Assessing Fire Risk Zones in Phrae Province, Northern Thailand, Using a MaxEnt Model

Abstract

This study aimed to investigate the physical factors influencing the occurrence of forest fires and to create a fire risk map of Phrae Province. Remote sensing and geographic information system (GIS) technology were applied for the analysis, focusing on seven factors: the digital elevation model (DEM); slope; Normalized Difference Vegetation Index (NDVI); aspect; and distances from people, water, and roads. All of these geographical factors can affect forest fires. This resulted in a MaxEnt (Maximum Entropy) model with an AUC (area under the curve) of 0.849, indicating its great prediction ability. The findings revealed that the variables influencing forest fire incidence were the DEM, NDVI, slope, distance from roads, distance from water, distance from communities, and aspect, in that order. Subsequently, a fire risk map for wildfires was developed by reclassifying the data into five levels—very low risk, low risk, medium risk, high risk, and very high risk—accounting for 341,395.54, 88,132.64, 76,162.41, 81,157.55, and 57,384.10 hectares or 52.99, 13.68, 11.82, 12.60, and 8.91% of the total area, respectively. The areas classified as very high risk, high risk, medium risk, and low risk included the Song, Long, and Rong Kwang Districts. The area with the lowest risk was Nong Muang Khai District.

1. Introduction

Forest fires are natural disasters that cause severe damage to natural resources, especially in Thailand, where the northern region [1] has a complex mountainous terrain and is mostly covered with forest areas. Forest fires in the northern region not only destroy these forest areas [2,3,4] but also contribute to air pollution problems, such as the presence of fine smoke dust (PM2.5), which have a serious detrimental impact on public health. Forest fires also have an impact on biodiversity [5,6,7]. The loss of wildlife habitats has long-term consequences for residents’ livelihoods and economies. The resulting pollution can also be transported to nearby places [8].

Phrae Province is more than 70% forest by area, characterized by basin terrain and surrounded by cities and agricultural areas. There is a culture there of burning pre-cultivation areas, which is not the ideal approach, and many areas are sheltered from the rain by the mountain ranges that surround Phrae Province [9]. The elements that determine the occurrence and spread of forest fires are broad and complicated, such as the terrain, land use, types of tree species, climate, relative humidity, temperature, and rainfall, along with human activities such as burning leaves or plant residues, burning to prepare agricultural land, and intentionally burning forests. Given all these elements, it is critical to assess and classify the areas at risk of forest fires when planning measures to prevent and reduce forest fire losses using today’s sophisticated science and technology [10,11,12]. As a result, geographic information systems (GISs) have emerged as a significant tool for gathering, maintaining, analysing, and visualizing spatial data on forest fire risk variables in the form of maps or spatial models [13].

For determining the risk of forest fire [14], the use of spatial models to identify and predict danger areas has received extensive attention [15]. The Maximum Entropy Model (MaxEnt) is one of the most popular and internationally accepted statistical models, with the advantage of allowing presence-only data to be used to analyse the links between occurrences (wildfires) and environmental variables, such as sea level elevation or slope [16,17]. MaxEnt can be used to develop a model of the probability of forest fires in connection with these characteristics, allowing for the accurate and efficient identification of regions at high risk of forest fire [18,19].

Combining a geographic information system with the MaxEnt model strengthens its ability to assess and analyse risk areas in more diverse and comprehensive dimensions [18,20,21]. Data from various sources, such as satellite data, topographic data, and weather data, can be used for integrated analyses [14,22]. Land use data, including historical forest fire statistics, can help to pinpoint the exact location and extent of forest fire-prone areas [18,19,20]. The findings of this enquiry can then be used to plan and manage forest fires, including aspects such as resource use, staff preparation, and public relations with the local people, to effectively reduce the damage or impact resulting from forest fires [23,24].

The purpose of this research is to study the application of geographic information systems in conjunction with MaxEnt models. To assess and classify the areas at risk of forest fires in Phrae Province, we collected and analysed data on related environmental factors; we then prepared a map showing these risk areas to support the relevant agencies in sustainably preventing and managing forest fires in the future.

