Predicting the Occurrence of Forest Fire in the Central-South Region of China

Abstract

Understanding the spatial and temporal patterns of forest fires, along with the key factors influencing their occurrence, and accurately forecasting these events are crucial for effective forest management. In the Central-South region of China, forest fires pose a significant threat to the ecological system, public safety, and economic stability. This study employs Geographic Information Systems (GISs) and the LightGBM (Light Gradient Boosting Machine) model to identify the determinants of forest fire incidents and develop a predictive model for the likelihood of forest fire occurrences, in addition to proposing a zoning strategy. The purpose of the study is to enhance our understanding of forest fire dynamics in the Central-South region of China and to provide actionable insights for mitigating the risks associated with such disasters. The findings reveal the following: (i) Spatially, fire incidents exhibit significant clustering and autocorrelation, highlighting areas with heightened likelihood. (ii) The Central-South Forest Fire Likelihood Prediction Model demonstrates high accuracy, reliability, and predictive capability, with performance metrics such as accuracy, precision, recall, and F1 scores exceeding 85% and AUC values above 89%, proving its effectiveness in forecasting the likelihood of forest fires and differentiating between fire scenarios. (iii) The likelihood of forest fires in the Central-South region of China varies across regions and seasons, with increased likelihood observed from March to May in specific provinces due to various factors, including weather conditions and leaf litter accumulation. Risks of localized fires are noted from June to August and from September to November in different areas, while certain regions continue to face heightened likelihood from December to February.

1. Introduction

Forests are vital in mitigating climate change as they absorb carbon dioxide, thereby reducing greenhouse gas emissions [1,2,3,4,5]. They also play a critical role in the water cycle by regulating precipitation and maintaining river and lake volumes [6,7]. Additionally, forests provide essential resources such as timber, medicinal plants, and food and are fundamental to many cultures and communities. Consequently, forest protection and management are imperative for sustaining Earth’s ecological balance and human well-being [8,9,10]. Forest fires pose significant threats by causing ecosystem destruction, biodiversity loss, increased atmospheric greenhouse gas emissions, soil degradation, and disruptions in the water cycle. Thus, implementing effective forest management and fire prevention strategies is crucial for preserving forests and maintaining environmental equilibrium [11,12,13,14,15].

Forest fire prediction plays a crucial role in forest management and environmental conservation [16,17]. It enables early warnings of fire risks, optimizes resource allocation, reduces economic losses, protects ecosystems and biodiversity, minimizes human casualties, and helps mitigate climate change. Effective prediction and response strategies can significantly reduce the impact of forest fires on both nature and humans [18,19]. As a significant research area within the realm of forest fire prevention [20,21], forest fire prediction has seen considerable advancements thanks to recent technological progress. The primary methods for predicting forest fires include those based on physical models and statistical analysis. Specifically, physical model-based prediction involves using real-time meteorological data, vegetation parameters, and terrain information to forecast the potential spread of fires under various conditions [22,23,24,25,26]. This method integrates complex environmental data to provide accurate and timely predictions, enhancing preparedness and response capabilities.

Continuous optimization and refinement of physical models are undertaken to enhance prediction accuracy. This includes incorporating more detailed combustion mechanisms and chemical reaction models to more accurately simulate the dynamics of fire spread, as well as utilizing high-resolution terrain and vegetation data to more precisely simulate the impact of surface conditions on fire occurrence [27,28]. Statistical analysis-based fire prediction employs mathematical models to assess the risk of fire occurrences by analyzing historical fire data, meteorological data, human activities, and other related factors. These models typically rely on probabilistic statistical principles to reveal correlations between fire occurrences and various factors [29,30,31,32]. To improve the accuracy of forest fire prediction, researchers continually refine and optimize models. For instance, the application of artificial intelligence and machine learning technologies enables the training and optimization of forest fire prediction models, improving their adaptability and accuracy [20,33,34,35,36,37,38,39]. These intelligent algorithms can process large datasets, learning patterns, and regularities hidden within the data to enhance prediction accuracy.

The LightGBM (Light Gradient Boosting Machine) is recognized as a highly efficient gradient boosting decision tree algorithm, which is widely applied in various machine learning challenges, notably in predicting forest fires [40,41,42]. This algorithm stands out for its ability to efficiently manage large datasets, seamlessly handle missing values and categorical features, and provide highly accurate predictions. Such capabilities make LightGBM particularly well-suited for tasks involving extensive analysis of environmental and meteorological data, which are common in forest fire prediction scenarios. Additionally, LightGBM supports parallel processing and GPU acceleration, further boosting its performance and adaptability. These features make it a favored choice among researchers and practitioners who require robust and efficient solutions for predictive modeling in complex datasets [43,44,45,46].

