Development of an Agent-Based Model to Evaluate Rural Public Policies in Medellín, Colombia

Abstract

Rural areas near large cities do not satisfy the food needs of the city’s population. In Medellín, Colombia, these areas satisfy only 2% of the city’s food needs, highlighting an urgent need to review and improve policies supporting agriculture. This study was conducted over a ten-year period since the release of the Medellín policy related to land use. The model uses agent-based modelling, geographic analysis and dichotomous variables, combining these structures to create a decision-making element and thus identify changes to examine in relation to current land use and detect properties with a potential for conversion to agricultural use. By evaluating post-processed geographic layers, land use in agricultural rural environments is prioritized, setting up clusters of homogeneous zones and finding new areas of rural influence. The implications of this study extend beyond Medellín, offering a model that can be applied to other regions facing similar challenges in agricultural productivity and land use. This research supports informed and effective decision-making in agricultural policy, contributing to improved food security and sustainable development. The results show that some properties are susceptible to policy changes and provide a framework for the revision of local regulations, serving as a support tool for decision-making in rural public policies by giving the local administration key factors to update in the current policies. The findings are relevant to local stakeholders, including policymakers and rural landowners, suggesting that several properties are susceptible to policy changes promoting agriculture and supporting informed decision-making in agricultural policy, contributing to food security and sustainable development. Also, this approach promotes efficient and sustainable agriculture, highlighting the importance of geographic analysis and agent-based modelling in policy planning and evaluation.

1. Introduction

Public policy plays a crucial role in shaping social realities and addressing collective needs. As defined by Torres-Melo and Santander [1], public policy is the result of collective action in the public sphere, involving a series of political transactions where the government’s role extends beyond local execution to ensure coordination and cooperation among key actors. This concept emphasizes two fundamental elements: the political and the public. In the public domain, societal goals are realized through governmental interventions or social actor participation, while the political aspect arises from the objectives set in response to societal needs [2]. Public policy thus represents a response to collectively constructed needs that capture governmental attention [3], reflecting the political choices and will around these collective objectives. However, despite these well-established principles, the challenges of adapting public policies to address the evolving environmental and economic contexts persist [4].

The agricultural sector requires public policies that can adapt to the diversity of agricultural practices, emerging rurality, and restructuring of public management. These policies are increasingly integrated into broader economic goals, focusing more on the spatial aspects than specific sectors. They aim to modernize crops, ensure food security, and generate resources to create a balance between urban and rural areas. However, as noted by Sy et al. [5], despite the advancements in public policies in this century, the agricultural sector still lacks the developments necessary to reach good levels. Current policies focus on fostering sustainable productivity growth, ensuring environmental protection, and adapting to climate change, particularly in the context of international markets and sustainable agri-food chains. This raises a key research question: how can agricultural policies be better aligned with specific environmental conditions to promote sustainability and economic efficiency? The hypothesis underlying this study posits that current policies may not fully account for regional environmental variability, leading to inefficiencies in soil use and reduced economic potential.

The integration of Geographical Information Systems (GISs) with Agent-Based Modelling (ABM) has emerged as a powerful tool for analysing these complex spatial dynamics and the policy impacts. Crooks et al. [6] highlight how this combination allows for a nuanced understanding of spatial patterns and processes, offering valuable insights into urban planning and environmental management. This approach is particularly relevant for evaluating agricultural policies in relation to specific environmental conditions, enabling an accurate assessment of soil use in alignment with the intended purposes and potential economic impacts. Recent studies, such as those by Xu et al. [7], show how geospatial technologies have been instrumental in shaping environmental policy responses, particularly in adapting to climate change. These advancements underscore the need for modern tools that can integrate spatial data into policy formulation, ensuring that land use decisions reflect both current and projected environmental realities.

In the context of sustainability and food security challenges, innovative approaches are required to ensure resilient and adaptable systems. Qazi and Al-Mhdawi [8] demonstrate the application of Bayesian Belief Networks to model sustainability and adaptation dynamics in global food security, emphasizing the importance of incorporating uncertainty and probabilistic reasoning into policy-making. This aligns with the need for adaptive strategies that can respond to the multifaceted risks posed by climate change, population growth, and resource depletion. Additionally, Castro and Sen [9] provide a view of how transformation policies must evolve in response to climate change, highlighting the growing intersection between environmental sustainability and development. The combination of these models can offer a robust framework for evaluating agricultural policies in a dynamically changing environment.

Rural transformation, a critical aspect of development, requires spatially differentiated approaches. Joachim et al. [10] explore such pathways for rural transformation in Pakistan, emphasizing the importance of tailored policies that account for regional disparities. Wang et al. [11] provide a comprehensive review of rural transformation studies, offering insights into definitions, measurements, and indicators that can guide research and policy formulation. These studies highlight the need for policies that recognize the heterogeneous nature of rural areas, where factors such as proximity to urban centres, access to markets, and land-use patterns significantly influence development trajectories. Incorporating machine learning methods in this context, as discussed by Gautron, Maillard et al. [12], can further enhance the precision of policy evaluation, particularly in optimizing land use for agricultural productivity.

