Next Article in Journal
Study on Seismic Response of Segmented Utility Tunnels Crossing Ground Fissures
Previous Article in Journal
Sex-Based Differences at Ventilatory Thresholds in Trained Runners
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Improving Wildfire Simulations via Geometric Primitive Analysis in Noisy Crowdsourced Data

by
Ioannis Karakonstantis
and
George Xylomenos
*
Department of Informatics, Athens University of Economics and Business, 10434 Athens, Greece
*
Author to whom correspondence should be addressed.
Appl. Sci. 2025, 15(16), 8844; https://doi.org/10.3390/app15168844
Submission received: 16 July 2025 / Revised: 7 August 2025 / Accepted: 8 August 2025 / Published: 11 August 2025
(This article belongs to the Section Computing and Artificial Intelligence)

Abstract

A key challenge in real-time wildfire simulation is data acquisition from dynamic sources, such as user-submitted data collected via mobile phones. Information obtained from firefighting personnel in the field, or even bystanders, typically outperforms pre-existing information in terms of its spatial and time resolution and can be used to execute more accurate fire simulations; these can be continuously updated as new data are added. However, combining data from users with heterogeneous knowledge backgrounds and biased conceptual barriers introduces additional distortion to what we know about an evolving wildfire. We examine the problem of resolving geometric ambiguity, where users submit duplicate or distorted spatial entries about a modeled wildfire, under real-time constraints. We argue that an optimization algorithm from the Ant Colony Optimization family is a strong candidate to tackle this problem, taking into account the nature of the submitted data and the limitations introduced by mobile phones.

1. Introduction

Wildfire simulation is a topic gaining increasing attention due to the growing threat to human lives and infrastructure from wildfires. The output of a wildfire simulator can guide firefighting authorities in their choice of appropriate countermeasures and suppression techniques, or in mobilizing the appropriate personnel and resources [1]. In addition, when wildfires reach inhabited areas, an evacuation plan [2] for the residents must be created and updated with the most up-to-date information about the situation. The unfortunate events of the devastating wildfire in Mati, Athens, in 2018, when an extremely fast-spreading wind-driven wildfire cut off hundreds of residents in their homes or to dead ends, leading to more than a hundred deaths, highlighted the importance of timely prediction of the behavior of a wildfire and the significance of an updated evacuation plan driven by up-to-date data.
A critical issue in simulating wildfires is that we often have to rely on incomplete or outdated information. For example, our knowledge about a wild forest at a specific time may be based on satellite observations or earlier aerial surveys that do not provide detailed information on the types and state of vegetation, and even this information may become outdated as the landscape evolves over time. When a wildfire is in progress, more detailed and up-to-date information is gathered by eyewitnesses, firefighters, drones, and aircraft converging in the area. Although such observations can help simulations to be updated with the most recent data, they are in themselves often inaccurate. For example, different firefighting crews may provide conflicting estimates of the fire front or where one type of vegetation stops and another one starts, due to their different vantage points and the uncertainty surrounding such an evolving phenomenon.
We argue that due to the fast-evolving nature of wildfires and the uncertainty surrounding them, the best way to predict their spread is to rely on mobile phone-based applications where local simulations are executed using existing data (e.g., aerial surveys and maps), data exchanged with other mobile phones, and data directly entered by the user. In general, when connectivity is limited, the situation changes in real time, and plans need to be updated on the fly; for example, when calculating evacuation routes due to wildfires, floods, or earthquakes, it makes sense to focus on computation at the edge, rather than at centralized cloud infrastructures [3].
Developing such an application relies on two key requirements. First, the necessary input data must be collected and processed, even in environments with poor network coverage. In the context of spatial data scarcity and network coverage, this involves gathering real-time information from nearby users via structured formats such as Geo-JSON files. Second, the geometric shapes submitted by users—which describe the wildfire environment’s characteristics—must be correlated both spatially and temporally. This correlation can then be used to filter duplicate data or outliers.
While any method of shape correlation could theoretically be employed and, indeed, many existing solutions to this problem have been proposed, we argue that an Ant Colony Optimization (ACO) inspired approach offers a distinct advantage: it introduces a degree of randomness into the solution components, thereby counterbalancing the inherent subjectivity in user-submitted data. This subjectivity arises from the fact that multiple individuals contribute to the dataset, inevitably introducing variability into their measurements or interpretations. Additionally, after the algorithm correlates two shapes, it reduces their spatial resolution and complexity by matching their corresponding edges, reducing the required storage and processing power. This approach ensures that the simulator will be fed by the most accurate and representative model of the wildfire environment, despite initial data limitations. In Figure 1, a flowchart of the envisioned local simulation mechanism and its components is given. The ACO algorithm is a crucial item in this process, since it pre-processes the raw data and ensures that only the necessary spatial data will reach the local wildfire simulator.
In this paper, we focus on the shape correlation problem, in the scope of a decentralized platform for wildfire simulation, fed by real-time crowdsourced data. This approach inevitably implies mixing user-submitted data of questionable quality and existing data. Not only may quantitative measurements be inconsistent, but lots of (nearly) duplicate data will be submitted, too. Simply aggregating data or interpolating missing data leads to a decrease in the quality of data and cumulative errors, while crucial information necessary for accurate simulation may be omitted. Additionally, submitted data (e.g., a forestry area) may have different spatial resolutions, or may be sampled with fewer points. To address these issues, the problem can be decomposed into solving the following subproblems:
  • Correlate user-submitted data regarding geometric primitives (e.g., matching similar shapes in the spatial and time domains), taking into account the human aspects of the submitters and the corresponding data time stamps.
  • Mitigate the impact of inconsistent and ambiguous user-submitted measurements on the simulation engine.
  • Simplify the user-submitted modeled environment to meet real-time constraints (e.g., remove duplicate entries) for the wildfire environment, to reduce the simulation time, latency, and network bandwidth required to exchange them.
  • Implicitly derive additional knowledge, extracting information from the changing nature of the ontologies involved in a wildfire environment.

