Forecast Zoning of Forest Fire Occurrence: A Case Study in Southern China

Abstract

Forest fires in the southern region of China pose significant threats to ecological balance, human safety, and socio-economic stability. Forecast zoning the occurrence of these fires is crucial for timely and effective response measures. This study employs the random forest algorithm and geospatial analysis, including kernel density and standard deviation ellipse methods, to predict forest fire occurrences. Historical fire data analysis reveals noteworthy findings: (i) Decreasing Trend in Forest Fires: The annual forest fire count in the southern region exhibits a decreasing trend from 2001 to 2019, indicating a gradual reduction in fire incidence. Spatial autocorrelation in fire point distribution is notably observed. (ii) Excellent Performance of Prediction Model: The constructed forest fire prediction model demonstrates outstanding performance metrics, achieving high accuracy, precision, recall, F1-scores, and AUC on the testing dataset. (iii) Seasonal Variations in High-Risk Areas: The probability of high-risk areas for forest fires in the southern region shows seasonal variations across different months. Notably, March to May sees increased risk in Guangxi, Guangdong, Hunan, and Fujian. June to August concentrates risk in Hunan and Jiangxi. September to November and December to February have distinct risk zones. These findings offer detailed insights into the seasonal variations of fire risk, providing a scientific basis for the prevention and control of forest fires in the southern region of China.

1. Introduction

Forests, serving as integral components of Earth’s ecosystems, play an indispensable and profoundly significant role, exerting crucial importance in maintaining ecological balance [1,2]. Forests serve as valuable habitats for biodiversity, providing a secure refuge and abundant food sources for numerous wildlife species. Additionally, forests play a crucial role in maintaining water cycles and protecting water resources, effectively reducing the occurrence of floods and droughts while regulating the natural water balance. More importantly, forests contribute to global climate regulation by absorbing and storing atmospheric carbon dioxide, aiding in slowing the pace of global climate change [3,4,5]. Furthermore, forests hold immense economic and social significance for human society, providing rich resources, creating employment opportunities, and serving as income sources for local communities, while offering tranquil recreational spaces [6,7,8]. Large-scale forest fires pose severe threats to the environment and society, causing extensive damage to forest ecosystems, harming wildlife habitats, disrupting biodiversity, accelerating soil erosion, leading to water source contamination, increasing the risk of floods and droughts, releasing harmful gases and particles, adversely affecting air quality, and causing catastrophic socio-economic losses [9,10,11,12,13,14,15,16,17].

The importance of forest fire prediction lies in protecting ecosystems, mitigating economic and social losses, and providing theoretical foundations for forestry management agencies to prevent forest fires, allocate firefighting resources rationally, and conduct effective fire prevention efforts.

Predicting forest fires, as a highly researched area, involves interdisciplinary cooperation across atmospheric science, ecology, geography, mathematics, and computer science [18,19,20,21]. Scholars and researchers globally have explored various methods and technologies for accurate fire prediction [10,11]. These models analyze factors like meteorological data, topography, and vegetation types, establishing mathematical relationships between fire risk levels and these factors. For instance, while statistical models, including logistic regression [22,23,24,25], Poisson regression [26,27,28], and Bayesian networks [29,30,31,32], prove robust in generating accurate outcomes, their effectiveness is contingent upon the richness and reliability of the underlying data. Acknowledging and addressing potential limitations related to data quantity, quality, and model assumptions is crucial for refining the precision and applicability of these models in diverse scenarios. Physical models, such as LANDFIRE [33,34], represent a forest fire prediction method based on physical processes and natural laws. This approach focuses on simulating and analyzing the actual physical processes of fires, emphasizing the dynamic behavior and spread mechanisms of fires compared to other models [35]; these models prioritize dynamic simulation, emphasizing intricate fire behavior and propagation mechanisms through simulations of actual fire processes. This emphasis results in detailed and precise fire simulation outcomes, contributing to their efficacy in fire risk assessments and emergency response planning. However, physical models come with inherent limitations. Their operation necessitates extensive parameters and computational resources, a requirement that may pose challenges in terms of estimation and acquisition. Achieving accurate model validation and calibration further adds to the complexity. Additionally, the applicability of physical models can be constrained by data limitations, particularly in regions where obtaining sufficient data for model operation and validation proves challenging.

Machine learning models have gained popularity in recent years for forest fire prediction, utilizing big data analysis and pattern recognition [20,36,37,38,39,40]. These models, including random forests and support vector machines [37,41,42,43,44,45,46], demonstrate significant advantages in handling diverse and large-scale data. Machine learning models offer advantages such as big data processing capabilities and computational efficiency, addressing limitations related to temporal dynamic understanding and data quality is crucial for enhancing their effectiveness in forest fire prediction applications. The random forest algorithm, as a typical representative of machine learning, has lower computational costs compared to other models. However, treating the occurrence of forest fires throughout the year as a holistic study [47,48], while neglecting monthly variations in fire risk, may lead to a limited understanding of fire incidents. This is because the occurrence and spread of forest fires are influenced by various factors, such as meteorological conditions and ecological elements, which may vary significantly between different months.