Advanced data science techniques, such as natural language processing (NLP), offer new avenues for evaluating agricultural and rural policies. Domínguez et al. [13] illustrate how NLP can be applied to social network data to assess policy effectiveness and public perception. This approach allows for the extraction of meaningful insights from large volumes of unstructured data, providing a more nuanced understanding of policy impacts and stakeholder sentiments. Spatial analysis is increasingly recognized as a critical component of policy-making in both rural and urban contexts. Anríquez et al. [14] apply a spatial analysis to refine the scale of the rural–urban landscape in Chile, demonstrating its relevance to policy formulation. This approach enables a more precise and context-specific understanding of the challenges and opportunities within the rural–urban continuum, supporting the creation of more sustainable and equitable development strategies.

This study aims to develop an agent-based model that evaluates soil policies in relation to specific environmental conditions. This approach will enable a precise assessment of whether soil use is aligned with its intended purpose and how surrounding environments might influence or alter its economic potential, contributing to the ongoing efforts to improve agricultural policy design and implementation. Additionally, research by Forsyth et al. [15] and Cavazos et al. [16] underscores the importance of refining soil policies in areas as a response to both local and global environmental challenges. The current implementation, with some adaptation, can be reproduced in other cities with similar public policies by replicating the dichotomous process for those policies.

2. Materials and Methods

2.1. Literature Review, Study Area, and Methodology

Agent-based simulation is a moderately used tool for evaluating agricultural policies. Although these studies are not primarily aimed at evaluating agricultural policies, they offer valuable information from the existing literature on the previous and post-evaluations of agricultural policies. Huber et al. [17], analysed 20 European agricultural ABMs with a focus on standing for the decision-making process. Depending on the approach, the models include different elements at the farm level, which leads to the omission of values, social interactions, norm consideration, and learning from farmers’ decision-making in ABMs in Europe. Therefore, they suggested that new models should improve the representation of farmers’ decision-making and better represent the perspective of agricultural systems.

Kremmydas et al. [18], conducted a systematic literature review of the peer-reviewed journal articles that focused on the use of ABM in the evaluation of agricultural policies. They concluded that ABMs for evaluating agricultural policies can complement conventional agricultural models (based on equilibrium), as agricultural systems can be easily modelled and can include interactions between farm and market information. This type of modelling, although surpassing conventional modelling methods, faces difficulties in adoption by modelers and presents limitations in applicability to large-scale evaluations.

Mora-Herrera et al. [19] conducted a study like that of Kremmydas et al. [18]. In this study, the time window was from 2009 to 2019. It concludes that since 2009, ABM has been increasingly used to model agricultural systems, as it has proven to be a useful tool for integrating qualitative and quantitative information and has been an analytical framework for examining the society–nature relationship of agricultural systems under an interdisciplinary and transdisciplinary approach at both micro and macro levels.

Public policies in the city of Medellín, Colombia are set up through local regulations, which are defined for periods of four to eight years. Currently, Medellín has exceeded this period, making it necessary to review its policies oriented towards agricultural production. The goal is to create a model that verifies public policies for the agricultural benefit of the city, considering that Medellín’s rural area satisfies only 3% to 5% of the city’s food needs [20]. The main area of research in this study is based on a hybrid proposal that contains machine learning and agent-based simulation. The first belongs to the area of artificial intelligence and statistics and the second belongs to the area of operations research and computing algorithms. The model uses the city’s rural public policies in geographically post-processed layers with dichotomous characteristics, prioritizing land use in rural agricultural environments and finding properties where the evaluation of neighbourhood composition finds, based on physical variables, if the land use should change to accommodate an agricultural environment.

Medellín’s public policies are based on the Land Use Planning (POT) [21], which establishes rules for the use of Medellín’s territory. The POT is a local regulation that determines the use of the land, the current version has been in effect since 2014, and the territory has changed in this period. The current model uses the most recent geographic information for the year 2023 and combines this data with the 2014 POT. The data are available online on the official website of Medellín City [22] and are available for download.

Simulation is a widely used technique in operations research because it is a flexible, powerful, and intuitive tool. In a matter of seconds or minutes, years of operation of a system can be simulated, while generating a series of statistical observations about the performance of the system over this period. Due to its exceptional versatility, simulation has been applied in a wide variety of areas. More so, its horizons continue to expand due to great advances in the development of simulation software [23].

Agent-based simulation is a technique used to model complex systems, especially in the social field, allowing for a representation closer to reality by incorporating individual behaviour and interactions [11,24,25,26]. This methodology is distinguished by the creation of models where the components of the system are represented explicitly and individually, facilitating greater realism and scientific rigor [27,28]. Agent-based systems are characterized by the autonomy, heterogeneity, and independence of their agents, enabling the analysis of complex phenomena through the study of interactions and the emergence of new properties [29,30]. In addition, the approach allows us to understand phenomena such as social segregation and migratory patterns through emergence, where local interactions can generate complex patterns not evident at the individual level [31,32]. The implementation of the ODD protocol supports the validation of the simulation models, promoting the generation of solution scenarios will be proposed, which will be tested in an agent-oriented programming environment for a subsequent analysis of the results [33,34,35].

Agent-based modelling, when combined with spatial data, offers solutions to a diverse range of problems, such as traffic congestion analysis, planning for the spread of diseases, or optimizing resources like electrical energy and wood, through the simulation of decisions and movements within a defined space. Agents in these models make decisions based on their location and can influence or change their environment, which is analysed with spatial modelling tools like GISs [36]. These models enable the exploration of the dynamics between observed spatial patterns and the processes that create them, with a focus on spatial representation and analysis. The integration of GISs with agent-based simulation enhances the analysis of spatial data and improves the understanding of complex phenomena through a more dynamic interaction between the data and the models.