Contributions

In this paper, we propose a novel shape matching algorithm inspired by Ant Colony Optimization (ACO), in which the choice of vertex correspondences is made probabilistically, guided by a pheromone deposition mechanism. Unlike traditional deterministic or exhaustive methods, our approach is particularly suited to the noisy, inconsistent nature of crowdsourced, hand-drawn polygons submitted by operators in the field. In the Results Section, we demonstrate that human-submitted shapes or the new shapes that appear during the evolution of the wildfire phenomenon exhibit implicit patterns, which our method can exploit, leading to improved matching performance and allowing future knowledge extraction in realistic scenarios.
Compared to other well-known and easy-to-implement [4,5] methods, our approach exploits the fact that we are expecting data to be entered via mobile phones, reflecting different vantage points, and evolving over time, as the wildfire evolves. For example, the Bounding Box method is very coarse in dealing with details and extremely sensitive to outliers. If two shapes share the same boundaries but different geometries, they are oblivious to the method. In the case of Hausdorff Distance, we expect performance to deteriorate in noisy, variably sampled data, especially under the influence of outliers.
Our method operates at the vertex level, allowing for fine-grained geometric matching. The probabilistic sampling of the solution search space downplays outlier influence by weighing heuristic information with popular solution components, offering better robustness to noise and inconsistent data points. This enables a non-exhaustive exploration of the search space, while preserving diversity (due to bias inherent by human-submitted inputs) and avoiding local optima.
Regarding input, our method does not require equal-sized or ordered point sets, allowing for partial matching. Hence, our method can cope with sampling irregularities and inconsistencies, a typical characteristic of crowdsourced data. Last but not least, the method is focused on geometric primitives related to wildfires, enabling us to extract additional patterns and knowledge in the time domain.
Finally, a previous ACO-based method presented in [6] is not suitable to directly query the crowdsourced dataset in its full extent; it needs modifications, in both the spatial and temporal domains. In our scheme, we have tuned the pheromone deposition mechanism to prioritize the influence of the most recent components of the wildfire environment since, in a fast-evolving phenomenon like a wildfire, they are more likely to reflect the real-world conditions.

2. Materials and Methods

2.1. A Motivating Example

Correlating two geometric primitives is a key requirement for the efficiency of our decentralized platform for wildfire simulation, as it helps merge user-submitted ontologies [7] (like vegetation areas, wildfire frontline, wildfire perimeter, etc.) describing the same wildfire feature into a single entity, limiting space, computing, and network overheads. Instead of relying on users to describe in an arbitrary manner each area’s characteristics, we adopt a predefined ontology that we have devised explicitly for this purpose [8], which provides a common frame of reference, shared terminology, and standardized rules. This approach addresses the diverse cognitive barriers among users coming from different backgrounds. By introducing an ontology, it becomes possible to reduce potential ambiguities in data definition—a strategy that has already been presented in the literature in a similar context [9,10]. As a result, user-submitted entries for the same area are expected to differ only in the shape of each submission, as each user draws the area as a polygon on a map. The resulting problem may sound trivial, considering the plethora of available solutions for shape matching given in [4]; however, when taking into consideration the variation in the data over time and the human nature of the submitters, the situation is more challenging. Intuitively, we would like to somehow weigh user choices and favor the most common or popular ones, assuming that these are more likely to be the ones with the highest quality.
It is also crucial, though, not to rule out any submitted data, since non-popular user-submitted data may be due to a rapid change in the conditions of the evolving phenomenon. Given that, we argue that an optimization technique deriving inspiration from the behavior of social insects is the perfect candidate. In the literature, a wide range of bio-inspired optimization methods have been proposed, modeling the collective behaviors of social animals such as ants, bees, birds, and fish, leading to algorithms like Ant Colony Optimization [11], Particle Swarm Optimization [12], Bat Algorithm [13], and Artificial Bee Colony [14]. It is quite easy to implement such a method both locally and centrally, while it is fed from time and spatially evolving human activity, and it is also relatively power-efficient. In particular, the behavior of ants has inspired a huge amount of optimization methods under the family of Ant Colony Optimization (ACO) algorithms [15]. Such algorithms have been successfully applied to a variety of problems, including shape matching [6,16,17]. However, the time dimension of user-submitted data and the impact of human activity on the submissions, which are critical in our context, have not yet been explored.
Figure 2 depicts six instances of the environment of a typical wildfire as submitted by the same number of users located in the field at time t1. The aim of the users is to gather as much information about the environment as possible. Usually, this information contains, but is not limited to, observations about areas with forestry fuels, burnt areas, locations of firelines, attack points, etc. Each observation submitted by its corresponding user is represented as a Geo-JSON entry containing time-stamped spatial information about an area in the wildfire’s environment (e.g., a forestry area, an unburned area, etc.), followed by some properties based on the observer’s perception (e.g., the type of vegetation contained in the specific area, its horizontal and vertical composition, etc.). At time t2, we see three possible instances of the output generated by the proposed algorithm, each produced locally by an equal number of distinct users. A closer examination of the grid containing the three distinct instances in t2 reveals that the geometric primitives differ from those submitted by the initial six users at t1. For example, the upper part of the forestry area (NW) was mapped differently by the users, and the algorithm generated an intermediate shape that lies between their respective inputs. The same applies to the upper left white rectangular shape containing a fuel-free rocky area. It is crucial for the decentralized simulation platform to avoid saturating users’ storage space with duplicate copies of the same geometric elements. The most recent version should have the greater probability to take precedence, ensuring that as time progresses, outdated or low-quality geographic data are gradually forgotten, while newer, higher-quality data are reinforced. Furthermore, each node in the network should only distribute to other nodes the final representation (the final grid instance in Figure 2) of the wildfire environment to minimize network congestion.