This study examines forest fire data from 2001 to 2019 in China’s Central-South region, using GIS technology to understand the complex dynamics of fire distribution, frequency, and severity in relation to environmental factors. The Central-South region is particularly vulnerable to forest fires due to its dense forestation and varied terrain. Our research employs a LightGBM-based model to accurately predict fire probabilities, aiming to improve early detection and preventive measures. The selection of this specific study period is driven by the availability of data and the need to assess long-term trends comprehensively. Drawing on insights from prior studies in this field, our analysis not only enriches existing knowledge but also proposes practical strategies for reducing forest fire occurrences and minimizing their effects on natural landscapes and human communities. By adopting a comprehensive approach, this study strives to provide a strong scientific foundation for effective forest fire management.

This study introduces a pioneering method for analyzing forest fires in China’s Central-South region by integrating GIS technology with a LightGBM-based prediction model. Opting for LightGBM due to its efficiency and accuracy with large datasets, the research significantly improves the precision of fire risk assessments and supports the development of targeted prevention strategies. The study’s objectives include evaluating the influences of climatic, geographical, and human factors on fire occurrences, enhancing the spatial resolution of risk predictions for better fire management, and providing decision-makers with a reliable tool to optimize resource allocation during fire crises. This approach not only boosts the efficacy of forest fire management but also sets a scientific and methodological standard for future fire prevention and mitigation strategies.

2. Resources and Methods

2.1. The Study Area

As illustrated in Figure 1, the Central-South region of China is a key area, distinguished by its abundant natural resources and strategic location. This region includes the provinces of Henan, Hubei, Hunan, Guangxi, Guangdong, and Hainan and features a diverse topography of plains, hills, mountains, and plateaus. The climate here is varied, marked by substantial rainfall and plentiful sunshine, creating an ideal environment for agriculture and forestry. The region is also enriched by major rivers such as the Yangtze, Yellow, and Pearl Rivers, which provide copious water resources for both local use and broader distribution.

Economically, the Central-South region is predominantly driven by the secondary and tertiary sectors, with manufacturing and tourism serving as the core industries. Guangdong Province, particularly the Pearl River Delta, is a leader in regional economic development, positioning it as one of China’s most economically advanced areas [47]. Henan, Hubei, and Hunan focus on agriculture and heavy industry, whereas Guangxi and Hainan are recognized for their tourism and tropical agriculture. The region boasts a high Gross Domestic Product (GDP) and per capita GDP, underscoring its crucial role in China’s economic growth. Its demographic is primarily Han Chinese, along with various ethnic minorities, reflecting a rich cultural tapestry.

2.2. Data Sources

As detailed in

Table 1. Overview of data sources employed in the research.

Classification	Data	Resolution	Source	References
Topographic	Slope/Elevation/Slope direction	1 km	https://www.resdc.cn (Accessed on 5 May 2023)	[33,48]
Climate	Average daily surface temperature/average daily relative humidity/daily maximum surface temperature, etc.	-	https://data.cma.cn (Accessed on 1 May 2023)	[18,49,50,51]
Vegetation	Fractional vegetation cover	-	https://www.resdc.cn (Accessed on 2 May 2023)	[52,53]
Social and human factors	Distance from road/Distance from residential area/Gross Domestic Product/Population	1:100,000,1:100,000, 1 km, 1 km,	https://www.resdc.cn (Accessed on 8 May 2023)	[33,54,55]

, the study adopted a comprehensive data collection and processing strategy to analyze forest fire occurrences effectively. The data were organized into four primary categories: topographic, climate, vegetation, and social and human factors—each playing a crucial role in understanding and predicting forest fires.

Fire Data Utilization: In the research, the Moderate Resolution Imaging Spectroradiometer (MODIS) dataset on forest fires, which comprises 18,705 identified fire occurrences, was utilized. This dataset is provided by National Aeronautics and Space Administration (NASA) and is accessible via NASA’s Earth Data portal [33]. Using ArcGIS 10.4, the researchers generated a balanced set of random points to represent unburned areas, assigning a label of ‘1’ to burned points and ‘0’ to unburned points. To ensure data integrity, these points were selected at a 1:1 ratio, adhering to principles of spatial and temporal randomness. The selection was based on the 2020 national land use data, specifically excluding water bodies, urban areas, or marine zones and focusing on forested regions. This approach produced a dataset that includes both burned and unburned points.

For robust model training and validation, the researchers adopted a conventional 70:30 split for the dataset, ensuring effective feature capture during training and accurate performance evaluation on the validation set. The dataset provides detailed information on forest fires, including occurrence dates, geographic coordinates, confidence levels, and brightness measurements, from 2001 to 2019, focusing on the southwestern region with a confidence level above 80%.

Topographic Data: This category includes crucial details such as terrain elevation and slope, which are essential for analyzing how the landscape affects forest fire behavior and spread. Despite the limitations in the availability of higher-resolution models like SRTM-1 arcsec or ALOS-2 DEM and their high computational demands, a 1 km resolution DEM was selected. This resolution ensures uniformity across various data sources and facilitates extensive regional analysis. Although this choice may reduce some terrain details, it allows for consistent and systematic analysis across the Central-South region, enhancing their understanding and prediction of forest fire dynamics.