2.2. ODD+D Protocol

2.2.1. Descriptions of the Model Following the ODD+D Protocol

The ODD+D (Overview, Design concepts, Details, and Dynamics) protocol is used in agent-based simulation to build and document agent models. Its components are the Overview, which provides a summary of the model and its purpose, the Design concepts, which identify the entities, attributes, and relationships of the model, the Details, which describe the specific rules and algorithms used in the model, and the Dynamics, which describes how the model runs and behaves over time. The present model was developed in the software NetLogo (version 6.4) [37]. NetLogo is a multi-agent programmable modelling software used primarily for social and natural phenomena. It allows the modelers to create models where agents interact with each other and with their environment with rules coded into the logo language. The software has a user-friendly interface for creating complex systems where agents interact with the terrain and the other properties. The software is one of the most cited software to develop agent-based simulations [38].

Purpose

Seventy percent of Medellín’s land is rural and is characterized as a peasant economy area [21]. It has 52 villages divided into five townships, where around 12,500 families live on around 16,000 farms, with an average area of 1.51 ha per family productive unit and produce 29,000 tons of food per year, according to the FAO [20]. This shows that it is important for the city because it can serve as an alternative to the supply of fresh or value-added food. Therefore, to achieve rural development, a comprehensive intervention is needed. The strategic lines of each township vary, but all aim to improve the peasant economy, despite the disparities in development levels. A simulation model was created to assess soil policies and their impact on property units, using lot-level aggregation to determine land assignment accuracy and its economic potential.

Entities, State Variables, and Scales

The primary entities in the model are the system, the environment, and the agents. The entities are hierarchical and have their relationship within the territory. The hierarchy of agents is detailed in Figure 1 and Figure 2 where the rural land of Medellín is divided into blocks and lots, the latter being the smallest territorial divisions of the city, each of which is assigned a motionless agent that perceives its surroundings and inherits all the physical properties of its location; these agents are indicated with houses in Figure 2.

Agents

The system handles relating the soil layers obtained from maps of the land-use plan (POT) and they are set up as the land policies and these properties are transferred to the agent who is in the centroid of each lot or parcel. The agents are created as objects related to the lot or parcel where it will have the ability to interact with its neighbours according to the initialization criteria of the model and according to its characteristics, whether residential or without constructions; it will also have the ability to adapt to its environment. The agent variables include the geographic unit code (CBML), summation and average of the public policy field sums, average lot area, and current lot usage. It also identifies whether the property’s vocation changes or if the batch was processed. The model uses K-means and Agglomerative Clustering to classify properties into clusters ranging from 8 to 15. Additional model variables (M) evaluate soil classification, equipment, hydrography, planning, protection, production, and land value, with the possibility of defining a new cluster if needed.

Environment

The model environment is a NetLogo grid based on a 30 × 30 patch mesh and is configured as a corner location with a start at the bottom left. This is related to the 380 km² of extension of Medellín. The patch-related environment links only a commune encoding, an identifier, and a colour. The commune describes the rural area, and urban areas are coded from 1 to 16. Rural areas are coded 50, 60, 70, 80, and 90. Using NetLogo’s GIS plug-in allows us to link geographic objects as patches to the model, and it is this relationship that supports the linking of an agent to each vector object in the model of the lots. The geographic elements of the model correspond to the lots in the rural commune (district). This information is available in an open data service called GeoMedellin [22] and POT [21]. The initialization of the model is shown in Figure 3.

Process Overview & Scheduling

The initialization of the model is based on three parts, cleaning and initialization of variables, identification and characterization of agents, and assigning agents to lots or parcels. With these procedures performed where the variables are initialized, the agents are created in each centroid of the patch, the x and y coordinates in the variables of each agent are assigned to a Boolean variable that defines whether the agent was scanned or not, and finally the POT data.

• In the decision variables of the model, the model allows you to control variables to create variations of the model the variables are described as follows: The variable Cluster Type (ClusterType) determines the model’s approach, choosing between the classification of clusters using either K-means or Hierarchical methods. The Number of Clusters (QuantityClusters) specifies the number of clusters the model will use, which can range from 8 to 15. The Minimum Area (MinimumArea) defines the smallest area that must be considered in the model; a larger area will result in fewer lots being included, while a smaller area will increase the number of lots. Finally, the Properties in a defined radius (PropertiesInRadiusPatch) variable sets the number of properties within a given radius for the model, with values ranging from 0.1 to 2.0.

After initialization, the model operates annually, where each step represents a year in which agents make decisions influenced by their surroundings. The model is structured into three primary components: (1) configuring agents and patches using geographic data to define classifications, (2) calculating the environment based on the number of agents within a specified radius, and (3) reporting changes according to the agents’ decisions. The initialization process involves setting up the environment, ensuring that any residual elements from previous simulations are cleared, and correctly initializing all entities. Geographic information for Medellín is loaded with rural data represented at the lot level, while urban data are excluded from the model.