2.2. The ACO Family of Algorithms

The ACO family of algorithms is inspired by the indirect communication mechanism deployed by ants, which characterizes their foraging behavior [18]. When a promising source of food is discovered by an ant, it lays a pheromone trail during its return trip to its nest. That pheromone trail is detectable by other members of the colony, guiding them towards the food source. If the food source is preferred by the ants, the trail between it and their nest will soon be reinforced by deposition of additional pheromone trails left by other members of the colony. On the other hand, if the food source has been exhausted or new, more favorable sources have been discovered, the existing pheromone trails will be left to evaporate.
ACO algorithms implement digital ants that explore the search space of complex optimization problems. Typically, when a promising solution is found, a “pheromone trail” is deposited to guide the rest of the “ants” to converge on that point of the search space. In order to escape from local optima, each ant selects its next move probabilistically according to pheromone levels and/or other parameters, so that any point of the search space can have a chance to be evaluated. The robustness and efficiency of ACO methods is well documented in the literature, so it is not uncommon to come across some ACO variant for any optimization problem, both discrete and continuous [19].

2.3. An Existing ACO-Based Solution

In our work, we extend the initial ACO-based framework for shape matching presented in [6] by introducing the concept of proximity, in both the spatial and time domains. For the specific problem of shape matching, the input is two sets of points I , J , which represent the vertices of the two shapes to be compared. Given a function π   s u c h   t h a t   i I , j J : π i j assigns the points of I to J, we try to find an optimal assignment that minimizes an objective function:
i I D ( R i , R π ( i ) )
where R i is a shape descriptor [6] and particularly a shape context [20] for point i, and D() is a distance function. Assuming that I J ,   π ( ) can be a non-injective, non-surjective function of i . The authors of the original algorithm [6] proposed a scheme to express the heuristic information utilizing the weighted sum of two metrics as a (multi-objective) distance function. These metrics are the shape descriptor distance and a proximity information function, measuring the degree of matched neighboring points.
Commonly, ACO instances are expressed as graph problems: we define the complete directed graph G = {V, E} where V = I J and each edge eij in the graph represents the matching between vertices i and j. Each digital ant starts from a randomly selected vertex, and in every step, it traverses an edge according to a computed probability. The probability of the k-th ant to select the edge connecting vertices i ϵ I and j ϵ J is given by Equation (2)
p i j k = α · τ i j + ( 1 a ) · η   i j l ϵ N i [ α · τ i l + ( 1 a ) · η   i l ]
where N i is the set of edges originating from vertex i and τ i j is the level of pheromone laid to edge i , j . The term η i j represents the heuristic information associated with edge i , j and α   ϵ   ( 0,1 ) is a user-specified parameter that weighs the heuristic information over the pheromone trails. The exact formulation of the heuristic function η i j is given in [6] and it is normalized to the range of (0,1). It may involve more than one metric, such as a Euclidean and a Geodesic distance function.
A pheromone matrix τ i j is initialized with an initial value μ and is updated at each execution step, ensuring that it cannot fall below a certain value τ . At each update, every value in the matrix is reduced by a constant amount (pheromone evaporation) except values associated with an edge i , j that is part of a matching conducted by ant k (pheromone deposition). Δ τ i j k is the amount of pheromone deposited at edge i , j by the kth ant; it is a function of δ , which is a user-specified parameter (deposition constant). Hence, the elements of the matrix can be computed according to Equation (3):
τ i j = τ i j + k = 1 m Δ τ i j k