Climate Data: Meteorological records, including temperature, humidity, and wind conditions, were meticulously analyzed. These climate factors are pivotal in understanding the environmental conditions that contribute to forest fires, as they directly influence both the likelihood and behavior of these events. Lightning activity data were not used because obtaining long-term, large-scale data of this nature is challenging and their quality and consistency are often questionable, complicating collection and analysis.

Vegetation Data: A detailed inventory of vegetation types and their extents was compiled, acknowledging their crucial role in determining fire vulnerability. Comprehensive datasets on forest coverage and assorted vegetation types were amassed to pinpoint areas more prone to fire occurrences, a susceptibility tied to the presence of combustible material.

Social and Human Factors: An array of socio-economic dimensions, including demographic profiles, economic indicators, population density, and specifics of residential zones, were scrutinized to gauge the potential impact of human activities on fire hazards. The factors considered included agricultural practices, unauthorized burning incidents, Gross Domestic Product (GDP) levels, and observance of significant holidays. There is a lack of long-term, large-scale, reliable data on lightning activity that would be necessary for a robust analysis of their impact on forest fires in the regions studied. Due to these data constraints, incorporating lightning data could lead to inaccurate modeling and skewed results. Therefore, it is believed that excluding this factor until reliable data become available is a prudent approach. The potential importance of lightning as a natural ignition source is acknowledged, and it is agreed that future research should incorporate it into the analysis as reliable data become available. Including such data would undoubtedly enrich the content of the research and enhance the understanding of the causes of forest fires.

The study utilized the variance inflation factor (VIF) method for diagnosing multicollinearity. The expression for the variance inflation factor is as follows:

$$VI{F}_i=\frac{1}{1-R_i^2},(i=\mathrm{1,2},\dots ,n),$$

(1)

when 0 < VIF < 10, there is no multicollinearity among variables; when 10 ≤ VIF < 100, there is a strong multicollinearity among variables; and when VIF ≥ 100, there is severe multicollinearity among variables. Through this diagnostic approach, it was concluded that the VIF values for all variables were below 10, indicating the absence of multicollinearity in the model. This finding further validates the reliability and accuracy of the study.

The approach entailed the synthesis of these varied data categories into a cohesive, uniform dataset. A thorough data cleansing process was undertaken to address challenges like missing entries, anomalies, and repetitions, safeguarding data integrity and accuracy. The culmination of this process was data normalization, which aligned disparate data forms and scales, making them commensurate and apt for inclusion in their predictive framework. This vital step ensured consistency across datasets, facilitating an exhaustive analysis of the diverse factors affecting forest fire dynamics.

2.3. Method

In this comprehensive study, Figure 2 plays a crucial role as an illustrative roadmap, delineating the intricate technical journey undertaken to explore the complex issues surrounding forest fires. This roadmap details the process of integrating diverse datasets, which offer multiple perspectives on forest fire phenomena, including detailed fire incidents, land use patterns, meteorological observations, socioeconomic factors, extensive vegetation descriptions, and terrain data. To ensure the comparability and analyzability of these varied data sources, advanced normalization techniques are applied, effectively minimizing amplitude variations among the datasets and achieving a unified data framework that supports consistent and balanced analysis.

Moving from data preparation, the study employs advanced data examination techniques. Kernel density analysis identifies regions with high concentrations of fire incidents, highlighting forest fire hotspots. Spatial autocorrelation analysis uncovers complex spatial relationships among fire events, offering insights into their interconnectedness. Additionally, the application of standard deviation ellipses outlines directional trends and the spread of fire incidents, enhancing the understanding of forest fire dispersion patterns. Building on these analytical insights, the study adopts the advanced Light Gradient Boosting Model (LightGBM) algorithm, leveraging machine learning to predict forest fire risks with remarkable accuracy. This predictive model integrates a wide array of factors, including historical fire data, meteorological conditions, land use, and socioeconomic factors, to create a comprehensive forecast of future fires.

2.3.1. Kernel Density Estimation

Kernel density estimation (KDE) employs a smoothing approach to delineate the distribution shape of data, making it particularly suitable for analyzing continuous data. This method involves placing a kernel—typically a Gaussian kernel—around each data point and weighting these points based on the kernel’s bandwidth. The result is a comprehensive density estimate that provides a graphical representation of data distribution [56,57,58]. In the context of forest fires, kernel density analysis transforms discrete forest fire occurrences into continuous density maps, offering a clear and intuitive depiction of the spatial distribution of forest fires. Importantly, this technique does not rely on predetermined distribution assumptions, allowing for flexible scaling of analysis to accommodate diverse distribution patterns. In forest fire management, KDE plays a crucial role in identifying areas at high risk, optimizing resource distribution, and uncovering underlying factors contributing to fire incidents. Consequently, it significantly enhances the effectiveness of strategies aimed at forest fire prevention and management. The formula of kernel density analysis is as follows [59]:

$$f\left(x\right)={\sum }_{i=1}^{n}k\left(\frac{x-x_i}{h}\right)$$

(2)