In the NetLogo environment, each geographic patch (lot) is represented as an agent. The model uses classifiers, such as K-means or Hierarchical methods, to subdivide these patches into clusters ranging from one to fifteen. The other clusters are assigned distinct colours, allowing for clear differentiation. The process includes defining each patch’s colour based on its cluster. The model creates agents within each geographic object (lot). This process begins by eliminating any existing agents to prevent interference from previous simulations, then the model evaluates each geographic element, identifies its centroid, and assigns data such as location coordinates, lot area, cluster value, and various policy-related metrics. These variables are crucial for assessing the agents’ behaviour as the model runs, allowing for a detailed analysis of how public policies and geographic factors influence decision-making processes.

The values from the calculated variables from the public policies of the POT, the Property Use Code, where we are only interested if it is 1 (residential) or 9 (an unbuilt lot), and the Boolean variables “Change” and “Scan” that start falsely only to be edited later by the model, create the sum of its public policy components. A patch with good policies should add up to 10 in value, comprising all the dichotomous variables, and an average closer to 1 is better, which we will define in the execution of the model. With this procedure, the agent already has all the necessary variables to be evaluated.

This model is an environment-based model; therefore, the next procedure handles defining the environment for each agent and if it is susceptible to change clusters. The procedure performed by each agent is as follows (these steps can be configurated in Figure 3):

• It is only allowed to evaluate those lots that have an acceptable environment for the change; therefore, if they comply with the public policy average which is greater than 0.8 [39], the sum of the columns is greater than 8, and the lot area is greater than the minimum lot area previously defined, then we proceed.
• We proceed to evaluate if the type of property is susceptible to change; for this, the lot must be residential or not built on.
• Then each patch is evaluated in its respective type of classification model according to the deciding variable “ClusterQuantity” where the model to be used is defined as either that of 8, 12, or 15 clusters to evaluate its surroundings.
• The code starts by checking whether the variable “ClusterQuantity” is equal to a value selected in the decision variable and continues to evaluate the procedure for that number of clusters.
• After the number of clusters, a “while” loop is defined that executes through the variable “i” that is defined as the value of the cluster from zero to j which is the maximum number of clusters per execution, and j varies from 8 to 15 while “i” varies from 1 to the value of “j”.
• The number of turtles that have a cluster value equal to “i” and that are within a radius defined by the PropertiesInRadius variable is counted. This count is stored in a local variable called CantTurtlesRadio.
• Then, the value of CantTurtlesRadio is added to the end of a list (vector) and then increment the variable “i” by 1.
• Once the “while” loop is over, the code calculates the maximum value in the list and stores it in the “q” time variable. It then finds the position of this maximum value within the list and stores it in the result variable “w”.
• The code then checks to see if “w” is different from the current cluster value. If so, the value of the NewCluster variable is used to update “w” and a cluster change is found and the Change? Variable is set to true.
• Finally, regardless of the result, all the turtles have the variable Scanned? in true.
• When all the agents (patches) are finished, the model is finished.

This procedure is described in the pseudocode below (Algorithm 1).

Algorithm 1. Cluster change pseudocode

IF QuantityClusters = [8,9,10,11,12,13,14,15] THEN   WHILE i < j DO     CantTurtlesRadio < −COUNTING TURTLES WITH CLUSTER = i in RADIUS     x < −x ADD CantTurtlesRadio     i < −i + 1   END WHILE     q < −MAXIMUM VALUE of x   w < −POSITION of q IN x     IF w ≠ CLUSTER THEN     NewCluster < −in     Changes? < −TRUE   END YES     Scanned? < −TRUE END YES

2.2.2. Design Concepts

Only those that contribute to the model are described, those that are not detailed do not affect the modelling or are not present.

Basic Principles

This agent-based simulation model is developed to address the problem of structuring the public policies detailed in the Land-Use Plan (POT). These policies are described through dichotomous variables obtained from geographic objects. The model is conceived as an extension of the geographic properties, including the location and classification of lots. This allows you to decide whether the classification and its later cluster change are correctly assigned or if specific adjustments are needed. Some farms may need a rural revitalization process to promote agricultural production.

Adaptation

The agents adapt to the related environments as described in the Process Overview & Scheduling Section, where the algorithm uses a vectorized function to validate the suitability of the agent in its cluster and decides to change it as soon as its environment does not favour its development if it can change, i.e. if it is a residential or unbuilt property (a lot).

Learning

The agents have a learning procedure based on success, in the Process Overview & Scheduling Section the algorithm is detailed, in which a list is created where all the favourable environments to the agent are saved, and then the agent makes a decision, based on its entire environment, to choose the best changes based on its experience (the whole list). This informed decision is based on what the agent learned and ensures the agent makes changes to the cluster that benefits it.

Detection

The agents have a detection procedure based on a variable defined by the user called the “PropertiesInRadiusPatch”, this variable establishes a radius defined by pixels detailed in the configuration image of the model where there is a horizontal extension of 27,303 m, where a patch equals 880 m on a side and a pixel equals 32 m on a side (all are sides for squares), and these metrics determine the way in which the number of properties is related to the radius of patches. Therefore, it is determined in units that the values 0.1, 0.5, 1.0, 1.5, and 2.0 are equivalent to 88 m, 440 m, 880 m, 1320 m, and 1760 m, which defines its surroundings. In this, it is evaluated how many properties exist with these conditions in that radius.

Stochasticity

The agents have a stochastic detection procedure provided by NetLogo where each agent is randomly scanned in a uniform manner where each agent changes and registers randomly, which allows real-time interaction between the elements.