2.4. Outline of the Proposed ACO Algorithm

Our proposed ACO algorithm departs from [5] in that, to adapt to the human nature of the submitters, we propose that pheromone trails should be boosted based on the influence of human-made decisions. A typical user entry is a time-stamped Geo-JSON file, such as the one shown in Figure 3.
A density-based clustering algorithm (like DBSCAN [21]) is applied to the time-stamped entries of the geometric primitives, to extract the time windows where changes occurred to the wildfire environment. These time windows form the corresponding C classes, as extracted by the clustering algorithm. The clustering process is required to create sets I , J , which form the graph required by ACO. The clustering input is the time stamps of the user-submitted data, and the output is the density-clustered time windows. For each resulting cluster (time window), a Geo-JSON file is probabilistically selected as set I , while set J is formed from the remaining Geo-JSON files. Sets I , J are fed to the ACO [6] algorithm, which constructs a matching between all points in I and their corresponding ones in set J. Hence, we end up with C instances (after the simplification process) of the same geometric primitive. Finally, we can monitor the speed of changes that occur inside the consecutive time windows, extracting information about fire-related metrics like the speed of spread, the direction, the expected emitted heat, etc.
A simplified outline of the proposed algorithm is provided below in Figure 4:
The process of coordinate simplification tries to deal with the inconsistent and ambiguous user-submitted measurements. Points corresponding to set I are matched with the most promising ones in set J, and their mean value is kept as the representative coordinate in their own cluster.
The speed metric is specific to the domain of the geometric primitive; it can express changes in the area or length of two shapes, etc. A simple expression for function calculateDiff() could be the one shown in Figure 5:
The perform_ACO(I,J) function is adapted from the original article in [6], and shown in Figure 6:
The ConstructMatching() function constructs a new feasible matching between two sets, taking into account a user-specified parameter (α) that controls the influence of pheromone intensity on the ant’s decision-making, which counterbalances the influence of the heuristic information.
The same principle applies to the spatial domain: it is not clear if two different entries from different users are describing the same object. Again, a density-based clustering of the geometric primitives belonging to the same time window can be conducted. The output will be geometric primitives that form promising candidates to be instances of the same object. The final step is to feed these into the ACO algorithm and decide if they refer to the same object according to their matching score.
Since the ACO algorithm returns a (best) matching (e.g., an estimate of the real evolution of a geometric primitive in time), we can also inductively conclude the rate of spread (on average) the geometric primitive is subject to. This can lead to further knowledge extraction, even for features for which we have no recent measurements. Supposing that the area in Figure 7 describes the burnt area during a wildfire and users submitted the areas of Figure 8 during the evolving situation, we can express the burn rate in time, even if that number is not known originally. Additionally, vegetation composition due to partially burnt areas, fire perimeter/area, or other related characteristics can be extracted too.

2.5. A Modified Pheromone Update Model

An additional dose of pheromone levels is introduced, according to the number of solution components, with a more recent date. This leads the algorithm to prefer constructing the solution with newer data, which are likely to contain more up-to-date information about the evolving phenomenon. The proposed modification is not intended to produce matching with better (lower) cost; rather, it “forces” the algorithm to increase the number of solution components in the matching with the most updated information.
This is achieved by the addition of the term λ in the pheromone deposition process described in [6]. The term λ expresses the number of solution components that are members of the newest submitted entry, multiplied by a coefficient that governs their influence against the traditional computation of pheromone deposition. A simple outline of the proposed modification is provided in Figure 9:
While the function evaluate_recent() can be expressed as shown in Figure 10.

2.6. An Illustrated Example

To examine the efficiency of the proposed method and its ability to generate a simplified version of the original raw user submissions, we asked several users with different knowledge backgrounds to draw several shapes on a map (via OpenStreetMaps) depicting a wildfire characteristic. The user group included a participant who had extensive training and expertise in firefighting methods. The area was located inside a wildland urban interface (WUI) [22,23] zone in East Attica. Figure 7 depicts a typical user submission of a forestry area consisting of 36 points in WGS 84 coordinates projected on the map. On the other hand, Figure 8 depicts the shape of the same forestry area (also expressed in WGS 84 coordinates) as submitted by four different users. It is clear that the level of detail (from 36 up to 45 vertices) and the effort to draw the exact details of the area vary from user to user.
The output of our algorithm is visualized in Figure 11, where the points of user B are matched with points collected from other user-submitted points, forming the most promising points to extract for the final “simplified” version of the geometric primitive.

