Satellite-Based Wildfire Monitoring Across the Fire Lifecycle Using Artificial Intelligence: Trends, Challenges, and Future Directions

Abstract

Wildfires are complex disasters that extend beyond vegetation combustion, producing cascading impacts on the atmosphere, soils, hydrology, ecosystems, and human society. Their risks and impacts are being amplified by climate change, highlighting the need to understand and monitor wildfires not as isolated events but as full-cycle phenomena that encompass pre-fire, active-fire, and post-fire stages. This study systematically reviews wildfire research published over the past decade that has used satellite remote sensing and artificial intelligence (AI) and synthesizes trends in data use and methodology from a lifecycle perspective. Relevant literature was retrieved from the Scopus database using search terms related to wildfire, satellite observations, and AI. The selected studies were classified into pre-fire, active-fire, and post-fire stages. They were compared in terms of study region, input data, reference data, input representation, and algorithmic approach. The results show that relevant studies have increased rapidly in recent years, with clear stage-specific differences in research objectives. Pre-fire studies mainly focus on wildfire occurrence prediction and risk assessment. Active-fire studies emphasize active fire and smoke detection, as well as the monitoring of fire fronts and spread dynamics. Post-fire studies concentrate on burned-area mapping, burn severity assessment, damage evaluation, and vegetation recovery analysis. Geographically, the literature is strongly concentrated in North America and Asia, with a substantial proportion of studies focused on a small number of countries, including the United States, China, Australia, Canada, Brazil, and South Korea. Although single-country studies still dominate, multi-regional and global-scale studies have increased in recent years. In terms of data use, optical satellite imagery remains the most widely used source, while the integration of thermal infrared, synthetic aperture radar, meteorological, topographic, and land-cover data has become increasingly common. Methodologically, machine learning and deep learning approaches increasingly supplement or replace conventional index-based and rule-based methods, with model architectures and input representations differing by wildfire stage and research objective. Nevertheless, regional bias in training data, data imbalance, limited model interpretability, and restricted cross-regional generalizability remain major challenges. Future wildfire monitoring research should therefore move toward integrated lifecycle frameworks that combine multi-source data fusion, regional scalability, physically informed learning, and interpretable AI.

1. Introduction

Wildfires are not limited to vegetation combustion; they are complex disasters that generate cascading impacts on the atmosphere, soils, hydrology, ecosystems, and human society (Bowman et al., 2009; Chuvieco et al., 2019; Pausas & Keeley, 2021; Potter et al., 2023). In recent years, wildfires have become an increasingly important environmental issue because climate change has intensified hot and dry conditions and increased the occurrence of extreme fire weather (Jolly et al., 2015; Abatzoglou & Williams, 2016; Senande-Rivera et al., 2022; Seneviratne et al., 2021). The Intergovernmental Panel on Climate Change (IPCC) has assessed that, in some regions, the likelihood of fire weather characterized by high temperature, dry conditions, and strong winds has already increased, and that such conditions are expected to occur more frequently as warming intensifies (Seneviratne et al., 2021). Recent research has also shown that climate change is an increasingly important driver of regional burned-area variability (Abatzoglou & Williams, 2016; Senande-Rivera et al., 2022; Cunningham et al., 2024). Wildfires can therefore no longer be treated merely as temporary and localized events. Instead, they should be regarded as long-term management challenges that are closely linked to the climate crisis (Bowman et al., 2017; Moritz et al., 2014; Pausas & Keeley, 2021; Intergovernmental Panel on Climate Change, 2023).