2.2.3. Initialization

The initial configuration of the model allows the combination of four variables and a fifth that allows the visualization of the environment of the agents if activated, the variables that are visualized in Figure 3 and are described below as follows:

ClusterType, options [K-means, Hierarchical], allows you to choose the Machine Learning classification model for the simulation.

ClusterQuantity, options [8,9,10,11,12,13,14,15], allows you to choose the amount of cluster for the simulation.

ShowTurtles, [On, Off] options, under the on choice, allows you to display the agents as houses in the centroid of each lot.

MinimumArea, options [100, 200, 500, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000], allows you to choose the amount of area of each patch so that it can be included in the model to be detailed when comparing its environment. If it is less than this value, it will not be included.

PropertiesInRadiusPatch, options [0.1, 0.5, 1, 1.5, 2], allow you to choose the number of properties in a measured radius at the patch distance described in the detection choice. Also, it allows you to validate how many properties are within a radius of proximity.

The individual execution of the model allows the manual configuration of these elements with the execution of the “Go” button, and allows the execution and export of the results with the “Export” button. Now the individual configuration is different from the generality of the model where a configuration is made when the generalized processes are run 90 times in the following configurations detailed below [40].

[“QuantityClusters” 8 12 15]

[“MinimumArea“ 100 1000 10,000]

[“PropertiesInRadiusPatch“ 0.1 0.5 1 1.5 2]

[“ClusterType” “K-means” “Hierarchic”]

These results cannot be processed directly in the computer’s RAM because they are geographical elements that generate a high load on the memory. Therefore, each run must be exported to a flat file, which is saved with the name of the configuration. Then, through a Python script, these results are joined with a data frame. The modeler chooses the number of clusters, minimum area, radius, and cluster type that maximizes changes per structure. In addition, geographical spaces are defined to show the quality of the public policies related to this space.

In other words, these lots with this configuration require an additional policy that helps increase the potential for change or administrative support to improve their participation in agricultural activities, assuming this is the situation, as seen in Figure 3 and Figure 4.

2.2.4. Input Data

The model does not have input data that stands for changes or variations over time. Each agent holds the information needed to make decisions or conduct variations, as described in the simulation in process. Each execution of the model is a year since the data aggregation processes are considered a cross-section of the cadastral information that refers to the data annually.

3. Results

The data for the model, as declared in Section 2.2, were downloaded from the official homepage of Alcaldia de Medellín, a government site where the POT is hosted. The POT was downloaded and processed using the QGIS software (version 3.38) [41] and merged using the Join Attributes by Location tool and using the political division layer as a base; every type or restriction represents a geographic layer used in this process. This procedure accumulates attributes by the location of each policy from the Land-Use Plan until a single layer with all the attributes is generated through the union of the earlier layers. In each layer, a dichotomous value is defined, if it helps the agricultural process, a TRUE value is assigned, if not a FALSE value is assigned. The purpose is to create a model of zonal grouping or classification, where the characteristics of the majority decide the behaviour of the other properties; in other words, if in a cluster or sector where the majority behaves or is oriented towards agricultural development, the other properties in that cluster can also show this behaviour.

The soil classification system for each layer described previously categorizes the diverse types of terrain based on their suitability for a specific use. For example, Suburban soils (soil for household constructions), are characterized by their proximity to existing infrastructure and good access to roads and utilities and are suitable for urban development, including housing, commercial, and industrial projects. Areas with readily available infrastructure and central locations make them the preferred choice for the location of public facilities, such as schools, hospitals, and police stations. Rural habitable centres with fertile soil and water resources are well suited for agricultural and residential use. Similarly, areas with good access to services were appointed to rural housing development. Finally, other soils not falling into these categories have the potential for various uses, such as agriculture, forestry, or conservation efforts, depending on their specific characteristics and environmental suitability.

The process of assigning dichotomous values to each layer of the POT was conducted on each of the integral elements of the rule, an example is described below using a flowchart as seen in Figure 5. This encompasses diverse land uses, including areas undergoing mining transition, agricultural production, agroforestry, agricultural livestock production, and forest production. Also, soils without a specific production classification are included in this category. It distinguishes subcategories such as Agroforestry, Agricultural Production, Agricultural Livestock Production, Forest Production, Areas in Mining Transition, and Not Applicable (N/A), and clearly indicates whether each subcategory aligns with the criteria of being an area for protection and production, offering a comprehensive overview of the diverse uses within this soil category (PP stands for Protection and Production). This is an example of how production areas are evaluated as dichotomous values when the soil is suitable for agricultural production and when it is not, assigning true and false values [21]. The process described previously is shown in Figure 5.

The model finds properties that have changed policy and sets up clusters of homogeneous zones in which new areas of rural influence are found. This model supports the evaluation of rural public policies about upcoming local regulations. This method was developed using the ODD protocol described in the next section. The Clustering of the soil is conducted through Machine Learning algorithms such as K-means and Agglomerative Clustering. These algorithms were applied to the geographical attributes to find the cluster to which each property was assigned or classified. K-means is a technique that works iteratively to assign each data point to one of the K clusters based on the distance between the points and centroids of the clusters. On the other hand, Agglomerative Clustering is an algorithm that starts with each data point in its own group and, at each step, merges the closest groups based on the distance between them, until all points are in a distinct group or in predefined K groups [42].

The purpose of this model is to quantify the number of cluster changes per property unit, using an agent-based model that evaluates the conditions related to geographic objects and their radius of influence. Next, the results of those properties that have presented changes in their cluster will be analysed, and the instances where the agents are comfortable in that cluster, but prefer to change within the radius of influence with respect to other properties that have a better classification are also analysed.