2. Materials and Methods

2.1. Study Area

Phrae Province is located in the northern section of Thailand. It is located at latitude 17.70 to 18.84 degrees north and longitude 99.58 to 100.32 degrees (Figure 1), around 155 m above sea level. Its distance from Bangkok is approximately 555 km by Highways 11 and 101 and 550 km by railway (to Denchai Railway Station). It covers around 6538.59 square kilometres, or 1.27 percent of the country’s total area. It is classed as a medium-sized provincial area with a breadth of around 59 km (measured from the easternmost part of the westernmost city in Long District) and a length of approximately 118 km (measured from the northern end of the second district to the southernmost point of Wang Chien District). Currently, Phrae Province is one of the North’s most important car traffic hubs, connecting to Nan Province. Phayao Province, Chiang Rai Province, Lampang Province, Lamphun Province, Chiang Mai Province, and Phrae Province can all be considered gateways to Lanna [25,26].

To the north, the region borders Phayao and Nan Provinces; to the east, Nan Province; to the west, Lampang Province; and to the south, Ut-taradit and Sukhothai Provinces. Phrae Province is bordered by mountains in all four cardinal directions. Doi Chang Pha Dan is the tallest mountain relative to the mean sea level. Generally, the flat area is between 120 and 200 m above mean sea level, while the city of Phrae is 161 m above mean sea level [27]. The most important river in Phrae Province is the Yom River, which originates in the Phee Pan Nam Mountains. Summer begins at the end of February and lasts until mid-May [28]. The hottest temperature ever recorded there was 46.0 degrees Celsius, in 2020, while the average maximum temperature in April is 37.3 degrees Celsius. The rainy season runs from mid-May until mid-October. The average yearly rainfall is between 1000 and 1500 mm. The lowest temperature ever recorded there was 4.6 degrees Celsius, on 2 January 1974, and the average lowest temperature in January is 14.4 degrees Celsius [29].

2.2. Data Preparation

Hotspot data for a total of 10 years from 2015 to 2024 were imported from the hotspot information system on the Department of Forestry website (https://wildfire.forest.go.th) accessed on 25 April 2024. [30]. We chose from 300 heat spots that appeared in the same pixel during 5 years.

Digital Elevation Model (DEM) data with a spatial resolution of 12.5 m × 12.5 m were obtained from the National Aeronautics and Space Administration (NASA) website (https://asf.alaska.edu) accessed on 25 April 2024. They were utilized to generate geographic data for the slope and aspect parameters [31] and tests for redundancy by Variance Inflation Factor (VIF) analysis. Furthermore, the Normalized Difference Vegetation Index (NDVI) data used in this study represents a seasonal composite of the time when there is the maximum leaf cover, making it ideal for assessing fuel load in the area. This technique was chosen since the research area’s forest cover is largely deciduous, with leaf density varying substantially across the seasons.

The use of high-resolution DEM data (12.5 m by 12.5 m) allows for a more precise description of the terrain and its influence on wildfire hazard. However, high spatial resolution can lead to overfitting, which occurs when the model is too closely matched to the training data, gathering noise or unnecessary information that may not translate well to other places. This potential issue highlights the need of balancing resolution with model robustness, ensuring that the model continues to give accurate predictions across diverse geographies.

Data on road distances (https://datagov.mot.go.th/organization/drr) accessed on 25 April 2024, information on distances from water sources (https://webgis.dwr.go.th/) accessed on 25 April 2024, and distances from residential locations were imported as spatial data in the form of point features at a scale of 1:50,000. The spatial data were then prepared in the form of lines (interpolation). The roads used include both main roads and secondary roads in the area, and interpolation was performed extending to the boundaries of the province.

Additional fires were randomly distributed around the region, and with the occurrence of recurrent fires, 300 locations were analyzed over the research area (Figure 2). These sites were chosen at random from pixels that saw hotspot occurrences each year, with the requirement that the same pixel had fires for at least 5 years, indicating that the region is a high-risk zone for fires (Figure 2).

Repetitive fire point data were exported in “CSV” file extension format for use as an import file for the MaxEnt program to model (https://wildfire.forest.go.th/firemap/index.html, accessed on 25 April 2024).

Data on 7 factors, namely, the numerical height model (DEM), slope; vegetation variance index (NDVI), slope direction (aspect), and distances from roads, waterways, and dwellings, were created in the “GRID” format by ArcGIS 10.6.

2.3. Maximum Entropy Model

MaxEnt is a machine learning algorithm that only considers existential data. This technique seeks the probability distribution that has the highest entropy or is closest to being constant. However, it is vulnerable to the constraints imposed by visible data. As a result, it avoids hypotheses that lack empirical support [32]. When predicting forest fires, MaxEnt assigns the likelihood of ignition to each site based on its consistency with the observed average environment at the point of combustion [33]. Each site is assigned a non-negative probability, and the sum of all assigned probabilities over the entire region is equal to 1.

