Assessing the Transformation of Armed Conflict Types: A Dynamic Approach

Abstract

Armed conflict is a dynamic social phenomenon, yet existing research often overlooks its evolving nature. We propose a method to simulate the dynamic transformations of armed conflicts. First, we enhanced the Spatial Conflict Dynamic Indicator (SCDi) by integrating conflict intensity and clustering, which allowed for the distinction of various conflict types. Second, we established transformation rules for the SCDi, quantifying five types of transformations: outbreak, stabilization, escalation, de-escalation, and maintaining peace. Using the random forest algorithm with multiple covariates, we simulated these transformations and analyzed the driving factors. The results reveal a global trend of polarization in armed conflicts over the past 20 years, with an increase in clustered/high-intensity (CH) and dispersed/low-intensity (DL) conflicts. Stable regions of ongoing conflict have emerged, notably in areas like Syria, the border of Afghanistan, and Nepal’s border region. New conflicts are more likely to arise near these zones. Various driving forces shape conflict transformations, with neighboring conflict scenarios acting as key catalysts. The capacity of a region to maintain peace largely depends on neighboring conflict dynamics, while local factors are more influential in other types of transformations. This study quantifies the dynamic process of conflict transformations and reveals detailed changes.

1. Introduction

Armed conflict, as a complex and dynamic social phenomenon, exerts profound impacts on the global economies, geopolitical stability, and human well-being. For instance, the Russia-Ukraine conflict has displaced millions and triggered substantial economic disruptions [1,2], while the Sudanese civil war forced nearly 4 million people to flee within just over 100 days [3]. These crises underscore the urgent need to understand and forecast the dynamic trends of armed conflict to inform effective risk mitigation and loss reduction strategies [4,5,6,7].

The study of armed conflict risk modeling has evolved significantly over the past few decades, driven by advances in data availability and computational methods. Early approaches predominantly employed statistical models to identify conflict predictors at the national level. For example, Goldstone et al. [8] utilized logistic regression to predict political instability and civil war onset, finding that political instability and economic deprivation were significant drivers. However, the model’s reliance on aggregated data overlooked subnational variations. Similarly, Buhaug and Rød [9] applied spatial regression techniques to examine the geographic clustering of conflicts, revealing the influence of terrain and border proximity on violence. Their approach improved the understanding of spatial dependencies but struggled with temporal dynamics. Cederman et al. [10] introduced a disaggregated regression framework to study ethnic conflicts, emphasizing the role of group-level inequalities. Despite its nuanced perspective, the model’s static nature limited its ability to capture conflict escalation over time. Raleigh and Hegre [11] further refined statistical approaches by incorporating disaggregated data into logit models, enabling finer temporal resolution. Yet, these models often assumed linear relationships, which constrained their ability to represent the complex, nonlinear interactions among conflict drivers such as resource scarcity or ethnic tensions [6,7]. These limitations underscore the need for more flexible and dynamic modeling frameworks to better understand the multifaceted nature of armed conflicts.

In response to the shortcomings of statistical models, researchers have increasingly turned to machine learning techniques to enhance the prediction and simulation of armed conflict risks. Machine learning methods excel at capturing nonlinear relationships and leveraging large, high-resolution datasets without requiring rigid theoretical assumptions [12]. For instance, Witmer et al. [13] employed random forest models to predict subnational conflict events in Africa, integrating socioeconomic, environmental, and political variables. Their model achieved high predictive accuracy but faced challenges in interpreting complex feature interactions. Similarly, Hegre et al. [14] used machine learning techniques, including boosting methods, to forecast global conflict risks through 2030, incorporating climate and demographic data to highlight the role of environmental stressors. While their approach offered robust long-term projections, it was less effective for short-term dynamics. Vesco et al. [15] applied neural networks to model conflict escalation globally, using high-frequency data to capture temporal fluctuations in violence. However, the computational complexity and lack of transparency in their model posed barriers to practical application. These studies, supported by advancements in computational power and geospatial data [16], have enabled unprecedented granularity in conflict risk assessments. Nevertheless, many machine learning models prioritize static risk mapping over dynamic forecasting, often failing to account for how conflicts evolve or spread over time. This gap highlights the need for integrated approaches that combine the predictive power of machine learning with the temporal and spatial dynamics of conflict processes.