Figure 6 shows the changes in the number of clusters chosen. In this case, the options are [“ClusterQuantity” 8 12 15]. The use of K-means allows for the most changes, with almost a 10% difference between the perceived changes using the Hierarchical model and the model using K-means. This graph allows us to find the model that registers the most changes and presents the 58,772 changes among all the possible configurations, compared to 48,580 for the model that uses a Hierarchical classifier. This difference of 10,192 changes leads us to choose the K-means model, which provides better results. Mora-Herrera et al., [19] in their work, detail the interdisciplinary and transdisciplinary structure and the approach using both micro and macro level models. In the present model, the classification is made at the micro level while the behaviours of cluster changes are interpreted with the macro level approach, where new clusters generate continuous blocks of classification. From this point on, we will only consider this model for the results.

As shown in

Table 1. Possible configurations for K-means.

MinimumArea (MA)	PropertiesInRadius (PRP)	QuantityClusters (CC)	ConfigurationModel
10,000	0.1	12	AM_10000|PRP_0.1|CC_12
10,000	0.1	15	AM_10000|PRP_0.1|CC_15
10,000	0.1	8	AM_10000|PRP_0.1|CC_8
10,000	0.5	12	AM_10000|PRP_0.5|CC_12
10,000	0.5	15	AM_10000|PRP_0.5|CC_15
10,000	0.5	8	AM_10000|PRP_0.5|CC_8
10,000	1.5	12	AM_10000|PRP_1.5|CC_12
10,000	1.5	15	AM_10000|PRP_1.5|CC_15
10,000	1.5	8	AM_10000|PRP_1.5|CC_8
10,000	1.0	12	AM_10000|PRP_1|CC_12
10,000	1.0	15	AM_10000|PRP_1|CC_15
10,000	1.0	8	AM_10000|PRP_1|CC_8
10,000	2.0	12	AM_10000|PRP_2|CC_12
10,000	2.0	15	AM_10000|PRP_2|CC_15
10,000	2.0	8	AM_10000|PRP_2|CC_8
1000	0.1	12	AM_1000|PRP_0.1|CC_12
1000	0.1	15	AM_1000|PRP_0.1|CC_15
1000	0.1	8	AM_1000|PRP_0.1|CC_8
1000	0.5	12	AM_1000|PRP_0.5|CC_12
1000	0.5	15	AM_1000|PRP_0.5|CC_15
1000	0.5	8	AM_1000|PRP_0.5|CC_8
1000	1.5	12	AM_1000|PRP_1.5|CC_12
1000	1.5	15	AM_1000|PRP_1.5|CC_15
1000	1.5	8	AM_1000|PRP_1.5|CC_8
1000	1.0	12	AM_1000|PRP_1|CC_12
1000	1.0	15	AM_1000|PRP_1|CC_15
1000	1.0	8	AM_1000|PRP_1|CC_8
1000	2.0	12	AM_1000|PRP_2|CC_12
1000	2.0	15	AM_1000|PRP_2|CC_15
1000	2.0	8	AM_1000|PRP_2|CC_8
100	0.1	12	AM_100|PRP_0.1|CC_12
100	0.1	15	AM_100|PRP_0.1|CC_15
100	0.1	8	AM_100|PRP_0.1|CC_8
100	0.5	12	AM_100|PRP_0.5|CC_12
100	0.5	15	AM_100|PRP_0.5|CC_15
100	0.5	8	AM_100|PRP_0.5|CC_8
100	1.5	12	AM_100|PRP_1.5|CC_12
100	1.5	15	AM_100|PRP_1.5|CC_15
100	1.5	8	AM_100|PRP_1.5|CC_8
100	1.0	12	AM_100|PRP_1|CC_12
100	1.0	15	AM_100|PRP_1|CC_15
100	1.0	8	AM_100|PRP_1|CC_8
100	2.0	12	AM_100|PRP_2|CC_12
100	2.0	15	AM_100|PRP_2|CC_15
100	2.0	8	AM_100|PRP_2|CC_8

, you have the 45 possible configurations for the K-means model; the model configuration field will identify which of the detailed configurations (45) presents the most changes.

Of the number of possible outcomes, we will detail the first 10 of the 45 possible configurations of the K-means model, which was chosen to report with because it had the greatest number of changes.

The first ten are presented in

Table 2. Ten best settings in terms of total changes.

Configuration	50	60	70	80	90	Total
AM_100|PRP_2|CC_15	199	1546	117	1064	282	3208
AM_100|PRP_1.5|CC_15	199	1545	117	1062	270	3193
AM_100|PRP_1|CC_15	193	1531	115	1057	257	3153
AM_100|PRP_0.5|CC_15	184	1491	95	1051	230	3051
AM_100|PRP_2|CC_12	135	1183	110	820	257	2505
AM_100|PRP_1.5|CC_12	135	1183	109	821	242	2490
AM_100|PRP_1|CC_12	133	1180	111	815	231	2470
AM_100|PRP_0.5|CC_12	124	1144	88	804	204	2364
AM_100|PRP_0.1|CC_15	65	928	32	830	99	1954
AM_1000|PRP_2|CC_15	196	897	107	276	271	1747

. The table shows that the preferred number of clusters is 15 and 12 in the configurations. The greater the number of clusters, the better the probability of change in these categories. However, looking at the first five in the fifth configuration with the most changes, you find one of twelve clusters. This configuration is repeated in three more configurations, reaching four out of ten possible configurations.