The model is constrained so that the average of each predictor over the predicted distribution is close to the empirical mean at the observed fire sites. MaxEnt selects the distribution with the greatest consistency, representing the least biased inference based on the available data [32].

In addition, MaxEnt can accommodate complex and nonlinear relationships between predictors and the appearance of fires. It enables different types of functions, such as linear functions and quadratic functions, improving the flexibility and expressiveness of the model. The MaxEnt application allows users to create possible distributions through various parameters that can be used for learning and to generate graphical results. At the same time, it provides accurate results based on AUC (area under the curve) measurements and the jackknife method, a new sampling technique that is useful for estimating bias and variance [33].

2.3.1. Model Customization

The MaxEnt model was run using dedicated software (version 3.4.4) to optimize its performance [34]. We evaluated the model’s performance using the AUC (area under the curve) for the training and the true skill statistic (TSS). All models were trained with a fixed random value (randomseed = true) to ensure that the results were completely reproducible, including the AUC values and response curves. Background points were sampled only within areas with valid fuel type data (e.g., non-blank fuel type values), representing areas where wildfires are ecologically feasible. This restriction is consistent with best practices for ecologically specific modelling, which recommend limiting the background to accessible areas [35] to avoid bias and inflated accuracy values [36]. Final predictions were also obscured in flammable areas of the forest, ensuring consistency between practice and projection.

The Digital Elevation Model (DEM), Slope, Normalized Difference Vegetation Index (NDVI), Aspect, Distance from roads, Distance from water sources, and Distance from communities were chosen for their direct correlation with forest fire incidence and ability to be accurately measured within the study area. We excluded wind speed, land use type, and man-made spark sources because precise data is impossible to acquire in all studied regions. Land use patterns can vary greatly, and there is occasionally a lack of trustworthy high-resolution data. This makes it difficult to make substantial improvements to the model.

Generally, the predictive performance of a MaxEnt model is evaluated using the receiver operating characteristic (ROC) curve and the AUC. The ROC graph shows the true positive rate versus the false positive rate, which is summarized as a single value between 0 and 1. A value of 0 indicates random prediction, while a value of 1 indicates perfect classification. In this investigation, logistic thresholds based on the tenth percentile training presence were used. To establish the 10% minimum threshold, the maxentResults.csv file was inspected, and the column labeled “10th percentile training presence logistic threshold” was chosen. In general, an AUC value below 0.6 indicates weak predictive ability, a value between 0.6 and 0.7 is considered poor, a value of 0.7 to 0.9 is considered moderate, and a score above 0.9 indicates high prediction accuracy [37].

2.3.2. Estimating Uncertainty

To quantify the predictive uncertainty associated with the parameter setting, we also calculated and generated a response curve for each predictive variable and used the jackknife method to highlight the relative influence of each variable [38,39,40,41]. The jackknife method is a binomial test that depends on the criteria, based on omissions and projected areas [33]. The raster was reclassified into five risk levels with the same range width: (0.0–0.2), (0.2–0.4), (0.4–0.6), (0.6–0.8), and (0.8–1.0) [42]. This classification simplifies interpretation and supports practical application in fire management planning.

Also, the linear coherence of DEM, slope, and side is very important since these things have a direct effect on how people utilize land and how close they go to it, both of which are important for figuring out how likely a fire is to happen. As a result, you must often monitor the model’s interactions to ensure that they are realistic.

Additionally, make the model less flexible. This new information will assist users in understanding how uncertain the model is, improving forecast accuracy, and demonstrating how well it performs in various scenarios.