As one of China’s “Three Major Forest Regions” [49], the Southern Forest Region is renowned for its vast forested areas. The issue of forest fires in this region has gained significant attention. What sets it apart is not the intensity of the fires but rather their distinctive characteristic of high frequency, primarily influenced by the interplay of geographical factors and human activities [44]. Therefore, forest fire prevention in the Southern Forest Region holds immense significance and challenges.

Past research in the Southern Forest Region of China has made significant strides, delving into its unique ecological environment, complex geographical factors, and the impact of various human activities on the carbon emissions from forest fires in this sensitive ecosystem [50]. Prior studies, while valuable, often lack detailed spatiotemporal analysis and actionable monthly predictions for fire risk in the Southern Forest Region. This study fills this gap by employing the random forest algorithm, offering a precise predictive model that considers climate, vegetation, topography, and human factors. Unlike previous approaches, this model supports decision-makers in formulating targeted fire prevention strategies with accurate resource allocation. We also propose specific prevention strategies grounded in the model’s predictions and real-world conditions, aiming to not only reduce fire risk but also maintain ecological balance, supporting sustainable development.

2. Resources and Methods

2.1. The Study Area

As shown in Figure 1, the Southern Forest Region is situated in China, extending from the Qinling Mountains to the south of the Huai River, eastward to the Yungui Plateau. It encompasses nine provinces, including Zhejiang, Anhui, Fujian, Jiangxi, Hubei, Hunan, Guangdong, Guangxi, and Hainan (Taiwan is not included due to data limitations). Geographically spanning central and eastern China, it covers a total land area of 1.386 million square kilometers. Influenced by temperate monsoon and subtropical monsoon climates, the region experiences a warm climate with abundant rainfall, creating highly favorable conditions for plant growth. Consequently, it has become a significant base for artificial forest growth in China [50]. The Southern Forest Region, characterized by vast land and abundant forest resources, stands as a crucial component of China’s forest ecosystems. However, due to climatic conditions and vegetation diversity, this region often becomes a hotspot for forest fires. These fires have the potential to cause significant harm to both ecosystems and socio-economic aspects. Therefore, effective forest fire prevention and management in the Southern Forest Region are of utmost importance.

2.2. Data Sources

As shown in

Table 1. Description of datasets used in this study.

Classification	Data	Resolution	Source	References
Fire hotspot data	Fire occurrence information	1 km	https://earthdata.nasa.gov/ (accessed on 15 August 2021)	[51]
Land-use data	Forest coverage data	1 km	https://www.resdc.cn (accessed on 1 August 2021)	[52]
Meteorological data	Daily average surface temperature/Daily average relative humidity/Daily maximum surface temperature/Accumulated precipitation from 20:00 to 20:00/Daily sunshine hours/Mean wind speed/Daily minimum relative humidity, etc.	-	https://data.cma.cn (accessed on 15 August 2020)	[10]
Economic and community	Gross Domestic Product, Population/Road grid/Residential area/Public holiday, etc.	1 km, 1 km, 1:100,000, 1:100,000	https://www.resdc.cn https://www.webmap.cn (accessed on 15 January 2022)	[10,11,52,53]
Vegetation	Fractional vegetation cover	1 km	https://www.resdc.cn (accessed on 15 January 2022)	[10]
Topographic	Slope/Altitude/ Orientation	1 km	https://www.resdc.cn (accessed on 15 January 2023)	[26,54]

, the data utilized in this study mainly encompass fire point data, land-use data, meteorological data, socio-economic data, vegetation data, and topographic data. Firstly, during the data collection phase, it is crucial to ensure the accuracy and completeness of the data. Fire point data can be obtained through satellite remote sensing or ground observations, requiring organization based on time and location to establish the spatiotemporal distribution of fire events. Land-use data, including forest coverage, assist in analyzing regions prone to fires. Meteorological data consider parameters such as temperature, humidity, and wind speed to understand the meteorological conditions’ impact on fires. Socio-economic data may include population and economic indicators to assess the influence of social factors on fire risk. Vegetation data describe the extent and types of vegetation significantly affect fire risk. Topographic data include elevation and slope, contributing to the analysis of topography’s impact on fire spread. Although lightning data are crucial for triggering wildfires, their reliability and comprehensiveness are lacking, preventing their inclusion in our predictive model. Human activities, ranging from agricultural practices to illegal burning, add complexity due to operational difficulties in collecting and quantifying data. Their impact is also intertwined with geographical, cultural, and socio-economic factors, making accurate prediction challenging. We indirectly account for human activities through factors like population data, residential areas, GDP, and special holidays. To ensure the reliability and consistency of the in-depth forest fire occurrence prediction data analysis in this study, we have undertaken a series of crucial steps involving the integration of diverse data sources, including meteorological, topographic, socio-economic, and vegetation data, to establish a unified spatiotemporal framework. During the data cleaning phase, emphasis was placed on addressing missing values, outliers, and duplicate entries to guarantee data integrity and accuracy. This process was pivotal in maintaining the overall quality of the dataset. We collected information from various sources, encompassing meteorological, topographic, socio-economic, and vegetation data, and integrated them to establish a unified spatiotemporal baseline. This step was undertaken to ensure consistency across different data sources, allowing for a comprehensive analysis within a cohesive framework. The integrated dataset facilitates a holistic understanding of the multifaceted factors influencing forest fire occurrences. In the final stage, normalization of the data was conducted. This involved standardizing various data types to ensure comparability during analysis. Normalization played a crucial role in eliminating the influence of different scales and units, enabling variables to be compared on the same magnitude. This normalized dataset provides a foundation for building robust and comparable predictive models. This study conducted normalization of the data for these purposes.

