Developing an Agent-Based Model to Mitigate Famine Risk in North Korea: Insights from the “Artificial North Korean Collective Farm” Model

Abstract

North Korea experienced a catastrophic famine in the mid-1990s that resulted in millions of deaths. This study aims to build an agent-based model to understand the risk of land degradation and famine in North Korea and explore potential solutions to mitigate this risk. The model concept reflects the general information of collective farms in North Korea, which was set in 1960, and the abstract of the social–ecological system of North Korean agriculture. The model comprises the agent, environment, and external factors. An agent is defined as a household that decides its labor allocation and land use by how much food it has. The environment is set to multiple layers that represent the soil environment of the farm, affected by agents’ decisions. The external factors reflect a land-use restriction policy, international trade (food aid), and climate change. We calibrated the model using data from North Korea and simulated various scenarios. Our findings suggest that the famine in the model would occur around 35 years later, almost reflecting the 1995 famine in the real world. The scenario-based simulations show that additional food aid or import could delay the famine, but fundamental solutions are needed to break the vicious circle of land degradation and declining food yield. Our study provides insights into the complex adaptive system of North Korean agriculture and the critical points of famine risk. Our agent-based model can be a useful tool for policymakers and researchers to explore potential solutions and inform risk management strategies.

1. Introduction

The rise and fall of civilizations have long been topics of interest across various academic fields, such as archaeology, history, physical geography, and paleoenvironmental studies. While these fields often rely on historical records, microfossils (e.g., pollen), and ice cores, paleoenvironmental data is often insufficient to provide a complete picture of past events. Agent-based modeling (ABM) is an approach that simulates complex social–ecological systems by considering various factors such as human behavior, environmental conditions, and their interactions, which can help address uncertainties in research [1,2].

The “Artificial Anasazi” study is an ABM approach that draws parallels between the ancient Anasazi society and North Korea. The Anasazi civilization, founded in Arizona, USA, collapsed for unknown reasons, and the “Artificial Anasazi” study applies ABM to reveal the specific complex systematic process and narrative of its decline [3,4]. This approach offers an alternative to overcome the lack of information in civilization studies and the resulting uncertainties. Other studies had attributed the Anasazi civilization’s collapse to a drought around AD 1200, but the ABM approach uncovered a more nuanced and complex explanation.

By simulating various factors such as environmental conditions, resource availability, and human behavior, the “Artificial Anasazi” study revealed how changes in the Anasazi society’s social and ecological systems interacted and contributed to its collapse. These findings highlight the importance of understanding the complex interdependence between human societies and the natural environment, as well as the potential value of ABM in uncovering new insights into past events.

The success of ABM in shedding light on the collapse of the Anasazi civilization [3,4] has inspired scholars to use this approach to study other societies that have experienced similar challenges due to limited and incomplete information. One such example is contemporary North Korea.

While the abundance of data and information in contemporary society has enabled new insights and discoveries, it is still challenging to study complex systems like civilizations with incomplete and limited information. This is particularly true in the case of North Korea, where information about the country remains scarce and often indirect. Despite this challenge, parallels can be drawn between North Korea and the Anasazi civilization, both of which faced similar challenges such as famine. By applying agent-based modeling (ABM), it is possible to gain a deeper understanding of the complex systematic processes underlying these societal challenges and their interactions with human behavior and environmental conditions. In this study, we applied ABM to investigate the societal challenges faced by North Korea and draw insights from the Anasazi civilization, highlighting the potential of this approach to overcome the limitations of traditional research methods.

North Korea, officially known as the Democratic People’s Republic of Korea, is situated in East Asia and the West Pacific, bordering South Korea to the south, and Northeast China and Primorsky Krai of Russia to the north. The country covers an area of 120,540 km² and has a population of approximately 25 million people (2022 estimates). North Korea’s steep topography, with around 70–80% of the country being mountainous, and its colder and drier climate, categorized as Dw in the Köppen climate classification, have resulted in lower population support capacity compared to neighboring countries [5]. North Korea experienced a devastating famine in the 1990s that resulted in the deaths of millions of people [6,7,8]. Furthermore, indicators related to human life, such as life expectancy and mortality, do not reach the level of the 1980s, before the famine. For instance, in the late 2010s, North Korean men in their 20s, who grew up during the famine, had an average height of 157 cm, which is 15 cm shorter than men in their 20s in genetically similar South Korea, China, and Japan [9]. Despite these challenges, North Korea has continued to develop its nuclear weapons program, leading to further international tensions and sanctions.

This study aims to investigate the impact of land degradation on food security in North Korea, drawing on lessons learned from the Anasazi civilization in the southwestern United States, where land degradation contributed to the civilization’s collapse. By comparing the two cases, this study seeks to identify potential strategies to address land degradation and food insecurity in North Korea. The causes of the famine are complex, including natural disasters, isolation due to the collapse of the Eastern Bloc, an inefficient economic system, and an inappropriate agricultural and land management policy known as “Juche Agriculture” and “Darakbak, [5]” Land degradation is believed to be a key factor in the food shortage, although there is no direct indicator of the land degradation trend in North Korea [9,10]. Instead, proxy data such as forest area calculated by statistics and satellite images could be an alternative [11].

This study aims to investigate the relationship between land degradation and food production decline as a key factor in North Korea’s great famine of the 1990s. Land degradation is defined as the decline in land productivity, whether ecologically or economically [12,13], and has played a crucial role in North Korea’s famine crisis. As an illustration, the country’s agricultural and economic policy failures have contributed to soil quality degradation, and land-use mismanagement has resulted in soil erosion and natural disasters such as landslides. The “vicious cycle of land degradation” has also contributed to the decline in agricultural productivity, leading to further land degradation as people use intensive land practices to compensate for the losses. This vicious cycle has played a pivotal role in the devastation, famine, and poverty experienced by North Korea.

Despite the significance of the link between land degradation and famine, both problems are highly complex, and little information is available in North Korea and elsewhere [14]. This study contributes to our understanding of the relationship between land degradation and food production decline in North Korea and highlights the need for further research and policy interventions to address these interconnected problems.