3. Results and Discussion

3.1. Wildfire Vulnerability Model

Model performance is routinely examined via testing and validation. However, the user does not have to establish that the model’s outputs properly represent reality. Here, the omission rate was taken as the proportion of test locations in a pixel that were predicted to be inappropriate for the risk area. The projected region was the proportion of all pixels expected to be at danger of wildfire [43]. Figure 3a depicts the average as red, the anticipated omission rate as black, and the omission rate for the model training sample as light blue. The omission rate was estimated using both the presence log from training and the test log [44]. Furthermore, the receiver operating characteristic (ROC) performance curve was investigated. The ROC curve shows the sensitivity (true-positive percentage). This is the absence of mistakes caused by omissions, as well as the fraction of observed absences that were incorrectly anticipated. The false-positive fraction (1–specificity) is a type of computing mistake. An AUC score of 0.50 shows that the model is close to random and a poor predictor, whereas a value of 1 represents ideal model accuracy. The model’s conclusions should be rigorously verified, given that the ecological gap extends beyond the geographical bounds of the area and does not include all areas at danger of wildfire. In this context, the area was examined to determine the frequency of forest fires and the presence of what was expected to make the model a reality. The AUC value of the training data was near to one (0.849), indicating that the model outperformed random prediction, and thus, validating the model (Figure 3b).

3.2. Environmental Factors Influencing Wildfires

Wildfires are frequently caused by fuel-related environmental conditions in remote locations, including mountains [45,46]. When the environmental variables were utilized alone, mean sea level elevation (pDEM) made the highest contribution to the model’s growth, at 81%, with the vegetation index (pNDVI) coming in second at 67% (Figure 4). A high vegetation index is crucial to this model because Phrae Province is primarily composed of deciduous forest areas, causing large amounts of fallen leaves to be present as fuel during the dry season. Additionally, highland areas with a high vegetation index are subject to fewer intrusion disturbances than easily accessible low-lying areas [47,48]. Furthermore, in line with the findings of earlier research, the model predicted that the regions in danger of wildfires would be determined by physical variables influencing human access [49,50,51].

The MaxEnt model’s response curve demonstrates the role of pDEM in simulating wildfire-risk zones for Phrae Province. It was discovered to be the biggest and most substantial contributor, accounting for 80.2%. According to the current study, the likelihood of fires is high in locations with difficult-to-reach high sea level elevations and resources where both flora and fauna are in high demand by intruders (

Table 1. Percent Contribution and Permutation importance of predictor variables.

Code	Variables	Percent Contribution	Permutation Importance
pDEM	Mean sea level elevation	80.2	78.5
pNDVI	Vegetation index	13.0	12.0
pRoad	Distance from road	3.3	3.2
pWater	Distance from water source	1.2	1.9
pLocation	Distance from urban	0.9	1.9
pAspect	Aspect	0.8	1.0
pSlope	Slope	0.7	1.5

and Figure 5A). The pNDVI vegetation index range of 0.2 to 0.4 shows that this area is most likely to have fires. There are a lot of trees that are ready to fall during the molting season, which provides fuel in the area. Researchers found that the area between 0 and 1000 m had the highest chance of fire because it is easy for people to get to and there are a lot of activities that can accidentally start fires, like using fire in farming areas where hunting fire points aren’t well controlled (Figure 5C). The distance from water sources was 1.2% (see Table 1). This is because it is far from the water supply, which is good for the activities of wild animals that people want to hunt, and there are a lot of wild vegetables that can be eaten. Location distance from urban areas contributed 0.9% (Table 1). As illustrated in Figure 5E, the range is from 10,000 to 12,000 m. Because the distance from the metropolitan region must be remote in order to avoid smuggling into the area to gain access to resources in the illegally limited area. The fire reaction curve to the slope shows that there are a lot of big forest fires in the north. However, this is not really different from the other directions because the slope must also be related to other factors (Figure 5F). The fire response curve to slope shows that slopes of less than 20% are where high forest fires happen, as seen in Figure 5G.

3.3. Identifying Areas at Risk of Wildfires

The MaxEnt model creates continuous raster images with values ranging from 0 to 1, representing areas in danger of wildfire. There are no set standards for selecting the criteria. The model’s performance is determined by the data utilized and the mapping purpose. As a result, it differs depending on the category. We determined the threshold values using several statistical metrics. These values were stored in a file called “maxentResults.csv”. The most widely utilized criterion was the logistical requirements for the minimal training status. The logistical criterion for the 10th-percentile training state was the same as the logistical requirement for the training state’s sensitivity and specificity [32,52]. In this study, the logistical criterion of the 10th-percentile training state was utilized to calculate the minimum threshold of 10%. The maxentResults.csv file was inspected, and a column entitled “Logistics Criteria of the 10th percentile training state” was chosen. The wildfire-prone regions were divided into five levels: very low risk, low risk, medium risk, high risk, and very high risk. These risk levels accounted for 341,395.54, 88,132.64, 76,162.41, 81,157.55, and 57,384.10 hectares or 52.99, 13.68, 11.82, 12.60, and 8.91% of the total area, respectively (

Table 2. Wildfire risk levels in Phrae Province.