2.3. Method

In this study, a comprehensive dataset encompassing fire point data, land-use information, meteorological records, socio-economic indicators, vegetation details, and topographic measurements was employed. To ensure consistency and balance among these diverse datasets, normalization techniques were applied, minimizing disparities in input and output magnitudes, equalizing the contributory rates of various factors, and resolving unit measurement inconsistencies. Following this, spatial autocorrelation and standard deviation ellipse methods were utilized for an in-depth examination of forest fires’ spatiotemporal patterns. Additionally, the random forest algorithm was employed to forecast and delineate forest fire occurrences, enhancing the comprehension of wildfire distribution and associated risks. Subsequently, a meticulous analysis of the regional delineation outcomes was conducted, providing pertinent suggestions to aid decision-makers in effectively managing and mitigating forest fire hazards. A visual representation of the entire research methodology is presented in Figure 2, and to ensure the model’s trustworthiness, an accuracy assessment of the predictive outcomes was undertaken.

2.3.1. Spatial Autocorrelation Analysis

“Spatial autocorrelation” is a statistical analysis method used to study the correlation or dependency between different locations in spatial data. Through spatial autocorrelation analysis, researchers can better understand the distribution characteristics of geographic data, reveal potential spatial patterns, and analyze the interactions between locations [55].

Global autocorrelation is an analytical method that measures whether the distribution of a phenomenon in an entire geographical area exhibits spatial patterns. It involves the correlations between all locations within the entire geographical area. Global autocorrelation typically employs statistical indicators (such as Moran’s Index) to determine whether the data show trends of clustering, dispersion, or randomness. If the data exhibit significant clustering or dispersion trends in the global autocorrelation analysis, it indicates the presence of some spatial pattern [56]. Local autocorrelation focuses on the correlation between a specific location and its neighboring locations to determine whether there is clustering or dispersion around that particular location.

The formulas are as follows [55,57,58]:

Global autocorrelation:

$$I=\frac{n{{\displaystyle \sum }}_{i=1}^n{{\displaystyle \sum }}_{j=1}^n{W}_{ij}\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{n{{\displaystyle \sum }}_{i=1}^n{{\displaystyle \sum }}_{j=1}^n{W}_{ij}{\left(x_i-\overline{x}\right)}^2},$$

(1)

In this formula, I is the global Moran’s I index, $n$ is the number of spatial units, $W_{ij}$ represents spatial weights between units $i$ and $j$, $x_i$ and $x_j$ are the values of variable $x$ for units $i$ and $j$, and $\overline{x}$ is the mean of variable $x$.

Local autocorrelation:

$$I^{\prime }=[n\left(x_i-\overline{x}\right){{\displaystyle \sum }}_{j=1}^n{W}_{ij}\left(x_j-\overline{x}\right)]/{{\displaystyle \sum }}_{i=1}^n{\left(x_i-\overline{x}\right)}^2,$$

(2)

In this formula, $I^{\prime }$ is the local Moran’s I index, $n$ is the number of spatial units, $W_{ij}$ represents spatial weights between units $i$ and $j$, $x_i$ is the value of variable $x$ for unit $i$, $\mathrm{and} \overline{x}$ is the mean of variable $x$.

Global autocorrelation is used to identify spatial patterns within an entire geographical area, while local autocorrelation is employed to study local variations and spatial heterogeneity in greater detail within a geographic region. Autocorrelation patterns include H–H (high–high), H–L (high–low), L–H (low–high), and L–L (low–low), which are used to describe the correlation between different locations in geographic space: H–H (high–high) pattern: In a geographic space, high-value locations cluster around each other, indicating mutual association in high-value areas, forming a high–high clustering phenomenon; H–L (high–low) pattern: In a geographic space, high-value locations cluster around low-value locations, indicating an association between high-value areas and low-value areas in adjacent regions, forming a high–low clustering phenomenon; L–H (low–high) pattern: In a geographic space, low-value locations cluster around high-value locations, representing an association between low-value areas and high-value areas in adjacent regions, forming a low–high clustering phenomenon; L–L (low–low) pattern: In a geographic space, low-value locations cluster around each other, indicating mutual association in low-value areas, forming a low–low clustering phenomenon.

2.3.2. Standard Deviation Ellipse

Standard deviation ellipse is a statistical graphic used for visualizing and describing the distribution of data. It is commonly employed to investigate the degree of dispersion and variability in data, particularly in multivariate data analysis, to understand the relationships between multiple variables [59]. The primary purpose of the standard deviation ellipse is to visualize the distribution characteristics of data, aiding observers in comprehending the variability of data and the relationships among multivariate variables [60].