In addition, the configuration that generates the most changes sets the minimum area to 100 m², which is contrary to the radius, which varies between the different configurations. Three of them have the longest distance of two units, while only two configurations have the second-largest distance of 1.5 units. The rest of the configurations have values equal to or less than one unit. This distribution of the radius among all possible configurations provides transparency and confidence to the model and the processes conducted. The changes are related to the variability of the configurations, proving that current policies have room for improvement. In addition, it provides an understanding that nearby properties can positively influence the sectional classification of properties. The properties change without the need to prioritize a defined radius of influence but considering their environment and changing, in detail, all its aspects. This makes it possible to classify the properties with room for improvement and action towards rural development.

4. Discussion

In the image in Figure 7, we can see that the properties with cluster changes for the first ten configurations have a linear relationship. For example, in Santa Elena on the right side of the map (eastern) the cluster of properties changes and are grouped according to their development and policies; additionally, although Santa Elena is recently considered a rural land oriented to residential development, the field of improvement is visualized in sectors with residential components for the agricultural field. Huber et al. [17], as in the present study, use farms as an object of study; however, in our case, each lot or territory corresponds to the structure that the author describes and relates to the decision of the agents in the model. On the other side of the city, in the southwestern sector of San Antonio de Prado, the sector related to the urban environment does not change, but its periphery is related to the improvement of environments towards agricultural production. And, in the other three townships, Altavista, Palmitas and San Cristóbal, most of the properties are visualized with cluster changes that are related to the proportions of the environment and the sizes of the properties that are defined there. Therefore, the agricultural potential of the city of Medellín, according to the new clusters and the potential for improvement with respect to current public policies, is distributed in these three locations. The administrative entities will comment on the improvements in this area of development in the face of an update to the Territorial Planning Plan.

The findings of this study emphasize the critical gap in the food supply chain from rural areas in Medellín, Colombia, where these areas only satisfy 2% of the city’s food needs. This highlights an urgent need to review and improve agricultural policies to enhance food security and sustainable land use. By using ABM and geographic analysis, this study shows properties with potential for agricultural change, setting up a framework for land-use prioritization. This approach is consistent with recent research using ABM and geographic methods for land-use optimization, as seen in the works by Zhao et al. [43], Malik et al. [44], and Bai et al. [45].

In comparison to Zhao et al. [43], who applied a data-driven ABM to assess land consolidation potential in rural areas, this study focuses on enhancing agricultural land use through policy-driven changes. Zhao’s work used cadastral data to evaluate land consolidation complexity, providing insights into land optimization. While our study goes beyond mere land consolidation by detecting properties specifically for agricultural use and the present model also applies cadastral data. Both approaches share a common goal: improving land use through modern, data-driven tools to enhance food security. However, while Zhao’s work emphasizes land consolidation, our focus is on agricultural development and its potential to satisfy urban food demands.

Additionally, Malik et al. [44] applied analytical and deep learning methods to assess agricultural vulnerability to climate change in Jammu, Kashmir, and Ladakh. Although the focus of their study was the impact of climate stress on agriculture, their method shares commonalities with our research in terms of analysing geographic and socio-economic data to inform agricultural policies. Malik et al. found high-risk areas for agricultural decline due to climate stress, while our study shows areas that can be enhanced through policy changes to increase agricultural productivity. Both approaches emphasize the importance of data-driven decision-making in the development of resilient agricultural systems.

Bai et al. [45] explored land-use changes in Inner Mongolia using multi-agent systems to simulate the effects of socio-economic and natural factors on farmland use. Like Bai et al., our study leverages ABM to simulate potential land-use changes based on geographic and socio-political factors. While Bai’s research focused on household behaviour and the abandonment of farmland, our research shows properties suitable for conversion to agricultural use. Both studies prove the ability of ABM in understanding the dynamics of land use and how targeted policy interventions can lead to sustainable outcomes.

Furthermore, Ralha et al. [46] used a multi-agent model to simulate land-use changes in the Brazilian Cerrado, focusing on environmental policies and land consolidation. Ralha’s approach, like ours, relied on ABM to simulate the impact of land-use policies on agricultural regions. However, while Ralha’s model was used to predict environmental outcomes, our model prioritizes agricultural productivity and land conversion, offering a framework for policymakers to revise local agricultural policies. Both studies emphasize the role of ABM in guiding policy decisions, especially in regions facing pressure from urbanization and environmental challenges.

The integration of agent-based modelling in this study aligns with the methodological approaches seen in other works. For instance, González-Méndez et al. [47] used ABM to model urban development planning based on human needs, highlighting how agents can account for complex social dynamics in land-use decisions. Our research applies a similar ABM framework but focuses on agricultural land use in peri-urban areas, simulating how policy changes can encourage agricultural expansion. Both studies highlight the versatility of ABM in addressing diverse land-use challenges, whether in urban or rural settings.

Another study, by Maciąg et al. [48], introduced a GIS-based algorithm to evaluate land consolidation using cadastral data, like our use of geographic analysis to assess land-use potential. Maciąg et al. focused on improving land consolidation processes in Poland, while our study emphasizes agricultural land conversion in Colombia. Both studies use GIS tools to inform land-use policies, but our research is distinct in its focus on increasing agricultural productivity to address food security in urban areas.

