively stable average of approximately 10% each.

In S2, the proportion of mature stands (age class VII or higher) decreased from the current 43% to 35% after six periods. Concurrently, the percentage of young forests increased significantly, reaching approximately 41%. In subsequent periods, all age classes except for mature stands (VII or higher) maintained a relatively stable average of approximately 11% each.

In S3, the proportion of stands in age class VII or higher decreased significantly, from the current 43% to 32% after six periods, representing a reduction of approximately 11%—the most substantial decrease among all scenarios. Harvesting activities primarily targeted older stands in age class VII or higher from the first to the third period, resulting in a more rapid improvement in the age class structure compared with that in other scenarios. The proportion of young forests increased markedly, reaching approximately 36% after the third period, comparable to that for S1. By the sixth period, the proportion of young forests stood at approximately 11%, which is 3% higher than in other scenarios, further emphasizing the shift toward a more balanced age class structure.

In S4, the proportion of stands in age class VII or higher decreased from the current 43% to 33% after six periods, similar to the decrease observed in S2. The proportion of young forests in the third period reached 42%, slightly higher (1%) than in S2. Although harvesting primarily targeted stands in age class VII or higher, a considerable number of sub-compartments in age class V were also selected for harvest in the first period.

As the scenarios progressed through each period, the increasing area of young forests signaled a notable improvement in the overall age class structure. This enhancement of the age class distribution through carefully crafted management plans can significantly contribute to both a stable timber supply and SFM practices [38].

3.4. Comparison of Expected Management Performance

We conducted a comparative analysis of the projected management performance of forest management scenarios across six consecutive 10-year periods (Figure 6). Upon evaluating the economic value of the remaining trees in each scenario, S3 (Carbon Storage Priority) emerged as the most economically advantageous, followed by S1 (Weighting Method), S2 (Timber Production Priority), and S4 (Water Source Replenishment Priority). S3 exhibited exceptional economic potential, with an expected value approximately 1.3-fold greater than that of the second-ranking S1 (KRW 775.7 billion compared with KRW 441.0 billion). While S2 yielded a slightly lower value of KRW 359.3 billion compared with that of S1, S4 was assessed to have by far the lowest economic value at KRW 107.5 billion.

In S1, the residual economic value commenced at KRW 329.2 billion in the initial period and increased to KRW 441.0 billion by the end of the sixth period. The economic value derived from harvesting was estimated at KRW 47.6 billion in the first period, declining to KRW 29.1 billion after six periods. A notable dip in economic value occurred in the fifth period, attributable to a lower harvested volume of 93,290 m³ compared with the average. The cumulative potential profit over the six periods amounted to KRW 27.1 billion. This outcome, achieved by applying balanced weights to the management objectives, demonstrates that relatively stable harvest volumes were maintained while simultaneously realizing viable economic value.

In S2, the residual economic value initially stood at KRW 267.9 billion in the first period, subsequently rising to KRW 359.3 billion by the end of the sixth period. The economic value derived from harvesting was estimated at approximately KRW 40.0 billion in the first period, which declined slightly to KRW 33.5 billion after six periods. This scenario exhibited similarities to S1 in terms of period-by-period analysis; however, it distinguished itself by maintaining a more consistent harvest volume of 905,994 m³, closely adhering to the prescribed constraint. This approach, which prioritized timber production, ensured a stable economic value across the six 10-year periods. Consequently, S2 generated a cumulative profit of KRW 23.9 billion over the course of six periods.

S3 demonstrated the most substantial residual economic value, commencing at KRW 576.2 billion in the initial period and escalating to KRW 775.7 billion by the conclusion of the sixth period. The economic value derived from harvesting in the first period was calculated at KRW 86.5 billion, reaching its zenith at KRW 129.5 billion during the fourth period. This remarkable economic performance can be attributed to the prioritization of carbon storage, which yielded high initial profits and culminated in an estimated total profit of KRW 54.4 billion throughout the six periods.