The formula is as follows [61]:

$$SD{E}_x=\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{(x_i-\overline{X})}^2}{n}},SD{E}_y=\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{(y_i-\overline{Y})}^2}{n}},$$

(3)

In this formula, $SD{E}_x$ and $SD{E}_y$ are the standard deviations of variables $x$ and $y$, respectively, $n$ is the number of observations, and $\overline{X}$ and $\overline{Y}$ are the means of variables $x$ and $y$, respectively.

$$\tan \theta =\frac{({{\displaystyle \sum }}_{i=1}^n{\tilde{x}}_i^2-{{\displaystyle \sum }}_{i=1}^n{\tilde{y}}_i^2)+\sqrt{{({{\displaystyle \sum }}_{i=1}^n{\tilde{x}}_i^2-{{\displaystyle \sum }}_{i=1}^n{\tilde{y}}_i^2)}^2+4{({{\displaystyle \sum }}_{i=1}^n{\tilde{x}}_i{\tilde{y}}_i)}^2}}{2{{\displaystyle \sum }}_{i=1}^n{\tilde{x}}_i{\tilde{y}}_i},$$

(4)

In this formula, $\tan \theta$ is the tangent of the rotation angle, and ${\tilde{x}}_i$ and ${\tilde{y}}_i$ are the rotated coordinates of points $i$ in the new coordinate system.

$${\sigma }_x=\sqrt{2}\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{\left({\tilde{x}}_i\cos \theta -{\tilde{y}}_i\sin \theta \right)}^2}{n}},$$

(5)

$${\sigma }_y=\sqrt{2}\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{\left({\tilde{x}}_i\sin \theta +{\tilde{y}}_i\cos \theta \right)}^2}{n}},$$

(6)

In this formula, ${\sigma }_x$ and ${\sigma }_y$ are the standard deviations of the rotated coordinates, and ${\tilde{x}}_i$ and ${\tilde{y}}_i$ are the rotated coordinates of points $i$ in the new coordinate system.

The standard deviation ellipse is a powerful spatial analysis tool that plays a crucial role in wildfire research. By employing the standard deviation ellipse, we can accurately detect spatial clustering or dispersion of wildfires. When the major axis of the standard deviation ellipse aligns with the distribution direction of a specific area, it may imply the potential risk of wildfire clustering in that region. Therefore, utilizing the analysis results of the standard deviation ellipse, we can implement appropriate preventive measures to address wildfire clustering, ensuring that timely and effective protective actions are taken to reduce the threats and losses posed by wildfires.

2.3.3. Random Forest Model

Random Forest is an ensemble learning method widely used for classification and regression problems [62]. It enhances prediction accuracy and stability by combining multiple decision tree models. Each decision tree within the Random Forest framework serves as an autonomous classification or regression model. The construction of these decision trees relies on training data; however, to promote diversity, Random Forest incorporates randomness during the development of each tree. When partitioning nodes within a decision tree, Random Forest considers a randomized subset of data features rather than the entire set. This approach mitigates correlation among decision trees, fostering model diversity [63,64].

In Random Forest, for classification problems, the final classification result is determined by voting on the predictions of each decision tree. For regression problems, the predictions of multiple decision trees are averaged. Due to the stochastic feature selection and the ensemble of multiple decision trees, Random Forest exhibits good resistance to overfitting and often requires minimal parameter tuning [65].

The occurrence of forest fires is influenced by various factors, including meteorological conditions, land use, and vegetation cover. Random Forest is effective in handling large amounts of multivariate data, making it suitable for complex forest fire prediction problems [66]. In random forests, decision trees typically use the Gini index or information gain to select split nodes. The Gini importance of a feature is calculated by comparing the change in the Gini index before and after each node split. This method assesses the contribution of a feature to reducing impurity.

In summary, the Random Forest model has become a powerful tool in forest fire management and prevention due to its various advantages. Its high accuracy, ability to handle multivariate data, resistance to interference, and feature importance analysis enable decision-makers to assess forest fire risks more accurately and comprehensively. This, in turn, allows for targeted preventive measures to be taken, maximizing the protection of ecosystems and human communities.

In this study, after multiple attempts and careful parameter adjustments, we successfully identified the optimal random forest model, comprising 85 decision trees. This meticulously chosen model, fine-tuned by adjusting the number of trees in the random forest, ensures outstanding performance in predicting forest fire occurrences. Consequently, this configuration is deemed the most suitable for addressing the research question. The model selection in this study has been thoughtful, aiming to provide an accurate and reliable solution for forest fire prediction.

2.3.4. Assessment Criteria

Accuracy, Precision, Recall, F1, and AUC are widely employed metrics in machine learning and statistics for assessing model performance. These metrics are utilized to evaluate the effectiveness of classification models. The formulas are as outlined below [10,11,67]:

$$\mathit{Accuracy}=(\mathit{TP}+\mathit{TN})/(\mathit{TP}+\mathit{FP}+\mathit{TN}+\mathit{FN}),$$

(7)

$$\mathit{Recall}=\mathit{TP}/(\mathit{TP}+\mathit{FN}),$$

(8)

$$\mathit{Precision}=\mathit{TP}/(\mathit{TP}+\mathit{FP}),$$

(9)

$$F1=2\times (\mathit{Precision}\times \mathit{Recall})/(\mathit{Precision}+\mathit{Recall}),$$

(10)

In a binary classification scenario, True Positive (TP) represents the instances correctly identified as positive by the model, True Negative (TN) denotes instances correctly identified as negative, False Positive (FP) signifies instances incorrectly labeled as positive, and False Negative (FN) indicates instances incorrectly labeled as negative. These metrics provide a basis for evaluating the accuracy and effectiveness of a classification model by quantifying correct and incorrect predictions in both positive and negative categories.