Additionally, the work of Stan and Sánchez-Azofeifa [49], which simulates land cover changes in the Edmonton–Calgary corridor under different government intervention scenarios, parallels our findings about the impact of policy-driven land-use changes. Like Stan’s study, which examined urban expansion and agricultural land protection, our research explores how policy interventions can transform rural land use to meet urban food needs. Both studies provide valuable insights into how policy reforms can reshape agricultural landscapes and contribute to sustainable development.

The present model is primarily an approximation based on the current public policy regulations and presents data spanning over 10 years of use. The model introduces new classifications for properties with policies that are not suitable for agricultural use. It is intended to apply the model in collaboration with the local government, and it is expected that land oriented towards territorial speculation, used to influence the buying and selling of lots, could be adjusted to an agricultural utilization strategy, or otherwise, a penalty could be applied. Additionally, large lots of land designated for residential use could be oriented towards self-sustainability under tax incentives based on the new policies.

In conclusion, the present work contributes to the growing body of literature on sustainable land use and agricultural policy by providing a flexible, transferable model for improving agricultural land use near urban areas. By incorporating geographic analysis and agent-based modelling, the research supports informed decision-making that can enhance food security and sustainable development. The results of this study, alongside the comparative insights from other works, emphasize the importance of policy-driven approaches to land-use planning.

Future research could further refine the model by incorporating climate variability and socio-economic factors, building on the methodologies used by Malik et al. [44] and Bai et al. [45]. This would create an even more comprehensive framework for addressing the complex challenges facing rural land use and agricultural productivity in rapidly urbanizing regions.

5. Conclusions

This study focuses on the development of an agent-based model to evaluate soil policies in Medellín, Colombia, with the goal of optimizing agricultural production. The critical finding of this study is the urgent need for revision and enhancement of agricultural policies, as Medellín currently satisfies only 2% of its food requirements locally. Using geographic analysis and dichotomous variables enables an effective evaluation of current land use and identification of properties, which could help us understand their different agricultural purposes.

By analysing post-processed geographic layers, the model assists in prioritizing land use in agricultural rural areas, establishing clusters of homogeneous zones, and pointing out new areas of rural influence. This approach provides a framework to guide future revisions of local regulations, supporting decision-makers in enhancing rural public policies. Consequently, the agent-based modelling approach promotes more efficient and sustainable agricultural practices, describing the importance of precise and localized policy evaluations to improving agricultural output and sustainability.

The use of ABM in this study has proven to be an effective tool for evaluating agricultural policies, particularly in the context of rural land-use changes. By simulating interactions between different agents within the agricultural system, the model captures the dynamic nature of policy interventions on land-use decisions. This approach builds on existing research by Huber et al. [17], and Mora-Herrera et al. [19], who highlight the importance of ABM in representing land decision-making and integrating complex variables such as geographic and policies data. In this study, ABM provides a comprehensive framework for exploring how targeted policy changes can optimize land use for agricultural purposes, particularly in the peri-urban areas of Medellín.

The integration of the GIS with ABM enhances the model’s ability to analyse spatial patterns and predict land-use changes. By combining geographic data with public policy regulations, the model allows for a more accurate assessment of rural properties that are suitable for agricultural development. This spatial analysis provides valuable insights into the dynamics of land use, highlighting areas where policy interventions can significantly improve agricultural productivity.

This study also demonstrates the flexibility of the model, as customizable parameters such as the number of clusters and minimum area allow for tailored analyses of different land-use scenarios. This adaptability is crucial for effective policy decision-making, as it enables policymakers to explore multiple outcomes and select the best strategies based on real-time data and evolving conditions. The ability to adjust these parameters enhances the model’s relevance for various geographic and socio-economic contexts, making it a valuable tool for supporting rural development beyond the Medellín region.

This study provides a clear pathway for policy recommendations aimed at improving agricultural productivity in Medellín’s rural areas. The findings suggest that certain properties are highly susceptible to reclassification and agricultural conversion, offering opportunities for public policy revisions that could significantly boost food security. Also, the model highlights areas for improvement, particularly in terms of reclassifying land to support sustainable agricultural practices. These insights offer a practical framework for policymakers to enhance rural development and address the growing food needs of the city.

As future work, the incorporation of other units of measurement to agricultural development environments is planned, such as the social and fiscal factor and its infrastructure. For example, for the social factor, the component of well-being and economic state of the people in rural sectors could be used and how this improves or affects agricultural performance would be evaluated. For the fiscal sector, the model could be developed around the equitable and well-distributed application of land taxes such as the property tax, which in Medellín is unified and allows its payment in four quarters without interest. Finally, for infrastructure development, the underdevelopment of road infrastructure in the three districts with the least number of access roads could be analysed, which coincidentally are the ones with the greatest amount of development, i.e., Altavista, Palmitas, and San Cristóbal.

Public policies in Colombia establish the use of land and link this use to compensation for its usage through a tax. The process conducted to create the public policies of the city of Medellín, which involves converting geographical values into dichotomous data as seen in Figure 5, can be applied to other territories, and then aggregated into geographical layers. To adapt the model to another territory, the actual procedure must be followed, prioritizing the characteristic of use or application. For the present exercise, agricultural use is encouraged, but other aspects can also be addressed.