3. Results

To test the performance of the proposed method, a test dataset was generated featuring a collection of 16 polygons in Geo-JSON format (Figure 12). The study region is located in the area of Kallitechnoupoli, east of Athens, an area known for its susceptibility to wildfires in recent history. Although the entire dataset is available for processing, only the subset of features that will eventually be correlated to the same or nearby areas needs to be analyzed using the described ACO method. The remaining portions of the dataset, which consist of uncorrelated geometric primitives, can be directly entered into the wildfire simulator, without requiring additional processing.
In the dataset, we included six instances of the same forestry area with different spatial resolutions, as submitted by an equal number of volunteer firefighters (Figure 13). As a baseline approach, we implemented in MATLAB (version R2025a).
A simple function that returns the weighted sum of the shape similarity and the area ratio of the tested shapes. The result is weighed and normalized to [0, 1]. Shape similarity is calculated by the Modified Hausdorff Distance (MHD) among the vertices [25,26], while area similarity is the area ratio of the two shapes. Additionally, a final distance threshold has to be met (in the region of 100–150 m for a typical area). For shapes that achieved a similarity above 0.6 (indicating a degree of correlation), we ran a comparison between the baseline and ACO methods. This filtering step removed Geo-JSON files that belonged to other areas from our dataset. The parameter selection for both the ACO and baseline methods (e.g., weight values) is based on experimental results and empirical observations. The parameters chosen are listed in Table 1. Although this test is considered a form of benchmarking, the methods are not directly comparable to each other, since the objectives differ, even though they aim to answer the same question: to what extent are two shapes similar? ACO constructs a feasible solution in a non-deterministic way, which is promising according to its objectives (e.g., favoring recent solution components) and tuning parameters. That solution is then processed to simplify the geometry. The benchmark only serves as a reference point of how the baseline method would perform if we did not care about the rest of the objectives covered in the paper.
We also asked the same group of volunteers to draw LineStrings representing the boundary of a past wildfire event. We interpreted these inputs as the active fireline for the same wildfire event, since it is not practical to acquire this information in a real-world scenario. In Figure 14, we depict their hand-drawn inputs. The baseline approach was modified accordingly to produce an output that combines both the modified Hausdorff distance (MHD) and the length ratio of the tested firelines. This output is calculated as a weighted sum of the two metrics and is normalized to the range [0, 1], ensuring a common comparison framework.
In Figure 15, we present the shape similarity scores obtained by both the baseline method and ACO. Even though the comparisons were conducted among all polygons representing the same area, we included cases where a clear bias was evident in how volunteer firefighters drew the area. It is important to note that the participants had comparable levels of experience. The similarity scores of both methods were found to be quite close. Although ACO generally achieves higher similarity scores than the baseline method in most cases, drawing definitive conclusions remains challenging due to the level of uncertainty and noise introduced by human submitters. While similarity metrics are fundamental to addressing our problem, their practical significance is not straightforward to determine. ACO demonstrates a clear efficacy by identifying optimal vertex correspondences between shapes, instead of using a simple interpolation between them. Consequently, it derives solution components directly from unprocessed user inputs, yielding more reliable outcomes.
The similarity results obtained for the different fireline entries can be seen in Figure 16, showing that our approach can also handle this kind of input (jagged lines as opposed to polygons). Given the significant uncertainty involved in accurately determining the spatial position of a wildfire front—arising from factors such as dense smoke, physical obstacles blocking clear view, limited terrain awareness, and cognitive limitations in understanding fire dynamics—along with the inherently rapid and unpredictable movement of active firelines, we argue that similarity metrics in this context offer limited interpretative value. Since the fire front is expected to shift continuously over time, strict geometric comparisons may not meaningfully reflect the evolving reality on the ground. For this reason, our research does not prioritize the correlation of firelines as the basis of our benchmarks. On the other hand, we consider an accurate mapping of the composition of forestry fuels as a far more vital input for accurate wildfire simulation.
In Figure 17, we present the similarity scores of both the baseline and modified ACO methods for the different levels of spatial resolutions of the same forestry area, obtained for the same number of users. The radar graph in Figure 18 depicts the output in the same six test instances, while presenting the area (in Km2) and number of edges of each test case. On average, our ACO method produces better similarity scores, making it a more consistent method, while maintaining a lower variability across different test cases. In contrast, the Baseline method works well in some cases, but it underperforms in others, due to its limited ability to account for the underlying topological features of the shapes. This difference is largely due to the design of ACO, which combines heuristic information with probabilistic decision making, allowing it to be less volatile and less susceptible to a wide variety of inputs.

4. Discussion

The shape matching algorithm running in each local instance of the mobile application is a critical piece of its overall flow (see Figure 1), as it allows user-submitted data entries to be added to the database only after their differences are reconciled and merged into fewer shapes; it is impossible to run an accurate simulation with contradictory data entries for the same area. Some input information is eliminated after the simplification process, since we only keep (and propagate) the results of the ACO algorithm, as opposed to the entire input dataset. This is necessary to compress the dataset that describes the environment of the simulation, removing any duplicate entries or outliers, while maintaining space and network efficiency.
We reiterate that the aim of our decentralized wildfire simulator is not to yield the most analytically accurate simulation possible, but to provide a simulation computed with the most updated information collected in the field due to the fast-evolving nature of a wildfire. A design compromise between spatial accuracy and transmission latency is essential for operational effectiveness. Device-to-device communications in wildfire-affected areas experience restricted throughput as a consequence of transient nodal encounters and not due to available bandwidth. This necessitates the adoption of the strategy to create highly lean data payloads—both in size and required processing complexity—to ensure successful network-wide distribution of the information.