Recent research has begun to address the dynamic nature of armed conflicts by modeling their escalation and de-escalation. Mueller et al. [17] developed a predictive framework based on conflict history and textual features extracted from news reports to forecast violence escalation. Their model demonstrated moderate success in identifying short-term trends but struggled with long-term predictions due to its reliance on historical data alone. In a more advanced approach, Vestby et al. [18] evaluated gradient boosting algorithms to predict subnational violence escalation and de-escalation, achieving exceptional performance in capturing localized trends. Despite these advances, current models rarely consider the multifaceted attributes of conflicts, such as their intensity (the severity of violence) and concentration (the spatial clustering of events) [19]. These attributes are critical for understanding the magnitude and geographic scope of conflicts, yet their integration into dynamic models remains limited. Furthermore, few studies adopt a spatiotemporal perspective, which is essential for capturing the evolution of conflicts across time and space [20].

The incorporation of diverse driving factors has enriched conflict modeling, reflecting the complex interplay of socioeconomic, political, and environmental conditions. Early studies identified low socioeconomic development [21,22], weak governmental capacity [23,24], and resource scarcity [25,26] as key drivers of conflict risk. More recently, researchers have explored additional factors, including harsh geographic conditions [27,28], ethnic discrimination [29,30], and climate change risks [31,32,33,34,35], which have gained significant attention due to their role in exacerbating resource competition and social tensions. While these factors have been incorporated into static models, their role in driving dynamic conflict transformations remains underexplored. Most studies analyze driving factors in isolation, neglecting their synergistic effects or spatiotemporal variations, which limits the accuracy of dynamic forecasting.

This study addresses these gaps by introducing an enhanced Spatial Conflict Dynamic Indicator (SCDi) that captures the intensity and concentration of armed conflicts to assess their evolution over the past 20 years. Specifically, we (1) refine the SCDi to classify conflict types based on their intensity and spatial clustering, providing a robust framework for analyzing global conflict patterns; (2) quantify transformations in conflict types using machine learning to simulate annual spatial probability distributions, revealing the complexity of conflict dynamics; and (3) analyze the synergistic effects of driving factors, including conflict-related, socioeconomic, geographic, and climatic variables, to identify the determinants of conflict transformations. By integrating spatiotemporal dynamics, conflict attributes, and diverse drivers, this study offers a novel approach to modeling armed conflict transformations. This framework provides decision-makers with actionable insights to mitigate conflict risks, prevent escalation, and formulate effective policy responses.

2. Materials and Methods

Figure 1 illustrates the conceptual definitions and analytical framework of the study. First, we defined the SCDi and its transformation types. Second, we modeled the SCDi transformation types by integrating various factors using the random forest algorithm. Finally, the study results encompassed the spatiotemporal characteristics of the SCDi and its transformations, the globally simulated transformation probabilities, and the importance of different features. By integrating multisource data and applying machine learning methods, this research aimed to provide a comprehensive and dynamic perspective for understanding and predicting conflict transformations worldwide, thereby supporting the development of more effective intervention and mitigation strategies.

2.1. Data

The data used in this study included armed conflict data, socioeconomic data, climate disaster data, and geographic data. The specific data are listed in

Table 1. Classification of collected variables and their abbreviations.

Type	Covariates	Abbreviation
Conflict-related	SCDi last year Total number of conflicts last year Fatalities of conflicts last year Total number of conflicts in the vicinity last year Fatalities of conflicts in the vicinity last year Total number of conflicts in 1989–1999 Fatalities of conflicts in 1989–1999 Number of years of conflict in 1989–1999 Total number of conflicts in the vicinity in 1989–1999 Fatalities of conflicts in the vicinity in 1989–1999 Number of years of conflict in the vicinity in 1989–1999 Total number of global conflicts last year Fatalities of global conflicts last year	SL NCL
Socioeconomic	Population Population change Gross domestic product (GDP) GDP change Human development index (HDI) Net migration Excluded ethnic groups Road density Urban proportion Farmland proportion Human footprint Peacekeeping operations records Peacekeeping operations records in the vicinity Critical infrastructure index Food price index Oil price index Frequency of major international sanctions	POP POPC GDP GDPC HDI NM EEG RD UP FLP HF POR PORV CISI FPI OPI SAC
Climate and disasters	Multi-hazard frequency Standard precipitation index Standard temperature index Mean annual temperature Mean annual precipitation	MF SPI STI AAT AP
Geographic	Longitude Latitude Elevation mean Elevation standard deviation Mountain coverage Travel times to cities Distance to nearest country anywhere Distance to own borders Distance to capital Distance to the coastline Forest proportion Normalized Difference Vegetation Index (NDVI)	LON LAT EM ESD MC TTC DNCT DOB DCA DNCS FRP NDVI

. To standardize the spatial resolution, this study employed PRIO-GRID as the fundamental spatial analysis unit. PRIO-GRID is a global geospatial grid dataset developed by the Peace Research Institute Oslo (PRIO), primarily used in social science research, particularly conflict studies [16]. Its core design involves dividing the Earth’s surface into grid cells of a uniform size of 0.5 degrees, standardizing the collection and analysis of geospatial data.