Wildfires should also be understood as part of a full-cycle mechanism that includes risk accumulation before ignition, detection and spread immediately after ignition, and damage assessment and recovery after suppression (Moritz et al., 2014; Alcasena et al., 2019; Jain et al., 2020; Wooster et al., 2021). Wildfire research and operational response are commonly divided into three stages: pre-fire, active-fire, and post-fire, each of which requires distinct observational objectives (Fig. 1; Moritz et al., 2014; Chuvieco et al., 2020; Wooster et al., 2021). In the pre-fire stage, ignition probability and wildfire risk must be assessed using information on fuel conditions, weather, topography, and drought status (Jolly et al., 2015; Sakr et al., 2010; Massetti et al., 2019; Alcasena et al., 2019). During the active-fire stage, the key tasks include rapid detection of ignition points, monitoring of fire movement, tracking of smoke and thermal signals, and observation of spread patterns (Matson & Holben, 1987; Giglio et al., 2003; Schroeder et al., 2014; Wooster et al., 2021; de Almeida Pereira et al., 2021). In the post-fire stage, it becomes important to quantify subsequent impacts, including burned area, burn severity, ecosystem disturbance, and vegetation recovery (Cocke et al., 2005; Key & Benson, 2006; Allen & Sorbel, 2008; Roy et al., 2019; Gibson et al., 2020; Quintano et al., 2020). An integrated understanding of wildfire, therefore, requires not isolated observations at a single time point but a monitoring system in which different forms of information are continuously collected, linked, and interpreted across all stages (Moritz et al., 2014; Alcasena et al., 2019; Jain et al., 2020; Wooster et al., 2021).

Fig. 1. Overview of wildfire monitoring across the fire lifecycle, including pre-fire, active-fire, and post-fire stages, along with their primary objectives and representative analysis targets.

Wildfire monitoring has traditionally relied on a variety of observation and assessment methods. Wildfire observers and watchtowers have long been used to visually identify early smoke or flames across large, forested areas (Alkhatib, 2014; San-Miguel-Ayanz et al., 2012). More recently, this approach has been supplemented by mountain surveillance cameras and video systems (Alkhatib, 2014; Park et al., 2020; Gaur et al., 2020; Geetha et al., 2021). In addition, the Composite Burn Index (CBI) has been widely used for post-fire field assessment. CBI is a representative field-based indicator for evaluating burn severity through direct on-site surveys of multiple strata, including soil, fuel, herbaceous vegetation, shrubs, and trees (Key & Benson, 1999, 2006; Cocke et al., 2005; Allen & Sorbel, 2008). However, these approaches share several limitations. Observer- and watchtower-based monitoring requires continuous human involvement and is strongly constrained by topography and line-of-sight conditions (Alkhatib, 2014; San-Miguel-Ayanz et al., 2012). Camera-based monitoring often detects fires only after smoke becomes visible and is affected by terrain obstruction, limited observation range, and weather or lighting conditions (Alkhatib, 2014; Park et al., 2020; Gaur et al., 2020; Geetha et al., 2021). Field-based assessment methods, such as CBI, provide detailed information but are difficult to apply repeatedly and consistently across large areas (Key & Benson, 2006; Allen & Sorbel, 2008; Roy et al., 2019). They also require substantial time and cost when rapid assessment of widespread damage is needed after large wildfires (Allen & Sorbel, 2008; Roy et al., 2019; Gibson et al., 2020; Chuvieco et al., 2020).