2.3.5. Monthly Forest Fire Prediction and Zones

In order to express the occurrence probability of forest fires more intuitively, we introduced a 5-level rating system for qualitative assessment of the danger level of forest fire occurrences [10,11]. Specifically: (i) Forest fire occurrence probability from 0 to 0.2 is rated as Level 1 risk, indicating a minimal likelihood of forest fire incidents; (ii) Forest fire occurrence probability from 0.2 to 0.4 is rated as Level 2 risk, signifying a low likelihood of forest fire incidents; (iii) Forest fire occurrence probability from 0.4 to 0.6 is rated as Level 3 risk, suggesting a moderate likelihood of forest fire incidents; (iv) Forest fire occurrence probability from 0.6 to 0.8 is rated as Level 4 risk, denoting a high likelihood of forest fire incidents; (v) Forest fire occurrence probability from 0.8 to 1 is rated as Level 5 risk, representing an extremely high likelihood of forest fire incidents.

3. Results

3.1. Changing Trends of Forest Fire Occurrence in Southern China

As depicted in Figure 3, the Southern Forest Region exhibits a notable downward trend in the number of forest fire points over interannual variations, reflecting sustained efforts and effective forest fire prevention measures. From 2001 to 2019, a significant shift in this trend can be discerned. Initially, the data from 2008 indicate a comparatively high number of forest fire points, reaching 6392. This elevation in numbers might be attributed to the influence of meteorological conditions and other environmental factors, leading to widespread fires. However, there has been a gradual decline in the number of forest fire points since then. Starting from 2004, the count decreased from 4845 to 2059 in 2015 and further diminished to 855 in 2019.

Regarding seasonal distribution, a higher concentration of forest fire points is observed during winter, spring, and autumn. Winter and spring seasons are typically characterized by dryness, rendering forests vulnerable to drought and associated risks, thus resulting in a comparatively higher number of fire incidents. Autumn might also experience an escalated risk of fires due to factors like wind and drought. Conversely, the summer season witnesses a relatively lower number of fire points. This could be attributed to increased precipitation, which aids in suppressing fire occurrences.

3.2. Autocorrelation Analysis Results of Forest Fire Occurrence in Southern China

As shown in Figure 4, the global autocorrelation analysis in the Southern Forest Region distinguishes city types based on their similarities and differences in socio-economic characteristics. The L–L (Low–Low)-type cities, totaling 63, are primarily concentrated in provinces like Hubei, Anhui, and Zhejiang, indicating regions with similar lower levels of certain attributes. Conversely, the 27 L–H (Low–High)-type cities, clustered in areas such as Hunan, Fujian, and Guangdong, demonstrate a mix of low and high values in specific socio-economic indicators. The H–L (High–Low)-type cities, though limited to three, signify unique areas where certain attributes are high while others are low. The most prevalent are the 48 H–H (High–High)-type cities, distributed across provinces like Guangxi, Guangdong, Fujian, Jiangxi, and Hunan, indicating regions with consistently high values across multiple indicators. Local autocorrelation analysis further refines these patterns, providing detailed insights into the specific spatial arrangements and co-occurrences of cities with similar socio-economic characteristics in the Southern Forest Region.

3.3. The Results of Standard Deviation Ellipse

As depicted in Figure 5 and summarized in

Table 2. Southern China forest fire occurrence oval parameters’ standard deviation.

Year	XStdDist (km)	YStdDist (km)	Shape_Length (km)	Shape_Area (km2)	Rotation (°)
2001	513.48	257.36	2488.88	415,116.33	45.31
2002	525.64	276.84	2581.94	457,117.00	50.21
2003	535.90	243.99	2536.59	410,730.43	50.66
2004	505.58	270.13	2493.36	429,021.98	51.09
2005	561.84	300.10	2770.59	529,661.20	48.89
2006	498.15	259.63	2439.90	406,283.93	59.67
2007	499.33	257.68	2439.12	404,185.56	54.99
2008	474.76	266.62	2375.20	397,641.16	55.94
2009	443.83	249.19	2220.26	347,424.36	59.17
2010	605.72	300.76	2928.86	572,277.90	64.45
2011	488.84	290.72	2488.72	446,441.48	55.08
2012	502.52	305.35	2575.90	482,033.26	50.54
2013	475.88	296.45	2459.14	443,168.63	47.73
2014	455.87	288.67	2368.55	413,385.56	41.94
2015	487.27	264.91	2414.90	405,497.93	53.57
2016	496.91	292.63	2522.03	456,787.78	43.75
2017	477.31	248.15	2336.24	372,067.93	53.05
2018	525.32	261.24	2541.10	431,094.48	54.44
2019	583.22	280.33	2796.93	513,585.51	42.88

, the generation of standard deviation ellipses for forest fires in the southern region spanning from 2001 to 2019 allows for a visual representation of spatial clustering patterns and temporal variations in these incidents. Given that over 90% of forest fires result from human activities, the ellipses predominantly encompass areas characterized by high population density, including provinces such as Guangdong, Guangxi, Hunan, Jiangxi, and Fujian. Particularly noteworthy is the substantial coverage over Guangdong, aligning with its status as a densely populated province in China. Consequently, in the context of forest fire prevention and control, it is advisable to intensify governance efforts specifically in these high-incidence regions, concurrently emphasizing the management of ignition sources and increasing public awareness of fire prevention.