Risk Level	Area (ha)	Percentage of Total Area
Very low risk	341,395.54	52.99
Low risk	88,132.64	13.68
Medium risk	76,162.41	11.82
High risk	81,157.55	12.60
Very high risk	57,384.10	8.91
Total area	644,232.24	100.00

and Figure 6). In Figure 5, green areas indicate the places with the lowest fire danger.

3.4. Impact on Management

Among all the districts, Song District has the greatest fire danger area, regardless of risk level, and its very-high-risk area covers 19,234.75 hectares. The very-high-risk zones in Long and Rong Kwang Districts are around the same size: 12,582.15 hectares and 10,992.62 hectares, respectively. The total area that is most likely to be affected is 25,117.56 hectares. In these three districts, fires might easily start due to the proximity of farms to woods. For example, farmers may incorrectly use fire to manage farm commodities or to hunt in the forests [53,54]. On the other hand, Nong Muang Khai District is the safest place at all risk levels, where accessibility to certain areas and vegetation density are key contributors to fire occurrence. Only an area of 874.89 hectares is regarded as being in very high danger, while an area of 9286.48 hectares is thought to be at low risk. Nong Muang Khai District is generally flat, with only a small area near forests; hence, there are fewer forest fires [4,54].

Table 3. Wildfire risk levels by district (ha).

Risk Level	Den Chai	Mueang Phrae	Rong Kwang	Long	Wang Chin	Song	Sung Men	Nong Muang Khai
Very low risk	23,390.26	59,033.74	37,655.86	64,476.01	66,593.11	59,098.11	21,861.96	9286.48
Low risk	10,117.41	8076.82	11,807.18	17,599.91	10,803.63	24,158.76	4339.73	1229.19
Medium risk	8168.84	3942.03	11,218.35	15,971.24	8583.93	22,636.41	4612.84	1028.78
High risk	7970.72	4815.62	14,395.64	16,063.35	7422.96	25,117.56	4275.22	1096.49
Very high risk	4335.62	3278.89	10,992.62	12,582.15	4061.89	19,234.75	2023.27	874.89
Total	53,982.85	79,147.10	86,069.67	126,692.67	97,465.52	150,245.59	37,113.02	13,515.83

shows that much of Song District is at high risk. This suggests that the area needs management and risk reduction that focuses on safety.

4. Conclusions

In this study, MaxEnt and GIS models were combined to develop a wildfire risk assessment model based on known wildfire locations and environmental factors, enabling accurate risk assessments and zoning of forest fires. This analysis was extensive and accurate. The MaxEnt Forest Fire Risk Assessment Model was developed with technical assistance from a GIS, and it used a 2015–2024 dataset of forest fire locations and seven environmental factors in Phrae Province to enable provincial forest fire risk assessments and zoning.

(1)

This study found that geographical factors were the most important environmental components for forest fires, ranked in descending order as follows: altitude, vegetation index, slope, distance from roads, distance from water sources, distance from communities, and slope direction. The areas where forest fires occurred were remote, high above sea level, high in slope, and densely vegetated areas.

(2)

The ways in which environmental variables influence forest fires are complicated and changeable. The response curves of forest fires to the seven specified environmental variables were complex and nonlinear. Forest fires are more likely to occur in distant places with dense vegetation since Phrae Province is covered in deciduous forest, providing plenty of fuel.

(3)

There is considerable geographical variance in the danger of forest fire across Phrae Province. The lowest-risk locations account for 52.99% of the total area, whereas the highest-risk areas account for 8.91%, which is a low figure. Most of the high-risk sites are spread throughout the Song Long and Rong Kwang District areas. This land is adjacent to forests, and agriculture is extensively practiced along the forest edge. Low-risk regions were found in Nong Muang Khi, where the low-lying riverside areas and few forest areas make forest fires less likely to occur and quick to control. However, the areas where forest fires occur are frequently subject to neglect and a lack of interest in controlling fires lit for agricultural purposes, which then spread to forests.

(4)

The MaxEnt model is flexible enough to work in other areas with similar environmental variables. It could also be beneficial for figuring out how likely a wildfire is to happen in nearby areas with similar topography and types of plants.