For these reasons, satellite-based observation has become increasingly important for full-cycle wildfire monitoring. Satellite data can repeatedly observe large areas, monitor mountainous and remote regions that are difficult to access, and provide spatially and temporally consistent observations through standardized sensors and processing frameworks (Matson & Holben, 1987; Giglio et al., 2018; Roy et al., 2019; Lizundia-Loiola et al., 2020; San-Miguel-Ayanz et al., 2012). These advantages make it possible to link pre-fire risk assessment, active-fire and smoke detection, post-fire burned area and burn severity mapping, and recovery monitoring within a single observational framework (Giglio et al., 2003; Giglio et al., 2018; Roy et al., 2019; Lizundia-Loiola et al., 2020; Chuvieco et al., 2020; Jain et al., 2020). Satellite observations are subject to limitations related to clouds, smoke, spatiotemporal resolution, and sensor characteristics (Schroeder et al., 2014; Giglio et al., 2018; Roy et al., 2019; Zhang et al., 2024). Nevertheless, they remain a fundamental basis for wildfire monitoring because they provide broad coverage, reproducibility, and quantitative observations that are difficult to obtain through field observation or camera-based monitoring alone (Matson & Holben, 1987; Bowman et al., 2009; Chuvieco et al., 2019; San-Miguel-Ayanz et al., 2012; Jain et al., 2020). To understand wildfires across their full lifecycle and analyze them in a comparable manner, it is therefore essential to examine an integrated monitoring framework centered on satellite remote sensing (Bowman et al., 2009; Roy et al., 2019; Chuvieco et al., 2020; Jain et al., 2020; Wooster et al., 2021).

Satellite-based wildfire monitoring has traditionally relied on physics-based, empirically derived, and rule-based approaches tailored to specific stages of the fire lifecycle (Perry, 1998; Stocks et al., 1989). In the pre-fire stage, satellite-derived indices have been widely used to characterize fuel conditions and assess spatial variability in wildfire risk. Vegetation indices, such as the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index, enable monitoring of vegetation greenness and biomass and serve as proxies for fuel availability (Huete et al., 2002; Tucker, 1979). Moisture-related indices, including the Normalized Difference Water Index (NDWI) and the Normalized Difference Moisture Index, provide proxy information on vegetation water content, which is closely linked to fuel moisture conditions (Dennison et al., 2005; Wilson & Sader, 2002). Owing to their physical interpretability and computational simplicity, these approaches have served as practical tools for large-scale risk assessment (Jolly et al., 2019). However, their capacity to capture fine-scale meteorological variability and complex terrain effects remains limited, constraining their applicability in heterogeneous environments (Coen, 2018; Srock et al., 2018).

During the active-fire stage, thermal anomaly detection using the mid-infrared (MIR) and thermal infrared (TIR) channels of satellite sensors plays a central role in identifying and monitoring fire activity (Csiszar et al., 2014; Kaufman et al., 1998; Robinson, 1991). Early approaches primarily relied on single- or multi-channel threshold-based techniques. These methods have since evolved into contextual active fire detection frameworks, in which the radiometric signal of a target pixel is compared with surrounding background pixels to improve detection robustness (Eva & Flasse, 1996; Flasse & Ceccato, 1996; Ichoku et al., 2003). Additional efforts have sought to quantify combustion intensity using Fire Radiative Power (FRP) and to analyze fire spread dynamics by integrating physical fire behavior models, such as the Rothermel model (Andrews, 2018; Wooster et al., 2005). Although these approaches enable near-real-time fire detection and monitoring, they remain limited by false detections caused by clouds, smoke, and sun glint, as well as by difficulties in detecting small or low-intensity fires because of the spatial resolution constraints of satellite sensors (Li et al., 2003; Wang et al., 2018).

In the post-fire stage, spectral index-based approaches that analyze changes in surface reflectance have become a standard framework for burn severity assessment (Brewer et al., 2005; Eidenshink et al., 2007). Among these, the differenced Normalized Burn Ratio (dNBR), which uses pre- and post-fire differences in near-infrared and shortwave infrared (SWIR) reflectance, has been widely adopted for quantifying vegetation loss and soil exposure (Cocke et al., 2005; Escuin et al., 2008; Key & Benson, 2006). However, conventional indices such as dNBR exhibit several limitations. Their sensitivity to pre-fire vegetation conditions and biomass can lead to inconsistencies across ecosystems. Variations in soil background and seasonal differences in image acquisition can also introduce uncertainty (Lutes et al., 2006). As a result, thresholds derived for a specific region are often difficult to generalize across different geographic and environmental conditions (Cansler & McKenzie, 2012; Hudak et al., 2007; Verbyla et al., 2008). To reduce dependence on pre-fire vegetation conditions, modified indices such as the Relative dNBR (RdNBR) and the Relativized Burn Ratio have been proposed (Miller & Thode, 2007; Parks et al., 2014). These indices normalize burn severity metrics with respect to pre-fire conditions, improving comparability across heterogeneous landscapes and extending their applicability to post-fire recovery monitoring (Parks et al., 2014; Soverel et al., 2010). Nevertheless, even these modified indices remain sensitive to phenological variability and acquisition timing, which limits their robustness under temporally inconsistent conditions.

