Introduction

The causes of forest fires are diverse, given the number of elements that intervene in them, such as available fuel, low humidity in the environment (^{Cárdenas-Salgado and Pizano, 2019}), and direct or indirect human interference, the latter being responsible for 75 to 96 % of forest fires (^{Hirschberger, 2016}). It should be taken into account that fires have the capacity to renew ecosystems (^{Manríquez, 2019}) and thus bring about their continuity.

According to Conafor records of 1970 to 2022, there have been 374 742 forest fires in Mexico, which have affected 14’829 944 hectares, the most affected entities being the State of Mexico, Mexico City, Michoacán, Chihuahua, and Jalisco (^{Conafor, 2022}).

The current climate variations have generated a series of changes in seasonal cycles, leading to a greater release of greenhouse gases, and the loss of natural areas responsible for reducing carbon concentration aggravates the problem (^{Semarnat, 2018}). Today, the rise in global temperatures and droughts have caused forest fires to increase in intensity, as observed in Europe, Australia, and the Western United States (^{Williams et al., 2019}; ^{Dupuy et al., 2020}; ^{Haque et al., 2021}). When a large number of medium-sized or large-scale forest fires occur, human populations are affected by the release of gases such as carbon monoxide, carbon dioxide, sulfur dioxide, and particulate matter, the latter being the most dangerous (^{Sandoval et al., 2019}; ^{Correa, 2020}).

In Mexico, forest fires tend to occur in times of low rainfall, particularly in spring (^{Alanís-Rodríguez et al., 2008}), and their number increases with such phenomena as droughts (^{Espinoza and González, 2019}). The main forms of fire management are related to the formation of fire control brigades (^{Aguilar et al., 2021}), as well as to controlled burn planning to reduce the concentration of available fuel that might generate a large-scale fire (^{Pérez-Salicrup et al., 2018}; ^{Rodríguez et al., 2020}).

There are different methods and technologies available for fire detection through the analysis of elements such as fire, smoke, and hot spots, among others (^{Ramos, 2010}). One of the most widely used is satellite technology, which can capture virtually any surface and can verify events such as smoke and fire; an example of this technology is the Early Warning System for Forest Fires (Sistema de Alerta Temprana de Incendios Forestales, SATIF) (^{Conabio, 2022}). Another type of mechanism for estimating fire danger are sensor networks (SN), i. e., electronic devices that are deployed in different areas to obtain such data as humidity, temperature, and wind speed, among others (^{Aakvaag and Frey, 2006}; ^{Cama et al., 2012}); the information extracted by the sensors is usually sent wirelessly to a control station.

Other detection systems, such as convolutional neural networks (CNN), usually employ artificial intelligence for efficient pattern recognition (^{Berzal, 2019}); they are one of the most common applications of deep learning (^{Wiatowski and Bölcskei, 2016}; ^{Zhou, 2020}) and are responsible for classifying objects in digital images. On the other hand, there are genetic algorithms, which are tools that represent the evolutionary and learning behaviors of living beings, allowing the generation of solutions to problems that require optimization processes, such as the process of calculating routes for the control of forest fires using drones (^{Wang et al., 2018}; ^{Zhou et al., 2019}; ^{Shaji, 2022}).

When forest fires grow rapidly, the number of personnel needed to contain them is often insufficient, and the associated costs can be high (^{Mendoza et al., 2012}). Today, the application of unmanned aerial vehicles (UAV) allows monitoring and control of forest fires, as can be seen in ^{Kinaneva et al. (2019)}, ^{Sungheetha and Sharma (2020)}, and ^{Li et al. (2022)}. ^{Madridano (2020)} proposes an architecture for coordinating drones to fight forest fires; similar proposals are presented by ^{Pérez-Sánchez et al. (2017)}, with the use of a geo-referenced system for drone use.