This study focuses on the feedback between land degradation and food production decline as a key factor in North Korea’s great famine of the 1990s. The “vicious cycle of land degradation,” in which agricultural productivity declines due to land degradation, and land degradation worsens as people use intensive land to make up for it, is an essential part of the problem of devastation, famine, and poverty. Moreover, land degradation and famine problems are highly complex, and little information is available about them in North Korea and other areas [14].

Similar to the “Artificial Anasazi” model, this study aims to create “The Artificial North Korea Collective Farm” agent-based model to comprehensively understand the complex interplay between land degradation and famine. Specifically, this study conceptualizes North Korea’s agriculture, which is marked by land degradation and food production decline, within the framework of “collective farms,” the basic agricultural landscape unit in North Korea. Based on this conceptualization, we developed an ABM that simulates the North Korean famine in the 1990s. By conducting simulations, we investigated how land degradation and declining food production contribute to starvation and assessed whether alternative policies could have mitigated the risk. The findings of this study are expected to offer insights that can inform policy interventions in North Korea and other countries facing similar challenges.

2. Materials and Methods

2.1. Concept of Model

Agent-based models of agricultural systems typically rely on on-site investigations, interviews, or behavioral experiments to describe the agricultural landscape and farmers in reality (e.g., [1]). However, due to the lack of available data and the impossibility of conducting field studies in North Korea, this study had to rely on limited information to build an agent-based model of an “Artificial North Korean Collective Farm”. To achieve this, the model was simplified and abstracted as much as possible, following the example of the “Artificial Anasazi” study.

The collective farm is a farming community in which multiple farmers hold their land as a cooperation (e.g., “Kolkhoze” in the Soviet Union) or a state-own agricultural firm (state farm, e.g., “Sovkhozy” in the Soviet Union). North Korea adopted this system in 1954 and changed all farms to the collective farm system in 1960 [15]. Despite the lack of data on North Korean collective farms, some studies have recorded general information [5,15,16]. A typical farm is between 550 and 750 hectares and is almost identical to ‘Ri’, the smallest administrative unit in the country. Each farm has a population of 300–400 households, 700–900 workers, and 1900–2000 dependents, and is organized into 5–10 workgroups, which are the basic unit of farming and yield distribution.

The “Artificial North Korean Collective Farm” concept (Figure 1) was developed by combining this information with abstraction and summarization. Unknown details and difficult-to-generalize attributes were addressed using abstraction and summarization techniques. For instance, the initial agricultural land cover pattern, which varies with each farm, was set to 30% rice field and 70% other farmland, based on a summary of North Korean land use [5]. The concept describes the shape and topography of the farm using simplified square and ski-slope shapes. The initial forest cover was added to reflect the degradation caused by the conversion to farmland. The positioning of the different land uses, with steep forests in the north, flat rice fields in the south, and villages located on the mountain edge, was illustrated using traditional knowledge for environmental adaptation strategies of Korean people, named “the location with back to the mountain and facing the water (배산임수 in Korean, 背山臨水 in Chinese letters) [10,17].

2.2. Model Formation

The ABM model was developed using NetLogo 6.2.0 [18] and based on the Land-use dynamic simulator (LUDAS) framework, which is a well-established ABM approach for agricultural systems. LUDAS has been applied in various regions, including the Hong-ha watershed of Vietnam [1,19,20], Inner Mongolia of China [21], the Gariwang-san region of Gangwon-do, South Korea [22,23,24], and Pyongannam-do of North Korea [25]. In this study, the ABM model was constructed based on the conceptual design and consists of three submodules: the environment, human agents, and external factors.

The environment submodule uses a multi-layered grid structure, comprised of 250 by 250 grids with each grid size of 100 m². Each layer represents processes that influence land degradation and land-use change. The human agent submodule consists of 400 households, their attributes, and decision-making processes, which influence land-use change and soil quality. The output of the environmental submodule is the soil quality, which affects food yield. Thus, this structure describes the feedback loop of land degradation and food yield, with each loop representing a year. Figure 2 depicts the ABM model’s structure and components.

2.2.1. Environment Submodule

The environment submodule, also known as the landscape-environmental system, models the environmental processes of land degradation and food yield in North Korea, with multiple layers (

Table 1. The layers and attributes of the environment submodule.

Classification		Layer	Variable Name in Model
Basic Information	Constant Value (Topographic Value)	Elevation (m)	p_elevation
		Slope (degree)	p_slope
		Upslope Contributing Area	p_as
		Surface Curvature	p_cur
	Changeable attribute	Rainfall (mm/year)	p_rainfall
		Mean Temperature (°C)	p_temp
Related to Submodules	Physical Environment	Soil Erodibility Factor (K)	p_kfactor
		Rainfall Erosivity Factor (R)	p_rfactor
		Land Management Factor (C)	p_cfactor
		Topographic Factor (LS)	p_lsfactor
		Land Management Factor (P)	p_pfactor
		USLE Result (A)	p_usle
		Normalized USLE Results	p_usle_es
	Chemical Environment	Chemical Soil Potential Index	p_soil-potential
	Soil Quality Index	Physical Soil Point	p_soil-physical-point
		Soil Quality Index	p_soil-quality
Environment to Human Agent	Food Yield	NDVI Estimates	p_ndvi
		Yield Potential (kg/100 m2)	p_yield-potential
		Yield Estimation (kg/100 m2)	p_yield-estimation
		Yield Amount (kg/100 m2)	p_yield

Source: Modified from [9]

). This subsystem consists of five submodules: biological, physical, and chemical soil environments, the soil quality index, and the potential of food yield that combines these indicators.

The physical soil environment submodule represents the process of physical soil erosion and adopts the universal soil loss equation (USLE) model, which is widely used throughout the world (Equation (1)):

A = R × K × LS × C × P

(1)

where A is the mean (annual) soil loss, R is the rainfall erosivity factor, K is the soil erodibility factor, LS is the topographical factor, and C and P are land management factors (

Table 2. Calculation Methods of USLE Factors in physical soil environment submodule.