The armed conflict event data used in this study were sourced from the Uppsala Conflict Data Program Georeferenced Event Dataset (UCDP GED) [36], which is renowned for its comprehensive documentation of organized violence and conflict events worldwide. From the UCDP GED, we extracted information on the location, date, and number of fatalities of conflict events, and based on these data, we derived the SCDi and conflict-related variables. The historical record of conflict in a region may lead to vulnerabilities in the social structure and political institutions, laying the groundwork for future armed conflicts [37,38]. At the same time, conflicts in a region also tend to spread to surrounding areas [39]. Hence, we derived short-term conflict variables for each grid cell and its adjacent cells (up to eight), including the previous year’s SCDi, the number of conflicts, and the resulting fatalities. Additionally, we extracted historical conflict data, including the number of conflicts, fatalities, and conflicts from 1989 to 1999. To reflect the global conflict situation, we also extracted the total number of global conflicts and fatalities by year.

In terms of socioeconomics, economic downturns and overpopulation have also been identified as key predictive variables for conflict [40,41]. At the same time, population movements and ethnic divisions are important channels through which conflicts spread across regions [42,43]. Additionally, infrastructure such as roads are strategic targets in conflicts [44]. Therefore, we compiled socioeconomic variables from various public datasets to reflect the socioeconomic conditions in the study area. These variables include population and its annual change [45], gross domestic product and its annual change [46], HDI [47], net migration [48], excluded ethnic groups [49,50], road density [51], human footprint [52], and critical infrastructure index [53]. Furthermore, we extracted the number of peacekeeping operations per grid cell and surrounding areas from the Geocoded Peacekeeping Operations Dataset [54]. Using the Historic Land Dynamics Assessment+ Global Land Use Change between 1960 and 2019 dataset [55], we extracted annual proportions of urban and agricultural land from 2000 to 2019. To capture the global economic context, we obtained the annual global food price index and oil price index from the Food and Agriculture Organization [56]. Furthermore, we extracted the annual frequency of international sanctions imposed on each country, primarily focusing on economic sanctions from G20 countries, the European Union, and the United Nations due to their significant influence, and allocated these to the corresponding grid cells of the respective countries [57].

An increasing number of scholars agree that climate change, in certain contexts, can increase the risk of armed conflict [32]. At the same time, regions frequently affected by natural disasters also overlap with areas prone to armed conflict [58]. To assess the impact of climate change on the likelihood of conflict, we extracted the annual average temperature and precipitation, along with their standardized indices, for each grid cell from 1970 to 1999 via the Climatic Research Unit database [59]. Additionally, we incorporated the multi-hazard frequency from the Center for Hazards and Risk Research (CHRR) at Columbia University [60] to understand the potential influence of natural disasters on conflict patterns.

Due to the importance of terrain in military strategy, terrain is a key factor influencing conflict [61]. The geographic location also affects the development of conflicts, such as the distance to political centers, capitals, and national borders [39,62]. Additionally, surface covers like forests provide concealed spaces and resources, increasing the likelihood of prolonged conflicts [63]. Therefore, we compiled variables from various public datasets to reflect the geographic location and topographical features in the study area. This included longitude, latitude, mean elevation, elevation standard deviation, mountain coverage, travel times to cities, distance to nearest country anywhere, distance to own borders, and distance to capital, as provided by the PRIO-GRID system. Furthermore, the study included distance to the coastline [64], the NDVI [65,66], and the forest proportion derived from the aforementioned land use data.

In total, we compiled a dataset of 47 variables covering four key dimensions: conflict-related, socioeconomic, climate and disasters, and geographic, as presented in Table 1. To mitigate potential issues arising from highly collinear features and to ensure robust variable importance assessments during modeling, we computed the Pearson correlation coefficient matrix for all variables. As shown in Figure 2, the vast majority of variable pairs (93.80%) exhibited correlation coefficients below 0.3, with only a small proportion (0.83%) exceeding 0.7. Based on this analysis, we excluded six variables—elevation standard deviation, urban proportion, total number of conflicts in the vicinity last year, critical infrastructure index, total number of global conflicts last year, and oil price index—due to high collinearity. The remaining 41 variables were retained for the modeling process.

2.2. Construction of the SCDi and Its Transformation Types

Walther et al. [50] constructed the SCDi on the basis of conflict concentration and conflict intensity. Conflict concentration measures the spatial heterogeneity of violence distribution, characterizing whether events demonstrate clustered patterns or dispersed arrangements across the subregion’s geographical terrain. Conflict intensity quantifies the localized density of violent incidents within a specified subregion, reflecting the magnitude of violent engagements. The conflict concentration quantifies whether events in a cell are clustered or dispersed using nearest-neighbor distances,

$${CC}_{i,j,t}={\displaystyle \frac{{\overline{d}}_{i,j,t}}{d_{exp}}},$$

(1)

where ${\overline{d}}_{i,j,t}$ is the mean observed distance between each event and its nearest neighbor in grid cell $\left(i,j\right)$ in year $t$ and $d_{exp}$ is the expected mean distance under spatial randomness. The formula for calculating $d_{exp}$ is defined as follows:

$$d_{exp}={\displaystyle \frac{0.5}{\sqrt{N_{i,j,t}/A_{i,j}}}},$$

(2)

where $N_{i,j,t}$ is the number of violent events in cell $\left(i,j\right)$ during year $t$ and $A_{i,j}$ is the area of cell $\left(i,j\right)$. If ${CC}_{i,j,t}$ is less than 1, the spatial distribution of events is classified as clustered, meaning events are closer together than expected under random distribution; otherwise, it is classified as dispersed. The conflict intensity measures the relative density of violent events per unit area within a grid cell:

$${CI}_{i,j,t}={\displaystyle \frac{N_{i,j,t}}{A_{i,j}}}.$$

(3)

A grid cell is classified as experiencing high-intensity conflict if its ${CI}_{i,j,t}$ exceeds 4 events per 50 × 50 km² grid unit; otherwise, it is categorized as low-intensity conflict. In this study, we redefined conflict intensity by referring to the “Classification of Fragility and Conflict Situations for World Bank Group Engagement [67]”. The final conflict intensity is calculated as follows:

$${CI}_{i,j,t}={\displaystyle \frac{D_{i,j,t}}{P_{i,j,t}}},$$

(4)

where $D_{i,j,t}$ is the number of conflict-related deaths in grid cell $\left(i,j\right)$ in year $t$ and $P_{i,j,t}$ is the total population in grid cell $\left(i,j\right)$ in year $t$. A gird cell is classified as experiencing high-intensity conflict if its ${CI}_{i,j,t}$ exceeds 10 deaths per 100,000 population; otherwise, it is categorized as low-intensity conflict. This approach directly relates to the impact of conflict on human life, providing a more accurate reflection of the lethality and destructiveness of conflicts than merely counting the number of conflict events.

On this basis, grid cells can be categorized as no conflict (NC), clustered/high-intensity (CH), clustered/low-intensity (CL), dispersed/high-intensity (DH), or dispersed/low-intensity (DL) cells. Finally, building on these five types, we classified the annual changes in conflict types across grid cells into five types (

Table 2. SCDi transformation rules.

Transformation	Meaning	Type
NC → NC	Remain non-conflict	Maintaining peace
NC → DL NC → DH NC → CL NC → CH	From non-conflict state to conflict state	Conflict outbreak
DL → DL DH → DH CL → CL CH → CH DH → CH CH → DH DL → CL CL → DL	Maintain the original conflict level	Conflict stabilization
DL → NC DH → NC CL → NC CH → NC DH → DL CH → DL DH → CL CH → CL	Reduce conflict level	Conflict de-escalation
DL → DH; DL → CH; CL → DH; CL → CH;	Raise conflict level	Conflict escalation

), using changes in the number of conflict-related deaths as a critical indicator for capturing conflict transformations [18]: conflict outbreak, conflict stabilization, conflict escalation, maintaining peace, and conflict de-escalation. The specific transformation rules are as follows: When a grid cell is marked as NC both in the previous year and the current year, it is considered to have maintained peace. When a grid cell transforms from the previous year’s NC state to one of the DL, DH, CL, or CH types, it indicates the outbreak of conflict. If a grid cell is in conflict but its intensity remains unchanged, it is regarded as the stabilization of the conflict. When a grid cell is in conflict, a transformation from high intensity to low intensity is considered a de-escalation of conflict. Conversely, when a grid cell is in conflict, a transformation from low intensity to high intensity is considered an escalation of conflict.

2.3. Modeling the SCDi Transformation

We employ random forests as the base model and use the one-vs-rest (OvR) strategy for multi-class classification. The one-vs-rest strategy is a method that transforms a multi-class classification problem into a series of binary classification problems. For a classification problem with N classes, the OvR method constructs N binary classifiers, with each classifier responsible for distinguishing one class from all the other classes. The advantage of this approach is that it allows the use of existing binary classification algorithms to handle multi-class problems without any modifications to the algorithms themselves. Moreover, it provides independent probability estimates for each class, and each binary model can also aid in subsequent model interpretation.

Random forest (RF) is an ensemble learning method consisting of multiple decision trees, which are aggregated by voting or averaging their predictions. Each decision tree in the RF is constructed via bootstrap samples with replacements from the original data. The combination of bootstrapping aggregation and random feature selection helps reduce algorithm variance and prevent overfitting. In our random forest (RF) model, we optimized the hyperparameters to ensure robust performance for our dataset. Specifically, we conducted a grid search over key parameters, including the number of decision trees and the minimum leaf population, using 5-fold cross-validation to evaluate model performance. The grid search tested combinations of the number of trees (50, 100, 150, 200) and minimum leaf population (1, 2, 5). The results identified 150 decision trees and a minimum leaf population of 2 as the optimal configuration, providing a balance between predictive accuracy and computational efficiency. These parameters were used for all subsequent model training and evaluation.

For each year, we constructed an annual sample by randomly selecting an equal number of samples for each type on the basis of the minimum count among all conflict risk transformation types. The samples from 2001 to 2019 were then aggregated for model training. We repeated this sampling and modeling process 20 times to obtain 20 RF models. For the 20 ensemble RF models, we employed tenfold cross-validation to evaluate the final classification performance.

The performance of classification algorithms in predicting target types was evaluated via three common evaluation metrics: precision, recall, and the F1 score [68]. For the classification performance of a multiclass model, a confusion matrix was constructed for comprehensive evaluation. The confusion matrix is a $5\times 5matrix$, with rows representing the true classes and columns representing the predicted classes. By counting the true class and model-predicted class for each sample in the test dataset, we populated the confusion matrix. To compute precision and recall, we first needed to determine the classification outcomes for each class.

True-positive (TP) refers to samples that are correctly predicted as a certain class; false-positive (FP) refers to samples that are incorrectly predicted as that class while belonging to other classes; false-negative (FN) refers to samples that are incorrectly predicted as other classes while belonging to that class.

Precision is defined as the ratio of the TP to the total number of samples predicted as that class via the following formula:

$$Precision={\displaystyle \frac{TP}{TP+FP}}.$$

(5)

Recall is defined as the ratio of TP to the total number of samples actually belonging to that class, using the following formula:

$$Recall={\displaystyle \frac{TP}{TP+FN}}.$$

(6)

Precision measures the accuracy of a model in predicting positive samples and quantifies how well it covers true-positive samples. Finally, after considering both precision and recall comprehensively, we used the F1 score as a metric for assessing classifier performance. The F1 score is the harmonic mean of precision and recall and is calculated using the following formula:

$$F1={\displaystyle \frac{2\times Precision\times Recall}{Precision+Recall}}$$

(7)

The F1 score ranges between 0 and 1, with higher values indicating better model performance.

Finally, we utilize out-of-bag (OOB) error to assess variable importance [69]. In a random forest model, each tree is trained on a bootstrap sample (in-bag sample) randomly drawn from the data, leaving about one-third of the data as out-of-bag samples. The OOB error is the prediction error rate calculated on these out-of-bag samples. To measure a variable’s importance, we randomly permute its values and calculate the OOB error rate again, then compute the difference between the original and permuted OOB error rates. A substantial increase in OOB error indicates higher importance of the variable, as it shows the model heavily relies on this variable for accurate predictions. We randomly permute each variable 100 times and calculate the mean decrease in the accuracy score of the model after permutation, which serves as the permutation importance of that variable, ensuring statistical significance. Finally, we normalize the importance scores of all variables obtained.