The term f(x) denotes the kernel density estimate calculated within the specified threshold interval, indicating the estimated density of occurrences per unit area. The variable n stands for the total number of forest fires occurring within this interval, providing a quantitative measure of fire incidents. The parameter h represents the predetermined search radius or bandwidth for the kernel density estimation window, which determines the scale of smoothing applied to the data. Lastly, the symbol k refers to the kernel function employed in the analysis, which is a mathematical function used to weight the data points within the search radius, thereby influencing the shape of the resulting density estimate.

2.3.2. Spatial Autocorrelation Analysis

Spatial autocorrelation, a pivotal concept in geography and statistics, is utilized to examine the degree of similarity or correlation across different geographic locations [60,61,62]. It delves into the presence of patterns, connections, or resemblances between proximate or remote locations within a geographic space, serving as a fundamental tool for analyzing the distribution, clustering, and spatial variations in geographical phenomena [63]. In the context of forest fire studies, spatial autocorrelation analysis offers several benefits. It unveils the geographic distribution patterns of fires, aids in the efficient management and allocation of resources, facilitates the generation of predictions and early warnings, optimizes the design of monitoring networks, and supports spatial decision-making processes aimed at mitigating fire risks and improving response measures. This analytical approach is indispensable in advancing the understanding, prevention, and management of forest fires, underscoring its essential role in forest fire research and control efforts.

The formulas are as follows [64]:

Global autocorrelation:

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

(3)

In this equation, I represents the global Moran’s I index, $n$ stands for the total number of spatial units, $W_{ij}$ denotes the spatial weights between units $i$ and $j$, $x_i$ and $x_j$ represent the values of variable $x$ for units $i$ and $j$, and $\bar{x}$ signifies the average or mean of variable x.

Local autocorrelation:

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

(4)

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 $i$,$\mathrm{a}\mathrm{n}\mathrm{d}\bar{x}$ is the mean of variable $x$.

In this equation, $I^{\prime }$ represents the local Moran’s I index, $n$ stands for the total number of spatial units, $W_{ij}$ denotes the spatial weights between units $i$ and $j$, $x_i$ signifies the value of variable $x$ for unit i, and $\bar{x}$ represents the mean or average of variable $x$.

Global and local autocorrelation analyses provide a sophisticated approach to examining spatial patterns over broad areas and within particular locales, respectively. These techniques categorize spatial relationships into four distinct configurations: High-High (H-H), where areas of high values are found in proximity to each other; High-Low (H-L), where areas of high values are adjacent to areas of low values; Low-High (L-H), where areas of low values are surrounded by areas of high values; and Low-Low (L-L), where areas of low values cluster together. This classification scheme enhances our comprehension of the associations between areas of similar or divergent values, uncovering trends of clustering or dispersion. Such understanding is vital for developing precise strategies in spatial planning and analysis, enabling a more nuanced management of geographical spaces based on their specific attributes.

2.3.3. Standard Deviation Ellipse

The standard deviation ellipse is a visualization tool used in multivariate statistical data analysis [65,66]. It constructs an ellipse with a specific shape and orientation by considering the standard deviation and covariance matrix of the data, reflecting the dispersion and correlation of data points [67]. This visualization tool is commonly employed for displaying data distributions, detecting outliers, and performing data clustering analysis. By examining the shape and orientation of the ellipse, it helps researchers gain a better understanding of the characteristics and structure of the dataset [68]. In the context of forest fires, the advantage of using standard deviation ellipses lies in their ability to visually depict the distribution of forest fire data, identify clusters of fire sources and anomalies, and provide valuable support for spatial planning and data analysis, ultimately enhancing our understanding of the spatial features of forest fires and improving risk management and response strategies.

The formula is as follows [66]:

$${SDE}_x=\sqrt{\frac{\sum _{i=1}^n {\left(x_i-\bar{X}\right)}^2}{n}},{SDE}_y=\sqrt{\frac{\sum _{i=1}^n {\left(y_i-\bar{Y}\right)}^2}{n}},$$

(5)

In this equation, $SD{E}_x$ and $SD{E}_y$ represent the standard deviations of the variables $x$ and $y$, while $n$ stands for the number of observations. Additionally, $\bar{X}$ and $\bar{Y}$ denote the averages or means of variables $x$ and $y$, respectively.

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

(6)

Within this mathematical expression, $\mathrm{t}\mathrm{a}\mathrm{n}\theta$ represents the tangent of the angle of rotation, whereas ${\tilde{x}}_i$ and ${\tilde{y}}_i$ signify the transformed or rotated coordinates of individual points $i$ within the updated coordinate system.

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

(7)

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

(8)

In this equation, ${\sigma }_x$ and ${\sigma }_y$ represent the standard deviations of the transformed coordinates, while ${\tilde{x}}_i$ and ${\tilde{y}}_i$ denote the coordinates of individual points i after rotation within the updated coordinate system.

2.3.4. Light Gradient Boosting Model

The Light Gradient Boosting Machine (LightGBM) is recognized as an efficient gradient boosting framework specifically designed to handle large datasets without compromising speed or accuracy [45]. This robust system incorporates two key advancements: Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB), strategically engineered to optimize computational efficiency and memory usage during training. GOSS operates by selectively retaining data points with substantial gradients while down-sampling others, effectively mitigating computational burden without compromising training accuracy. This targeted approach ensures that high-gradient instances remain influential in the learning process, enhancing efficiency without sacrificing fidelity. On the other hand, EFB focuses on reducing feature dimensionality by bundling exclusive features—those that do not concurrently exhibit values. By consolidating such features, EFB minimizes redundancy, thereby streamlining computational overhead and memory consumption. These innovative techniques empower LightGBM to offer accelerated training speeds and decreased memory requirements when handling large-scale datasets, surpassing traditional gradient boosting methods in efficiency while maintaining, or even improving, model performance [69,70,71].

As for determining the hyperparameters of a machine learning model, experts suggest employing techniques such as grid search, random search, or Bayesian optimization. These methods involve systematically exploring the hyperparameter space to find the combination that optimizes model performance. Additionally, techniques like cross-validation can be used to evaluate model performance across different hyperparameter settings, helping to identify the most suitable configuration.

The final prediction is obtained by summing the predictions of all individual trees:

$$F\left(x\right)={\sum }_{t=1}^T{\alpha }_t{h}_t\left(x\right)$$

(9)

where ${\alpha }_t$ is the learning rate, controlling the contribution of each tree to the final prediction, and $T$ is the total number of trees in the ensemble.

2.3.5. Evaluation Indicators

In the fields of machine learning and statistical analysis, a comprehensive set of metrics such as Accuracy, Precision, Recall, F1 score, and AUC (Area Under the Curve) are crucial for assessing the performance of classification models. These metrics serve as essential benchmarks that evaluate how effectively a model can classify data into the correct categories. Each metric provides a different perspective on the model’s performance, enabling a holistic assessment of its classification accuracy. These metrics are defined as follows [33,72]:

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

(10)

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

(11)

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

(12)

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

(13)

In binary classification tasks, such as evaluating a forest fire prediction model, the classification outcomes are critically analyzed using four key measures: True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN). True Positives refer to the instances where the model correctly predicts actual fire incidents. True Negatives denote the accurate identification of non-fire incidents, where the model rightly predicts the absence of a fire. Conversely, False Positives occur when non-fire incidents are incorrectly classified as fires, indicating an error of overprediction. False Negatives represent actual fire incidents that the model fails to detect, showcasing an error of underprediction.

These distinctions are crucial for evaluating the accuracy and effectiveness of the forest fire prediction model. They provide a comprehensive measure of the model’s ability to distinguish accurately between actual and non-existent forest fires. Analyzing these outcomes offers insights into the model’s precision and reliability, enabling a better understanding of its capacity to forecast potential fire outbreaks effectively. Such an evaluation helps in refining the model’s predictive algorithms and enhancing its overall performance in forest fire management strategies.

3. Results

3.1. Forest Fire Kernel Density Analysis in the Central and Southern Regions

As shown in Figure 3, regions within the Central-South area, particularly in Guangdong Province (notably Heyuan, Meizhou, Shaoguan) and Hunan Province (including Yongzhou, Binzhou, Hezhou), exhibit high forest fire kernel density due to a combination of factors including fluctuating climate conditions, abundant vegetation, complex geographic landscapes, and human activities. The region’s monsoon climate, characterized by wet summers and dry winters, often leads to vegetation desiccation, which increases susceptibility to forest fires. Moreover, the lush vegetation provides a rich fuel source, while the rugged mountainous and hilly terrain facilitates the spread of fires. Human-induced factors, such as indiscriminate burning of waste and the misuse of open fires, also significantly increase the risk of forest fires. To mitigate the wildfire threat, it is imperative to adopt preventive strategies, enhance monitoring and early warning systems, and foster public awareness. These measures are crucial for the effective management and reduction in forest fires in the region, helping to protect both the environment and local communities.