Factor	Calculation Method
Rainfall erosivity factor (R)	Toxopeus equation adopted in the Korean Peninsula [26,27]            R = 38.5 + 0.35 × P P: Annual Precipitation
Soil erodibility factor (K)	The random number ranged by calculated data Pyongannam-do of North Korea [25] Note: soil particle information is needed. However, there are limited data and no way to effectively estimate.
Land management factor (C,P)	Agricultural land: Table of research in the Korean Peninsula [28] Mountainous Region: Adopted Europe model by NDVI [29,30]            $C=e^{\left(-\alpha \times \frac{NDVI}{\beta -NDVI}\right)}$ α,β: parameters (normally 1)
Topographical factor (LS)	Adopted Europe mode [29,30]         $L=1.4{(\frac{As}{22.13})}^{0.4}, S={(\frac{Sin\beta }{0.0896})}^{}$

Source: Modified from [9]

).

The chemical soil environment submodule is designed to reflect the assumption that more intensive land use leads to more soil chemical consumption. If the human agent in the model invests more time or work into the farmland, the soil’s available chemicals for growing crops (e.g., organic matter) decrease (Equation (2)):

$$\begin{array}{c}\mathrm{If} F_{yield} > F_{yield-potential}, S{c}_1=S{c}_0-\alpha \frac{F_{yield}}{F_{yield-potential}}\\ \mathrm{If} \mathrm{not} (F_{yield}\le F_{yield-potential},), S{c}_1=S{c}_0\end{array}$$

(2)

where F_yield is the annual crop yield, F_{yield-potential} is the annual crop yield potential (see “potential of food yield”), Sc₁ is soil quality in the year, Sc₀ is soil quality of the previous year ($0\le S{c}_x\le 0.5)$, and α is a parameter (1 in this model). Suppose the human agent in the model puts in more work. In that case, the yield will be larger than the yield potential (see next). Consequently, human decisions regarding labor allocation determine the soil’s chemical quality in the model.

The biological soil environment submodule represents the ecological potential of the soil. The normalized difference vegetation index (NDVI) obtained by remote sensing can indicate this potential in areas where information is scarce, such as North Korea [31]. NDVI can also calculate food yield (e.g., [32]) and soil erosion (e.g., [29,30]) in this context. Therefore, this study adopted NDVI estimation models by land use (

Table 3. Linear regression results for NDVI estimation in different land-use types: Coefficients and coefficient of determination (R-squared) as biological soil environment submodules.

Land Use	R2	p-Value	NDVI Formula 1,2
Rice	0.158	<0.001	Ln[1.464 − 0.019 Ln (Slope) − 0.026 Ln (Elevation) + 0.215 Ln(Precipitation) − 0.023(Temperature)]
Non-rice farm	0.142	<0.001	Ln[2.115 + 0.005(Slope) − 0.01 Ln (As) + 0.02 Ln(Elevation) + 0.002(Temperature)]
Forest	0.289	<0.001	7111.178 − 30.361 Log(As) + 83.839 Ln(Slope) + 213.554 Ln(Elevation) + 36.004(Temperature) − 1.107(Precipitation)

Source: Modified from [25]. ¹ Abbreviation: As-Upslope Contributing Area, ² Data sources: United States Geological Service (USGS), eMODIS (NDVI), USGS SRTM 90 m (Slope, Elevation, As), Korea Meteorological Administration (Temperature, Precipitation).

).

The soil quality index submodule is an integrated physical and chemical index indicator, calculated as the sum of the results of the physical and biological submodules. Since the outputs of the biological submodule use the crop yield (potential), the soil quality index does not include these outputs to prevent double counting. The USLE results were normalized from 0 to 0.5, based on previous experience in the Korean Peninsula [27].

The potential of the food yield submodule is based on satellite image analysis of North Korea, which compares NDVI data and food yield statistics [25,32]. The formula in this module varies with land use (

Table 4. Linear regression results for Food yield Potential (F_{yield-potential}) in different land-use types: Coefficients and coefficient of determination (R-squared).

Land Use	R2	p-value	Fyield-potential Formula 1
Rice	0.449	0.048	1137.209NDVI−601.416
Non-rice farm	0.555	0.021	0.66NDVI + 0.641

Source: Modified from [25,32]. ¹ Data sources: United States Geological Service (USGS), eMODIS (NDVI, Korea Meteorological Administration (Food yield as).

).

The R-square of each model adjusts the biological and food yield potential submodules to reflect the uncertainty. In order words, these results ($Y_{equation}$) were reflected only in the range corresponding to r squared ($Y_{estimate}$). Random numbers ($Y_{random,}$, but range in $\xi$) were substituted for the remaining ranges $(1-R^2)$. The final results were designed to be derived by combining them (Equation (3)).

$$\begin{array}{c}{Y}_m=Y_{estimate}+Y_{random}\\ Y_{estimate}=Y_{equation}+R^2\\ Y_{random}=\xi +(1-R^2)\end{array}$$

(3)

In this model, annual food yield is calculated based on the results of these submodules and the impact of human decisions. Specifically, the yearly food yield equation takes into account each land’s food yield potential, soil quality, and labor force input (Equation (4)):

$$\begin{array}{c}{F}_i^{yield}=\alpha F_i^{yield-potential}\times \beta S{Q}_i\times \gamma L_i\\ L_i=\frac{L{T}_i}{L{T}_{reference}}\end{array}$$

(4)

where $F_i^{yield}$ is the annual crop yield, $F_i^{yield-potential}$ is the annual crop yield potential, $S{Q}_i$ is the soil quality index, $L_i$ is the workforce index calculated by the ratio of annual labor time ($L{T}_i$) to reference labor time ($L{T}_{reference}$), the mean annual labor input by crop, and α, β, γ is a constant (1 in this model).

2.2.2. Human Submodule

Table 5. The attributes of each agent in the human submodule.