Despite their long-standing utility, conventional approaches are inherently constrained by a limited number of input variables and predefined representations of physical processes, making it difficult to capture the complex interactions that govern wildfire dynamics fully. Accordingly, artificial intelligence (AI)-based approaches have been adopted rapidly in recent wildfire research to improve monitoring accuracy and efficiency (Jain et al., 2020). Machine learning and deep learning techniques can learn complex nonlinear relationships from large-scale datasets. In particular, the integration of multi-modal data, including satellite observations, meteorological variables, topography, land cover, and anthropogenic information, enables more comprehensive and data-driven analyses across the wildfire lifecycle (Ghali & Akhloufi, 2023; Zakari et al., 2025).

AI-based approaches can also reduce the subjectivity and regional dependency of conventional threshold-based methods by enabling automated feature extraction and data-driven parameter optimization (Abdollahi & Pradhan, 2023; Jain et al., 2020; Kondylatos et al., 2022). These capabilities can support more robust model performance under complex surface conditions and varying meteorological environments. Recent studies have applied a wide range of AI models to wildfire monitoring, including traditional machine learning techniques such as Random Forest (RF) and Support Vector Machine (SVM), as well as deep learning approaches. Convolutional Neural Networks (CNNs) have been widely used for learning spatial patterns, while Transformer-based architectures have increasingly been adopted to capture temporal dynamics and long-range dependencies (Ghali & Akhloufi, 2023; Kang et al., 2023; Singh et al., 2025).

Despite this growing body of research, existing studies often focus on specific stages of the wildfire lifecycle or on particular methodological approaches, resulting in a fragmented understanding of wildfire monitoring systems. Previous review papers have also tended to focus on individual wildfire stages or specific applications, such as wildfire risk prediction, active fire detection, smoke monitoring, burned area mapping, burn severity assessment, or post-fire recovery. Although these reviews have provided valuable insights into each component, relatively few studies have examined how satellite observations and AI methods differ across the full wildfire lifecycle within a single review framework. This stage-specific focus has limited a comprehensive understanding of how research targets, data sources, reference information, input representations, and model architectures are selected and adapted for different wildfire monitoring objectives.

To address this gap, this study provides a comprehensive review of AI-based satellite wildfire monitoring across the pre-fire, active-fire, and post-fire stages. This review synthesizes studies from the three major wildfire stages within a single paper and compares their stage-wise characteristics. Specifically, we organize the reviewed studies according to wildfire stage and comparatively analyze stage-wise differences in research targets, study regions, satellite sensors, auxiliary and reference data, input representations, and AI model architectures. Through this stage-wise synthesis, we identify major research trends, methodological limitations, and future research needs across the wildfire monitoring literature. In doing so, this study aims to contribute to the development of more accurate, scalable, and reliable wildfire monitoring and management systems.

2. Materials and Methods

The literature reviewed in this study was collected from Scopus using a search query designed to identify wildfire-related studies that incorporated satellite data and AI-based approaches. The search covered English-language journal articles published from 2016 to 2025. The query was applied to titles, abstracts, and keywords using the terms “wildfire,” “satellite,” and “artificial intelligence,” “machine learning,” or “deep learning.” The final Scopus search query was as follows: TITLE-ABS-KEY (“wildfire” AND “satellite” AND (“artificial intelligence” OR “machine learning” OR “deep learning”)) AND PUBYEAR > 2015 AND PUBYEAR <2026 AND LIMIT-TO(SRCTYPE,“j”) AND LIMIT-TO (DOCTYPE, “ar”) AND LIMIT-TO (LANGUAGE, “English”).

Studies were included when they met the following criteria: (1) they addressed wildfire-related monitoring, prediction, detection, mapping, or assessment; (2) they used satellite remote sensing data as a main or supporting data source; and (3) they applied AI-based methods, including machine learning or deep learning. Studies were excluded when they were not directly related to wildfires, did not use satellite data, did not apply AI-based methods, were review papers, conference papers, book chapters, or were not written in English. After title, abstract, and full-text screening, 157 studies were finally selected for review.

The selected literature was classified according to the primary wildfire lifecycle stage addressed by each study: pre-fire, active-fire, and post-fire. Studies covering more than one stage were retained as cross-stage cases. The studies were then analyzed comparatively using a structured framework that recorded the study area, wildfire-related target information, satellite platform and sensor, auxiliary data, reference data, input representation, model architecture, and model name. In this review, data analysis considered not only satellite observations, but also additional information used to support modeling. Auxiliary data included climatic, topographic, ecological, hydrological, and anthropogenic variables. Reference data used for training and validation were derived from satellite products, in situ observations, aerial data, land-based cameras, and simulation outputs. The analytical framework also recorded how model inputs were represented, distinguishing pixel-based, image-based, mixed image-and-pixel, and object-based approaches, and grouped the applied methods into major architectural families, such as tree-based, kernel-based, and neural network-based models.

3. Results

3.1. Trends by Wildfire Stage

Wildfire monitoring studies that integrate satellite remote sensing and AI over the past decade (2016–2025) show clear differences across wildfire stages in terms of research focus and analysis targets (Fig. 2). Overall, the post-fire stage accounts for the largest proportion of studies (43.8%), followed by the active-fire stage (36.4%) and the pre-fire stage (19.8%). This pattern indicates that recent research has been more strongly oriented toward post-fire damage assessment and impact analysis, whereas comparatively less attention has been devoted to pre-fire prediction.

Fig. 2. Distribution of artificial intelligence-based satellite wildfire monitoring studies over the past decade (2016–2025), classified by wildfire stage and primary analysis targets. The inner ring illustrates the relative proportion of studies in the pre-fire, active-fire, and post-fire stages, while the outer ring provides a detailed breakdown of research targets within each stage.

In the pre-fire stage, research is primarily concentrated on wildfire risk assessment and occurrence prediction. Other targets, including burn severity prediction and burned area prediction, are explored to a much lesser extent. This pattern suggests that pre-fire studies remain strongly focused on estimating ignition likelihood and spatial susceptibility using combinations of meteorological, vegetation, topographic, and anthropogenic information (Banerjee, 2022; Jaque Castillo et al., 2021).

In the active-fire stage, active fire detection clearly dominates, supported by the widespread use of TIR and MIR satellite observations (Kang et al., 2022; Kang et al., 2023; Zhang et al., 2025b; Zheng et al., 2024). Fire spread and smoke analysis have also received increasing attention, suggesting a gradual expansion from simple fire occurrence detection toward the monitoring of fire dynamics and atmospheric impacts (Chaturvedi et al., 2025; Zhang et al., 2021a). However, burned area progression during active-fire events remains relatively underrepresented, likely because this task requires temporal continuity and spatial tracking beyond instantaneous fire detection.

The post-fire stage shows the highest level of research activity, with burned area mapping and burn severity assessment emerging as the primary tasks (Hu et al., 2023b; Montorio et al., 2020). This dominance reflects the suitability of high-resolution optical satellite data, such as Landsat and Sentinel-2, for quantifying wildfire extent and damage intensity. In addition to these core tasks, a range of other applications, including recovery monitoring, emissions estimation, and structural damage assessment, has been explored (Dixon et al., 2023; Soleimani et al., 2024; Yan et al., 2024). Nevertheless, long-term monitoring tasks, particularly those related to ecosystem recovery, remain relatively underrepresented despite their importance for understanding post-fire resilience and landscape dynamics.