On the other hand, there are other types of proposals for drone coordination models, as shown in ^{Casbeer et al. (2005)}, ^{Harikumar et al. (2018)}, and ^{Momeni et al. (2022)}, to show optimal routes, and for coordinating of this type of devices without wasting battery power, as this is an element that conditions their mobility. The papers by ^{Chowdhury et al. (2019)}, ^{Shao et al. (2021)}, and ^{Sun et al. (2022)} analyze proposals for route calculation. This type of device requires route planning, due to the complexity of the distances, as well as to the number of elements available for monitoring (^{Wu et al., 2020}; ^{Dinh et al., 2021}; ^{Saeed et al., 2022}). One of the algorithms that stand out for route planning in complex environments is the ant colony algorithm (ACO). High-impact works, such as those by ^{Yang et al. (2020)}, ^{Wang and Han (2021)}, and ^{Stodola et al. (2022)}, highlight the implementation of the ACO because it is very proficient in solving the routing problem based on the traveler’s problem, as shown in the study by ^{Chaudhari and Thakkar (2019)}, compared to other algorithms like particle swarm optimization (PSO), artificial bee colony (ABC), and others.

Based on the above information, the objective of this research is to develop a proposal for the theoretical prevention and detection of forest fires with UAVs whose displacement is planned automatically using the ACO algorithm, as these devices can incorporate fire detection sensors and fire extinguishing mechanisms, and have flight capabilities that allow them to move through different types of terrain. For this reason, the implementation of UAV devices in the State of Mexico along routes defined through an ACO will help solve the currently existing issues in the detection and containment of forest fires.

Materials and Methods

The State of Mexico is the state in which historically there have been the highest number of forest fires in the country. This state is located in central Mexico; its coordinates are 18°25' and 20°17' N, and 98°33' and 100°28' W; its geography exhibits a cover of temperate forests of approximately 62 % (^{Ceballos et al., 2008}), and these forests are more susceptible to wildfires. Taking into account the above, a proposal has been designed for the generation of routes dedicated to the prevention and detection of forest fires using UAVs. The basis of this work is a procedure that consists in identifying geographic locations where forest fires have occurred, reordering these locations in order to obtain the shortest route, and finally, displaying the route proposed by the ACO.

An ACO is a recurrent algorithm for solving problems like that of the traveler; it consists in simulating the behavior of ants to calculate the probability of a path by creating a pheromone trail and indicating the route to be taken (^{Goss et al., 1989}; ^{Dorigo et al., 1996}). In this algorithm, the ants are represented by simple computational agents (computer program) which take an edge (path); the pheromones (information related to the previously explored paths) are then selected; once the agent finishes the task, the result is evaluated in order to subsequently modify the level of pheromones in the trail and determine the best possible route, favoring the shortest edges (also known as the shortest path) with the largest amount of pheromones, the best agent being the one that can update the trail (^{Merkle and Middendorf, 2000}). Figure 1 shows an example of a colony of ants, which illustrates the process of travel to find the best route in search of a target (O) and return to a base point (B), in this case, it is observed that after the generation of the traces, the one that represents the shortest distance to be covered is chosen at the end. For the purpose of the proposal, the algorithm will allow the UAV to find the shortest distance route in the fire verification process.

Figure 1 Example of the ACO process for distance travel. 

Each agent moves from state x to state y, corresponding to a temporary solution. An agent k is in charge of computing a set of ak(x) of feasible routes from its current state in each iteration, moving probabilistically. For an agent pxyk, the transition from state x to state y depends on the combination of the values of the nxy (movement); it is computed by a heuristic indicating the feasibility of the trace, and the level of the τxy trace of the movement, which indicates its competence in the previous cycle. Agent kth moves from state x to state y with a probability estimated using Equation (1) (^{Gambardella and Dorigo, 1996}).

Where:

τxy = Indicates the amount of pheromones used in the transition from state x to state y; 0 ≤α is the parameter that controls the influence of τxy ηxy = Indicates the desirability of the xy state transition

1dxy = m, where d is the distance

β≥1 = Parameter in charge of controlling the influence of ηxy

When all the agents complete a solution, the traces are updated with τxy←1-pτxy+∑k∆τxyk. In the process of evaluating the edge distances, Equation (2) is applied to determine the distance between two points, as it is an important element in determining the routes along which the agents will move.

Where:

d = Distance

A = First point

B = Second point

xn = Abscissae of the different points

yn = Ordinates of the different points

Equation 2 is used to calculate the distances in order to determine the edges that the agents are traveling and the shortest distance between the total distance of the route that is being planned.