5. Conclusions and Future Work

Early-warning systems aiming at populations that live in a wildfire’s path overlook two fundamental operational challenges: the need for continuously updated fire data and the likelihood of impaired network infrastructure. We argue that incorporating our method across the scenarios described earlier will enhance their expected outcomes and increase their robustness to dynamic data and varying network conditions.
In this paper, we proposed a novel method for shape matching tailored to the challenges arising in a wildfire environment. We chose a method inspired by the ACO framework due to its robust and flexible nature. Its efficiency was demonstrated via a test scenario, in which we examined six instances of hand-drawn forestry areas, as submitted by volunteer firefighters. These instances were identified as the most spatially correlated elements within a dataset consisting of 16 polygons (Figure 12). Their high degree of correlation suggests that they represent closely matching geographic features, making them particularly relevant for analysis and optimization using the ACO method. Focusing on these correlated instances allows for more targeted modeling of areas where changes occur, while reducing computational and space overheads when they are relayed to the network. The method achieved higher similarity scores on average than a baseline method on polygons, while keeping its variation smaller compared to baseline. After a successful shape matching, the method constructed a simplified version of the input geometry, selecting the most promising edges of the polygons involved, to minimize the spatial resolution, processing power, storage, and network bandwidth required. These factors are crucial for the next stage, in which we want to process the data describing the wildfire environment in a decentralized wildfire simulator. Similar results were observed in scenarios where LineStrings representing firelines were evaluated, further demonstrating the robustness and general applicability of the proposed method.
We are considering various directions for future work. One issue that we are considering is how to handle outliers more effectively and how to treat submissions from heterogeneous users, distinguishing between expert and untrained submitters. We also need to test our approach for smoothing out short-term fluctuations and mitigating temporal noise in crowdsourced inputs by prioritizing more recent, but not necessarily reliable, data, in real-world scenarios. Another issue is the further exploitation of meaningful information about the nature, behavior, and even risks the submitters may be facing. As soon as user-submitted data are streamed to mobile nodes, mobility patterns can be exported from both the submitted data themselves as well as from the place and time-node encounters that are conducted. Rapid and unpredicted changes in mobility patterns or the formation of swarms, consisting of moving users, may be the outcome of an uncontrolled fire outbreak leading to evacuation processes or to activities linked with modified suppression tactics. These approaches seem promising and have gained attention in the literature [2,27].

Author Contributions

Conceptualization, I.K. and G.X.; methodology, I.K.; software, I.K.; validation, I.K.; writing—original draft preparation, I.K.; writing—review and editing, I.K. and G.X.; visualization, I.K.; supervision, G.X. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Only test data were created for this article; these data are available from the authors upon request.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
ACOAnt Colony Optimization
MHDModified Hausdorff
WUIWildland–Urban Interface