Upon analyzing the shifting center of the ellipse over time, a discernible northward trend in the central position of forest fire occurrences in the southern region emerges. In 2001 and 2007, the center was situated in Guangdong, whereas in 2013 and 2019, it shifted to Hunan, with the displacement range being relatively modest. The change in the azimuth angle of the ellipse from 2000 to 2019 remained minimal, fluctuating between 42.87° and 54.99°. This indicates that the long axis of the ellipse consistently aligns in the southwest–northeast direction, corresponding to the spatial distribution pattern of forest fire incidents. The short semi-axis fluctuates within the range of 257–296 km, peaking in 2013, while the long axis is concentrated within 499–583 km, displaying notable variations in 2019. Examining the trends of the short and long axes reveals a relatively stable pattern in the east–west direction and an expanding trend in the north–south direction. The changing flattening ratio suggests that the east–west expansion in fire-prone areas was more pronounced in 2013 than the north–south expansion, whereas in 2019, the east–west expansion trend was weaker compared to the north–south expansion trend.

These findings offer valuable insights into the spatial and temporal dynamics of forest fires in the southern region from 2001 to 2019, contributing to informed decision-making for the development of effective prevention and control strategies.

3.4. Assessment of Predictive Accuracy for Forest Fire Occurrence in Southern China

Figure 6 presents the model’s performance metrics, including accuracy, precision, recall, F1-score, and AUC. For the testing dataset, the accuracy is 95.87%, precision is 95.35%, recall reaches 96.98%, the F1-score is 96.16%, and the AUC measures at 98.97%. Remarkably, the model utilized by the research institute exhibits exceptional accuracy in large-scale studies, surpassing comparable models [68]. Moreover, it is imperative to delve into the practical implications of these performance metrics. Discussing how these results translate into real-world scenarios and the potential benefits they offer in terms of decision-making, resource allocation, or problem-solving is crucial.

3.5. Assessment of Factors’ Importance in Forest Fire Occurrence in Southern China

As shown in Figure 7, the assessment of the significance of forest fires in the southern forested areas of China reveals that temperature, vegetation coverage, and humidity are the most crucial factors influencing the occurrence of fires. The rise in temperature increases the risk of forest fires. In warm and dry weather conditions, moisture in plants and soil evaporates rapidly, making the forest dry and prone to ignition. Furthermore, high temperatures accelerate the decomposition of combustible materials, releasing more flammable gases, and thereby increasing the likelihood of fire occurrences. Changes in vegetation coverage also impact forest fires. Dense vegetation provides more fuel for fires, intensifying their severity. On the other hand, a decrease in vegetation coverage may lead to exposed ground, making the spread of fires easier.

Humidity is another vital factor influencing the occurrence of forest fires. Low humidity causes rapid evaporation of moisture in plants and soil, contributing to the dry and flammable nature of the forest. In addition to these three key factors, GDP and population density also affect forest fires. GDP, as a critical indicator of economic activity, plays a significant role in southern China’s forest fires. With the growth of GDP, expanded economic activities contribute to increased natural resource development. Infrastructure projects, agricultural expansion, and industrial processes associated with economic growth may damage the forest environment, raising the risk of fires. Economic development also influences land-use patterns, potentially exacerbating fire incidents. Population density is another crucial factor affecting forest fires in southern China. Higher population density, associated with increased human habitation and activities, increases the number and distribution of fire sources. Daily life, agricultural practices, and industrial activities become potential ignition sources, while high population density may lead to excessive use and depletion of forest resources, further elevating the risk of fire occurrences.

3.6. Monthly Forest Fire Prediction Zones in Southern China

As shown in Figure 8, the distribution of high-risk areas in the Southern Forest Region exhibits seasonal variations in different months, reflecting the seasonal characteristics and the impact of climate conditions on fire risk. The following is a more detailed description of high-risk areas in different months:

(i)

March to May: High-risk areas are mainly distributed in Hezhou, Baise, and Fangchenggang in the Guangxi Zhuang Autonomous Region; Huizhou and Meizhou in Guangdong Province; as well as Hengyang and Shaoyang in Hunan Province; and Fuzhou in Fujian Province. During this period, these areas may face an increased fire risk as vegetation becomes drier and more flammable, intensifying the potential for wildfires.

(ii)

June to August: This period is characterized by high temperatures and dry conditions, with high-risk areas concentrated in Hengyang and Shaoyang in Hunan and Jiujiang in Jiangxi. The combination of high temperatures and dry weather conditions makes vegetation more prone to ignition, leading to a faster spread of wildfires.