In S4, the residual economic value in the initial period was KRW 108.2 billion, which marginally decreased (0.64%) to KRW 107.5 billion after six periods. The economic value derived from harvesting was estimated at KRW 15.5 billion in the first period, diminishing to KRW 6.8 billion by the end of the sixth period. By prioritizing water source recharge, a stable, albeit low, economic value was maintained, yielding a cumulative profit of KRW 7.5 billion over the course of six periods.

The four management scenarios exhibited differences in economic performance and harvest volume, aligning with their respective objectives. Notably, S3 (Carbon Storage Priority) demonstrated the highest economic performance among all scenarios. The outcomes suggest that prioritizing carbon storage in forest management planning for the study site is not only an economically viable strategy but also offers significant public benefits. This conclusion is further supported by Birdsey et al. [39], who found that carbon storage strategies enhance long-term economic returns through ecosystem service payments while providing ecological benefits. Similarly, Mason et al. [40] highlighted the growing importance of carbon-focused strategies in achieving both higher economic gains and environmental sustainability within forest management practices.

4. Conclusions

This study successfully developed a novel forest management model using a reinforcement learning algorithm tailored to national forest management objectives. The model was applied to a real-world forest site to assess and compare its predicted outcomes with those of the existing management plan. By incorporating both economic and public interest functions into the objectives, the study utilized GIS spatial data and administrative records to analyze forest status, determine management-eligible areas, and assign appropriate weights to various objectives. A 60-year long-term management plan was created using a Python-based Q-learning algorithm, which forecasted timber production and management performance by parcel and over time.

The model’s application in the Bongpyeong pilot forest management zone in Pyeongchang County, South Korea, demonstrated its potential to balance economic returns with public interest goals. Among the four forest management scenarios evaluated, S3 (Carbon Storage Priority) stood out as the most effective, showing the highest economic performance while adhering to sustainable forest management principles. These results suggest that emphasizing carbon storage in forest management can significantly boost long-term economic returns and public benefits.

Reinforcement-learning-based models in forestry aim to optimize sustainable forest management (SFM) by maximizing economic value while maintaining balanced age class structures and ensuring young forest regeneration. This study developed the model using static data, leveraging key principles of reinforcement learning. The adaptive nature of reinforcement learning allows it to respond dynamically to environmental changes, positioning it as a viable tool for addressing issues like climate change, landslides, and pest infestations in future forest management plans.

The primary aim of national forests is to secure a stable timber supply while fulfilling various public interest functions. By applying the reinforcement learning-based model to the Bongpyeong pilot zone, prioritizing carbon storage proved to be the most beneficial strategy, particularly in improving the age class structure and ensuring a steady timber supply. As a result, this study confirmed the feasibility of developing comprehensive forest management plans that effectively balance economic objectives with public interest considerations.

Additionally, the study explored the potential of establishing forest management strategies that sustain timber production while integrating multiple public interest functions. The methodologies and findings presented here can serve as foundational data for developing strategic forest management plans that address various social demands, such as implementing sustainable cyclic management systems and adapting to climate change.

However, this study was exploratory in nature, and the profit rate was used as an economic criterion to provide a comparable and quantifiable metric for evaluation. While this was useful for comparing different objectives, we recognize that it may not fully reflect the long-term ecological trade-offs between economic returns and carbon storage. Future research should explore alternative economic metrics, such as net present value (NPV), to more comprehensively evaluate these trade-offs and offer different forest management strategies.

The model’s reliance on static data may not fully capture real-time forest dynamics or external factors like climate change, pest outbreaks, or unexpected natural disturbances. Future research should focus on improving the model’s integration of real-time data to enhance its adaptability to dynamic forest conditions. Furthermore, while this study applied the model to a specific forest management zone, broader validation across different forest types and regions is necessary to evaluate its general applicability.

Despite these limitations, this study makes a valuable contribution by introducing a reinforcement-learning-based approach to forest management. Its adaptive nature allows the model to tackle complex management challenges while optimizing both economic and ecological outcomes. This approach offers forest managers a practical solution to achieving sustainable management practices in a rapidly changing environment and could serve as a foundation for more comprehensive management strategies in the future. Moreover, the methodologies and findings of this study provide crucial data for developing forest management plans that meet diverse social needs, including the implementation of sustainable cyclic management systems and responses to climate change.