Classification		Attribute Name	Variable Name in Model
Basic Information of an Agent		Number of household members (P)	h_people
		Number of working age in a household	h_labor_pop
		Number of non-working age in a household	h_non_labor_pop
		Household location	h_xcor
			h_ycor
		Work group (0~10)	h_group
Considerable attributes of Agents	Food Allocation	Gross household food yield (A)	h_yield
		Government and Farm Authorities Proportion of A	h_yield_gov
		Household Proportion of A (B)	h_yield_mine
		Rationing by Government (C)	h_food_gov
		B + C	h_food
		Food Allocation per capita (B + C/P)	h_food-houshold
		Food Pressure	h_food-pressure
Related to Decision-making Process	Labor Allocation	The labor time of Agents	h_time
		Redundancy of Labor time	h_redundancy
	Agent leaving Procedure	The year the agent feels hungry	h_hunger-time
		Whether Agent is hungry or not (h_food_pressure > 4)	h_hunger

Source: Modified from [9]

presents the characteristics of the human agents in the model. The agents are assumed to be aware of their environmental conditions, including the environment submodule and other human agents, and their actions can affect and be affected by the entire system.

The model includes three submodules for linking human agents’ decisions and their environment: land allocation, food allocation, and food pressure. The land allocation rule is based on the simple principle of ‘equal distribution’. The food allocation rule is based on economic distribution research conducted in North Korean cooperative farms [15,33]. Finally, the food pressure index is determined by the food yield model result, previous land allocation, and population of the agents (Equation (5)):

$$A{F}_{pressure}=\frac{Pop\times Fl}{A{F}_{yield}+A{F}_{aid}}$$

(5)

The food pressure of a household agent is calculated using Equation (5), where AF_pressure represents the food pressure of an agent, AF_yield is the portion of annual crop yield for a human agent, and AF_aid is the portion of yearly crop import or aid for a human agent, resulting from food allocation. Pop represents the agent population, and Fl is the annual minimum food requirement per capita (164.25-kg crop). If an agent’s AF_pressure is larger than 1, they will suffer from hunger.

Due to data limitations, agents’ decision-making processes are based on a simple assumption: if human agents need more food, they change their labor or land-use strategy. The decision-making process of household agents consists of two submodules: labor force allocation and land-use change. If their food pressure exceeds 1, they will allocate their labor force and change their land-use strategy. On the other hand, they will maintain their previous plan if their food pressure is below 1.

The labor force allocation submodule is based on the assumption that a laborer can work 300 days a year and 8 h in one day. Additionally, the minimum annual input labor time for an average crop per 100 m² in the rice field is 36 h, and in other farms, it is 48 h, according to the results of test simulations. Initially, the labor time was set at 6 h for a 100 m² rice field and 8 h for the same area in other farms, based on statistics from South Korea (no such data exist in North Korea). However, this time setting did not produce any changes. Therefore, we increased the labor time and found appropriate values for simulating North Korea.

If an agent’s food pressure is below 1, they will allocate their time based on the second assumption (first year) or the same as last year (after the second year). If an agent’s food pressure exceeds 1, they will reallocate their time based on the allocated soil quality. First, they will reduce or eliminate their time invested in lower-productivity land. Second, they will increase their extra time for higher productivity land (

Table 6. Decision making strategies and soil quality thresholds of human agent in labor force allocation submodule.

Strategies	Labor time Allocation	Soil Quality Thresholds
Reduce or eliminate Time	Eliminate time	SQi < 0.5 ± 0.05 Fyield-potential ≤ 0
	Reduce 20%	0.5 ± 0.05 ≤ SQi < 0.6 ± 0.05
	Reduce 10%	0.6 ± 0.05 ≤ SQi < 0.7 ± 0.05
Maintain the status quo	No Change	0.7 ± 0.05 ≤ SQi < 0.9 ± 0.05
Increase	Increase 20%	0.9 ± 0.05 ≤ SQi

Source: Modified from [9,10,25].

). The workforce index in Equation (3) is calculated by dividing labor time by the initial labor time based on the second assumption.

The land-use change submodule is activated when the food pressure of household agents rises above 1, and they have additional time or land without any labor time. The first option of the land-use change strategy is to convert bare land or forest to farmland. If the soil quality index of such land is above 0.5, rice fields (slope below 5 degrees) or other farmland can be established. The second option of the land-use change strategy is to convert farmland to bare land, where nobody inputs their time.

2.2.3. External Factors, Model Settings, and Model Outputs

The model formation stage includes considerations of external factors, model settings, and model outputs that could affect North Korea’s land degradation. For this model, three external policy factors were selected and constructed: climate change, change of food aid reflecting international relations, and land-use restriction change. The factor of climate change was based on the Intergovernmental Panel on Climate Change (IPCC) (2013) Representative Concentration Pathways (RCP) 2.6~8.5 scenarios and applied to temperature and precipitation shifts (both mean and standard deviation) [34]. For example, in the RCP 8.5 scenario, where emissions continue to rise throughout the 21st century, the mean annual temperature will increase by 2 °C (0.4 standard deviations) 20 years later, and precipitation will increase by 4% of temperature rise. This factor mainly affects environment submodules in the model, such as soil erosion and biological soil environment.

The factor of change in food aid reflects how much food can be aided or imported per year and is based on the population support of the entire model. For instance, a 100% amount setting of this factor indicates that 100% of additional food to support the population would be supplied outside of the model (foreign). On the contrary, the model will be a closed system if this factor setting is 0%. As a result, this factor would connect to the agent’s food pressure.

The factor of land-use restriction change is related to the forest clearing to expand the farm in North Korea. From the 1960s to early 1980s, the supreme leader of North Korea, Kim Il-sung, and his government encouraged farmland expansion to mountains and coasts to overcome harsh environments for harvesting [5,35]. However, his governing instructions changed in the late 1980s, especially mountain terracing. He stated that farming on a steep slope is dangerous and requires restriction [35]. The North Korean authorities started to worry about soil erosion and landslide risk caused by forest clearing. As a result, a steep slope land-use policy was established in late 1980 [35]. However, this policy was too late to mitigate risk and could not work due to food and energy shortages. In this context, this scenario is designed to permit deforestation on the steep-sloped forest (15 degrees) or not. Additionally, reflecting reality, households can ignore this policy if they experience extreme food pressure (above 3 in this model).

Although these external factors were set, the climate change and land-use restriction factors rarely affected the test results, while the food-aid factor could change the model’s outcome.

The model’s outputs include the number of households, land cover change, food production, soil quality index, and food pressure (

Table 7. Summary of significant model outputs and their descriptions.

Variable Name	Description
Number of Households	Based on the assumption of the model when agents with severe food pressure (>3) for more than three years are eliminated. At the end of each simulation, the number of households is counted, and the result is derived. The simulation is stopped when more than half of the agents disappear.
Land-Use Change (Land Cover Change)	Check the trend of changes in mountainous areas, agricultural lands (rice fields and others), and empty fields. The city area is set in the model but does not change, so it is not checked.
Food Yield	The total sum of food production of households and land in the model. Unit: Kilograms/Model or household.
Soil Quality Index	Average of soil quality indicators in the model derived from environmental submodels. There is no unit. If it is close to 0, it means low soil quality. If it is close to 1, it means high soil quality.
Food Pressure Index	The degree to which households want food (or feel hungry). Less than one if the annual minimum food requirement (165 kg of grain per capita) is met, and two is defined as half of the minimum food requirement.

). The model is designed to stop when the number of households decreases to more than half from the initial setting. A household is eliminated if it has experienced hunger (severe food pressure, more than three years), so the number of households can be said to be the result of the model’s collapse. Land cover change and soil quality indicators can be seen as results of land degradation. Changes in food pressure and production are indicators of food problems and famine risks. Soil quality index and food pressure can be meaningful as internal variables of the model that can additionally confirm environmental processes and agents’ decision-making in the model.

The operation of the model is divided into initial setting (setup) and execution (go). In the initial setting, the model was set according to the initial information of the “artificial North Korean collective farm” design and model. Most of the external environment variables are placed in the initial setting stage. However, in the case of climate change variables, they change according to the flow of the model execution stage. In the execution stage of the model, the feedback process of the environment and actors described above is driven. As explained earlier, it is defined as one year elapsed each time it is executed. The implementation of the model was set to 100 years, or the point at which North Korea’s land degradation and food problems were maximized. It was set that household agents who have experienced severe food pressure for more than three years are removed from the model (“leave from the virtual cooperative farm”) to hypothesize the point at which North Korea’s land degradation and food shortages were maximized. When the number of breakaway household agents reaches half of the initially set number of actors, it is assumed that the model collapses and is set to end. The simulation was performed 100 times for each case.

3. Results

3.1. Replicability of the Great Famine in the 1990s

The agent-based model used in this study has limitations in terms of its application of field data, making it challenging to verify and correct the model’s assumptions. To overcome this limitation, we designed a model that reproduces the great famine of the 1990s by fitting the North Korean agricultural environment, land devastation, and food production decline from the 1960s to the 1990s. The model’s starting point was set in the 1960s when North Korea completed the collective farm system, and the external environmental conditions were set to those of North Korea in the late 20th century. These conditions included climate change as the current emissions scenario, land-use restriction policy with a gradient of more than 15 degrees set by the North Korean authorities, and food support, which was provided at a reduced rate of 1,000,000 tons per year to account for the population.

During the initial construction of the model, unintended results were derived from the relationship between randomly assigned variables. However, several test simulations were conducted to identify the significant impact of the annual standard labor time (LT_reference) of each crop (land cover) on the results. Subsequently, each random variable was set, and the simulation was conducted by assuming the initial conditions starting in 1960. The results of the simulation were presented on the endpoint of the simulation, land cover change, factors of land degradation (e.g., soil quality index), and factors of food production (e.g., food pressure).

Figure 3 depicts the distribution of simulation endpoints in a box chart format. The average time at which the simulations stopped was found to be 36.12 years, with most simulations ending between 35 to 38 years. This suggests that a famine-like situation similar to the one in North Korea occurred approximately 35 years after the simulation’s starting point of 1960, indicating that the food shortage resulting from land degradation began in earnest in the late 1980s and early 1990s, with the famine actually occurring in the mid-1990s.

While this result follows the factual events to some extent, it should be noted that events such as natural disasters and economic sanctions experienced by North Korea could not be incorporated into the model due to limitations in its construction. As highlighted in the introduction, such factors could be critical in causing famine. However, we excluded this to prevent unnecessary randomness. Therefore, this model does not account for unexpected events but provides insight into the background conditions such as climate change and agricultural policies that led to the famine. In explaining the great famine in North Korea in 1990, the limitations of the cooperative farm system, Juche farming methods, and rigid social structure are more critical than temporary natural disasters or economic sanctions. This finding underscores how the system’s internal processes, such as the farm system and agricultural policy, could result from a closed system problem caused by limited food or energy supplies.

The collapse or suspension of the model from the early 1990s is similar to the early warning signals of reduced food production in the early 1990s, albeit somewhat later than the indicators related to land degradation. As the number of agents was determined by “food pressure,” the agents’ number and endpoint can indicate the occurrence of a great famine.

The left side of Figure 4a,b shows the trend of land use and cover change in the simulation results. From the beginning of the simulation to the 20th year, there were no significant changes observed except for the addition of bare areas and the gradual expansion of mountainous regions. In the 20th to 30th years of simulation, it can be observed that land cover changes have increased rapidly, such as more land clearing. After 30 years, the simulation results show different trends, and it can be confirmed through the results that the simulation ends after most of the mountainous areas have been cleared and the bare area has increased rapidly.