3.2. Results of Autocorrelation Analysis on Forest Fire Occurrences in Central and Southern China region

Figure 4 illustrates that within the Central-South region, a total of 33 cities exhibit global autocorrelation characteristics of High-High (H_H), predominantly found in Hunan Province (notably Shaoyang, Hengyang, Yongzhou), the Guangxi Zhuang Autonomous Region (such as Guilin, Liuzhou, Hechi), and Guangdong Province (including Heyuan, Qingyuan, Zhaoqing). Additionally, four cities display local autocorrelation patterns of High-Low (H_L), located in Hubei Province (Suizhou and Huanggang) and Guangdong Province (Jiangmen and Zhongshan). Furthermore, 12 cities are identified with Low-High (L-H) features, spread across Hunan Province (Yiyang, Changsha, Loudi) and Guangdong Province (Guangzhou, Foshan). The remaining cities are either grouped under Low-Low (L_L) or categorized within non-significant regions. Notably, 18 cities with local H_H autocorrelation predominantly emerge in Hunan Province (Hengyang, Shaoyang, Yongzhou), Guangdong Province (Shaoguan, Heyuan, Meizhou), and the Guangxi Zhuang Autonomous Region (Hezhou, Guilin, Wuzhou), with only Guangzhou exhibiting an L-H pattern, while the others are classified as L_L or within non-significant zones.

The diversity in global and local autocorrelation characteristics among cities impacted by forest fires in the Central-South region underscores the complex interplay of factors such as climatic conditions, geographical landscapes, human interventions, forest management practices, and emergency response efficiencies. This convergence of elements not only influences the distribution and behavior of fires but also delineates the autocorrelation patterns observed. A comprehensive understanding of these contributing factors is essential for devising effective strategies to prevent and mitigate fire incidents, ultimately safeguarding the region against the adverse impacts of forest fires.

3.3. The Results of Standard Deviation Ellipse for the Forest Fires

Figure 5 and

Table 2. The standard deviation parameters of forest fire occurrence ellipses in the Central and Southern China region.

Year	XStdDist (km)	YStdDist (km)	Shape_Leng (km)	Shape_Area (km2)	Oblateness	Rotation
2001	372.534	268.997	2028.542	314,802.413	1.385	45.072
2002	388.413	296.779	2162.196	362,120.052	1.309	56.898
2003	373.526	258.979	2003.366	303,885.299	1.442	55.733
2004	381.818	285.181	2106.417	342,059.692	1.339	60.849
2005	436.113	317.943	2383.464	435,584.544	1.372	45.146
2006	376.658	271.229	2048.857	320,927.371	1.389	73.757
2007	370.653	286.349	2072.507	333,418.207	1.294	48.246
2008	303.707	380.991	2157.867	363,492.458	0.797	35.858
2009	345.373	287.833	1993.357	312,288.287	1.200	62.898
2010	491.176	339.273	2630.753	523,491.125	1.448	60.522
2011	348.712	398.867	2351.203	436,941.327	0.874	35.829
2012	297.208	422.206	2277.155	394,192.353	0.704	24.360
2013	308.922	461.787	2445.076	448,137.993	0.669	13.721
2014	300.586	424.630	2294.988	400,961.846	0.708	18.729
2015	415.751	283.334	2215.941	370,044.820	1.467	67.391
2016	317.519	470.922	2500.411	469,722.244	0.674	25.864
2017	275.339	369.203	2035.607	319,343.775	0.746	42.487
2018	431.219	305.251	2330.600	413,502.468	1.413	46.869
2019	276.814	562.527	2713.722	489,146.723	0.492	21.242

present an analysis covering the period from 2001 to 2019, highlighting a notable northward shift in the centroid of forest fire incidents within the Central-South region. This transition occurred primarily at the confluence of the Hunan, Guangdong, and Guangxi provinces, particularly affecting Zhaoqing and Qingyuan in Guangdong, Yongzhou and Chenzhou in Hunan, and Guilin and Hezhou in Guangxi. Initially, from 2001 to 2009, the centroid was predominantly located in Guangdong Province, indicating a high frequency of forest fires there and underscoring the considerable fire-related risks and challenges in the area. After 2009, the centroid exhibited variable shifts, predominantly oscillating between the borders of the Hunan and Guangxi provinces, except for 2011 and 2015 when it primarily returned to Guangdong. These shifts suggest that forest fire risks and patterns are subject to various influencing factors, including climatic fluctuations, topographical changes, variations in forest coverage, and human interventions.

The migration and distribution trends of this fire centroid provide essential insights into the evolving dynamics of forest fires in the region. They underscore the spatial trends in forest fire occurrences and hint at possible links to ecological changes, modifications in forest management practices, and the effectiveness of forest fire response measures. Consequently, a comprehensive examination of these shifts is crucial for formulating robust forest fire prevention and management strategies, aiming to mitigate fire-induced damages and contribute to the conservation and rehabilitation of the ecological environment.

3.4. Evaluation of Forecast Precision for Forest Fires in Southern China

Figure 6 highlights the exemplary performance of the Central-South forest fire prediction model on both training and validation datasets, showcasing its accuracy, dependability, and strong predictive prowess. In the training phase, the model achieved an impressive accuracy rate of 85.71%, precision of 87.5%, recall of 86.67%, and an F1 score of 87.08%, indicating its proficient and balanced approach in predicting forest fires. Moreover, an AUC value of 90.21% accentuates its exceptional ability to differentiate between fire occurrences and non-occurrences.