References

  1. Freire, J.G.; DaCamara, C.C. Using cellular automata to simulate wildfire propagation and to assist in fire management. Nat. Hazards Earth Syst. Sci. 2019, 19, 169–179. [Google Scholar] [CrossRef]
  2. Mitchell, H.; Gwynne, S.; Ronchi, E.; Kalogeropoulos, N.; Rein, G. Integrating wildfire spread and evacuation times to design safe triggers: Application to two rural communities using PERIL model. Saf. Sci. 2023, 157, 105914. [Google Scholar] [CrossRef]
  3. Xylomenos, G.; Pavlou, G.; Psaras, I.; Karakonstantis, I. Named Functions at the Edge. In Proceedings of the 2019 IEEE Symposium on Computers and Communications (ISCC), Barcelona, Spain, 29 June–03 July 2019; IEEE: Piscataway, NJ, USA, 2019; pp. 1–6. [Google Scholar] [CrossRef]
  4. Loncaric, S. A survey of shape analysis techniques. Pattern Recognit. 1998, 31, 983–1001. [Google Scholar] [CrossRef]
  5. Yang, M.; Kpalma, K.; Ronsin, J. A Survey of Shape Feature Extraction Techniques. In Pattern Recognition; Yin, P.-Y., Ed.; IN-TECH: London, UK, 2008; pp. 43–90. [Google Scholar]
  6. Van Kaick, O.; Hamarneh, G.; Zhang, H.; Wighton, P. Contour Correspondence via Ant Colony Optimization. In Proceedings of the 15th Pacific Conference on Computer Graphics and Applications (PG’07), Maui, HI, USA, 29 October–2 November 2007; IEEE: Piscataway, NJ, USA, 2007; pp. 271–280. [Google Scholar] [CrossRef]
  7. Kavouras, M.; Darra, A.; Kokla, M.; Kontaxaki, S.; Panopoulos, G.; Tomai, E. Geographic Information Science—An Integrated Approach and Special Topics; Association of Greek Academic Libraries: Athens, Greece, 2015. [Google Scholar]
  8. Karakonstantis, I.; Xylomenos, G. An ontology for dynamic wildfire simulations. In Proceedings of the 2025 IEEE Conference on Standards for Communications and Networking (CSCN), Bologna, Italy, 15–17 September 2025. [Google Scholar]
  9. Kalabokidis, K.; Athanasis, N.; Vaitis, M. OntoFire: An ontology-based geo-portal for wildfires. Nat. Hazards Earth Syst. Sci. 2011, 11, 3157–3170. [Google Scholar] [CrossRef]
  10. Masa, P.; Meditskos, G.; Kintzios, S.; Vrochidis, S.; Kompatsiaris, I. Ontology-based Modelling and Reasoning for Forest Fire Emergencies in Resilient Societies. In Proceedings of the 12th Hellenic Conference on Artificial Intelligence, Corfu Greece, 7–9 September 2022; ACM: New York, NY, USA, 2022; pp. 1–9. [Google Scholar] [CrossRef]
  11. Wang, C.; Zhang, S.; Ma, T.; Xiao, Y.; Chen, M.Z.; Wang, L. Swarm intelligence: A survey of model classification and applications. Chin. J. Aeronaut. 2025, 38, 102982. [Google Scholar] [CrossRef]
  12. Lei, J.; Yang, C.; Zhang, H.; Liu, C.; Yan, D.; Xiao, G.; He, Z.; Chen, Z.; Yu, T. Radiation shielding optimization design research based on bare-bones particle swarm optimization algorithm. Nucl. Eng. Technol. 2023, 55, 2215–2221. [Google Scholar] [CrossRef]
  13. Karakonstantis, I.; Vlachos, A. Bat algorithm applied to continuous constrained optimization problems. J. Inf. Optim. Sci. 2021, 42, 57–75. [Google Scholar] [CrossRef]
  14. Sharma, T.K.; Abraham, A. Artificial bee colony with enhanced food locations for solving mechanical engineering design problems. J. Ambient Intell. Humaniz. Comput. 2020, 11, 267–290. [Google Scholar] [CrossRef]
  15. Dorigo, M.; Stützle, T. Ant Colony Optimization; The MIT Press: Cambridge, MA, USA, 2004. [Google Scholar] [CrossRef]
  16. Saadi, Y.; Sari, E.N.; Herawan, T. An Improved Ant Colony Matching by Using Discrete Curve Evolution. In Information and Communication Technology; Linawati, Mahendra, M.S., Neuhold, E.J., Tjoa, A.M., You, I., Eds.; Lecture Notes in Computer Science; Springer: Berlin/Heidelberg, Germany, 2014; Volume 8407, pp. 248–256. [Google Scholar] [CrossRef]
  17. Zhang, J.; Zhang, L.; Saucan, E. A Novel Ant Colony Shape Matching Algorithm Based on the Gromov-Wasserstein Distance. In Proceedings of the 2020 8th International Conference on Digital Home (ICDH), Dalian, China, 19–20 September 2020; IEEE: Piscataway, NJ, USA, 2020; pp. 159–165. [Google Scholar] [CrossRef]
  18. Mailleux, A.-C.; Detrain, C.; Deneubourg, J.-L. Triggering and persistence of trail-laying in foragers of the ant Lasius niger. J. Insect Physiol. 2005, 51, 297–304. [Google Scholar] [CrossRef] [PubMed]
  19. Karakonstantis, I.; Vlachos, A. The ant colony optimisation solving continuous problems. Int. J. Comput. Intell. Stud. 2013, 2, 350. [Google Scholar] [CrossRef]
  20. Belongie, S.; Malik, J.; Puzicha, J. Shape matching and object recognition using shape contexts. IEEE Trans. Pattern Anal. Mach. Intell. 2002, 24, 509–522. [Google Scholar] [CrossRef]
  21. Schubert, E.; Sander, J.; Ester, M.; Kriegel, H.P.; Xu, X. DBSCAN Revisited, Revisited: Why and How You Should (Still) Use DBSCAN. ACM Trans. Database Syst. 2017, 42, 1–21. [Google Scholar] [CrossRef]
  22. Cohen, J. The wildland-urban interface fire problem. Fremontia 2010, 38, 16–22. [Google Scholar]
  23. Mell, W.E.; Manzello, S.L.; Maranghides, A.; Butry, D.; Rehm, R.G. The wildland–urban interface fire problem–current approaches and research needs. Int. J. Wildland Fire 2010, 19, 238–251. [Google Scholar] [CrossRef]
  24. van Kaick, O.; Contour Correspondence via Ant Colony Optimization. MATLAB Central File Exchange. Available online: https://www.mathworks.com/matlabcentral/fileexchange/24094-contour-correspondence-via-ant-colony-optimization (accessed on 9 August 2025).
  25. Alt, H.; Behrends, B.; Blömer, J. Approximate matching of polygonal shapes. Ann. Math. Artif. Intell. 1995, 13, 251–265. [Google Scholar] [CrossRef]
  26. Dubuisson, M.-P.; Jain, A.K. A modified Hausdorff distance for object matching. In Proceedings of the 12th International Conference on Pattern Recognition, Jerusalem, Israel, 9–13 October 1994; IEEE Computer Society Press: Los Alamitos, CA, USA, 1994; Volume 1, pp. 566–568. [Google Scholar] [CrossRef]
  27. Cova, T.J.; Dennison, P.E.; Kim, T.H.; Moritz, M.A. Setting wildfire evacuation trigger points using fire spread modeling and GIS. Trans. GIS 2005, 9, 603–617. [Google Scholar] [CrossRef]
Figure 1. A mobile-phone-powered wildfire simulator that exchanges data with nearby users via Geo-JSON files to address the lack of spatial data required to accurately model the environment.
Figure 1. A mobile-phone-powered wildfire simulator that exchanges data with nearby users via Geo-JSON files to address the lack of spatial data required to accurately model the environment.