3. Results

3.1. Spatiotemporal Characteristics of Armed Conflict Type Transformation

We divided the world into 52,259 grids for analysis, of which 48,116 grids did not experience any conflicts from 2000 to 2019. Therefore, this section focuses only on the remaining 4143 grids for analysis. On the basis of the SCDi calculations, we obtained the global SCDi types for 2000 and 2019, as shown in Figure 3a,b. Figure 3a shows that within these 4143 grid areas, approximately 5.91% exhibited CH armed conflicts, whereas approximately 2.82% of the regions had DH armed conflict characteristics. These regions were primarily concentrated in Colombia, Afghanistan and Central Africa. Furthermore, approximately 7.02% of the regions presented DL armed conflict features, whereas approximately 4.39% presented CL characteristics. These sporadic areas were scattered across India, Nepal, Colombia, Algeria and some countries in Central Africa. No armed conflicts occurred in the remaining regions during this year. Figure 3b shows the conflict characteristics in 2019, which underwent significant changes compared with those in 2000. Specifically, CH conflicts and DH conflicts were predominantly concentrated in Afghanistan, Syria, Mexico, and some countries in Central Africa, accounting for 8.59% and 2.94% of the 4143 grids, respectively. CL conflicts and DL conflicts were distributed mainly in India, Iraq, Mexico, and some countries in Central Africa, representing 6.90% and 11.44% of the research area, respectively.

According to the SCDi transformation rule, we obtained the evolving characteristics of conflicts from 2000 to 2019. Figure 3c shows that out of the 4143 grids under consideration, approximately 24.5% were identified as onset areas in 2019, predominantly concentrated in Afghanistan, Syria, Iraq, Yemen, and certain countries within Central Africa, such as Nigeria and Somalia. Approximately 16.8% of the regions experienced a reduction in conflicts and were primarily distributed across India, Nepal, and Central Africa. Conflict escalation areas accounted for only 1.3%, whereas conflict stabilization areas constituted approximately 2.1%, both of which were sporadically scattered within the conflict outbreak zones. The remaining regions exhibited a continued state of peace compared with the conditions observed in 2000.

From the perspective of the SCDi, Figure 3d illustrates the regions that underwent transformations between 2000 and 2019. This analysis focuses on the SCDi transformations of grid cells that experienced conflict in either 2000 or 2019. The main transformations observed include NC cells transforming to different types of conflict cells, such as DL (20.94%), CH (16.19%), CL (11.78%), and DH (6.41%), as well as different types of conflict cells transforming to NC cells, such as DL (12.04%), CH (8.46%), CL (7.20%), and DH (4.93%). There were also cases of mutual transformations or persistence of the same SCDi types, but their combined proportion was only 12.04%. These findings indicate significant changes in global geopolitical patterns over the past two decades, with a notable redistribution of conflict hotspots. An annual analysis can reveal the increasingly intricate transformational dynamics of armed conflicts observed over the past two decades. The findings from this annual examination are presented in Figures S1–S39, Texts S1–S4 and Tables S1 and S2, which demonstrate that armed conflicts, especially CH conflicts, have shown an overall increasing trend over the past two decades. Additionally, the outbreak, escalation, stabilization and de-escalation of conflicts present distinct spatial clustering characteristics.

3.2. Simulation Results of the Dynamic Transformations of SCDi Types

Based on the annual conflict transformation model on the global grid, the probability distribution of the annual transformation of SCDi types, compared with the previous year, was simulated via the random forest method coupled with covariates. Initially, the reliability of the simulation results was assessed using precision, recall and F1 score (

Table 3. Accuracy evaluation of simulation of the transformation characteristics of armed conflicts.

Types of Armed Conflict Transformation	Precision	Recall	F1-Score
Maintaining peace	0.918	0.878	0.897
Conflict outbreak	0.884	0.921	0.902
Conflict stabilization	0.669	0.598	0.631
Conflict de-escalation	0.544	0.545	0.544
Conflict escalation	0.587	0.648	0.615

), indicating that the model’s simulation results were credible. The model demonstrated significantly higher accuracy in simulating maintaining peace and conflict outbreaks, outperforming other types of conflict. The performance and analysis of other mainstream machine learning models, such as Support Vector Machine (SVM), Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Multi-Layer Perceptron (MLP), on the same training and testing datasets are presented in Table S1 and Text S5. Considering both performance and interpretability, random forest was determined to be the optimal model.

Figure 4 shows the spatial probability distributions of different types of conflict transformations in 2001 and 2019. The results indicate that there was a complementary trend between maintaining peace and conflict outbreaks, with both transformations having a relatively high global proportion. In regions experiencing conflicts, new conflicts were more likely to be triggered in surrounding areas. There were relatively few areas where conflicts had weakened or escalated, with these occurrences being sporadic and mainly concentrated in conflict-prone regions. When comparing the results from 2001 and 2019, it was observed that the number of regions maintaining peace decreased, whereas areas of armed conflict outbreaks in Afghanistan and the Middle East had a noticeable weakening trend, primarily due to ongoing armed conflicts in these regions. In Africa, areas of armed conflict outbreaks were increasingly concentrated in central regions. The spatial probability distributions of the armed conflict transformation types on an annual basis can be found in Figures S40–S58.

3.3. Potential Driving Factors of Different Transformation of Armed Conflict Types

Through dynamic simulation of armed conflict risk transformations, we identified the key driving factors influencing various types of SCDi transformations. As illustrated in Figure 5a, conflict-related factors had the most significant impact on every form of conflict transformation. Specifically, their impact on the five transformation types, ranked in descending order, is as follows: conflict outbreak (60.60%), maintaining peace (47.38%), conflict stabilization (42.48%), conflict escalation (39.87%), and conflict de-escalation (33.96%). In contrast, socioeconomic, climate and disasters, and geographic variables contribute less significantly, each typically accounting for 10–20% of the influence. Notably, the contributions of these three variable types peak during conflict de-escalation (20.31%, 23.72%, and 22.01%, respectively) and are lowest during conflict outbreak (13.55%, 11.92%, and 18.39%, respectively).

Figure 4b–f presents the top 10 variables for each transformation outcome, ranked by their decrease in accuracy score. Across all outcomes, conflict-related variables—namely SL, NCL, NCVL, FCVL, and FCL—consistently dominate the rankings, with SL emerging as the most critical predictor for every SCDi transformation type. This finding underscores that conflict dynamics serve as the primary drivers of all SCDi transformations. Socioeconomic variables, such as HDI, POP, GDP, and FPI, frequently appear in the top 10 across most outcomes but exhibit significantly lower importance scores (typically below 0.01). Their influence is most pronounced in conflict outbreak, stabilization, and maintaining peace, while their impact on conflict escalation and de-escalation is relatively minor. The HDI, developed by the United Nations, is a composite measure of life expectancy, education, and per capita income, reflecting a country’s overall social and economic development. By prioritizing human well-being, HDI serves as a key metric in assessing its influence on conflict dynamics. Geographic variables (LAT, LON) and climate-related variables (AAT, SPI, STI) also feature in the top 10 but with minimal importance scores (below 0.005), indicating that spatial and environmental factors play a secondary role in predicting conflict transformation outcomes. The presence of latitude and longitude (LAT, LON) in the top 10 for conflict outbreak, stabilization, de-escalation, and maintaining peace suggests their continued contribution to the model, potentially capturing spatial dependencies in conflict transformation patterns.

4. Discussion

Armed conflict is a dynamic social phenomenon. In this study, we comprehensively considered the concentration and intensity of armed conflicts and constructed the SCDi. By establishing transformation rules for the SCDi, we defined different types of transformations in armed conflicts to characterize their dynamic changes. Furthermore, by employing the random forest algorithm, we simulated the probability distribution of each type of transformation in global armed conflicts over the past 20 years and identified the potential driving factors behind different conflict transformations. This research offers a novel perspective for understanding the evolving trends in armed conflicts.

Numerous conflicts occur worldwide each year. However, their intensity and concentration vary, leading to varying societal impacts. We developed the SCDi, which classifies these conflicts into five types: CH, CL, DH, DL and NC. These findings demonstrate that four types of conflict (CH, CL, DH, DL) have exhibited an upwards trajectory over the past two decades, suggesting an increasing degree of global conflict risk. In particular, the rising trends are most evident for CH and DL, indicating a polarized development where both concentrated major confrontations and scattered minor clashes are on the rise. On the one hand, globalization increases tensions and fractious relations among states, increasing the likelihood of armed conflicts [20,70]. On the other hand, civilians increasingly play complex roles as both victims and perpetrators, leading to a rise in dispersed/low-intensity conflicts [71,72]. The relationship between the spatial patterns of armed conflicts—whether concentrated in dense clusters or dispersed across regions—and their escalation and de-escalation dynamics is complex, as shown by recent studies on conflict diffusion. Using high-resolution data and criminology-inspired micro-diffusion models, research identifies two key mechanisms: relocation diffusion, where violence shifts geographically without intensifying, and escalation diffusion, where conflict expands in scope and intensity [73]. Concentrated conflicts, like urban riots, escalate rapidly due to the contagious spread of violence in tight areas [74], while dispersed conflicts, such as Sahel insurgencies, persist at low intensity but resist de-escalation due to their diffuse nature [75]. Future research should refine spatiotemporal models to better distinguish concentration-driven versus dispersion-driven pathways, improving predictions of escalation risks and de-escalation opportunities.