The model’s efficacy is further validated by its performance on the validation set, with scores of 84.38% in accuracy, 86.25% in precision, 85.71% in recall, an F1 score of 86.02%, and an AUC of 89.79%. These metrics collectively demonstrate the model’s remarkable ability to generalize across various conditions, ensuring dependable predictions of forest fire incidents.

In essence, Figure 5 underlines the efficiency and reliability of the Central-South forest fire prediction model in forecasting forest fires, highlighting its vital contribution towards the proactive prevention and management of such disasters. The model’s stellar performance not only illustrates its capability in navigating the complexities of forest fire risk assessment but also reinforces its critical role in advancing forest management practices and protecting ecosystems.

3.5. Predicting Monthly Forest Fires in the Central and Southern Regions of China

As delineated in Figure 7 and

Table 3. Key fire prevention areas for forest fires in different months in Central and Southern China.

Timeframe	Province	Regions/Cities
March–May	Guangdong, Guangxi Zhuang Autonomous Region, Hunan, Hubei, Hainan	Meizhou, Chaozhou, Jieyang Wuzhou, Guilin, Baise Hengyang, Loudi, Yongzhou, Huangshi, Xianning, Dongfang
June–August	Guangdong, Hubei	Heyuan, Huangshi
September–November	Guangdong, Guangxi Zhuang Autonomous Region, Hunan	Meizhou, Heyuan, Shaoguan, Nanning, Hezhou, Yulin, Binzhou, Yongzhou, Hengyang
December–February	Guangdong, Guangxi, Hunan, Hubei	Shaoguan, Qingyuan, Zhaoqing Several regions Hengyang, Binzhou, Yongzhou, Huangshi

, the dynamic shifts in fire hazard zones within Central and Southern China across different months underscore the impact of seasonal variations and climatic conditions on wildfire risks. A refined overview of the temporal distribution of these high-risk zones reveals:

(i)

March–May: This period witnesses elevated forest fire risks in specific regions of Guangdong Province (Meizhou, Chaozhou, Jieyang), the Guangxi Zhuang Autonomous Region (Wuzhou, Guilin, Baise), Hunan Province (Hengyang, Loudi, Yongzhou), Hubei Province (Huangshi, Xianning), and Dongfang City in Hainan Province. With the climate warming yet remaining dry and with minimal rainfall, the dead vegetation from winter becomes prime fuel for fires. Further exacerbating the risk are agricultural practices like burning crop residue, increased tourism activities, and the misuse of fires outdoors.

(ii)

June–August: Most of the Central-South region experiences low forest fire risk due to the rainy season, which increases humidity. However, some areas in Guangdong’s Heyuan and Hubei’s Huangshi face high risks, potentially due to uneven rainfall distribution or localized drought conditions.

(iii)

September–November: The risk of forest fires increases again in areas like Meizhou, Heyuan, and Shaoguan in Guangdong Province; Nanning, Hezhou, and Yulin in the Guangxi Zhuang Autonomous Region; and Binzhou, Yongzhou, and Hengyang in Hunan Province. As autumn progresses, temperatures drop and humidity decreases, while fallen leaves provide new fuel for fires.

(iv)

December–February: This timeframe marks a high-risk phase for areas like Shaoguan, Qingyuan, and Zhaoqing in Guangdong; several regions in Guangxi; and Hengyang, Binzhou, and Yongzhou in Hunan, along with Huangshi in Hubei. Despite cooler weather, the dry atmosphere and lack of moisture elevate fire risks. Dry vegetation and the accumulation of flammable material, along with human activities such as land clearing for agriculture, intensify the potential for fires.

Regarding the evaluation of the forest fire prediction model, extensive training and validation processes have already been conducted, demonstrating the model’s accuracy, dependability, and strong predictive prowess across various conditions. Furthermore, it is confirmed that the predicted risk maps align with the historical distribution and real-world occurrence of forest fire risks in the Central-South region.

4. Discussion and Conclusions

4.1. Discussion

This study marks a significant advancement in predictive modeling for forest fire occurrences within the Central-South region, leveraging state-of-the-art integrated learning models and a comprehensive dataset. This innovative methodology meticulously analyzes the combined effects of climate change, human interventions, and geographic specifics on forest fire risk. The approach builds upon the foundational insights highlighted in seminal works [73], enhancing understanding of these critical factors. The research extends beyond corroborating existing knowledge; it offers a detailed, quantitative exploration of the complex interplay among various factors influencing forest fires, thereby establishing a new benchmark for precision in forest fire management strategies. The findings not only align with previous research but also enhance it, providing a more nuanced understanding of these dynamics [18,19]. The research advances beyond traditional predictive models (physical and statistical) by implementing the LightGBM algorithm, enhancing the ability to manage and analyze large datasets efficiently [40,41,42,43,44,45,46].