Regarding land use, the model shows a clear trend of land degradation from the 1960s, but the pattern varies somewhat by period and can be primarily divided into three categories. First, from the beginning of the simulation to around 27–28 years (1960-mid 1980s), the mountainous area was gradually converted to farmland. As mentioned in the introduction, farmland expansion policies were implemented and encouraged to supplement the low yield during this period. Second, it is the period of rapid farmland conversion in mountainous areas from around 27–28 until just before the end of the model. From about 27–28 years (1987–1988), forest clearing increased rapidly rather than gradually. This period was when North Korea considered mountain clearing dangerous and earnestly implemented a slope restriction policy. Through this, it can be seen that the land reclamation pressure is why the slope policy ended without effect. Third, the rapid expansion of the bare area occurred before the end of the simulation. This result refers to the sudden increase in land, which is in a state where the productivity of land has rapidly decreased, and no meaningful yield has been obtained. North Korean defectors who experienced the “Arduous March” period mention that the quality of soil and productivity of many cooperative farms during this period were at this level.

These pattern changes can be interpreted as time lags. The internal aspects of the model, the design of the land-use regulation policy, and the model’s feedback can explain the factors that cause this time lag. Firstly, land-use regulation scenarios are designed to be ignored by agents when food pressure is high. The decrease in agricultural productivity may have been gradual and exponential, but the following results can be seen as the rapid reclamation of mountainous areas beyond the reference point set by food pressure and the immediate abandonment of farmland.

The right of Figure 4a shows the food production and soil quality. Changes in food production and food pressure show a moderate trend until the 20th year, and then a transition to a decrease in food production and an increase in food pressure around the 20th year. From 30 years until the model’s collapse, the trend changes again, confirming that the decreasing and increasing trends boost dramatically. These results can be interpreted as a result similar to the change in land use. In the case of the soil quality index, it is somewhat different, but it gradually decreases without significant change until just before the end of the simulation, and unlike other results, it is not easy to find a noticeable inflection point.

In the 1960s, North Korea’s food production increased and continued to do so until the early 1990s. However, the ration system’s instability in the 1970s and 1980s, as noted by Lee [36], and the continued emphasis on “natural transformation” and cultivation from the late 1970s suggest unstable food supply and demand, gradually increasing food pressure. This result aligns with the observed increase in land degradation. Considering the reliability issues with North Korean food production statistics [6,7,36], the degree of food pressure rather than the amount of food production is a significant variable in explaining land degradation and food production decline.

The gradual change in soil quality indicators is meaningful, reflecting continuous decision-making between environmental submodels and agents within the model and serving as a symbol of land degradation in reality. Physical soil erosion is closely related to changes in land use, while chemical potential degradation is related to changes in land labor input by actors. Since there was little change in land use until 30 years elapsed, it can be considered to change entirely based on the decrease in chemical potential until then. The soil quality index at this time showed a nearly linear trend, and the displacement (standard deviation) was not significant, indicating that actors gradually felt food shortages after the model’s start. As a result, they increased the strategy of changing labor input to the land, causing gradual land degradation. The accumulation of this gradual land degradation can be seen as a potential cause of rapid change after 30 years. Since there was a change in the mountainous area after 30 years, the physical part of the soil quality index was likely significantly affected. Hence, there was a slight difference in the soil quality index’s results in the model’s later part.

3.2. Result of Applying Food-Aid Scienaros

Based on various test simulations, the external factor of change in food aid was applied to the environmental change scenario to evaluate how much the specific environmental change scenario reduces the land degradation problem and the risk of great famine in North Korea. The scenario of external food support is that external food support or income is additionally supplied at a specific rate at the level at which households in the model can survive. The sensitivity of the ratio increased by 12.5% from 0% and was set to 100%. The simulation endpoint (Figure 5) tends to increase with the food support rate, and the simulation lasts nearly twice as long as the control group (0%) when an additional payment of up to 50% of the required amount is made. However, no scenario could be maintained until the final endpoint (100 years), and the increase in the endpoint of the simulation has been stagnant since the 75% support, indicating that the collapse of North Korea’s soil ecological system could not be fundamentally prevented with additional food support alone.

In the test simulation results for the 1960s to 1990s, the change in land use appeared once every 20 years (1980) after the simulation and showed a rapid change in 30 years (1990). The effect of additional food support appears as a delay in the model collapse, but the time lag from the start of the model’s collapse to the end of actual failure does not appear significant. The mechanism that the great famine occurred as feedback from the process of land degradation to reduced food production, which was previously argued through similar results, is also sufficiently confirmed in this result. Therefore, it can be interpreted as a result that supports the claim of clearly working in “(artificial) North Korean cooperative farms”. Still, it can also be interpreted as a result that shows that food aid does not fundamentally change this feedback process.

In the test simulation results for the 1960s to 1990s, changes in the inflection point that can be identified in land-use changes were shown in changes in food production and food pressure in the model but were not well represented in soil quality indicators (Figure 6). The time lag effect is present in all indicators, but the time delay effect of food production–food pressure-related indicators and soil quality indicators appears differently. As with changes in land use, food production and food pressure showed a time delay effect by pushing back the inflection point. On the other hand, the soil quality index shows a time lag as a change in the slope of the trend.

4. Discussion

4.1. Interpretation and Limitation of the Famine in Complex Adaptive System Context

As described in Section 3.1, the trend of land degradation and food production decline in North Korea can be observed to have progressed through three stages, as supported by both simulation results and North Korean reality, since the completion of the cooperative farm system in the 1960s. These stages can be interpreted as phases towards emergence in a complex adaptive system context, as shown in

Table 8. Exploring the relationship between simulation results and real-world phenomena through a complex adaptive system lens.

Step	Land-Use Change Example (Figure 4b)	Phenomenon		Interpretation in Complex Adaptive System Context
		In Simulation	In Real-World
1		0–20 years later Few changes and event (Household and Environment) Soil Quality ↓ Gradually	AD 1960–1980 Completion of Collective Farm System (1960) Start of “A war to improve nature (in socialism context)”: e.g., mountain terracing “Juche Policy”: e.g., Maize-Centered, dense cultivation No significant record of Land Degradation	Equilibrium Set the structure of the feedback loop between land degradation and food production
2		20–30 years later Household action ↑ → Land-Use Change ↑ Significant deforestation for farmland increase	AD 1980–1990 Instability example: reduction food ration system [36], prohibit terracing (Contrary to before 1980, but not working) [35]	The edge of Chaos Critical Transition or Threshold Detecting Early Warning Signal
3		30 years later–end Rapid expansion of abandoned area “Almost Destroyed”	AD 1990–1995 Extreme natural Disaster Extreme famine	Collapse (New)Equilibrium

.