Based on the SCDi transformation types, this study used the random forest algorithm to simulate the spatial probability distribution of each type of armed conflict transformation from 2001 to 2019. Compared with previous studies, this research reveals more details about the dynamics of armed conflicts. Previous studies have treated conflicts as a classification problem and have evaluated the spatial distribution of armed conflict risk without considering their intensity and clustering characteristics [34,76,77]. As a result, they have failed to distinguish between varying levels of conflict intensity and have overlooked dynamic changes in these conflicts. This study represents an attempt to simulate dynamic changes in armed conflicts, successfully uncovering details and patterns in these changes. The findings confirm both the clustering nature and diffusion characteristics of conflicts by revealing that regions with persistent conflicts gradually establish stable areas, predominantly concentrated in Syria, the Afghanistan border area, the Nepal border area, etc., while new conflicts are more likely to emerge around these regions [34,76,77].

The study of potential driving factors for armed conflicts has long been a prominent topic, with a focus on socioeconomic [78,79], resource environmental [25,80], and climatic disaster factors [32,81,82]. However, few studies have examined the driving factors behind conflict transformation. This study analyzes the driving factors of different types of conflict transformation. The findings reveal that different types of conflict transformation are influenced by distinct potential drivers. Nevertheless, it is clear that the prevailing conflict situation in local and neighboring regions remains the primary catalyst for conflict transformation, consistent with existing literature [38,83,84]. Interestingly, whether an area can sustain peace is shaped predominantly by the state of conflicts in adjacent areas. In contrast, for other dynamic transformations of conflicts, local conflict dynamics exert greater influence.

The present study provides a novel quantification of conflict transformation dynamics, offering the first simulation of probability distributions for various types of conflict transformation. Furthermore, it analyzed disparities in potential driving factors and impacts across these different types. However, there are areas where this study can be improved upon. While most current conflict prediction studies have primarily focused on local and regional drivers of conflict, the integration of global political factors into such models remains at an exploratory stage. First, data availability poses a significant constraint. High-resolution spatiotemporal data on global political decisions—such as international sanctions, foreign interventions, or shifts in diplomatic relations—are often difficult to systematically obtain or quantify. Second, the mechanistic complexity of how cross-national political dynamics interact with local conflict processes is inherently high. These interactions are often nonlinear and require more advanced, cross-scale modeling approaches to capture their full effect. Despite these challenges, acknowledging the role of global political influences is important for future model development. As data infrastructures and computational methods evolve, incorporating such factors could enhance the explanatory and predictive power of conflict models, especially in capturing sudden shifts or externally driven conflict escalations.

5. Conclusions

In this study, we proposed a method to simulate the dynamic transformations of armed conflicts. Firstly, we constructed the SCDi to identify different types of armed conflicts and analyzed the changes in global conflict patterns over the past 20 years. The results indicated a polarization in global armed conflicts over the past two decades, with a significant increase in both CH and DL conflicts. Subsequently, by establishing the SCDi transformation rules, we quantified the transformation types of armed conflicts, including conflict outbreak, conflict stabilization, conflict escalation, conflict de-escalation, and maintaining peace. Analysis of conflict transformation types over the past 20 years revealed a significant decrease in regions maintaining peace, while regions experiencing conflict escalation and stabilization continued to increase, indicating an increasingly dire global armed conflict situation.

Lastly, using machine learning methods coupled with covariates, we simulated the annual probability distribution of different conflict transformation types at a spatial scale and analyzed the potential driving factors behind these transformations. The results revealed that new conflicts often occur in the vicinity of conflict hotspots, shedding light on the clustering and diffusion characteristics of conflicts. Furthermore, compared to external factors, conflict-related factors emerged as the primary drivers of conflict transformations. The findings of this study offer a novel perspective for the study of conflict dynamic simulations.