(iii)

September to November: In the autumn season, high-risk areas are mainly distributed in Qinzhou, Nanning, and Wuzhou in the Guangxi Zhuang Autonomous Region; Heyuan, Zhaoqing, and Meizhou in Guangdong Province; Sanming, Fuzhou, and Nanping in Fujian Province; Zhuzhou and Hengyang in Hunan Province; as well as Ji’an and Ganzhou in Jiangxi. During this period, there may be an increased fire risk due to the dryness of autumn and the accumulation of flammable vegetation caused by autumn winds.

(iv)

December to February: In the cold winter, high-risk areas are mainly located in Hezhou and Nanning in the Guangxi Zhuang Autonomous Region; Qingyuan, Heyuan, Huizhou, and Shaoguan in Guangdong Province; Fuzhou, Sanming, and Nanping in Fujian Province; Ganzhou and Fuzhou in Jiangxi Province; as well as Shaoyang and Hengyang in Hunan Province. These areas face higher fire risks during this period as low temperatures and dry weather conditions may promote the occurrence of wildfires.

4. Discussion and Conclusions

4.1. Discussion

The currently widely adopted Forest Fire Weather Index (FWI) system plays a crucial role in assessing the potential occurrence risk of forest fires. However, its relatively simple design, focusing primarily on meteorological factors, calls for a more comprehensive evaluation of fire risk that considers multiple key elements to better address practical needs [69].

Firstly, while the significance of meteorological factors is undeniable, as they directly influence fire occurrence, spread, and suppression, it is crucial to recognize that meteorological factors are just one among many elements constituting fire risk. Other factors, such as topographical features, vegetation types and conditions, soil moisture, human activities, and the history of past fires, have profound impacts on the potential risk of fires. Therefore, a more comprehensive fire risk assessment framework must be established by considering these factors collectively.

Secondly, a comprehensive fire risk assessment system should integrate data from multiple sources and at various scales. Satellite observations, ground monitoring, remote sensing technologies, and other data sources can provide more comprehensive information on aspects such as vegetation conditions, soil moisture, and meteorological changes. By integrating these diverse data, we can more accurately determine fire-prone areas and adjust fire prevention strategies promptly. Moreover, the application of advanced data processing technologies, such as machine learning and artificial intelligence, holds the potential to enhance the accuracy and efficiency of fire risk assessment.

Compared to other studies [70,71,72], our research adopts a multi-tiered approach to comprehensively address the issue of fire risk in the Southern Forest Region. By utilizing MODIS fire point data and incorporating multiple factors such as meteorological conditions, topographical features, and vegetation types, this study can more comprehensively and accurately predict and identify potential fire-prone areas. We employed the random forest algorithm, a powerful machine learning tool capable of uncovering potential patterns and trends from extensive data. By constructing a forest fire prediction model for the Southern Forest Region, we gain a better understanding of the patterns and influencing factors of forest fires, aiding in more effective responses to potential risks.

In addition to model construction, zoning work has been conducted, dividing the Southern Forest Region into different levels of fire occurrence zones for a more targeted development of fire prevention strategies and resource allocation. This helps improve the efficiency and accuracy of fire responses, protecting natural ecosystems and minimizing damage to communities and the environment. Furthermore, spatial analysis methods such as spatial autocorrelation and standard deviation ellipse have been applied to further investigate the patterns of fire distribution and spread. These methods help identify fire risk hotspots and potential paths of fire spread, providing more scientific foundations for disaster management and emergency response. This study conducted zoning work and applied spatial analysis methods for these purposes.

However, this study has some limitations. Firstly, the availability and quality of data may impact the accuracy of the model. Therefore, further improvements in data collection and processing methods will contribute to enhancing the model’s performance. Secondly, the predictive ability of the model may be influenced by future climate [73,74,75] and socio-economic changes [76,77], necessitating consideration of long-term prediction uncertainties. The issue of forest fires in the Southern Forest Region is highly complex, influenced by various factors, and requires comprehensive preventive measures. To effectively address this challenge, measures must be implemented on multiple fronts, including early warning systems, regulations, education, firebreaks, and the application of new technologies. The goal of these comprehensive measures is to minimize potential fire risks, protecting the crucial ecosystem of the Southern Forest Region. In the prevention and control of forest fires, the establishment of early warning systems, meteorological monitoring stations, and fire observation towers plays a crucial role [78].

These facilities not only help monitor meteorological conditions and fire risk levels but also enable the early detection of potential fire risks, allowing local authorities to take timely preventive measures, such as increased patrols and advance notification of residents. The establishment of strict regulations is also a cornerstone of forest fire prevention and control. These regulations specify the time and location for outdoor burning and prohibit illegal logging and deforestation, ensuring the sustainable utilization of forest resources. Through a robust legal system, violations can be curtailed, enhancing the effectiveness of fire prevention. In addition to policy and technological means, public education and awareness campaigns are equally vital [79].

By raising public awareness of fire prevention, educating people on how to use fire safely, manage waste properly, and deeply understand the hazards of forest fires, active participation in public education and awareness activities can play a positive role in fire prevention. This engagement fosters a societal atmosphere of collective effort to protect forest resources. Fire prevention engineering, including the establishment of firebreaks, water sources, and equipment, is also a crucial aspect of preventing and controlling forest fires. These engineering measures can respond rapidly to fires, effectively curb the spread of fires, and reduce the damage caused by fires. Overall, for the protection of the valuable ecosystem of the Southern Forest Region, comprehensive preventive measures are indispensable. Establishing early warning systems, enacting regulations, conducting education and awareness campaigns, creating firebreaks, and employing modern technology are all key elements. However, ongoing monitoring, education, and public participation should also be part of long-term efforts to ensure the sustainable protection of this ecological treasure. Only through comprehensive and sustained efforts can we better protect the ecological environment of the Southern Forest Region, reducing the occurrence and impact of fires.