Table 1 shows the pseudocode that represents the general form of the ACO, itemizing the basic functions, such as the generation of solutions, the agents involved, and, finally, the update of the pheromones, being these the solution space used by the model. In each cycle, the routes that were most successful for the agents are chosen to determine the order of the points to be used for the route.

The route calculation model for forest fire detection is displayed in 4 main stages, which are as follows:

1.- Data filtering. This is the main stage of the model; it allows to extract the data from the file showing the geographic positions where a forest fire has occurred in the State of Mexico and it selects those municipalities where there have been at least three forest fires, taking into account that, with locations where only two forest fires or less have occurred, the total distance obtained will be the same if the ACO is not applied, as there will be only one combination for calculating the route.

2.- Listing of positions. Geographic localities are first grouped by municipality, in alphabetical order; then, they are assigned a numerical index ranging from 1 to n, in chronological order of the occurrence of the fire, for the purpose of estimating the shortest route for the UAV to traverse when reordering with the ACO.

3.- Reordering by ACO. The geographic localities are analyzed by the ACO, municipality by municipality, to reorder the localities according to the shortest route that a UAV will be able to execute.

4.- Presentation of results. Two evaluation scenarios are formed: A and B. The former corresponds to the distance that would be obtained if the route were calculated according to the extraction order of the localities, while the latter scenario applies the ordering with ACO. The results are then compared to calculate the percentage of difference between these and verify the minimization of the travel distance. When the solutions are obtained, a map is generated showing, by means of markers, the fires previously recorded. The order of the geographic localities is then presented by the new positions acquired according to their indexes, e.g., a set of indexes [1 2 3] may change to [2 1 3]; this new order must be sent to the UAV to determine its route. Finally, the theoretical total distance of each route is shown.

Figure 2 consists of two maps; the map in Figure 2a shows the geographic localities that have been extracted from a municipality where forest fires have occurred, Figure 2b shows the route to be taken by a UAV for the verification of a possible fire according to the result obtained by the ACO. The corresponding order of the path shown in Figure 2b is [7, 1, 3, 2, 5, 6, 0, 4].

(a) Map with the geographic localities presented in the map for Jilotepec; (b) Route calculated by the ACO for Jilotepec. Inicio = Start.

Figure 2 Geographic coordinates extracted automaticallyt. 

The proposal is designed to enter data in real time for the calculation of routes in time intervals of minutes to avoid manual calculation.

Experimental conditions

During the experimental phase, Conafor records on forest fires that occurred in the State of Mexico during the year 2022 were obtained; they included only those municipalities where at least three fires occurred in different geographic locations. Two scenarios have been proposed for the experimentation process: A and B. In scenario A, the distance by municipality is calculated as the geographical locations where forest fires occurred were identified; in scenario B, the ACO was applied in each municipality in order to obtain the shortest route with respect to the identified localities. In both scenarios, the total distance of the route is displayed; the reordering of the localities is shown exclusively in scenario B, and, finally, a map is generated with markers indicating, by municipality, the localities where forest fires have previously occurred. The results obtained from each scenario are compared using a statistical average to verify the reduction in the route travel distances by ACO. Figure 3 shows an example of the two experimental scenarios.

(a) Scenario A; (b) Scenario B.

Figure 3 Experimental proposal for distance traversal using ACO. 

The proposal was developed using Python programming software and executed on a Mac Book Air^® device with an M1 processor and 8 GB of RAM. The total number of agents generated for the use of ACO were 1 000.

Results and Discussion

Table 2 shows the results obtained from scenarios A and B, for 51 selected municipalities. As may be seen, the routes generated in scenarios A and B display notable contrasts between each municipality per scenario, with the most efficient cases including more than 16 localities where forest fires have occurred; as can be seen in the case of the municipalities of Acambay, Atlacomulco, Ixtapaluca, and Ocuilan, with verified geographic localities of 45, 18, 49 and 75, respectively. The results of minimization of distances between routes show an average of 54 %, with a maximum of 93 % in the best of cases. The latter percentage was observed with the data obtained in Ocuilan, while in such municipalities as Xalatlaco and Xonacatlán, which include three locations where forest fires occurred, the reduction in the distance was only slightly above 2.5 %.