Enabled by this groundbreaking model, monitoring and real-time management of forest fires are poised for revolution. Equipped with cutting-edge technologies such as satellite-based remote sensing and the BeiDou navigation system, complemented by robust multi-network communications and mobile command capabilities, the model delivers unparalleled precision in tracking critical fire-related variables including temperature, humidity, wind patterns, and vegetation health. This comprehensive monitoring not only enables rapid responses from command centers and rescue teams during emergencies but also supports a proactive approach to forest fire management across the Central-South region. By pinpointing high-risk zones and deploying targeted prevention strategies, the model acts as a cornerstone for safeguarding China’s rich and diverse forest ecosystems, ensuring ecological stability while protecting lives and property.

Looking ahead, the research agenda is ambitiously set to explore the impact of extreme climate phenomena, specifically focusing on ENSO cycles and their significant influence on forest fire dynamics [74,75,76]. This initiative aims to further refine predictive capabilities, enhancing accuracy in forecasting and thereby improving fire management strategies. Faced with the challenge of acquiring long-duration, high-precision, and spatially consistent data across extensive areas, future efforts include utilizing higher-resolution VIIRS fire detection data to sharpen analytical precision [77]. The limitations of the study are acknowledged, particularly the challenges related to data availability and resolution. These constraints can impact the generalizability of the model outside the Central-South region of China, as regional variations in climate and vegetation require tailored adaptations of the predictive models.

Furthermore, the commitment to delineating nuances among different types of forest fires—whether ignited by lightning or human activities—motivates the exploration of more comprehensive datasets, integrating higher resolution data and a broader array of fire descriptors [78,79,80,81]. This holistic approach not only promises a more realistic simulation of forest fire dynamics but also advances predictive models capable of distinguishing between various fire types, providing deeper insights and more effective strategies for fire prevention and management. Addressing natural causes of forest fires, such as lightning, is a potential area for future development. Currently, the model does not differentiate between fires caused by human activities and those initiated by natural events like lightning. Enhancing the model to detect and predict lightning-induced fires could significantly refine strategies for fire prevention and management. The methodologies developed in this study hold promise for broader applications. Adapting the predictive model to suit different ecological and climatic conditions could bolster forest fire management efforts globally. This would involve recalibrating the model to account for local variations, ensuring its effectiveness across diverse global landscapes.

While the study has achieved notable milestones in the predictive modeling of forest fire occurrences, it stands at the frontier of a field replete with complexity and variability. The road ahead demands steadfast dedication to interdisciplinary collaboration and relentless innovation. Through the continuous enhancement of predictive models and their validation through empirical fieldwork, the aim is to significantly reduce the impact of forest fires. This mission transcends academic endeavors; it represents a commitment to preserving the natural environment and safeguarding human well-being amid the challenges of a changing global climate. Moreover, the research, while initially focused locally, paves the way for assessing the applicability of developed solutions on a global scale. This expansion could not only affirm the universal applicability of the models but also amplify global efforts in forest fire management, contributing to a safer, more resilient world. Thus, the journey, grounded in rigorous analysis and driven by a vision for a better future, continues with zeal, guided by the beacon of knowledge represented by the references and the potential for the findings to influence forest fire management practices worldwide.

4.2. Conclusions

This research meticulously examines the spatial and temporal patterns of forest fires in China’s Central-South region, utilizing a synergistic approach that integrates Geographic Information Systems (GISs) with advanced machine learning methodologies like LightGBM. The outcomes offer profound insights into the dynamics of forest fire occurrences, emphasizing not just the concentration of fires but also their interconnectivity—a key factor in understanding their distribution and potential spread within this vulnerable area.

(i)

Technological Integration: The study highlights the effectiveness of combining GISs with machine learning techniques to unravel the complex patterns of forest fires. This integration proves crucial in enhancing our understanding of the causative factors behind these events, showcasing the power of technological synergy in environmental science

(ii)

Model Performance: The Central-South forest fire prediction model stands out for its accuracy and reliability. With robust performance metrics, it effectively forecasts forest fire occurrences and differentiates between fire types, thereby playing a vital role in forest fire forecasting and risk management.

(iii)

Seasonal and Regional Variabilities: Our analysis reveals significant seasonal and regional variations in forest fire risk, identifying specific times and locations where risks are heightened. These insights are critical for the strategic allocation of resources and the development of targeted fire prevention protocols, underscoring the need for tailored fire management strategies.

(iv)

Holistic Fire Management Approach: The findings advocate for a holistic approach to forest fire management. By integrating state-of-the-art technology with a detailed understanding of the environmental and temporal factors influencing fire risk, the study paves the way for the development of more sophisticated and effective forest fire mitigation strategies.

In conclusion, through detailed data analysis and the application of cutting-edge technologies, this research significantly advances our understanding and management of forest fires. It provides a comprehensive framework for future environmental science endeavors, particularly in enhancing forest fire prevention and management strategies based on nuanced regional and temporal risk factors.