While acknowledging the importance of lightning data in igniting wildfires, the current limitation lies in the lack of reliable and comprehensive records within our specific geographic area and timeframe, hindering the inclusion of such data in our predictive model. Adding to the complexity, human activities manifest in diverse forms, including agricultural practices, indiscriminate disposal of cigarette butts, camping, and illegal burning, all potential triggers for wildfire incidents. The comprehensive and accurate collection and quantification of these activity data pose significant challenges. Moreover, the intricate connection between the impact of human activities and geographical conditions, cultural backgrounds, and socio-economic situations adds layers of complexity to accurately predict their effects.

In our present study, we attempt to indirectly consider the influence of human activities by introducing additional factors such as population data, residential area information, Gross Domestic Product (GDP), and special holidays. Our approach encompasses multiple key factors in forest fire prediction, leveraging the random forest algorithm to provide a competitive advantage in forecasting forest fires. It is crucial to note that different research methods and datasets may yield varying results. Thus, the conclusions drawn here only reflect predictive outcomes under specific conditions, and alternative studies might arrive at different conclusions in different environments with distinct data.

Our future research plans involve continual refinement of our model’s performance to achieve more accurate predictions of forest fire occurrence probabilities. This entails considering additional factors that influence fire risk, such as human activities and changes in land use. We also aim to explore finer spatial and temporal resolutions to precisely identify and predict potential fire-prone areas. Furthermore, long-term monitoring and model validation are imperative to deepen our understanding of the evolution and trends of fire risk. This approach will assist in addressing forthcoming challenges in fire risk and formulating more effective strategies for fire management and prevention.

In the international context, our forest fire prediction method can be applied to other regions with forest cover, with a wide range of potential uses. Firstly, the method can help countries manage and protect their forest resources more effectively. By accurately predicting the occurrence of forest fires, governments and relevant institutions can take timely preventive measures to reduce the damage of fires to ecosystems and human communities. This is crucial for protecting biodiversity, maintaining ecological balance, and promoting sustainable development. Secondly, our method can support cross-border cooperation and resource sharing. By sharing predictive data and experience, different countries can jointly address forest fire issues and optimize resource utilization. This cooperation can promote knowledge exchange and technological innovation, and promote the development of forest fire prevention around the world. In addition, this method can also provide support for international disaster management. Forest fires often have transboundary impacts, and through transnational cooperation and coordination, large-scale fires can be better addressed and controlled. Our prediction tool can provide international disaster management agencies with information on the likelihood and risk level of fires, which can help develop effective emergency plans and resource allocation strategies.

4.2. Conclusions

This study presents a comprehensive examination of forest fires in Southern China, emphasizing their profound implications for ecological balance, human safety, and socio-economic stability. The pressing need to predict forest fires, particularly in a region renowned for its biodiversity and vital ecosystem services, is paramount for enabling timely responses and fostering sustainable development. Through the strategic integration of the random forest algorithm with geospatial analysis techniques such as kernel density and standard deviation ellipse, our research offers robust predictions of forest fire occurrences in the Southern Forest Region. An in-depth analysis of historical fire data spanning from 2001 to 2019 reveals a commendable decline in fire incidents, which could be attributed to various factors such as enhanced management strategies, increased public awareness, and climatic influences.

(i)

Efficacy of the Prediction Model: The forest fire prediction model developed in this research exhibits outstanding performance metrics, underscoring the efficacy of the employed methodologies and algorithms. This scientific foundation paves the way for future research and the development of more precise forest fire risk prediction tools.

(ii)

Seasonal Variations and Forest Fire Occurrence: Our investigation identifies high-risk areas that exhibit seasonal variations, providing a nuanced understanding of the influence of climatic conditions on fire risk. The drier vegetation observed from March to May, followed by elevated risks during the hot and dry months from June to August, underscores the distinct challenges each season poses.

(iii)

Analysis of Historical and Current Trends: A thorough examination of historical fire records highlights a notable reduction in fire incidents, potentially linked to improved management strategies, heightened public awareness, and the impacts of climate change. The continued monitoring of these trends is paramount for evaluating the effectiveness of existing measures and pinpointing areas that require further attention.

(iv)

Contributions to Sustainable Development and Environmental Protection: In essence, our findings contribute significantly to environmental conservation and hold far-reaching implications for the sustainable development of the Southern Forest Region. The practical implementation of insights derived from this research can guide the formulation of effective policies and strategies geared towards preventing and controlling forest fires, thereby safeguarding the health of the region’s ecosystems and the well-being of its inhabitants.

In summation, our research underscores the importance of sustained monitoring, scientific prediction, and targeted intervention in mitigating forest fires in the Southern Forest Region. By integrating advanced analytical methodologies with geospatial techniques, we provide a powerful tool for informed decision-making in support of ecological and socio-economic sustainable development in the region.