Figure 4 shows a graph of the most relevant results, where it is possible to compare the reduction in the route traveled by each municipality. The most notable being Valle de Bravo, going from a total route in scenario A of 1 056.56 km to 125.01 km in scenario B. For the previous case, the advantage of the ACO over the calculation of routes that are generated without a distance minimization process is clearly observed. On the other hand, it is observed that in sites where there are three localities with distances of a few kilometers, the reduction is less than 5 %. In municipalities with three localities, but with greater distance between them, better results can be obtained in the minimization process, as observed in the cases of Calimaya and Malinalco.

Figure 4 Comparative graph of mileage reduction when applying the ACO. 

The largest percentage differences with respect to the distance traveled in the two scenarios for the analyzed municipalities were found in Ocuilan, Valle de Bravo, and Morelos, while the smallest decrease was obtained in the routes of Villa Guerrero, Xonacatlán, and Jilotzingo. The difference between a greater or lesser reduction in the distance of the route generated depends on the number of locations, the distance between these, and the probability of a greater number of existing returns.

Figure 5a shows the geographic localities in which forest fires have occurred in the municipality of Amecameca, which are indicated by markers in the form of blue balloons; these contain a reference name (name of the municipality and index of the locality) to provide a spatial reference of where the forest fire has occurred previously. Figure 5b shows the route calculated sequentially, while Figure 5c shows the route calculated with ACO.

(a) Map with aggregate coordinates; (b) Sequentially calculated route; (c) Route calculated by ACO.

Figure 5 Areas where forest fires have occurred, municipality of Amecameca. 

Based on the experimental results thus obtained, it is possible to verify that the ACO offers a viable solution for reducing route travel distances, estimating similar values in the minimization of distances to those cited in the works of ^{Zhou et al. (2019)} and ^{Wu et al. (2020)}. Compared to works involving drones flying in forested areas like those of ^{Sungheetha and Sharma (2020)} and ^{Li et al. (2022)} ―based on UAV devices with cameras or IoT devices deployed on an area of land, but without applying an analysis of the routes in order to reduce the travel distance―, it has the advantage of allowing, through the use of previous fire records, to generate routes to carry out the process of preventing and detecting forest fires, as well as to optimize the energy expenditure in the route of the UAV units.

During the development of the experimental scenarios A and B, no excessive consumption of computational resources such as RAM memory and CPU time was observed; however, the existence an internet connection error was apparent when calculating the distance of the routes to cover all the municipalities, which proved insufficient, as well as when using the data from the selected library, in this case, GeoPy. Therefore, it was necessary to carry out the evaluation by groups of 10 in 10 municipalities in order to avoid this particular result. The operational characteristics of the software allow it to process more data inputs than those used in the experimental phase to obtain data from all the municipalities in the country, as well as to provide real-time solutions, considering that the routes can remain in force for up to 5 years, as long as there are no drastic changes in the areas analyzed, due to changes in the composition or in the precipitation cycle and weather conditions.

Conclusions

Computational evolutionary algorithms, such as ant colony-based genetic algorithms, provide solutions to optimization problems, specifically for the generation of routes that can be traversed in the shortest possible distance by unmanned aerial vehicles. During the experiments, it was observed that despite the large number of points that have been extracted in different municipalities, the tendency was always oriented towards a reduction with respect to the sequential route initially proposed, obtaining in the best case a 93 % reduction of the initial distance, which indicates that, in regions where a large number of fires are detected in a seemingly random fashion, the gain in route planning is highly significant, with a mean of 54 % of all the municipalities analyzed, as well as when the points are distant (more than 10 kilometers).

While it is true that in Mexico there are groups for the prevention and control of forest fires, resources such as staff and time available are limited. Therefore, implementing UAV systems with route generation through ACO for the prevention and detection of fires at strategic points will allow the development of this type of activity in an efficient and effective way in coordination with human personnel. Since the calculation of routes involves a high degree of automation, the amount of data entered can be greater than hundreds of geographic locations, making it highly flexible to contemplate the entry of new locations, which can be implemented in real time to obtain new routes according to the needs that arise in specific contexts.