The first stage, which lasted from the initial stage to 20 years in the simulation and from 1960 to 1980 in reality, can be interpreted as a state of equilibrium in which no major changes are expected. During this period, both North and South Korea had similar levels of economic development and agricultural productivity. For instance, according to the United Nations Statistical Division, in 1973, North Korea’s Gross Domestic Income (GDP) per capita was USD 475, which is approximately 118.8% of South Korea’s GDP per capita (USD 400)) [37]. Although North Korea had problematic policies and decisions during this period, the system was able to maintain equilibrium.

The second stage, from the 1980s to the early 1990s in reality (20th–30th years in simulation), could be characterized as the “edge of chaos” era in the context of complex adaptive systems. During this period, the North Korean agricultural system underwent significant changes that pushed it far from equilibrium. The shift in policy, including ration control [36] and the prohibition of mountain terracing [35], suggests that policymakers recognized the system’s unstable state. However, despite these efforts, the catastrophic famine that would follow remained unforeseen.

The third stage, which occurred in the 1990s (after 30 years in simulation), witnessed the rapid collapse of both the simulated and the real world, and can be characterized as an emergence stage in a complex adaptive system context. Many experts at the time predicted a crisis in North Korea due to political circumstances, such as the collapse of the Soviet Bloc, and natural disasters. However, despite the imminent disaster, no one could have predicted the severity of the famine that occurred [9].

The primary mechanism for the collapse of the equilibrium state can be attributed to the feedback loop between land degradation and declining food production. As illustrated in Figure 7A, the pressure on food supply from the population leads to intensive farming and farming expansion, resulting in a decline in soil quality and subsequently, a reduction in food production. This, in turn, leads to an increase in food pressure. The feedback loop is the core of the ABM model used in this study (Section 2.2), and comparing the results of the ABM model with the actual situation, it can be concluded that the model worked well as intended. Furthermore, the study results support the hypothesis that the famine in North Korea was caused by a complex adaptive system, with feedback from land degradation and reduced food production being the key cause.

The limitations of this study emphasize the need for further research to consider external factors that may have contributed to the famine in North Korea. While this study highlights the role of internal processes such as land degradation and reduced food production, there are opposing views that suggest the cause of the famine was a failure to allocate available food resources rather than a failure to produce enough food [9,38,39]. Furthermore, external factors such as economic isolation and natural disasters cannot be ignored in a complex adaptive system approach, as they may serve as triggers of emergence. Although meaningful results were derived without considering external factors, future research should incorporate them to provide a more comprehensive understanding of the famine in North Korea.

The study also attempted to consider external factors through food-aid scenarios in Section 3.2, which showed limited effectiveness in mitigating the risk of famine. Breaking the feedback loop between land degradation and declining food production was identified as a key process in the study, and further research could explore how external factors could be utilized to intervene in this feedback loop. Overall, the results of this study provide valuable insights into the complex adaptive system that contributed to the North Korean famine, but future research should consider external factors to provide a more comprehensive understanding of the phenomenon.

Firstly, we will discuss how to overcome the feedback loop, based on the examples observed in the food-aid scenarios discussed in Section 3.2. As mentioned earlier, the feedback loop between land degradation and declining food production was found to be the primary mechanism of the equilibrium-to-collapse procedure in the ABM model. The food-aid scenarios showed that breaking this feedback loop requires more than just providing food aid. It requires a comprehensive approach that addresses the underlying causes of land degradation and declining food production, such as improving soil quality, reducing intensive farming practices, and promoting sustainable land-use policies.

In addition, future research should also take into account external factors that may trigger or exacerbate the feedback loop. This could include the effects of economic sanctions, climate change, and natural disasters on the North Korean agricultural system. A more comprehensive model that incorporates these external factors would provide a more accurate understanding of the complex adaptive system in North Korea and help policymakers develop more effective strategies to prevent future famines.

Overall, this study provides valuable insights into the causes of the North Korean famine from a complex adaptive system perspective. It highlights the importance of understanding the feedback loops and underlying causes of land degradation and declining food production in preventing future famines. However, further research is needed to fully understand the external factors that may exacerbate the feedback loop and to develop more comprehensive strategies for preventing future famines in North Korea.

4.2. How to Overcome Feedback Loop

The first simulation results presented in Section 3.1 summarize North Korea’s agricultural environment from the 1960s to the 1990s in a collective farm, demonstrating the consequences of the feedback or the vicious cycle between land degradation and food production decline in North Korea. Specifically, this feedback can be illustrated in the form of Figure 7A. Starting with the decrease in food production, the reduction in food production leads land users to feel the need to increase food production, that is, food pressure. As a result, residents use the land intensively for increased food production, which leads to a decline in soil quality. This decline in soil quality leads to a further decrease in food production, completing the feedback structure.

Finding a method to break this feedback loop, the study tested the external food supply scenario, described in detail in the second result (Section 4.2). The simulation demonstrated that increasing external food supply could slow the famine threat. However, while the slope of the factors, increasing food pressure and decreasing soil quality, was reduced in Figure 7B, the feedback structure was not eliminated. In other words, the feedback could not be fundamentally resolved.

The study proposes other ways of escaping or dissipating feedback, such as implementing a more sustainable agricultural structure, introducing an external supply of energy in addition to food, or introducing a more efficient and motivating system to replace the current collective farm system. However, we are still exploring how to implement these alternatives effectively in the model.

Despite these efforts, it is possible that a conclusion may be drawn that the feedback cannot be fundamentally prevented. Specifically, it may be concluded that the process of land degradation and declining food production which caused the great famine in North Korea in the 1990s is difficult to overcome due to the unique vulnerability of the Korean Peninsula. This pessimistic outlook is due to the historical background of the Korean Peninsula.

In the late 1600s, several great famines occurred on the Korean Peninsula, with hundreds of thousands to millions of victims expected. Historians and paleoclimatologists suggest that the main cause was the decline in agricultural production due to climate disasters caused by drought, cold weather, and other sudden changes in weather during the Little Ice Age, which had a global impact at the time [40]. However, the population of the Korean Peninsula at that time was rapidly increasing, and forests were rapidly decreasing. Additionally, the peninsula was diplomatically isolated, with no active exchanges with neighboring countries or external forces such as Western countries. Despite the famines that followed at the time, diplomatic efforts or international trade to overcome them were not active, and even small attempts led to significant controversy [41].

Many similarities exist between the great famine of the 17th century that occurred throughout the Korean Peninsula and the great famine of the late 20th century in North Korea in terms of isolation, land degradation centered on deforestation, and climate disasters as the leading causes of famine. However, the core structure of famine implementation, a vicious cycle in which these factors reduce land productivity, increasing the need for increased food production and accelerating land use which in turn deteriorates land productivity, is also similar. Therefore, it may lead to a pessimistic conclusion that the Korean Peninsula cannot be sustainable by itself in terms of population and natural environment structure since the late Joseon Dynasty and can only be sustained with external energy and food supplies.

Therefore, under this view of the Korean Peninsula, the conclusion that the external food supply confirmed in Section 3.2 delays the occurrence of famine in North Korea can be seen as a meaningful conclusion. Even if it is difficult to conclude that the Korean Peninsula can be self-sufficient or sustainable at present, the provision of external food and energy supplies can be a valuable solution in terms of buying time to explore alternative solutions. In order to truly break the feedback loop between land degradation and declining food production, further exploration of alternative solutions is necessary. These may include more sustainable agricultural practices, more efficient and motivating farming systems, and more active diplomatic efforts to facilitate international trade and collaboration. By continuing to investigate and implement these alternatives, it may be possible to achieve a more sustainable future for North Korea and the Korean Peninsula as a whole.

5. Conclusions

This study aimed to investigate the causes and processes of the North Korean Famine through a complex adaptive system framework. Our hypothesis was that land degradation and declining food production were key factors contributing to the famine. Similar to the “Artificial Anasazi” studies, we utilized an agent-based modeling (ABM) approach to deal with data shortage and complexity. By creating an “Artificial North Korean Cooperative Farm” and building an ABM model, we were able to address our research question. The following section highlights the key findings of our study.

• The simulations based on baseline conditions consistently showed that the North Korean Famine occurred around 35 years after 1960, the starting point. This suggests that even with limited and summarized information on cooperative farms and North Korea, the complex adaptive system implementation of North Korea’s great famine and land degradation problems in the 1990s could be sufficiently implemented.
• While external food support could delay the worsening of land degradation and the risk of famine, the simulation results revealed that fundamental improvement was needed to prevent a crisis. Although supplying at least 50% of the population’s minimum food needs delayed the emergence of famine beyond the mid-1990s projected by the baseline conditions simulations, it did not fundamentally prevent land degradation and the threat of a great famine.
• The study found that feedbacks between land degradation and food production decline are key complex adaptive system processes that cause emergent phenomena and feedbacks leading to the North Korean Famine. The reduction in food production has been steadily and progressively degrading the quality of the land, resulting in a repetition of decreasing food yield. This feedback framework was robust and not easily broken, as confirmed by scenario-based simulations. As such, crisis prevention efforts must be strengthened and improved to prevent such catastrophic events.

The primary objective of this study was to simulate the sudden impact of land degradation and food production decline in North Korea that leads to a great famine, rather than predicting the long-term trend of these factors. However, one major limitation of this study is that the model dataset only covers the period from the 1960s to the 1990s, and therefore may not be entirely suitable for predicting future events after the 2000s. Therefore, the results of this study should be interpreted in the context of developing “risk mitigation strategies” to prevent similar crises in the future. This means that the study could be used to compare the differences between the simulation results and actual events in the past, and to substitute new policy elements for those used at that time.

In light of the findings of this study, future research will prioritize the development of additional modifications, supplements, and scenario applications. These will be aimed at establishing a risk avoidance strategy for the problem of land degradation and food production reduction in North Korea, as well as addressing broader issues in the agricultural and social–ecological systems of the Korean Peninsula. To achieve these goals, the research will focus on two main strategies.

The first strategy will be to develop and improve the accuracy of the model by collecting and integrating additional data sources that reflect changes in the North Korean agricultural system and the country’s socio-economic conditions. This will include data on changes in agricultural technology, government policies, and the impact of external factors such as climate change and international sanctions. By incorporating these factors into the model, we can gain a more comprehensive understanding of the complex interactions between land degradation, food production, and external factors, and develop more effective risk avoidance strategies.

The second strategy will be to expand the scope of the study beyond the agricultural system to examine the broader social–ecological system of the Korean Peninsula. This will include analyzing the impact of political, economic, and social factors on the agricultural system and its sustainability. By taking a more holistic approach, we can identify the root causes of land degradation and food production decline in North Korea and develop more effective solutions to address these challenges. This will require collaboration with experts from a range of disciplines, including political science, economics, and sociology. By working together, we can develop a more comprehensive understanding of the complex social–ecological system of the Korean Peninsula and develop effective solutions to promote sustainability and resilience.

In conclusion, this study presents a model that highlights the negative feedback between land degradation and reduced food production in North Korea’s closed agricultural system. One limitation of the study is its focus on the North Korean agricultural system and its potential applicability to other countries facing similar challenges. To address this limitation, a third strategy can be pursued to adapt the model to agricultural systems in socially or ecologically different regions. However, it is important to recognize that North Korea’s social characteristics are unique, and few countries will be ecologically similar to North Korea. Therefore, the first strategy should be to classify North Korea socially and ecologically and develop a concrete model. The second strategy is to expand the model to the Korean Peninsula and its surrounding areas. By doing so, this research can contribute to sustainability research in global agriculture and food security. It is hoped that this study will provide insights for policymakers and practitioners to develop effective strategies for sustainable agriculture in North Korea and beyond.