Applsci 15 08844 g001
Figure 2. Formation of a simplified model after consecutively merging user-submitted data via the ACO algorithm. Users marked the forestry areas, shown as green outlines, and fuel-free rocky areas, shown as white rectangles. In this example, the ambiguous data eliminated are marked in red in the final step. The final output can then be redistributed to the network.
Figure 2. Formation of a simplified model after consecutively merging user-submitted data via the ACO algorithm. Users marked the forestry areas, shown as green outlines, and fuel-free rocky areas, shown as white rectangles. In this example, the ambiguous data eliminated are marked in red in the final step. The final output can then be redistributed to the network.
Applsci 15 08844 g002
Figure 3. Simplified user input in Geo-JSON format representing a grove in East Attica.
Figure 3. Simplified user input in Geo-JSON format representing a grove in East Attica.
Applsci 15 08844 g003
Figure 4. Outline of proposed algorithm.
Figure 4. Outline of proposed algorithm.
Applsci 15 08844 g004
Figure 5. Function calculateDiff().
Figure 5. Function calculateDiff().
Applsci 15 08844 g005
Figure 6. Function perform_ACO().
Figure 6. Function perform_ACO().
Applsci 15 08844 g006
Figure 7. A forestry area inside a wildland urban interface (WUI) zone in East Attica, as submitted by a user.
Figure 7. A forestry area inside a wildland urban interface (WUI) zone in East Attica, as submitted by a user.
Applsci 15 08844 g007
Figure 8. User-submitted descriptions of the forestry area of Figure 7, submitted by four other individuals (WGS 84 coordinates).
Figure 8. User-submitted descriptions of the forestry area of Figure 7, submitted by four other individuals (WGS 84 coordinates).
Applsci 15 08844 g008
Figure 9. The modified pheromone update function.
Figure 9. The modified pheromone update function.
Applsci 15 08844 g009
Figure 10. Function evaluate_recent().
Figure 10. Function evaluate_recent().
Applsci 15 08844 g010
Figure 11. Optimal match constructed by our algorithm: (a) Match of user B against inputs collected from the remaining users (normalized coordinates). Generated by modifying the code provided in [24]. (b) Final shape as generated by ACO, sampling the most promising user inputs to generate an optimal match.
Figure 11. Optimal match constructed by our algorithm: (a) Match of user B against inputs collected from the remaining users (normalized coordinates). Generated by modifying the code provided in [24]. (b) Final shape as generated by ACO, sampling the most promising user inputs to generate an optimal match.
Applsci 15 08844 g011
Figure 12. Test dataset consisting of 16 forestry areas in Kallitechnioupoli, East Attica, Greece. Each user-submitted shape is marked with a different shade of dark green.
Figure 12. Test dataset consisting of 16 forestry areas in Kallitechnioupoli, East Attica, Greece. Each user-submitted shape is marked with a different shade of dark green.
Applsci 15 08844 g012
Figure 13. Five noisy instances of the same forestry area with different resolutions, marked with dark green. The bottom right panel shows how they overlap, using a different shade of dark green for each user-submitted area.
Figure 13. Five noisy instances of the same forestry area with different resolutions, marked with dark green. The bottom right panel shows how they overlap, using a different shade of dark green for each user-submitted area.
Applsci 15 08844 g013
Figure 14. A noisy representation of a fireline obtained from the same group of volunteers. Each user-submitted line is marked with a different color.
Figure 14. A noisy representation of a fireline obtained from the same group of volunteers. Each user-submitted line is marked with a different color.
Applsci 15 08844 g014
Figure 15. Similarity scores obtained from the modified ACO and baseline methods for a forestry area.
Figure 15. Similarity scores obtained from the modified ACO and baseline methods for a forestry area.
Applsci 15 08844 g015
Figure 16. Similarity scores obtained from the modified ACO and baseline methods for a fireline.
Figure 16. Similarity scores obtained from the modified ACO and baseline methods for a fireline.
Applsci 15 08844 g016
Figure 17. Similarity scores of baseline and modified ACO methods for the six instances listed in Figure 15.
Figure 17. Similarity scores of baseline and modified ACO methods for the six instances listed in Figure 15.
Applsci 15 08844 g017
Figure 18. Performance evaluation obtained for the six instances. The number of edges and the area of the corresponding shapes are labeled accordingly.
Figure 18. Performance evaluation obtained for the six instances. The number of edges and the area of the corresponding shapes are labeled accordingly.
Applsci 15 08844 g018
Table 1. Parameter Selection.
Table 1. Parameter Selection.
ParameterValueImpact
Baseline Method
MHD similarity weight0.7favors shape matching over area matching
Area similarity weight0.3favors area matching over shape matching
ACO
Number of iterations1000exploration of search space/speed of convergence
Number of ants1exploration of search space
Pheromone influence (α)0.3exploration of search space
Evaporation rate0.01exploration of search space
Minimum pheromone0.1exploration of search space
Recent coefficient0.05increases pheromone deposition in favor of newer solutions
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Karakonstantis, I.; Xylomenos, G. Improving Wildfire Simulations via Geometric Primitive Analysis in Noisy Crowdsourced Data. Appl. Sci. 2025, 15, 8844. https://doi.org/10.3390/app15168844

AMA Style

Karakonstantis I, Xylomenos G. Improving Wildfire Simulations via Geometric Primitive Analysis in Noisy Crowdsourced Data. Applied Sciences. 2025; 15(16):8844. https://doi.org/10.3390/app15168844

Chicago/Turabian Style

Karakonstantis, Ioannis, and George Xylomenos. 2025. "Improving Wildfire Simulations via Geometric Primitive Analysis in Noisy Crowdsourced Data" Applied Sciences 15, no. 16: 8844. https://doi.org/10.3390/app15168844

APA Style

Karakonstantis, I., & Xylomenos, G. (2025). Improving Wildfire Simulations via Geometric Primitive Analysis in Noisy Crowdsourced Data. Applied Sciences, 15(16), 8844. https://doi.org/10.3390/app15168844

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop