Advancements in Artificial Intelligence Applications for Forest Fire Prediction

Abstract

In recent years, the increasingly significant impacts of climate change and human activities on the environment have led to more frequent occurrences of extreme events such as forest fires. The recurrent wildfires pose severe threats to ecological environments and human life safety. Consequently, forest fire prediction has become a current research hotspot, where accurate forecasting technologies are crucial for reducing ecological and economic losses, improving forest fire management efficiency, and ensuring personnel safety and property security. To enhance comprehensive understanding of wildfire prediction research, this paper systematically reviews studies since 2015, focusing on two key aspects: datasets with related tools and prediction algorithms. We categorized the literature into three categories: statistical analysis and physical models, traditional machine learning methods, and deep learning approaches. Additionally, this review summarizes the data types and open-source datasets used in the selected literature. The paper further outlines current challenges and future directions, including exploring wildfire risk data management and multimodal deep learning, investigating self-supervised learning models, improving model interpretability and developing explainable models, integrating physics-informed models with machine learning, and constructing digital twin technology for real-time wildfire simulation and fire scenario analysis. This study aims to provide valuable support for forest natural resource management and enhanced environmental protection through the application of remote sensing technologies and artificial intelligence algorithms.

1. Introduction

Forests cover approximately 4 billion hectares (30% of global land area) and play a crucial role in preserving biodiversity [1], offering habitats for numerous species [2], and regulating the regional climate [3]. They function as carbon sinks [4], sequestering and storing substantial amounts of carbon dioxide [5], thereby mitigating the effects of climate change [6]. Additionally, forests play a role in water cycle regulation, soil conservation, and erosion prevention [7]. The preservation of forests is essential for maintaining global biodiversity, and it may also improve microclimatic conditions in adjacent urban and natural ecosystems [8]. In recent years, climate change and human activities have significantly impacted the environment, leading to more frequent and severe extreme events such as wildfires [9,10]. Wildfires have been especially damaging, harming ecosystems and infrastructure, and posing risks to human life [11]. Studies on fire are becoming increasingly relevant, especially in the context of climate change. Trends of temperature increases point to an increase in the occurrence of this disturbance, and consequently an increase in greenhouse gas contribution in the atmosphere [12]. As temperatures continue to rise, wildland fires are expected to become even more frequent and severe, posing a major threat to human settlements, ecosystems, and the environment [13,14]. Accurate wildland fire prediction technologies are thus imperative to mitigate damages, enhancing fire management efficiency and safety [15,16,17].

To enhance the predictive capacity for forest fires and mitigate environmental and socio-economic risks associated with wildfire damage, researchers have conducted extensive investigations into forest fire prediction methodologies. In the domain of wildfire risk forecasting, early studies predominantly employed statistical analysis and physical models to elucidate wildfire dynamics. By extracting patterns from historical wildfire datasets or integrating physical mechanisms of fire propagation, these approaches have provided critical theoretical foundations for predicting fire occurrence trends, probabilities, and scales. Regression analysis models operate by identifying relationships between environmental variables and fire incidence. Su et al. [18] applied negative binomial (NB) and Geographically Weighted Negative Binomial Regression (GWNBR) models to analyze fire-driving factors in the Greater Khingan Range for fire occurrence prediction. Prasertsri et al. [19] achieved regional fire risk forecasting (R² > 82%) using geographically weighted regression (GWR). Dorji et al. [20] integrated Geographic Information System (GIS)-based logistic regression (LR) with Frequency Ratio (FR) models to develop wildfire susceptibility maps, with prediction accuracies reaching 88.3% for LR and 85.5% for FR, demonstrating the effectiveness of both models in wildfire risk assessment. Probabilistic models exhibit superior performance in handling uncertain and extreme events for wildfire risk prediction. Bayesian belief networks (BBNs) [21] and extreme value theory [22] demonstrated exceptional predictive capability (Area Under Curve (AUC) = 0.986) in applications across Iran and the French Mediterranean region. Physical models facilitate the mechanistic understanding of extreme fire events through the simulation of fire–atmosphere interactions. A coupled computational fluid dynamics and weather forecasting model [23] enabled bidirectional feedback simulation of cross-scale fire evolution, advancing comprehension of extreme wildfire mechanisms such as pyrocumulonimbus (pyroCb) events [24]. Climate models excel in long-term trend forecasting: Goss et al. [25] projected significant increases in extreme autumn wildfire risks in California under climate change scenarios, while a dynamic prediction system [26] achieved interannual wildfire probability forecasting (10–45 months) through the assimilation of oceanic and soil moisture data. Statistical models offer advantages in data-driven analysis and high interpretability of fire drivers, though their performance depends on the completeness of historical fire records. Physical models demonstrate strengths in mechanistic interpretation and the prediction of large-scale extreme fire events, albeit with higher computational complexity. However, both approaches exhibit limitations in estimating fire occurrence probabilities due to the inherent unpredictability and hazardous nature of wildfires [27].

With the advancement of artificial intelligence, the integration of multi-source heterogeneous data—such as satellite imagery, climate data, topography information, meteorology data, vegetation information, anthropogenic activities, wildfire records, road and river information, and economic level—with AI algorithms has significantly enhanced the accuracy of forest fire prediction. These models iteratively improve their performance through training, thereby strengthening their capacity to forecast wildfire risks while reducing false alarms. To address the limitations of statistical and physical models, researchers have introduced machine learning techniques (excluding deep learning in this context) into forest fire prediction studies. By uncovering nonlinear relationships within multidimensional environmental data, machine learning has markedly improved prediction precision and dynamic adaptability in wildfire forecasting. Artificial Neural Networks (ANNs) simulate nonlinear interactions to explore complex associations between forest fires and environmental variables. Razavi et al. [28] revealed that in temperate forests, a 0.1 °C temperature increase elevates fire probability by 4.75%, identifying total pressure, wind speed, the Keetch–Byram Drought Index, and relative humidity as critical sensitive parameters [29]. Pérez-Sánchez et al. [30] demonstrated the potential of ANNs in predicting wildfire burn areas, achieving a success rate exceeding 70%. Ensemble learning methods improve model robustness through feature optimization and algorithmic fusion. Wang et al. [31] developed Cost-Sensitive Boosting, which outperformed traditional approaches in short-term prediction by incorporating cluster-based resampling. Zhou et al. [32] achieved effective wildfire prediction (AUC = 0.83) using Categorical Boosting (CatBoost) combined with Principal Component Analysis (PCA) for dimensionality reduction. Random Forest (RF) demonstrated over 90% accuracy in analyzing driving factors on the Mongolian Plateau [33], while Jaafari et al. [34] integrated Adaptive Neuro-Fuzzy Inference Systems (ANFIS) with genetic algorithms for high-resolution spatial prediction (AUC = 0.92). Janizadeh et al. [35] proposed a hybrid model integrating Golden Jackal Optimization (GJO) with LightGBM, achieving superior performance (AUC = 0.931) in Hawaiian wildfire prediction.

Probabilistic machine learning models excel in quantifying uncertainty and predicting extreme fire events. Charizanos et al. [36] utilized Bayesian prior information to forecast wildfire susceptibility (AUC > 0.8), whereas Joseph et al. [37] developed a spatiotemporal zero-inflated negative binomial model that captured 99% of fire counts and 93% of burn areas, enabling cross-regional extreme event warnings. Bugallo et al. [38] validated a zero-inflated negative binomial mixture model at 95% confidence in Spain, providing technical support for wildfire prevention strategies. Hybrid multi-algorithm frameworks address limitations of single methods through complementary integration. Tehrany et al. [39] combined entropy indices with logistic regression (AUC = 88.51%), while Mohajane et al. [40] significantly improved prediction accuracy in northern Morocco (AUC = 0.989) using a Frequency Ratio–Random Forest (FR-RF) model. Tran et al. [41] optimized parameters via the Black Widow Algorithm in their Black Widow Optimization (BWO)–Extreme Gradient Boosting (XGBoost) model (AUC = 0.9269), outperforming conventional XGBoost (AUC = 0.9164). Machine learning methods offer advantages in interpretability, deploy ability, and higher accuracy compared to physical and statistical models. However, they often neglect deeper inter-factor relationships. To further advance prediction accuracy, researchers have begun incorporating deep learning techniques into wildfire forecasting systems.

Deep learning enhances forest fire prediction accuracy and adaptability by capturing nonlinear characteristics of fire risk factors and integrating multi-source data. Convolutional Neural Networks (CNNs) optimize fire detection and risk assessment through spatial feature extraction and multi-scale modeling. Ding et al. [42] trained a fully connected CNN using Himawari-8 satellite imagery, achieving wildfire localization accuracy exceeding 80%. Shams Eddin et al. [43] proposed a Location-Aware Adaptive Normalization (LOAN) layer to dynamically adjust the geographical sensitivity of features (Precision = 83.71%). Pais et al. [44] developed a CNN model based on land cover raster data, achieving 100% accuracy in high-risk zone identification. Yu et al. [45] improved the AUC to 0.9487 through Cellular Automata-Based CCNN (CA-Based CCNN) and active learning. Kanwal et al. [46] combined CNN-2D with Random ForestANFIS to attain 96.91% prediction accuracy, while Zhang et al. [47] demonstrated superior performance of a contextual CNN-2D model leveraging neighborhood information over traditional MLP architectures. Zou et al. [48] developed an attention-based model supporting 8-day temporal resolution predictions, and Prapas et al. [49] maintained AUC > 97.8% at a 64-day lead time using a global image segmentation framework. Casallas et al. [50] integrated 3D-CNN with the Weather Research and Forecasting (WRF) model to generate daily risk maps. The integration of CNNs with Recurrent Neural Networks (RNNs) overcomes limitations of single-model approaches. Ji et al. [51] proposed an SLA-ConvLSTM model that incorporates static location data and global climatic teleconnections to enhance spatial specificity. Deng et al. [52] combined 3D-CNN with Convolutional LSTM (ConvLSTM) to achieve daily scale predictions (AUC = 0.901). Bhowmik et al. [53] developed a ULSTM network capable of processing 37 million multimodal data points with >97% accuracy. Zhang et al. [54] implemented a CNN-2D-LSTM framework for global monthly burn area forecasting, enabling seasonal risk assessments. He et al. [55] optimized a ConvLSTM wildfire prediction model through channel-spatial attention mechanisms and vision transformer modules, achieving AUC = 97.90%. Comparative analyses by Ramayanti et al. [56] revealed CNN’s superior performance (AUC = 0.879) over LSTM and FR methods in wildfire susceptibility mapping.

The objective of this study is to further explore current advancements and synthesize diverse methodologies to deepen the understanding of wildfire risk prediction mechanisms. This paper categorizes wildfire risk prediction methodologies into three classes: (1) statistical and physics-based models, (2) machine learning-based approaches, and (3) deep learning-based frameworks. By synthesizing and critically evaluating state-of-the-art research, this review aims to provide actionable insights and future directions for advancing wildfire risk prediction systems. The findings are intended to serve as a reference for researchers and policymakers in designing robust, adaptive frameworks for wildfire mitigation and management. The remainder of this paper is structured as follows: Section 2 describes the methodological approach to the collection and classification of relevant reference sources. Section 3 presents three themes: (1) forest fire prediction studies based on statistical analysis and physical models; (2) machine learning-based approaches for wildfire forecasting; and (3) deep learning methodologies for fire risk prediction. It further discusses commonly used data types in wildfire prediction research and access methods for open-source datasets. Section 4 offers an analysis of the literature on forest fire prediction, and Section 5 concludes the study with key findings and future directions.

2. Materials and Methods

In this study, we review research on forest fire risk prediction conducted between 2015 and 2024. A systematic literature search was performed using the Web of Science, Scopus, and Google Scholar databases with keywords including “Prediction of Wildfire”, or “Prediction of Forestry Fire”, or “Wildfire Probability”, or “Wildfire Susceptibility”, or “Forestry Fire Probability”, or “Forestry Fire Susceptibility”. The publication trends (Web of Science database), illustrated in Figure 1, reveal a steady increase in wildfire risk prediction studies since 2018, with sustained high publication output over the past four years, underscoring the prominence of this research domain. Through a systematic review of the selected literature, we analyze widely adopted methodologies for early wildfire prediction to elucidate current research trends.

Through a comprehensive review and analysis of the selected literature, the technical methods employed in these studies were systematically categorized and summarized. The proportional distribution of these methodological approaches across the reviewed studies is visualized in Figure 2. Specifically, this paper categorizes wildfire risk prediction methodologies into three classes: statistical and physics-based models, machine learning-based approaches, and deep learning-based frameworks. By synthesizing and critically evaluating the latest research advancements, this study aims to provide actionable insights and future directions for advancing wildfire risk prediction systems.

3. Results

3.1. Forest Fire Prediction Using Statistical and Physics Models

This section provides a systematic introduction to common statistical analytical methods employed in forest fire risk prediction research, including regression analysis, probabilistic models, fire danger index-based analyses, meteorological model-driven predictions, and other supplementary approaches. The current literature on wildfire risk prediction predominantly incorporates statistical methods for analyzing fire-driving factors. However, this discussion focuses exclusively on studies that construct forest fire risk assessment models using statistical analytical frameworks. The technical classification of forest fire prediction research based on statistical analysis and physics modeling is illustrated in Figure 3.

3.1.1. Forest Fire Prediction Based on Regression Analysis

Matsoukis et al. [57] applied simple linear regression lines to separately estimate actual air temperature (T) and relative humidity (RH) data in mountainous regions. During the most fire-prone periods (both seasonal and diurnal) across the study years, 97% of the estimated meteorological forest fire risk indices showed classification consistency with those derived from actual T/RH measurements, demonstrating satisfactory performance. Brys et al. [58] proposed a novel methodology for wildfire risk assessment and behavior prediction leveraging open geospatial data and ontologies. This framework integrates a spatially weighted index model with multi-criteria analysis, significantly improving data utility through semantic and spatial relationships among wildfire resources. To support decision-making, the approach generates scenario-based analyses linked to open hotspot datasets. Beccari et al. [59] developed a wildfire risk assessment model by integrating the Forest FWI with land use information. Empirical validation demonstrated the model’s operational effectiveness in generating spatially explicit fire occurrence probability maps at 1.5 km resolution, highlighting its utility for high-resolution risk mapping applications. Joshi et al. [60] developed a Generalized Linear Model (GLM) framework incorporating binary logistic regression to predict forest fire risk by integrating satellite-detected ignition events (dependent variable) with environmental predictor variables. The model demonstrated robust predictive performance, achieving an AUC = 0.92 and ACC = 0.89. Pan et al. [61] constructed a spatial logistic model for forest fire risk based on the spatial distribution of forest fires and their influencing factors. The study demonstrated a good fit between the distribution of forest fires and their impact factors (p < 0.05). The AUC value was 0.757, and a probability distribution map of forest fires was obtained through hierarchical calculations. The forest fire risk zones in Shanxi Province were classified into zero, low, moderate, high, and extremely high-risk categories. Ríos-Pena et al. [62] applied a bivariate Structured Additive Regression (STAR) model to wildfire occurrence modeling in Galicia. The bivariate STAR framework extends classical logistic regression and generalized additive models by enabling flexible nonlinear effect modeling while directly incorporating spatiotemporal variables, thereby elucidating potential interdependencies among predictors. Mohammadi et al. [63] introduced a novel approach to identify climatic variables influencing wildfire burn area. Combining multivariate linear regression with nonlinear variable transformations, they derived a nonlinear wildfire size function. The results indicated that increased air temperature reduces moisture content, desiccating fuels and expanding burn areas, whereas higher relative humidity coupled with reduced wind speed diminishes fire spread. To improve seasonal wildfire forecasting, Yin et al. [64] developed a multivariate linear regression-based method for predicting spring burn areas in Western Siberia one season in advance (R² coefficient = 0.64).

Zhang et al. [65] utilized the Moderate Resolution Imaging Spectroradiometer (MODIS) hotspot database and logistic regression models to investigate the probability of at least one wildfire occurrence within 1 km² grid cells across a 1,030,000 km² region in south-eastern Australia, aiming to predict future fire events. Their analysis revealed that wildfires are most likely to occur in mountainous areas, forests, savannas, and regions with high vegetation cover, while grasslands and shrublands exhibited lower susceptibility. Lan et al. [66] developed a spatial wildfire susceptibility prediction model using logistic regression, incorporating climatic, vegetative, topographic, anthropogenic, and geographic predictors. The model demonstrated strong alignment with historical wildfire data, achieving an overall simulation probability of 81.6%. Papakosta et al. [67] developed a mesoscale probabilistic daily fire prediction model for the European Mediterranean region (Cyprus) using Poisson regression. The framework relies on readily available spatiotemporal data, including weather conditions, land cover, and anthropogenic presence. While the FWI partially captured weather-related fire risk, its standalone predictive capacity was limited. To forecast wildfire scale, Yang et al. [68] constructed multiple multivariate linear regression (MLR) models linking burned area (BA) to concurrent/preceding drought indices and temporal trends across seven regions of mainland China (2001–2019). The burned area was estimated using MODIS data, while drought conditions were quantified via the Standardized Precipitation Evapotranspiration Index (SPEI) and the Self-Calibrating Palmer Drought Severity Index (SC-PDSI). Kouassi et al. [69] employed generalized additive models (GAMs) and Spearman correlation to evaluate the relationships between climatic variables and wildfire occurrence. Their methodology integrated MODIS-derived monthly burned area data with trend analyses using seasonal Kendall and Sen’s slope tests.

3.1.2. Forest Fire Prediction Based on Probabilistic Model

Chávez et al. [70] proposed a probabilistic methodological framework to assess extreme drought conditions preceding wildfire occurrences in Central Chile, a region experiencing unprecedented prolonged drought. This innovative model evaluates not only meteorological and vegetation index anomalies that promote extreme events but also quantifies the rarity or extremity of such conditions. Gheshlaghi et al. [71] mapped forest fire susceptibility in Nowshahr County, Iran, using a hybrid approach integrating the Index of Entropy (IOE) with Fuzzy Membership Values, FR, and Information Value. Validation results demonstrated that the proposed model achieved robust predictive performance, yielding an AUC value of 0.890. Baranovskiy et al. [72] proposed a deterministic–probabilistic approach employing the Single Program Multiple Data (SPMD) computational model for parallelized forest fire risk estimation. The methodology was validated through implementation in the Timiryazevskiy forest area of Tomsk Region, Russia. Representative results demonstrating forest fire risk assessment capabilities in managed woodland ecosystems were systematically presented. Boadi et al. [73] implemented empirical probability distribution models to fire count data and conducted a comparative analysis with theoretical probability distributions derived from stochastic processes. The study demonstrated that fire frequency and associated casualties (fire-induced fatalities) in Ghana exhibit optimal modeling alignment with the negative binomial distribution for count data analysis. Jaafari et al. [74] investigated the comparative effectiveness of Frequency Ratio and its ensemble integration with Shannon’s Entropy (SE) methodologies for wildfire susceptibility prediction in Iran’s Chaharmahal-Bakhtiari Province. Validation metrics revealed distinct predictive competencies, with the standalone FR model attaining 85.16% prediction accuracy compared to the hybrid FR-SE ensemble’s 82.93% success rate, highlighting methodological trade-offs between statistical simplicity and entropy-enhanced spatial pattern recognition. Parisien et al. [75] conducted a forest fire susceptibility prediction for the Minudasht forest in Golestan Province, Iran, using remote sensing and GIS data. They employed the SE and FR models to map the forest fire susceptibility in the study area. The experiments showed that the AUC values for SE and FR were 83.16% and 79.85%, respectively, with standard errors of 0.044 and 0.047.

Mohammed et al. [76] utilized data mining methods based on the evidential belief function (EBF) and Weight of Evidence (WOE) models to map fire risk in Mariwan County, Kurdistan Province, Iran. The results indicated that the WOE model achieved a slightly higher AUC value (0.896) compared to the EBF model (0.886), suggesting its marginally superior performance. Bashari et al. [21] employed BBNs to address uncertainty in wildfire prediction. By constructing an influence diagram to identify factors affecting wildfires in Iran’s arid and semi-arid regions, they populated probabilistic relationships to generate the BBN. Experimental results demonstrated high predictive accuracy, with a Receiver Operating Characteristic (ROC) area under the curve (AUC) of 0.986. To enhance forecasting efficiency, Ujjwal et al. [77] integrated probabilistic risk metrics with spatial clustering for effective wildfire management. Their probabilistic risk index quantifies fire spread probabilities under given weather conditions from ignition points. By clustering Tasmania’s fire ignition locations into three risk clusters, they reduced computational demands by several orders of magnitude while maintaining ~95% accuracy in high- and low-risk classifications, validating the feasibility of combining probabilistic metrics with spatial clustering. For wildfire scale prediction, Koh et al. [22] developed a joint model integrating extreme value theory and point processes within a Bayesian hierarchical framework. Applied to summer wildfire data (1995–2018) in the French Mediterranean Basin, the model simultaneously characterized fire occurrence intensity and burn area distribution. Castel-Clavera et al. [78] evaluated wildfire occurrence and size using the Firelihood framework, a probabilistic Bayesian model of fire activity. Their findings emphasized that supplementing the FWI with spatial and seasonal effects significantly improved predictive performance. The refined probabilistic model substantially enhanced multiple aspects of wildfire activity forecasting in Mediterranean regions. Liu et al. [79] conducted a province-wide statistical analysis of 6519 forest fire events spanning 1969–2013, employing spatiotemporal decomposition and Kruskal–Wallis tests. The results revealed significant monthly, interannual, and regional variability in fire frequency and burned areas, with spatial autocorrelation indicating clustered fire regimes in southeastern ecoregions.

3.1.3. Forest Fire Prediction Based on Analysis of Fire Risk Index

To predict forest fire risk, Zubieta et al. [80] investigated Andean vegetation fire danger using the Global Vegetation Moisture Index (GVMI). An analysis of daily precipitation and temperature observations from the PISCO gridded dataset (2002–2016) revealed statistically significant associations between climatic parameters (e.g., cumulative precipitation, dry-day frequency, and heatwave frequency) and conditions conducive to increased wildfire occurrence. Their findings demonstrated that reduced vegetation moisture content, as estimated by GVMI during drought onset and rainy season transitions, effectively identifies potential wildfire ignition conditions. Artés et al. [81] proposed the Extreme Fire Behavior Index (EFBI), a novel fire danger metric complementing traditional indices. EFBI quantifies the likelihood of wildfire–atmosphere interactions that may trigger deep moist convection and pyroconvective instability. The results indicated its potential utility in identifying convection-driven fires and enhancing global fire danger rating systems. Chen et al. [82] developed a short-term wildfire risk assessment method based on MODIS data. Similarly, Stefanidou et al. [83] introduced the Mid-Term Fire Danger Index (MFDI), integrating MODIS time-series-derived dry fuel connectivity metrics with biophysical and topographic variables to achieve accurate 8-day-ahead ignition risk predictions. Abdo et al. [84] developed a spatial distribution map of forest fire susceptibility in the Al-Draikich region of western Syria by integrating NDVI, Normalized Difference Moisture Index (NDMI), Topographic Wetness Index (TWI), and other relevant factors using FR and Analytic Hierarchy Process (AHP) techniques. The experimental results demonstrated that the FR method achieved the highest predictive accuracy, with an AUC value of 0.864. Abedi et al. [85] developed a forest fire susceptibility map for the Noshahr forests in Northern Iran by employing a GIS-based Analytic Network Process (ANP) as a Multi-Criteria Decision Analysis (MCDA) method. The model integrated nine input parameters: slope, aspect, elevation, land cover, NDVI, annual precipitation, annual temperature, distance to settlements, and distance to roads. The validation results demonstrated that the proposed framework achieved satisfactory predictive accuracy, with an AUC value of 0.783. Mann et al. [86] developed a fire risk prediction model that integrates estimates of biophysical indicators associated with plant communities and human influences at each forecast time step. Busico et al. [87] developed a forest fire risk assessment methodology integrating GIS technology with the AHP. Validation experiments demonstrated strong agreement between the observed and predicted fire occurrences (R² = 0.79), indicating robust model performance in wildfire pattern replication. Ghorbanzadeh et al. [88] developed a wildfire risk prediction framework integrating the ANP as the central component of an MCDA methodology, systematically evaluating 12 critical factors influencing wildfire susceptibility mapping. Validation results demonstrated that the ANP-optimized susceptibility model achieved 84% predictive accuracy, highlighting the methodological efficacy of combining network-based decision structures with multi-factor environmental diagnostics.

Babu et al. [89] developed a fire danger prediction model incorporating three critical parameters: the Modified Normalized Difference Fire Index (MNDFI), Perpendicular Moisture Index (PMI), and potential land surface temperature. Validation experiments demonstrated robust model performance, achieving a classification accuracy of 87.31%. Kayet et al. [90] employed a Frequency Ratio model and AHP framework to assess forest fire risk in the Melghat Tiger Reserve, Central India. Experimental validation revealed that the FR model achieved significantly superior performance, with an overall accuracy of 81.3%, outperforming the AHP model (accuracy: 79.3%). Kumari et al. [91] developed a forest fire susceptibility prediction model through geoinformatics-based MCDA employing the AHP. The results delineated five distinct risk zones: very low fire danger (180 km², 14.85%), low fire danger (234 km², 19.30%), moderate fire danger (269.73 km², 22.16%), high fire danger (299.36 km², 24.59%), and extreme fire danger (232.56 km², 19.10%). Holdrege et al. [92] combined wildfire observations with climatic and vegetation data to construct a statistical model for annual wildfire probability across the sagebrush biome. The model highlighted climate’s predictive role, with highest wildfire probabilities forecasted in regions characterized by low summer precipitation ratios, moderate annual precipitation, and elevated temperatures. Yu et al. [93] proposed a wildfire prediction and risk assessment framework leveraging multiple fire danger indices (FDIs) integrating weather and fuel conditions. Using high-resolution (4 km) climate and fuel datasets (1984–2019), they evaluated four FDIs across the U.S., demonstrating that higher FDIs correlated with larger total wildfire extents, particularly at broader spatial scales. Surface soil moisture emerged as a critical wildfire predictor, offering advantages in fuel moisture estimation and ignition likelihood assessment.

Alizadeh et al. [94] integrated land–atmosphere coupling and soil moisture memory to predict wildfires in the western United States. By analyzing dynamic interactions among soil moisture, vegetation moisture, atmospheric aridity, fuel load, and post-fire recovery, they identified that soil moisture anomalies ~5 months pre-ignition enhance biomass growth, priming fire-prone vegetation. Subsequent concurrent reductions in soil moisture, vegetation desiccation, and atmospheric drying collectively drive ignition events. O et al. [95] analyzed 9840 global large wildfire events (2001–2018) alongside surface soil moisture and biomass data, revealing divergent soil moisture–fire relationships: in arid regions, above-average soil moisture supported biomass growth for fuel accumulation, whereas in humid zones, fires typically followed dry soil anomalies that reduced environmental moisture barriers. Sharma et al. [96] advanced soil moisture-based wildfire prediction through coordinated in situ measurement networks. Their findings underscored that field-based soil moisture data, when integrated with remote sensing, outperformed standalone remote sensing in wildfire risk forecasting, offering enhanced predictive capability for fire management applications.

To predict wildfire burn area, Vissio et al. [97] proposed a simple empirical data-driven model linking burned area to climatic drivers, comparing correlations derived from direct surface soil moisture measurements with drought indices such as the Standardized Precipitation Index (SPI) and SPEI. Their results demonstrated superior performance of SPEI-based models in arid regions and the Emilia-Romagna area. Farahmand et al. [98] introduced the Fire Danger from Earth Observations (FDEO) framework, leveraging satellite data across the contiguous United States (CONUS) for two-month-ahead wildfire danger forecasts. Integrating satellite observations of land cover type, vapor pressure deficit, surface soil moisture, and the Enhanced Vegetation Index (EVI) with the U.S. Forest Service (USFS)-validated Fire Program Analysis (FPA) database, they developed spatially gridded probabilistic predictions of fire danger, defined as deviations in expected burned area from baseline conditions. The model achieved a 52% overall accuracy improvement over 2004–2013 records, rising to 75% during peak fire seasons. Turco et al. [99] analyzed climatic drivers of June–October burned area in Portugal (1980–2017) using SPEI, SPI, and the Standardized Soil Moisture Index (SSI). Their findings indicated that climate change alone would escalate BA values that lack mitigating factors. He et al. [100] developed a vegetation type-specific threshold-based method for fire severity prediction, achieving accuracies of 0.64 and 0.76 in extreme and high-severity classifications, respectively. Bieniek et al. [101] evaluated spring snow disappearance dates (1959–2020) and associated synoptic-scale atmospheric drivers in Alaska to assess linkages to summer wildfire seasons. The largest fire seasons, accounting for 56%–95% of historical burned area, frequently followed early regional snow disappearance. April–May snowpack exhibited the strongest correlation with daily maximum temperatures, with weaker associations thereafter. In some years, early snow loss and sustained warm anomalies extended into late summer, amplifying fire danger indices—a pattern observed during high-burn area years. Urrutia-Jalabert et al. [102] investigated multidecadal relationships between wildfire occurrence and climatic variability in Central and South-Central Chile. Their analysis revealed that fire activity in Central Chile correlates strongly with above-average winter precipitation in the preceding year followed by spring-to-summer drying conditions.

3.1.4. Forest Fire Prediction Based on Meteorological Models

Climate models, though more costly and complex than traditional tools, offer distinct advantages in analyzing fire risk factors and enabling large-scale wildfire risk prediction. Paschalidou et al. [103] employed COST-733 classification to link wildfire occurrences in Greece with synoptic weather circulation patterns. Their analysis revealed that most wildfires could be attributed to a limited set of atmospheric configurations, reflecting the region’s fire-prone weather climatology. All eight classification schemes demonstrated that Greece’s most hazardous fire conditions arise from high-pressure systems over Northern/northwestern Greece combined with low-pressure systems in the Eastern Mediterranean—a pattern closely associated with the Etesian winds over the Aegean Sea. Coen et al. [23] implemented a two-way coupling of computational fluid dynamics (CFD) models, incorporating weather prediction frameworks with modules simulating fire spread and heat release, to simulate fire–atmosphere interactions across three orders of magnitude in scale. These advanced weather–fire coupled modeling systems, integrated with field-derived fuel data, real-time fire detection, and aerial/satellite remote sensing, are addressing critical scientific questions. While their dynamic computational demands exceed traditional tools, they have generated groundbreaking insights, particularly into the mechanisms underlying extreme wildfire events. Goss et al. [25] developed a climate model-based approach to project wildfire risks, using simulations to identify meteorological drivers of extreme autumn wildfires in California and assess their attribution to anthropogenic climate change. Their analysis projects that ongoing climate change will amplify extreme fire weather days by the century’s end.

Farfán et al. [104] developed a climatic probability model for wildfire occurrence, systematically examining the modulating influence of ENSO phases on vapor pressure deficit (VPD)-intensified meteorological conditions in Guanajuato state, Central Mexico’s semi-arid region. The study identified an elevated wildfire risk in southern sectors of Guanajuato, with fire-prone atmospheric states mechanistically linked to ENSO-driven VPD anomalies through coupled ocean–atmosphere dynamics. Grünig et al. [105] identified statistically significant positive correlations between summer VPD and wildfire magnitude metrics, with coefficients of determination (R²) of 0.19 for maximum fire size and 0.12 for peak burn severity, demonstrating the climatic control of atmospheric aridity on extreme wildfire behavior. Virgilio et al. [24] evaluated future shifts in pyrocumulonimbus (PyroCb) formation risks in south-eastern Australia using high-resolution regional climate projections. By analyzing coincident high atmospheric instability (C-Haines index) and near-surface fire weather conditions, they found that observed PyroCb events predominantly occur in forested, rugged terrains under extreme C-Haines scenarios. Lindley et al. [106] conducted a preliminary analysis of the Red Flag Threat Index (RFTI) from the Texas Tech University real-time weather prediction system, comparing its outputs with wildfire activity observed via GOES-16 during four wildfire outbreaks in the southern Great Plains. Visual comparisons of RFTI forecasts and satellite-detected fire evolution—presented in static and animated formats—validated the model’s predictive capacity. Their findings advocate for expanding fire weather-specific parameters in high-resolution numerical weather prediction systems to enhance wildfire forecasting.

To address the temporal resolution and latency limitations of RFTI forecasting, Jones et al. [107] developed a gridded RFTI using a five-year high-resolution (3 km, hourly) reanalysis product—the Real-Time Mesoscale Analysis (RTMA). Leveraging RTMA data enabled the creation of continuous wind and humidity climatology fields for the extended western contiguous United States (W-CONUS) domain, facilitating broader RFTI applicability. Their framework integrates these climatological datasets with forecasts from the Warn-on-Forecast System (WoFS), which provides short-term (0–6 h) probabilistic predictions of high-impact weather events at regional scales. Dacre et al. [108] proposed a novel prototype probabilistic daily wildfire warning system for Chile, combining state-of-the-art ensemble weather forecasts with satellite-derived fire and vegetation data to predict, visualize, and communicate wildfire risk likelihoods. This probabilistic approach quantifies uncertainties, empowering users to make informed risk management decisions. Low-Level Thermal Ridges (LLTRs), identified as common meteorological features associated with wildfires in the southern Great Plains’ grassland-dominated fuel landscapes, were analyzed by Lindley et al. [109]. Through detailed meteorological examination of 11 widespread wildfire events (2006–2014), they derived composite atmospheric profiles and conceptual models from LLTRs’ upwind fire behavior, enhancing predictive capabilities for wildfire outbreaks in the region. Zhao et al. [110] investigated wildfire dynamics in China’s monsoon-influenced regions, where climatic variability driven by multiple atmospheric systems leads to significant spatial heterogeneity in fire activity. Their study identified a circulation index, defined as the pressure differential between positive and negative fire–pressure correlation centers, which outperformed local weather variables in monthly large-fire prediction. Bowman et al. [111] proposed an “ecological drift” model to interpret wildfire trends, hypothesizing that vegetation patterns are fire-regulated and that anomalous shifts in fire frequency drive rapid landscape-scale vegetation reorganization. Contrary to model predictions, their findings revealed that short-term fire risk in sedgelands decreased for at least five years due to substantially reduced fuel loads.

Chikamoto et al. [26] developed a novel multi-year dynamic prediction system that integrates a state-of-the-art Earth System Model (ESM) with three-dimensional ocean data assimilation. By prescribing external radiative forcing, the system simulates observed low-frequency variability in precipitation, soil moisture, and wildfire probability, achieving close alignment with observational records and reanalysis data. This framework demonstrated robust performance in forecasting wildfire probability and drought conditions at lead times of 10–23 months and 10–45 months. Additionally, physics-based wildfire prediction methodologies have advanced. Graciano et al. [112] proposed a risk prediction model grounded in physical and conceptual frameworks, incorporating Monte Carlo simulations. Their approach employs probabilistic, stochastic, and physical techniques to refine sub-model parameterizations, demonstrating its utility in supporting decision-makers to implement preventive measures for wildfire mitigation. Pan et al. [113] introduced a physics-informed statistical–dynamical (SD) model for wildfire prediction, which exhibited the highest operational predictive skill (0.58) among tested methods and maintained stability via K-Fold cross-validation. By selecting relevant meteorological variables and optimizing dynamic forecast fields, the model enables accurate regional wildfire predictions. This methodology shows significant potential for seasonal wildfire forecasting and future projections under climate change scenarios.

3.2. Forest Fire Prediction Based on Machine Learning

Machine learning (ML)-based approaches are increasingly being applied to forest fire prediction research. ML techniques demonstrate a favorable balance between predictive performance and computational efficiency, providing distinct advantages in resource-constrained scenarios. Furthermore, their enhanced interpretability facilitates researchers’ understanding of fire risk factor contributions, thereby guiding the formulation and implementation of targeted fire prevention strategies. This section reviews ML-based forest fire prediction studies, emphasizing the accuracy and comparative analysis of methodologies, input variables, and their relative importance in wildfire events. A critical evaluation of the strengths and limitations of different ML models is also provided. The technical classification of forest fire prediction research based on machine learning is illustrated in Figure 4.

3.2.1. Forest Fire Prediction Based on Artificial Neural Networks

Razavi et al. [28] developed an ANN-based model to predict potential fire ignition points, focusing on non-anthropogenic wildfire causes and daily fire occurrence probability. Their results indicated that in temperate forests, a 0.1 °C temperature increase relative to the 7-day pre-fire average elevated the model’s output probability (>0.8) by 4.75%. Conversely, in boreal forests, temperature increases below 1 °C relative to the 7-day average signaled negligible wildfire likelihood. Temperature-independent variables exhibited greater predictive significance in temperate forests compared to boreal ecosystems. Sadatrazavi et al. [114] constructed an ANN model using meteorological parameters and a U.S. wildfire database (1992–2015), analyzing parameter sensitivity and influence. The input variables included temperature, relative humidity, total pressure, evaporation, soil moisture, snowpack, Keetch–Byram Drought Index (KBDI), precipitation, wind speed, and NDVI. Sensitivity analysis identified total pressure, KBDI, and relative humidity as critical predictors in temperate forests, whereas relative humidity and total pressure dominated in boreal forests. Impact magnitude rankings further highlighted total pressure, wind speed, and humidity as top drivers for temperate zones, versus KBDI, humidity, pressure, wind, and precipitation for boreal regions. Kantarcioglu et al. [115] evaluated forest fire susceptibility in Istanbul Province and the Thrace Region, Turkey, using ANNs. The results demonstrated that the ANN model achieved superior predictive performance, with an AUC of 0.94.

Ntinopoulos et al. [116] proposed a wildfire prediction framework integrating ANN and Radial Basis Function (RBF) networks, leveraging data from the FWI, Fosberg Fire Weather Index (FFWI), NDVI, and NDMI. To enhance vegetation-driven fire risk assessment, they introduced the vegetation-augmented FWI_veg index, optimized via Genetic Algorithms. The ANN model, applied to the 2018 Mati wildfire in Attica, Greece, demonstrated the efficacy of combining FWI_veg and traditional indices in AI-driven fire prediction. Aljumah et al. [117] designed an energy-efficient Internet of Things (IoT) architecture for early wildfire detection, utilizing fog-cloud computing and Jaccard similarity analysis to preprocess sensor data. A reduced-dimensional Forest Fire Index (FFI) was derived at the fog layer, followed by ANN-based susceptibility modeling. Self-organizing mapping techniques spatially visualized wildfire vulnerability. Nassa et al. [118] employed a MLP and the Advanced Relative Operating Characteristic method to address the shortcomings of existing approaches, such as false alarms, untimely notifications, and inadequate network coverage. The experiments revealed that the model achieved an accuracy rate of 90%, which was higher than that of fuzzy logic and the average consensus algorithm. For burn area prediction, Pérez-Sánchez et al. [30] employed a two-stage ANN approach to classify and forecast wildfires exceeding 30,000 m² across 39.41 km² of burned land (2000–2014). The model achieved success rates of 60%–70%, varying by regional conditions.

3.2.2. Forest Fire Prediction Based on Ensemble Learning

Jodhani et al. [119] developed a wildfire susceptibility assessment framework by synthesizing environmental variables and machine learning techniques. Variables including aspect, elevation, NDVI, drainage density, precipitation, and temperature were analyzed to characterize landscape features and evaluate fire susceptibility. The RF regression model was utilized to predict land surface temperature (LST) based on these parameters. Output maps visualized the geographic distribution of Normalized Burn Ratio (NBR), Differenced Normalized Burn Ratio (dNBR), and LST forecasts. Sánchez et al. [120] developed a wildfire risk prediction model using Random Forest algorithms, leveraging a decade-long dataset (2010–2019) of burned areas in Santa Cruz Province coupled with predictor variables encompassing topographic variables, land cover classifications, and ecoregion delineations. The validated model demonstrated robust predictive capability, achieving an AUC of 0.8, which signifies substantial discrimination power in fire risk forecasting. Jang et al. [121] developed a three-stage algorithmic framework for forest fire detection integrating (1) threshold-based segmentation, (2) machine learning classification, and (3) post-processing optimization, utilizing Himawari-8 geostationary satellite data over the Korean Peninsula. Experimental validation demonstrated a 93% detection probability for wildfire events, establishing this multi-scale approach as an effective operational solution for real-time fire monitoring in Northeast Asia’s complex terrain. Latifah et al. [122] developed a Random Forest model integrating climate variables and satellite-derived fire hotspot data to predict forest fire occurrences in Borneo. Spatiotemporal variability analysis demonstrated the model’s robust capability in capturing fire patterns.

To enhance prediction efficacy, Jaafari et al. [123] proposed a spatially explicit wildfire probability model integrating ANFIS with Genetic Algorithms (GAs) and Firefly Algorithms (FAs), achieving an AUC of 0.92. Janizadeh et al. [35] developed hybrid wildfire susceptibility prediction models using metaheuristic algorithms—Golden Jackal Optimization, Pelican Optimization Algorithm (POA), and Zebra Optimization Algorithm (ZOA)—coupled with Light Gradient Boosting Machine (LightGBM). The ZOA-LightGBM hybrid exhibited the highest accuracy for Kauai (AUC = 0.9314) and Molokai (AUC = 0.858) in Hawaii. Peng et al. [124] identified critical wildfire drivers among 14 variables (climate, vegetation, anthropogenic, and topographic) through Pearson correlation analysis and Random Forest, subsequently applying logistic regression for prediction. The model achieved an AUC of 0.944. Nikolaychuk et al. [125] employed a Random Forest machine learning method to map wildfire susceptibility in Irkutsk Oblast, Russia, using remote sensing, meteorological, forestry, and socio-economic data (2017–2020). Their classification framework achieved robust performance metrics: accuracy = 0.89, F1 score = 0.88, and AUC = 0.96. To address limitations of coarse-resolution remote sensing data, Radočaj et al. [126] proposed a wildfire prediction method based on MODIS 250 m imagery, featuring a two-tiered classification: (1) supervised machine learning for generalized land cover categories and (2) unsupervised classification for vegetation subtypes. The Random Forest classifier using Landsat 8 bands and indices achieved 83.6% accuracy in tier one. Subtype-fuel model matching in FARSITE (4.0) software yielded a peak 2-day average prediction accuracy of 0.906. While 250 m resolution enables high spatiotemporal wildfire growth prediction, limitations such as wildfire overprediction and smoke interference in MODIS imagery require further resolution.

Qiu et al. [127] investigated wildfire drivers by developing a Random Forest-based machine learning model incorporating 23 climatic and surface variables to analyze spatial and seasonal variability in wildfire occurrence and extent across California. Using Shapley Additive Explanations (SHAP), they quantified driver importance, revealing precipitation and soil moisture as dominant factors for large/extreme wildfires in summer (37% of total burned area) and autumn (63%), while elevation primarily influenced small-fire burned area (15%–58%) across seasons. Wind emerged as a critical summer driver, contributing 41% to large/extreme fire spread. Tang et al. [128] proposed a machine learning framework to assess wildfire susceptibility and seasonal ecosystem threats in China’s Greater Khingan Range. A Random Forest model evaluated wildfire susceptibility, while the dynamic equivalent coefficient factor method calculated ecosystem service value (ESV) at the grid level. The results highlighted significant seasonal variations in susceptibility, with 2019 wildfire-induced potential losses far exceeding regional GDP, establishing a critical link between wildfire risk and ecosystem service valuation. Valdez et al. [29] integrated wildfire events (2010–2015) with anthropogenic variables (e.g., proximity to infrastructure, settlements, and roads) using Random Forest. Their analysis identified dry fuel conditions, low precipitation, and proximity to unpaved/paved roads as primary drivers. High and extreme susceptibility zones covered 15% of the study area, concentrated in central and eastern regions. Yue et al. [129] employed Light Gradient Boosting Machine (LGBM) to construct wildfire susceptibility models for aggregated and individual land use types, enhancing the understanding of spatial risk distribution and land use-specific vulnerabilities. In a follow-up study, Yue et al. [130] applied LGBM with SHAP analysis to assess seasonal wildfire risk in Nanning, China, using a Remote Sensing Ecological Index (RSEI) derived from greenness, heat, wetness, and dryness metrics. Wildfires predominantly occurred in spring and winter, with minimal likelihood in summer and autumn. McNorton et al. [131] employed vegetation characteristic models, weather forecasts, and data-driven machine learning approaches to construct a global daily fire probability (PoF) model operating at a resolution of approximately 9 km across multiple delivery cycles. The PoF model outperforms existing indices by providing accurate forecasts of fire activity up to 10 days in advance, and in some cases, even up to 30 days. Wang et al. [132] developed an XGBoost-based model to predict monthly burned area in the contiguous U.S. at 0.25° × 0.25° resolution, integrating meteorological, land surface, socio-economic, and large-scale circulation pattern predictors. Incorporating circulation indices improved temporal correlations by 14%–44% regionally. The Energy Release Component (ERC) was identified as a key driver for large burns in the western U.S. (14%–27% contribution), whereas no dominant factor emerged for eastern regions.

3.2.3. Forest Fire Prediction Based on Probabilistic Models

Integrating prior information into Bayesian models holds significant potential for accurate wildfire probability prediction. Charizanos et al. [36] developed a Bayesian wildfire probability model utilizing NDVI data from Google Earth Engine, incorporating both informative and non-informative prior distributions to address mediator effects among predictors. Simulations identified forest vulnerability as the primary predictor of wildfire likelihood. D’Arcy et al. [133] proposed a zero-inflated negative binomial model for wildfire prediction, selecting spatial neighborhoods for missing observations and fitting marginal models to non-missing data. Wildfire counts followed a zero-inflated negative binomial distribution, while burned area was modeled using nonparametric and parametric techniques for the body and tail of the distribution, respectively. Yang et al. [134] constructed a Maximum Entropy (MaxEnt)-based wildfire risk assessment model using remote sensing-derived fire hotspots and 12 environmental variables (topography, climate, vegetation, and anthropogenic factors). Meteorological conditions (54.64% importance, 43.03% contribution) and vegetation status (25.40% importance, 34.69% contribution) dominated wildfire drivers, with forest density exhibiting a U-shaped relationship to fire probability—both excessively high and low densities increased risk. Naseer et al. [135] analyzed the influence of environmental and social variables on the occurrence of forest fires using the Maxent. They examined historical MODIS active fire data from 2000 to 2022 to establish the relationship between forest fire likelihood and environmental conditions. The experiments demonstrated that the AUC value for fire probability was 0.833, indicating good model performance.

Bugallo et al. [38] implemented zero-inflated negative binomial mixed models for wildfire prediction in Spain, achieving 95% confidence in retrospective validation, thereby offering actionable insights for preventive strategies. Joseph et al. [37] integrated wildfire records with meteorological and housing data in a spatiotemporal Bayesian framework to predict extreme wildfire events in the contiguous U.S. Their zero-inflated negative binomial model for fire counts and lognormal model for burn area achieved 99% coverage for fire counts and 93% for burn area. Martín et al. [136] applied MaxEnt to analyze intra-annual wildfire patterns and triggers in Northeastern Spain, using wildfire data (2008–2011) alongside GIS and remote sensing variables. By stratifying data into four seasons and further dividing into weekdays/non-weekdays, they created eight event scenarios, with model performance consistently exceeding AUC = 0.8. Kaur et al. [137] designed an IoT-fog-cloud collaborative framework for real-time wildfire monitoring and prediction. The system employs k-nearest neighbor (k-NN) classifiers to categorize forest terrain into wildfire susceptibility classes and Holt–Winters forecasting to project future risk levels. Validation results demonstrated high accuracy (93.97%), specificity (92.35%), sensitivity (93.01%), and precision (91.24%) in susceptibility assessment.

3.2.4. Forest Fire Prediction Based on a Multi-Algorithm Hybrid Model

To enhance wildfire prediction accuracy, researchers have increasingly adopted hybrid algorithmic approaches for model construction. Malik et al. [138] proposed two data-driven machine learning methods based on Random Forest models to predict wildfire risk near Monticello and Winters, California. The combined model utilized spatial and temporal parameters as a single integrated dataset for training and prediction, while the ensemble model processed separate parameter sets later stacked into a unified framework. Experimental results demonstrated superior accuracy for the combined model (92%) compared to the ensemble approach when using spatial–temporal data, with regional parameters achieving high ignition risk prediction accuracy. Tehrany et al. [39] modeled annual wildfire occurrences in Australia’s Brisbane River catchment by integrating entropy index (IoE), evidential belief function (EBF), and logistic regression. The EBF-LR hybrid achieved the highest accuracy (AUC = 88.51%). Achu et al. [139] proposed a weighted ensemble approach to assess forest fire susceptibility in the Wayanad district of Kerala State, India, by leveraging outputs from multiple machine learning techniques. Validation results demonstrated that the integrated model achieved superior predictive performance compared to individual ML methods, yielding an AUC value of 0.890. Al-Fugara et al. [140] employed SVR, ANFIS, and two metaheuristic models—Whale Optimization Algorithm (WOA) and Simulated Annealing—to map wildfire risk in the Jerash Governorate of Jordan. The results demonstrated that the SVR-based hybrid models (with AUC values of 0.965 and 0.949) outperformed the ANFIS-based hybrid models (0.904 and 0.894) in predictive performance. Bui et al. [141] proposed and comparatively evaluated three novel hybrid metaheuristic approaches—Biogeography-Based Optimization (BBO), Gravitational Search Algorithm (GSA), and Grey Wolf Optimizer (GWO)—integrated with a neural network classifier for forest fire modeling. Empirical results demonstrated that the AUC values for BBO (0.9515), GWO (0.9509), and GSA (0.9398) outperformed the conventional backpropagation-based neural network classifier, which achieved an AUC of 0.9271. Janiec et al. [142] utilized MaxEnt and Random Forest algorithms to model forest fire risk prediction. The results demonstrated that human activities exerted a significant influence on regional wildfire risk despite low population density, with anthropogenic factors exhibiting stronger correlations to wildfire occurrence than climatic or topographic variables. Janizadeh et al. [143] employed a GLM integrated with four ensemble methods—Partial Least Squares (PLS), Boosting, Bagging, and Bayesian frameworks—to assess forest fire susceptibility in the Chalus Rood Watershed, Mazandaran Province, Iran. The experimental evaluation revealed distinct predictive performances across hybrid models, with AUC values of 0.79 (GLM), 0.75 (PLS-GLM), 0.81 (Boosted-GLM), 0.84 (Bagging–GLM), and 0.85 (Bayesian–GLM), respectively.

Mabdeh et al. [144] developed four hybrid models, namely SVR-GA, SVR-SFLA, ANFIS-GA, and ANFIS-SFLA, and compared their performance in predicting forest fire risk. The results indicated that hybrid algorithms based on SVR outperformed those based on ANFIS, achieving a better AUC value of 0.97. Mohajane et al. [40] developed five hybrid machine learning algorithms—Frequency Ratio–Multilayer Perceptron (FR-MLP), FR–Logistic Regression (FRLR), FR–Classification and Regression Tree (FR-CART), FR–Support Vector Machine (FR-SVM), and FR–Random Forest—for wildfire susceptibility mapping in northern Morocco. The FR-RF model exhibited optimal performance (AUC = 0.989). Nur et al. [145] integrated GIS and machine learning to assess wildfire susceptibility in Sydney, with the hybrid SVR-PSO (Support Vector Regression–Particle Swarm Optimization) model yielding optimal performance (AUC = 0.882, RMSE = 0.006). Coughlan et al. [146] employed ML methods to develop lightning–ignition prediction models using decision trees, AdaBoost, and Random Forest. Both ensemble methods (RF and AdaBoost) achieved 78% out-of-sample accuracy. Jiménez-Ruano et al. [147] proposed a wildfire prediction framework based on multinomial Generalized Linear Models (GLMs) and GAM, achieving AUC values of 0.84 and 0.89 for natural and anthropogenic ignition predictions, respectively.

Researchers have implemented wildfire risk stratification and spatial mapping to enhance predictive analysis. Chen et al. [148] developed a wildfire probability model using Random Forest and XGBoost, demonstrating that incorporating dynamic time-lagged weather variables improved model accuracy. Li et al. [149] evaluated wildfire drivers via RF and simulated ignition probability using ANFIS, optimized with Particle Swarm Optimization (PSO) and Genetic Algorithms. Hybrid ANFIS models (PSO-ANFIS and GA-ANFIS) outperformed standalone ANFIS by mitigating overfitting during wildfire pattern learning. PSO-ANFIS exhibited superior classification accuracy, identifying high-risk zones predominantly in northeastern study areas—specifically the grasslands and forests of Mongolia’s Dornod Province, Russia’s Buryatia and Chita regions, and northeastern Inner Mongolia, China. These probabilistic maps provide reliable estimates for wildfire hazard management in adjacent regions. Zhang et al. [150] assessed wildfire risk distribution using a Backpropagation Neural Network (BPNN) with feature selection. Filter and wrapper methods generated five feature subsets, while Genetic Algorithms optimized network parameters. The refined BPNN model classified risk levels, with medium-to-extreme risk zones encompassing 90.26% of new fire incidents, validating its applicability. Hai et al. [151] proposed a hybrid machine learning approach integrating Multivariate Adaptive Regression Splines (MARS) with Cat Swarm Optimization (CSO). Utilizing a GIS database of 11 predictors and 262 fire locations (2013–2018), the MARS-CSO model outperformed Model Trees, Reduced Error Pruning Trees, and standalone MARS. The resultant risk map classified 20% of the study area as very low risk and 40% as extremely high risk, demonstrating robust spatial discriminative capability.

3.3. Forest Fire Prediction Based on Deep Learning

In recent years, the rapid advancement of deep learning has prompted its increasing application in forest fire prediction research. Deep learning algorithms enhance prediction accuracy by integrating multi-source heterogeneous data (e.g., satellite imagery, climate data, topography information, meteorology data, vegetation information, anthropogenic activities, wildfire records, road and river information, and economic level). However, the black-box nature of deep learning frameworks limits interpretability, hindering the elucidation of critical fire-driving factors. This section evaluates the accuracy, advantages, and limitations of deep learning-based wildfire prediction methodologies, with a focused analysis on model input variables (fire-driving factors) and their relative contributions to predictive outcomes. The technical classification of forest fire prediction research based on deep learning is illustrated in Figure 5.

3.3.1. Forest Fire Prediction Based on Convolutional Neural Networks

To predict forest wildfire risk, Ding et al. [42] trained and tested a fully connected CNN using over 5000 images from Himawari-8 satellites to identify wildfire locations and intensities. The proposed CNN achieved a detection accuracy exceeding 80%, which significantly outperformed other machine learning algorithms. Pais et al. [44] developed a CNN-based model for wildfire risk assessment using land cover raster data, leveraging interactions, continuity, and frequency among land cover types to identify ignition risks. The model demonstrated optimal performance in high-risk zones (accuracy = 100%), followed by moderate-risk zone (93%) and low-risk zone (91%) categories. Yu et al. [45] enhanced feature representation via a coordinate attention (CA)-based continuous CNN (CCNN), integrating positional information into network channels. To reduce annotation effort, active learning selected high-confidence samples for training, achieving 91.7% accuracy and AUC = 0.9487. Kanwal et al. [46] combined CNNs for feature extraction with machine learning techniques, optimizing a Random Forest model to attain 96.91% accuracy. Zhang et al. [47] proposed a global wildfire susceptibility model using CNN-2D and multilayer perceptron (MLP) architectures trained on seasonal data (2003–2016). The context-aware CNN-2D, leveraging neighborhood information, achieved superior precision. Sankaran et al. [152] implemented a deep learning-based framework with adaptive linear internal embedding (ALIE-FE) for feature extraction and a bio-inspired squirrel search optimization algorithm (BI-SSOA) for fitness optimization, attaining 98.99% classification accuracy. To enhance operational efficiency, Lelis et al. [153] designed a deep learning-driven computational system for wildfire risk assessment using unmanned aerial vehicle (UAV)-collected data. Integrating multi-sensor spatiotemporal data (temperature, humidity, vegetation), the system enables high-resolution risk estimation and proactive mission planning through autonomous drone missions. Pham et al. [154] assessed forest fire susceptibility by integrating remote sensing data from UAVs with SPOT satellite imagery, employing a hybrid model of the Artificial Bee Colony Adaptive Neuro-Fuzzy Inference System (ABC-ANFIS). The results indicated that a majority of the forests in the region were exposed to high to extremely high fire risks.

To enhance the temporal resolution of wildfire prediction, Zou et al. [48] proposed an attention-based deep learning modeling framework capable of generating 8-day temporal resolution wildfire risk forecasts. Prapas et al. [49] established an open-access global database containing seasonal and sub-seasonal wildfire drivers (climate, vegetation, oceanic indices, anthropogenic variables) alongside historical burned area and wildfire emission data (2001–2021). A deep learning model trained on this dataset treated global wildfire prediction as an image segmentation task, achieving high accuracy in forecasting burn area presence at lead times of 8 days (AUC = 97.54%), 16 days (97.56%), 32 days (97.81%), and 64 days (97.82%). John Ray Bergado et al. [155] designed AllConvNet, a deep fully convolutional network generating daily wildfire burn probability maps for 7-day forecasts. Son et al. [156] developed a super-resolution Convolutional Neural Network (SISR)-based wildfire risk prediction model by integrating the Climate Forecast System Version 2 (CFSv2) with single-image super-resolution deep learning. Evaluated across the 2018–2019 fire seasons, the CFS-SR model improved fire weather prediction accuracy with a 7-day lead time and enhanced spatial resolution to 4 km. Casallas et al. [50] implemented an early warning system (EWS) combining a 3D Convolutional Neural Network for satellite data analysis and a convolutional network for bias correction of Weather Research and Forecasting model outputs. Daily fire weather indices were quantified and fused with land cover classifications via a Naïve-Bayes classifier to estimate wildfire probabilities, enabling the construction of daily risk maps for targeted regions.

3.3.2. Forest Fire Prediction Based on Convolutional Neural Networks and Recurrent Neural Networks

Climate change is exacerbating wildfire hazards, posing significant threats to human livelihoods. Deep learning (DL) has been widely adopted in wildfire prediction research; however, conventional DL approaches often neglect intrinsic distinctions between static locational elements and dynamic variables while failing to integrate global teleconnections and Earth system characteristics during model training. To address these limitations, Üstek et al. [157] leveraged unsupervised learning techniques for wildfire prediction without ground truth data. Their framework employed two unsupervised strategies: (1) deep autoencoders to derive latent features for clustering, and (2) fully convolutional (FC) and Long Short-Term Memory (LSTM) autoencoders for input reconstruction. Experimental results demonstrated moderate performance (accuracy = 0.71, F1 score = 0.74, MCC = 0.42). Ji et al. [51] proposed a Static Location-Aware ConvLSTM (SLA-ConvLSTM) model that synergizes static geographic attributes with global climatic teleconnections. Framing wildfire prediction as an image segmentation task, the model quantifies location-specific dynamic variable impacts, with results highlighting enhanced predictive capability through spatiotemporal feature interaction and static location encoding. Prapas et al. [158] evaluated multiple DL models capturing spatial, temporal, or spatiotemporal contexts against a Random Forest baseline. ConvLSTM leveraging spatiotemporal dependencies achieved superior performance (AUC = 0.926), enabling nationwide daily fire danger mapping at higher spatial resolutions than operational solutions. Most DL-based wildfire models focus on spatial feature extraction. Deng et al. [52] bridged this gap by integrating 3D-CNN with ConvLSTM for multi-source spatiotemporal feature fusion. Through multicollinearity and weight analysis, redundant predictors were pruned. Unlike prior methods, daily weather forecasts were incorporated as inputs, refining temporal resolution from annual/seasonal to daily scales. The hybrid model achieved robust performance (AUC = 0.901, accuracy = 0.912), demonstrating the efficacy of high-resolution temporal modeling in wildfire prediction.

To enhance the efficiency of wildfire prediction, He et al. [55] optimized the ConvLSTM wildfire prediction model by incorporating channel and spatial attention mechanisms along with vision transformer patterns. They utilized a multi-source dataset comprising satellite-monitored wildfire products from 2012 to 2022, various factors related to terrestrial and human activities, simulated meteorological elements, and high-resolution vegetation images. Experimental results demonstrated accuracy, Kappa coefficient, and AUC values of 92.79%, 84.48%, and 97.90%, respectively, based on the ROC curve. Ramayanti et al. [56] generated wildfire susceptibility maps using Frequency Ratio, CNN, and LSTM-based deep learning. Experimental outcomes indicated that CNN slightly outperformed FR and LSTM, with AUC values of 0.879, 0.877, and 0.870, respectively. Bhowmik et al. [53] developed a multimodal wildfire prediction and early warning system based on a novel spatiotemporal machine learning architecture. They created a comprehensive wildfire database containing over 37 million data points, including historical wildfire data, environmental and meteorological sensor data from the Environmental Protection Agency, and geological data. The data were augmented into 2.53 km × 2.53 km square grids to overcome the limitations of sensor network coverage. A new U-Convolutional Long Short-Term Memory (ULSTM) neural network was developed to extract key spatial and temporal features from the dataset, particularly addressing the spatial nature of wildfire locations and the temporal progression of wildfire evolution. Experiments showed that the final ULSTM network architecture, trained using data from 2012 to 2017, achieved a wildfire prediction accuracy of >97% in 2018.

Fire size prediction models can assist fire rescue personnel in adopting appropriate measures based on early predictions of fire scale, thereby minimizing the losses caused by fires. Zhang et al. [54] proposed a novel hybrid Deep Neural Network, CNN-2D-LSTM, which employs a Deep Neural Network (DNN) approach to establish a dynamic prediction model for global wildfire burned area. This model integrates LSTM with 2D-CNN. Using this model, monthly spatiotemporal prediction maps of global fire areas were generated. The CNN-2D-LSTM model, based on convolutional–recurrent networks, can predict global burned areas one month ahead and can be extended to seasonal predictions of regional and global fire risks. To address the high computational resource demand of the Jules-Inferno model, Cheng et al. [159] constructed two data-driven models based on deep learning techniques to replace the Jules-Inferno model and accelerate global wildfire predictions. The first surrogate model uses CAE-LSTM, which applies Convolutional Autoencoders (CAEs) and LSTM to reduce data dimensionality and sequentially makes predictions in the reduced latent space. The second surrogate model employs a ConvLSTM network, which combines CNN and LSTM within a single network structure. Experiments demonstrated that the proposed models excelled in both computational efficiency and prediction accuracy (with an Absolute Error Percentage (AEP) less than 0.3% and a Structural Similarity Index Measure (SSIM) exceeding 98%, compared to the outputs of Jules-Inferno).

3.3.3. Forest Fire Prediction Based on Other Deep Learning Approaches

Liang et al. [160] proposed a model to predict the scale of forest fires in Alberta, Canada. This model uses meteorological factors as input values and employs Backpropagation Neural Network (BPNN), RNN, and LSTM to establish the prediction model. Among them, LSTM achieved the highest classification accuracy, reaching 90.9%. Seo et al. [161] developed FireDL, a model for predicting natural wildfires. In FireDL, the LSTM algorithm is utilized to predict the occurrence and duration of fires using datasets on lightning, vegetation, and climate. Experiments showed that FireDL performed strongly in predicting large fires (100,000 hectares), with a correlation coefficient of 0.72. Chen et al. [162] proposed a forest fire risk classification and prediction model by combining the Canadian Forest FWI system with LSTM networks and Random Forest models. The prediction accuracy of this method was 87.5%. Zhu et al. [163] introduced a novel network architecture that combines the SimAM and FANs networks. Additionally, to provide more unique information in the input data for the full attention network, SimFAN uses the SimAM module to redistribute the weights of the input data, enabling the full attention network to focus more on critical information in the input data. Experiments demonstrated that this model achieved an accuracy of 98.94%, capable of accurately predicting global wildfire risks. Hang et al. [164] developed an innovative artificial intelligence framework integrating Explainable Artificial Intelligence (XAI) techniques for forest fire susceptibility prediction in the Nainital District. The methodology synergistically combines multiple ensemble learners (AdaBoost, GBM, XGBoost, and Random Forest) with a DNN meta-learner within a stacked ensemble architecture. Experimental validation revealed exceptional predictive performance, achieving an AUC of 0.94, substantiating the framework’s capability to balance predictive power with interpretability in wildfire risk modeling. Mishra et al. [165] analyzed forest fire risks using the Mann–Kendall trend test and two machine learning approaches: MaxEnt and DNN. The forest fire vulnerability risks derived from both methods were categorized into four levels: very high, high, low, and very low. Although the burned areas obtained from both models were comparable, the DNN exhibited slightly superior performance compared to the MaxEnt model.

3.4. Forest Fire Prediction Dataset

The occurrence of forest fires is typically triggered by a combination of various factors. Common factors related to forest fires include meteorological conditions, terrain features, vegetation types, soil conditions, human activities, socio-economic factors, and so on. Therefore, the dataset for studying forest fire risk prediction is composed of forest fire data records and the influencing factors associated with forest fires. This chapter will describe the data types related to forest fire risk prediction and commonly used open-source data resources. Common data resources are presented in Appendix A.

3.4.1. Data Type

• Meteorological factors

Temperature [32,166]: High temperatures accelerate the evaporation of moisture from combustible materials (such as vegetation, dead branches, and leaves), reducing their water content and making them more susceptible to ignition. An increase in surface temperature may lead to spontaneous combustion (e.g., of piled-up coal or hay). Relative Humidity [30,128]: Low humidity (typically below 30%) significantly decreases the moisture content of vegetation and soil, creating a flammable environment. High humidity may inhibit combustion, but rapid drying after brief rainfall can instead increase the risk due to methane release from decaying grass. Wind Speed [29,33,167]: Strong winds accelerate the spread of fire, dispersing the fire source and expanding the burning area. They provide more oxygen, enhancing combustion intensity and forming flying embers (burning materials blown to distant locations that ignite new fires). Winds also change the direction of the fire field, increasing the difficulty of firefighting. Precipitation [117]: Short-term heavy rainfall can reduce the fire danger level, but thunderstorms may accompany lightning, triggering fires. After prolonged droughts, vegetation dries up, forming a large amount of combustible material. Scarcity of rain and snow in winter may lead to an increase in forest fires the following spring. Lightning [168]: Dry thunderstorms (thunderstorms without rainfall) are a significant contributor to forest fires, especially in remote and uninhabited areas.

2.

Vegetation factors

Vegetation Types [124]: Coniferous forests (such as pine and spruce), which contain resin and volatile oils, have a low ignition point and burn intensely (e.g., pine forest fires in the Mediterranean region). Deciduous forests (such as oak and maple), with higher leaf water content, are less flammable, but the accumulation of fallen leaves can increase the risk of surface fires. Herbaceous plants [66] (such as hay and sedges) are easily ignited during dry seasons and can spread fire rapidly (e.g., prairie fires in North America). Vegetation Moisture Content: Live vegetation is difficult to ignite when its moisture content exceeds 100%, whereas dead vegetation with a moisture content below 10% is highly combustible. During dry seasons, herbaceous plants rapidly dehydrate, easily accumulating combustible material to create a flammable environment. Fuel Load [31]: The greater the accumulation of combustible material per unit area of forest, the higher the fire risk and intensity.

3.

Topographical factors

Slope [169]: The slope primarily affects the spread of fire. When flames burn uphill, the heat radiation preheats the combustible materials above the slope, leading to an increased spread speed of the fire line along the slope. Aspect [170]: South-facing slopes receive more sunshine, causing vegetation to dry quickly and reducing the moisture content of combustibles. North-facing slopes have lower evaporation rates and often retain more shade-tolerant, moist vegetation (such as mosses and ferns), resulting in a lower combustion risk. Elevation [171,172]: In low-elevation areas (<1000 m), higher temperatures and lower humidity accelerate litter decomposition, with fires typically occurring frequently but on a small scale. In mid-elevation areas (1000–2500 m), coniferous forests are dense, with accumulations of dead branches, making them prone to high-intensity crown fires (e.g., fires in the Canadian Rockies). In high-elevation areas (>2500 m), low temperatures inhibit combustion, but exposed arid zones resulting from glacier retreat may create new fire hazard areas. River Sources [151,173]: Rivers serve as natural aquatic barriers effectively impeding the spread of forest fires. Areas near rivers usually have higher humidity, with higher moisture content in the air and soil, which helps reduce the drying of combustible materials and thus decreases the likelihood of fire occurrence.

4.

Soil Conditions

Soil Moisture [159,174]: Low soil moisture (<20%) reduces water uptake by plant roots, leading to rapid dehydration of surface vegetation and increased flammability. Deep soil drought (e.g., groundwater level decreases by more than 1 m) may cause deep-rooted plants to wilt, forming a large amount of dead fuel. Additionally, dry soil accelerates microbial decomposition, releasing CO₂ and reducing the water-holding capacity of humus, thus forming a dry and combustible layer. Soil Type [173]: Sandy soils have rapid drainage and poor water retention capacity, making vegetation prone to drought. They have low thermal conductivity, causing the surface layer to rapidly heat up during fires but accumulating less heat in deeper layers. Clay and loam soils have strong water-retention properties, delaying vegetation drying (e.g., lower fire risk in clay areas of tropical rainforests). After fires, they are prone to hardening, inhibiting plant regeneration and indirectly increasing the accumulation of combustible material on bare surfaces in the future. Organic Soils (Peat Soils): When organic matter content exceeds 30%, they can spontaneously combust during droughts and continue burning for months. The combustion depth can reach 3 m underground, requiring significant water injection or artificial rainfall for extinguishment.

5.

Human factors

Human-caused ignition sources include agricultural burning and land clearing, campfires, discarded cigarettes, and power transmission line failures. Human activities near roads are relatively frequent, increasing the probability of human-caused ignition sources and potentially triggering fires [31,66]. Other factors include monoculture plantations, the introduction of invasive species, and the overexploitation of water resources [39,147]. The completeness of fire prevention facilities (such as firebreaks and fire lanes) is also crucial. Additionally, fire monitoring and emergency response capabilities (such as lookout towers and drone patrols) play a significant role.

3.4.2. Technical Means

• Remote Sensing Satellite

Remote sensing has played a pivotal role in detecting and monitoring forest fires due to its ability to swiftly identify and respond to their occurrence. This technology enables the generation of maps delineating the extent and propagation patterns of fires, which is crucial for forest fire monitoring. Most satellites are equipped with the Multi-angle Imaging SpectroRadiometer (MISR). MIRS, a sensor primarily provided by the Earth Observing System (EOS) project of the National Aeronautics and Space Administration (NASA, USA), employs nine fixed cameras to observe the Earth’s surface from different angles. The most commonly used satellites for this purpose are MODIS Terra, VIIRS, Fengyun, Landsat 7/8, and Sentinel 2A/2B, with detailed descriptions provided in

Table 1. Common Satellite Descriptions [175,176,177].

Satellite	Spatial Coverage	Spatial Resolution	Revisit Time
MODIS	Global	500 m/1 km	24 h
AVHRR	Global	1.09 km	12 h
VIIRS	Global	375/750 m	12 h
Landsat-7/8/9	Global	30/60/100 m	8–16 days
Sentinel-1 A/B	Global	10 m	6 days
Sentinel-2 A/B	Global	10/20 m	3–5 days
Sentinel-3 A/B	Global	1 km	24 h
FengYun-3C	Global	1 km	6 days
MSG-SEVIRI	Regional	1 km/3 km	15 min
GOES-16 and 18	Regional	500 m/1 km/2 km	15 min
HJ-1B (WVC/IRMSS)	Regional	30 m/150–300 m	4 days
Himawari-8/9 AHI	Regional	0.5–1 km	10 min

. These satellites also aid in identifying burned areas, classifying burn severity, and estimating the total burned area. By conducting these measurements across numerous fires, the location, duration, size, temperature, and power output of the fires can be determined, data that were previously difficult to obtain. However, remote sensing satellites are significantly affected by environmental factors, with clouds and weather conditions posing significant challenges to fire detection.

2.

The Internet of Things Technology

IoT utilizes a vast network of sensors within the study area through Wireless Sensor Networks (WSNs) to monitor and record data related to forest fire occurrences, including temperature, humidity, soil conditions, and wind direction [117]. The data collected by these sensors can be used to predict the occurrence of fires and the potential hazards they may pose to human settlements and wildlife habitats. One of the primary advantages of WSN systems is their ability to operate autonomously without human intervention [178]. In recent years, the use of WSNs in fire detection and management has become increasingly important due to their capacity to collect accurate, real-time data to support decision-making processes during emergencies. However, the high deployment costs and limited durability of sensors restrict the practical application of IoT in large-scale forest fire monitoring, generally restricting its application to auxiliary small-area deployments. Currently, some researchers have combined unmanned aerial vehicles with IoT and WSNs for detecting forest fires instead of directly deploying sensors in the forest, thereby enhancing their practicality [179,180]. With sufficient funding, UAV-based IoT forest fire monitoring offers faster and more reliable forest fire detection compared to satellite remote sensing technology.

4. Discussion

4.1. Based on Statistical Analysis Methods and Physical Models

Statistical analysis models (such as regression analysis, probabilistic models, and fire danger indices) demonstrate robust data-driven capabilities and flexibility in forest fire prediction. Regression models (e.g., negative binomial regression [18], geographically weighted regression [19], and logistic regression [20,62,65]) effectively capture spatial heterogeneity and local characteristics by quantifying the nonlinear relationships between driving factors such as meteorology, topography, vegetation, and human activities and fire occurrences. For instance, geographically weighted techniques significantly enhance model prediction accuracy by introducing spatial weights (e.g., with an R² of 82% for the GWNBR model [19]), while Bayesian networks and extreme value theory models (e.g., BBN [21,22] and the Firelihood framework [78]) address data uncertainty and extreme events through a probabilistic framework, providing dynamic decision support for risk assessment. Fire danger indices (e.g., EFBI [81], MFDI [83], and FDEO [98]) achieve multi-scale dynamic assessments of fire danger by integrating real-time observational data on climate, soil moisture, and fuel conditions, particularly highlighting the correlation between soil moisture anomalies and fires in arid and humid regions. However, these methods rely on the quality and coverage of historical data, have limited predictive capabilities for unknown extreme events, and some models (e.g., logistic regression) still require manual feature engineering for modeling complex nonlinear relationships.

Physics-based models, such as climate models [25] and computational fluid dynamics models [23], provide breakthrough tools for studying the mechanisms of extreme wildfire events by simulating fire–atmosphere interactions and energy transfer processes. For example, bidirectionally coupled weather–fire models can dynamically simulate the feedback effects between fire spread and atmospheric circulation, revealing the formation conditions of extreme fire-related weather phenomena (such as PyroCb thunderstorms) [23,24]. Climate models, like the multi-year dynamic prediction system proposed by Chikamoto et al. [26], achieve wildfire probability predictions across seasonal to interannual scales by assimilating oceanic and atmospheric data and exhibit excellent performance in simulating low-frequency climate variability. However, these models are computationally intensive, rely on high-resolution meteorological data and accurate fuel parameterization, and have limited real-time predictive capabilities. Additionally, physical models are highly sensitive to initial conditions, and errors can accumulate during simulations, necessitating dynamic calibration using satellite remote sensing data (e.g., MODIS) and field observations.

The integration of statistical and physical models is a critical strategy for enhancing the accuracy of forest fire prediction. Statistical methods (such as regression and fire danger indices) excel at leveraging historical data and machine learning techniques to rapidly generate probabilistic predictions, making them suitable for short-term and local-scale fire risk assessments. In contrast, physical models (e.g., climate-coupled systems) reveal the long-term relationships between fire and climate from a mechanistic perspective, supporting scenario simulations of extreme events. For instance, the statistical–dynamical model proposed by Pan et al. [113] achieved high accuracy in seasonal fire prediction (with a skill score of 0.58) by combining dynamic predictions of meteorological variables with statistical optimization. Future research can further explore the following directions: (1) utilizing physical models to generate high-resolution meteorological fields to drive dynamic updates of statistical model parameters; (2) integrating multi-source observational data (such as soil moisture and vegetation water content) with model outputs through data assimilation techniques (e.g., Bayesian hierarchical models); and (3) developing hybrid frameworks (such as AI-augmented physical models) to enhance model generalization capabilities while reducing computational costs. Additionally, interdisciplinary collaboration (e.g., between ecology and climate science) will drive the evolution of fire prediction from single-factor analysis to multi-scale coupled systems, providing more reliable decision-making support for disaster prevention and control. Models and statistical method-based approaches exhibit good accuracy in predicting large-scale and long-term forest fire trends; we summarized the literature with clear and accurate prediction results in

Table 2. Statistical analysis methods.

Ref.	Methodology	Data Type	Study Area	Results
[20]	GIS-based logistic regression and Frequency Ratio models	MODIS hotspot data; environmental factors; climatic factors; anthropogenic factors	Western Bhutan	success rates = 88.3%
[19]	Geographic weighted regression model	Digital elevation model; slope; aspect; NDVI; NBR; LST; distance roads	Kalasin, Thailand	$R^2$ > 82%
[21]	Bayesian Belief Networks	Wildfire data; topographical factors; climate data; human factors;	Iran	ROC = 0.986
[181]	Dempster–Shafer-based evidential belief function and the multivariate logistic regression	Historical fire data; topography; climate; anthropogenic	Iran	AUC = 0.864
[182]	Statistical/probabilistic Weight of Evidence model and a knowledge-based Analytical Hierarchy Process	Slope, aspect, altitude, NDVI, annual rainfall, wind speed, land use, and proximity to rivers, roads, and human settlements	Huichang County, China	AUCsuccess rate = 0.94; AUCprediction rate = 0.91
[98]	Predictive model	Land cover type, vapor pressure deficit, surface soil moisture, and the Enhanced Vegetation Index, fire database	U.S.	Accuracy = 75%
[100]	A vegetation-type-specific fire severity classification	Antecedent drought conditions, fire weather, and topography	South-eastern Australia	$R^2$ = 0.89
[183]	Fuel hazard models based on environmental variables (environmental model)	Fire data; fuel data	South-eastern Australia	Accuracy = 41%–47%
[84]	Frequency Ratio and Analytic Hierarchy Process (AHP) techniques	Slope, aspect, curvature, elevation, NDVI, NDMI, TWI, rainfall, temperature, wind speed, TWI, and distance to settlements, rivers and roads	Western region of Syria	AUC = 0.864
[85]	Multi-Criteria Decision Analysis method	Slope, slope aspect, altitude, land cover, normalized difference vegetation index, annual rainfall, annual temperature, distance to settlements, and distance to road	Noshahr Forests (North Iran)	AUC = 0.783
[71]	Index of Entropy with Fuzzy Membership Value, Frequency Ratio, and Information Value	The spatial database incorporated the inventory of forest fire and conditioning factors	Nowshahr County in Iran	AUC = 0.890
[87]	Geographic Information System and the Analytic Hierarchy Process	Vegetation, morphology, climate, and proximity to human activities	Southern Italy	R2 = 0.79
[88]	Spatial multi-criteria decision-making, analytical network process	Topographic, climatologic, vegetation coverage and anthropological indicators	Northern area of Iran	Accuracy = 84%
[74]	Frequency Ratio, Shannon’s Entropy	Topography, climate, and human activities	Iran	success rate = 85.16%
[60]	Generalized Linear Model, binary logistic regression	Fire points and various independent variables	Chure region	AUC = 0.92
[89]	Statistical Analysis	Modified Normalized Difference Fire Index, Perpendicular Moisture Index	Uttarakhand Himalaya	Accuracy = 87.31%
[90]	Frequency Ratio, Analytic Hierarchy Process	Altitude, slope, aspect, topographic position index, NDVI, rainfall, air temperature, land surface temperature, wind speed, distance to settlements	India	Accuracy = 79.3%
[76]	Evidential belief function, Weight of Evidence	Slope percentage, slope direction, altitude, distance from rivers, distance from roads, distance from settlements, land use, slope curvature, rainfall, and maximum annual temperature	Iran	AUC = 0.896
[75]	Shannon’s Entropy, Frequency Ratio	NDVI, land use, Landsat data	Iran	AUC = 0.832

.

4.2. Based on Machine Learning

Advancements in modern computing technology have facilitated the creation of data-centric algorithms and methodologies capable of processing vast amounts of data to derive precise solutions for complex problems. Increasingly, machine learning approaches are being applied to research on forest fire prediction. Compared to traditional statistical analysis and physical models, ML techniques have emerged as core instruments for predicting fire risk and scale due to their abilities to integrate multi-source heterogeneous data, autonomously mine nonlinear relationships, and dynamically update data. By integrating meteorological factors (temperature, humidity, wind speed), surface characteristics (soil moisture, vegetation water content), human activities (population density, transportation networks), and real-time IoT monitoring data, ML models can construct high-dimensional feature spaces, significantly enhancing prediction accuracy and spatiotemporal resolution. Researchers have utilized fire indices such as the Forest FWI, the FFWI, NDVI, and NDMI, in conjunction with fire historical records and ML to predict forest fire risks [63,184,185,186,187]. To improve the accuracy of forest fire risk predictions, researchers have incorporated inputs such as temperature, relative humidity, total pressure, evaporation, soil moisture, total precipitation, wind speed, and vegetation data into ML models, resulting in enhanced accuracy for forest fire risk prediction using such methods [28,33,114,119,127]. Since most forest fires are closely related to human activities, some researchers have included human activity data in the inputs of ML models to improve the prediction accuracy of forest fire risks [66,124,125,169]. To obtain real-time environmental data of forests, Aljumah et al. [117] and Kaur et al. [137,188] proposed a forest fire risk prediction model based on the IoT architecture, which acquires real-time environmental data of forests through IoT systems to provide more precise data inputs for ML models, thereby enhancing the temporal resolution and accuracy of predictions.

Fire size prediction can assist forest protection personnel in adopting appropriate response measures based on the predicted scale from fire prediction models, enabling them to make preparations and minimize the losses caused by fires. To improve forest fire size prediction, Yazici et al. [189] proposed a prediction method that utilizes ANNs, Decision Tree Regression (DTR), Gaussian Process Regression (GPR), and Support Vector Regression (SVR) to predict the size of the burned area in forest fires. The research results indicate that the proposed method is reliable and can be used as a tool for predicting burned areas in different countries. Pérez-Sánchez et al. [30] employed ANNs to predict the size of forest fire burn areas. This model predicts burn areas for fires larger than 30,000 m², with a success rate higher than 60%–70%. Mohammadi et al. [63] combined MLR with nonlinear variable transformations to develop a forest wildfire burn area prediction model. This method enhances wildfire size prediction by incorporating wind speed variables. To improve the accuracy of forest burn size prediction, Han et al. [186] adopted Random Forest and ANN algorithms to develop a new model for predicting potential wildfire severity in the Upper Colorado River Basin (UCRB) in the United States. Experiments showed that the average prediction error of this model was 11.2%. To enhance the prediction area resolution, Wang et al. [132] proposed a wildfire prediction model based on XGBoost, which combines predictors of local meteorology, land surface characteristics, and socio-economic variables to predict monthly burned areas in adjacent regions of the United States at a grid cell resolution of 0.25° × 0.25°. Additionally, addressing the issue of missing data in partial observations, Cisneros et al. [190] proposed a statistical and machine learning-based method for spatial prediction of extreme wildfire frequency and size. This method can handle large datasets, including those with missing observations.

Table 3. Machine learning.

Ref.	Data Type	Study Area	Results
[30]	wildfire records; meteorological data; Fire Weather Indices	Southern Spain	Accuracy (60%~80%)
[117]	IoT sensors data (temperature, wind speed, humidity, and precipitation measure)	/	Accuracy = 96.78%
[116]	Normalized Difference Vegetation Index; Normalized Difference Moisture Index; Fire Weather Index	Greece	Accuracy = 75.8/76% (ignition/no ignition state)
[191]	Climate data; collect data	/	Accuracy = 86.0%
[124]	Wildfire Spot Data; meteorological data; NDVI data and vegetation type; topography, road network, and population;	Sichuan Province, China	AUC = 0.944, OA = 87.28%, TPR = 0.829, TS = 0.723
[126]	Landsat 8 data with level-2A; MODIS data; Shuttle Radar Topography data; weather data	Korcula island	Accuracy = 90.6%
[33]	Wildfire data; land cover data; climate data; vegetation cover; topography data	The Mongolian Plateau	Accuracy > 90%
[128]	Recording of historical fire; topographic, vegetation, climatic, and anthropogenic factors;	China	AUC > 0.95
[29]	Historical wildfires; remote sensing data; geographical information systems data; climatological data	Honduras	AUC = 0.87
[167]	Wildfires data; land cover maps; Open Street Map data; climate data; NDVI	Italian peninsula in the Mediterranean	AUC = 81.3%
[125]	Topographic; land cover; climatic; social; historical wildfires	Irkutsk Oblast	Accuracy = 0.89; AUC = 0.96
[132]	Wildfire data; predictor variables	US	IoA = 0.71
[32]	Vegetation factors, human factors, surface temperature, terrain factors, and meteorological factors; Fire Point Information	Yunnan Province, China	Accuracy = 0.82; AUC = 0.83
[31]	Meteorology; human activity; topography; fuel; geography; historical wildfire data	California, USA	Balanced Accuracy = 73.56%
[129]	Historical wildfire dataset; susceptibility conditioning factors; vulnerability factors	China	Accuracy = 0.932
[130]	Historical wildfire dataset; nonseasonal factors; seasonal factors; ecological vulnerability factors	Nanning, China	Precision = spring (0.982) > autumn (0.972) > summer (0.971) > winter (0.952) > (0.924)
[35]	Wildfire data; topographic; meteorological; anthropogenic; environmental	Kaua’i and Moloka’i Islands, Hawaii	AUC = 0.9314
[34]	Wildfire historical locations; predictor variables	Northern Iran	Accuracy = 0.99
[123]	Historical fire locations; explanatory variables	Iran	AUCsuccess rate = 0.92; AUCprediction rate = 0.91
[66]	Climate, vegetation, topographical, human activities, and location; forest burn scar data	Yunnan Province, China	AUC = 0.882~0.890
[63]	Burned area data; climate data; water vapor pressure	Central Bohemian Region	R2 = 0.761
[36]	NDVI data; exploration of predictors; wildfire data	Australian	Accuracy = 0.969
[137]	IoT sensors data; digital elevation maps	/	Accuracy = 93.97%
[136]	Wildfire data; independent variables	Spain	AUC > 0.8
[192]	Fire inventory map; explanatory variables;	Vietnam	AUC = 0.96
[169]	Meteorological factors; topographical factors; ecological factors; in situ factors; anthropogenic factors; wildfire data	Sikkim Himalaya	Accuracy = 0.914
[193]	Meteorological data; environmental data; burn area data	Southwestern Spain	G-means = 0.72~0.76
[194]	Topography; environment; climate; social economic; wildfire information	Australia	Accuracy = 96%
[170]	Forest fire data; topographical factors; climatic factors; other factors	Southern Iran	AUC = 88.2%
[185]	Forest fire dataset; climatic conditions; burned area; FWI	Northwest Portugal	AUC = 0.8052
[186]	Wildfire information; weather conditions; soil and land properties	Southwestern United States	Accuracy = 72%
[195]	Meteorological factors; fuel characteristics; topography; anthropogenic factors; regional fire history	Global	Accuracy = 0.88
[196]	Fire Weather Index system; Forest Fire Danger Index; wildfire features; weather features; LFMC features; social features	Western Australia	Accuracy = 99%
[197]	Fire dataset; meteorological datasets; ocean climate indices;	Global	IOA = 0.82, 0.82, 0.8, 0.75, 0.56
[198]	Inventory dataset; hazard-related factors; vulnerability-related factors	Queensland Australia	Prediction rates = 89.21%
[187]	Historical wildfire dataset; susceptibility conditioning factors; ecological and urban vulnerability factors	China	Accuracy = 0.863
[171]	Wildfire data; meteorological factors; human factors; topographical and vegetation factors; socio-economic	Southwest China	Accuracy = 95.0%
[172]	Fire history; meteorological factors; topographical and vegetation factors; socio-economic	California	F1-Score = 0.689
[199]	Wildfire data; topographical, meteorological/hydrological, vegetation, and anthropological factors	Northern Iran	AUC = 94.71%
[200]	Meteorologically related variables; vegetation-related variables; landform-related variables; wildfire data	Southern China	Accuracy = 94.23%
[201]	Historical fire data; topographic variables; weather variables; fuel variables	Great Xing’an Mountain Region, China	F1-Score = 0.923
[166]	Wildfire data; vegetation data; land surface temperature; thermal anomalies	Canada	Accuracy = 98.32%
[202]	Climate; topography; anthropogenic; vegetation; forest fires data	Northeast China	Accuracy = 85.2%
[203]	Historical fire location; conditioning factors	Algeria	Accuracy = 0.867
[173]	Topographic, weather, soil type, land use, and proximity to settlements, roads, and rivers;	Iran	Accuracy = 99.8%
[204]	Wildfire inventory data; conditioning factors	Iran	Accuracy = 88%
[205]	Meteorological factor; wildfire data; topographic factor; vapor pressure deficit	Israel	Accuracy = 0.67%
[138]	Fire history; weather; vegetation; powerline; terrain	Northern California	Accuracy = 92%
[148]	Wildfire occurrence dataset; dynamic step weather variables; fuel variables; topography and infrastructure variables	Sichuan Province, China	AUC = 0.98
[149]	Burned area data; climate; topography; vegetation; anthropogenic factors	China–Mongolia–Russia Cross-Border Area	Prediction Rate = 0.835
[39]	Historical fire; meteorological data; environmental data; human factors	Brisbane catchment, Australia	Accuracy = 88.51%
[150]	Physiography; meteorology; anthropology; wildfire data	South of China	Accuracy = 81.6%
[151]	Fire history; topographical factors; environmental data; socio-economic factors	Iran	AUC > 0.8
[40]	Fire information; topography; anthropogenic activities; climate; vegetation	North of Morocco	AUC = 0.989
[41]	Wildfire data; methodology	Hawai’i	AUC = 0.9269
[145]	Historical Fire; topographical; meteorological; environmental; anthropological	Australia	AUC = 0.882
[146]	Lightning–ignition database; environmental predictors	Australia	Accuracy = 78%
[147]	Terrain; fuel condition; human accessibility; weather; ignition seasonality	Montana, US	AUC = 0.84/0.89 (Natural/human-caused ignitions)
[139]	279 forest fire locations; Sentinel 2A satellite images; NASA fire archives; and field visits; fire-influencing factors	Southern Western Ghats, India	AUC = 0.890
[140]	109 fire locations; 14 relevant factors	Jerash Province, Jordan	AUC = 0.965
[141]	Historic fire locations; topological, climatic, and socio-economic data	Dak Nong, Vietnam	AUC = 0.9515
[120]	Topographic variables; land cover classifications; ecoregion delineations	Santa Cruz, Bolivia	AUC = 0.8
[121]	Himawari-8 geostationary satellite data	South Korea	Success rate = 93%
[143]	108 historical forest fire events; climatic; vegetation variables	Iran	AUC = 0.85
[115]	Anthropogenic; topographic; vegetation; hydrological factors	Türkiye	AUC = 0.94
[144]	Forest fire locations, fourteen factors	Jordan	AUC = 0.97
[135]	Temperature, wind, precipitation, elevation, slope, aspect, population density	Pakistan	AUC = 0.833
[118]	IoT sensors data; historical fire	/	Accuracy = 90%

presents descriptions of machine learning-based forest fire risk prediction schemes with accurate prediction results.

4.3. Based on Deep Learning

Forest fire risk prediction models based on deep learning can incorporate multi-source heterogeneous data as input, possessing robust data integration processing capabilities and achieving higher prediction accuracy. When training deep learning models with relatively simple data, the performance of the models tends to be poorer. For instance, Ding et al. [42] used over 5000 Himawari-8 satellite images to train a model based on Convolutional Neural Networks, achieving an accuracy rate of approximately 80%; Üstek et al. [157] constructed a model using unsupervised learning techniques without considering actual forest ground conditions. This model can reduce the number of input parameters but has a relatively low accuracy rate of 71%. To improve prediction accuracy, some researchers have enhanced the predictive power of their models by increasing the dimensionality of input data. For example, the accuracy of wildfire risk prediction was improved by incorporating land cover grids [44]; seasonal classification information was added to enhance the accuracy of wildfire prediction models [47]; and by using various variables related to seasonal and sub-seasonal fire drivers, along with historical burned area and wildfire emission data for model training, one model achieved an accuracy rate of up to 97.54% when predicting eight days in advance [49]. Additionally, some researchers have improved model accuracy by refining training methods. Sankaran et al. [152] employed feature extraction based on the adaptive linear internal embedding algorithm (ALIE-FE) and selected the extracted features through Recursive Wrapper-based feature subset selection. To estimate the optimal fitness function and improve prediction accuracy, the BI-SSOA was used for optimization. Liang et al. [160] combined BPNN, Recurrent Neural Networks, and LSTM to establish a prediction model for improved predictive performance. Chen et al. [162] integrated the Canadian Forest FWI system with LSTM networks and Random Forest models to build a prediction model aimed at enhancing predictive effectiveness. Zhu et al. [163] proposed a new network by combining SimAM and FANs, achieving a prediction accuracy rate of 98.94% for their model.

To enhance the timeliness and practicality of forest fire risk predictions, models proposed in the literature [48,49,155,156] offer forest fire risk predictions with the best temporal resolution of 7–8 days. To improve the temporal resolution of forest fire prediction models, researchers have combined CNNs with LSTM to enhance both the accuracy and temporal resolution of the models [52,158]. LSTM has the advantage of capturing long-term temporal dependencies between meteorological data (such as temperature, humidity, and wind speed) and fire risk factors, which can improve the model’s prediction temporal resolution. When combined with CNNs, the overall effect is very promising. Fire size prediction can assist fire rescue personnel in making early preparations and adopting appropriate response measures based on the predicted size of the fire in its early stages, thereby minimizing the losses caused by the fire. To improve forest fire size prediction, Zhang et al. [54] proposed a dynamic prediction model for global wildfire burned area based on a new hybrid Deep Neural Network (CNN-2D-LSTM). This model was used to generate monthly spatiotemporal prediction maps of global fire areas. The CNN-2D-LSTM model, based on convolutional–recurrent networks, can predict global burned areas one month in advance and can be extended to seasonal predictions of regional and global fire risks. Compared to other forest fire prediction methods, deep learning-based models exhibit higher accuracy and temporal resolution. However, deep learning models require substantial computational resources and storage capacity, and their interpretability remains limited. Researchers have utilized the SHAP model [130] to interpret the results of deep learning-based forest risk prediction models.

Table 4. Deep learning.

Ref.	Dataset	Study Area	Results
[43]	Historical wildfire data; topographical data; meteorological data	Global	Precision = 83.71%
[45]	Historical wildfire data; topographical data; meteorological factors	Chongqing in China	Accuracy = 91.7%, AUC = 94.87%
[46]	Historical wildfire data; topography; weather; fuel	South of Pakistan	Accuracy = 98.71%
[47]	Building binary ignition maps (BIMs); building explanatory variable maps (EVMs)	Global	Precision = 89.18% (MAM), 91.97% (JJA), 91.50% (SON), 95.19% (DJF)
[155]	Fire history records dataset; topographical databases; land cover/use; MODIS images; lightning imaging sensor data; meteorological data	Victoria is the southeastern state of Australia	Accuracy = 78.57%
[48]	Observational fire data; terrain data; fuel data; weather data	Anderson Forest	PR-AUC = 98%; PR-AUC = 24%
[42]	Himawari-8 satellite	Australian	Accuracy > 80%
[44]	Wildfire data; land cover data	South-Central Chile	accuracyhigh-risk = 100%; accuracymedium-risk = 93%; accuracylow-risk = 91%
[49]	Climate; vegetation; land cover; human presence; wildfire emissions	Global	AUC8 = 0.9754; AUC16 = 0.9756; AUC32 = 0.9781; AUC64 = 0.9782;
[152]	FIRMS; aspect, elevation, and slope; climate	Australia	Accuracy = 98.99%
[206]	historical data on forest fires; meteorological data; multispectral and SAR satellite imagery	Europe	F1 score = 0.87; IoU = 0.77;
[51]	Local/global variables; static variables; climatic indices; burned areas	Global	Accuracy = 95.8%
[158]	Daily weather data; satellite variables; roads density; population density; land cover; topography variables; historical burned areas; EFFIS burned areas	Greece	AUC = 0.926
[52]	Ignition; vegetation/remote sensing; topography; climatic; human factor	Yunnan Province, China	AUC = 0.901; Accuracy = 0.912
[55]	Historical hotspots; terrestrial information; terrestrial; anthropogenic; topographical	The six eastern provinces of China	Accuracy = 92.79%
[56]	Satellite data; wildfire inventory map; topographical; meteorological; land use; NDVI, NBR; anthropological	Maui Island, Hawaii	AUC = 0.879
[53]	Historical wildfire data; vegetation data; geological data; environmental and meteorological data	California	Accuracy = 97.1%
[54]	Meteorology datasets; biophysical predictors; burned area data set	Global	IOA = 0.9
[159]	Global temperature; vegetation density; soil moisture; previous forecasts	Global	AEP < 0.3%; SSIM > 98%
[157]	Historical weather; vegetation index; wildfires dataset	Global	Accuracy = 0.71
[174]	Daily weather data; satellite variables; soil moisture index; human activity; historical burned areas; elevation and slope	Greece	Precision = 92.3% (2020), 95% (2021)
[160]	Fire occurrence records; weather data	Alberta, Canada	Accuracy = 90.9%
[162]	Fire occurrence record; Fire Weather Index	Northeastern China	Accuracy = 87.5%
[207]	Fire historical records; meteorological, topographic, and landcover/vegetation factors	Victoria, Australia	Accuracy = 93.38%
[208]	Fuel conditions; climate factors; ignition; anthropogenic suppression	Global	Accuracy > 90%
[163]	Climate; vegetation; ocean indices; human activity variables; fire historical records	Global	Accuracy = 98.94%
[164]	Annual rainfall; evapotranspiration; distance from roads across	Western Himalaya	AUC = 0.94

presents descriptions of deep learning-based forest fire risk prediction models with accurate prediction results.

5. Conclusions and Future Works

5.1. Conclusions

In recent years, the increasingly significant impacts of climate change and human activities on the environment have led to more frequent occurrences of extreme events such as forest fires. The recurrent wildfires pose severe threats to ecological systems and human life safety. Consequently, forest fire prediction has become a critical research focus, where accurate forecasting technologies are essential for mitigating ecological and economic losses, improving wildfire management efficiency, and ensuring personnel safety and property security. To advance comprehensive understanding of wildfire prediction research, this paper systematically reviews the literature from the past decade on forest fire prediction. These studies typically utilize satellite remote sensing data, climate data, fire risk indices, vegetation data, human activity factors, topographic data, and road information, applying statistical analysis, physical models, and artificial intelligence techniques for wildfire prediction. This review emphasizes two key aspects: datasets with related tools and wildfire prediction algorithms. The selected literature was classified into three categories for analysis: statistical analysis and physical models, traditional machine learning methods, and deep learning approaches.

Additionally, this paper summarizes the data types and open-source datasets employed in the reviewed studies. Statistical methods exhibit simplicity, direct applicability, and reasonable accuracy. But deep learning and machine learning models outperform traditional approaches and achieve more precise wildfire predictions by leveraging multi-source data (including vegetation, topography, weather, infrastructure, climate, historical fire records, and socio-economic factors). However, deep learning models show limitations in interpreting correlations among fire risk factors, despite their high predictive accuracy, due to inherent interpretability challenges. While machine learning offers better interpretability, its insufficient capacity to uncover deep data relationships results in lower prediction accuracy compared to deep learning.

5.2. Future Works

Future research should focus on the following directions:

(1).

Enhancing model resolution and dynamic updates: Utilize physical models to generate high-resolution meteorological fields, driving the dynamic parameter updates of statistical models, thereby improving AI-based prediction accuracy through high-resolution data.

(2).

Multi-source data fusion: Integrate diverse observational data (e.g., soil moisture, vegetation water content) with model outputs using data assimilation techniques such as Bayesian hierarchical models.

(3).

Hybrid framework development: Design hybrid frameworks (e.g., AI-enhanced physical models) to reduce computational costs while enhancing model generalizability.

(4).

Interdisciplinary collaboration: Foster cross-disciplinary cooperation (e.g., between ecology and climate science) to advance fire prediction from single-factor analyses to multi-scale coupled systems, providing more reliable decision-making support for disaster prevention.

Although digital twins are well established in industrial applications, their potential in forest fire research remains largely unexplored. Digital twin technology, capable of simulating forest dynamics by integrating multi-source heterogeneous data, holds great potential for the precise prediction of fire ignition, spread, and behavior. This represents a promising frontier for future investigation.