3.2. Trends by Study Region

The geographical distribution of the studies reviewed here shows that AI-based wildfire research is strongly concentrated in a small number of regions rather than being evenly distributed across the globe (Fig. 3). Most studies are concentrated in North America and Asia. In other regions, research is more unevenly distributed and tends to be concentrated in a small number of countries rather than being broadly distributed across continents. This pattern is particularly evident in South America and Oceania, where Brazil and Australia account for much of the research, whereas studies explicitly focused on Africa remain extremely limited. This distribution suggests that wildfire research has not accumulated through comprehensive global coverage of wildfire phenomena. Instead, it has developed selectively according to region-specific disaster experience, data availability, and research capacity.

Fig. 3. Geographic distribution of studies included in this review at the country level. Countries are shaded by the number of studies, highlighting a strong concentration of research in North America and Asia and a more uneven distribution in other regions.

At the national level, the literature shows a strong concentration in a small number of countries (Fig. 3). The United States accounts for the largest share, followed by China, Australia, Canada, Brazil, and South Korea. A small number of leading countries, therefore, account for a substantial proportion of all single-country studies, indicating that geographic concentration is clearly observable in the actual distribution of the reviewed literature. When temporal changes in study extent are considered, however, the spatial scope of AI-based wildfire research appears to have gradually expanded in recent years (Fig. 4). Although single-country studies remained dominant throughout the study period, the number of studies adopting broader spatial scopes, including single-continent, multi-regional, and global approaches, increased in recent years (Fig. 4). This pattern suggests that AI research on wildfires has gradually expanded beyond individual regional case studies toward models and datasets that can be applied across multiple regions.

Fig. 4. Annual distribution of studies by spatial extent. Studies were classified as single-country, single-continent, multi-regional, or global, showing that although single-country studies remained dominant throughout the study period, broader spatial scales became increasingly common in recent years.

3.3. Trends by Data Used

Wildfire monitoring studies exhibit clear stage-dependent differences in the selection and use of satellite sensors, reflecting how the spatiotemporal characteristics of observations are aligned with the objectives of each wildfire stage (Fig. 5). When considered together with the sensor specifications summarized in Table 1, these patterns indicate that sensor choice is not arbitrary but strategically determined by the requirements of pre-fire prediction, active-fire monitoring, and post-fire assessment. In particular, sensors with frequent temporal observations are more commonly used for risk assessment and active fire detection, whereas high-resolution optical and radar sensors are more frequently used for post-fire damage assessment.

Fig. 5. Stage-wise distribution of satellite sensor usage in artificial intelligence-based wildfire monitoring studies. The number of publications for each sensor is presented across pre-fire, active-fire, and post-fire stages, highlighting differencesin sensor utilization according to research objectives.

Table 1. Specifications of major satellite sensors used in artificial intelligence-based wildfire monitoring studies

In the pre-fire stage, the dominance of Terra/Moderate Resolution Imaging Spectroradiometer (MODIS) and Aqua/MODIS indicates a tendency to use satellite sensors with high temporal resolution and global coverage. These characteristics are commonly associated with large-scale wildfire risk assessment (Engström et al., 2022; Zhang et al., 2022). The additional use of Landsat-8 and Sentinel-2 further indicates that higher-resolution imagery is selectively incorporated to refine spatial variability in wildfire risk.

In the active-fire stage, the dominant use of MODIS and Visible Infrared Imaging Radiometer Suite (VIIRS) sensors highlights the central role of polar-orbiting satellite observations in active fire detection. These sensors, equipped with TIR capabilities, are widely used to identify fire hotspots and monitor fire activity (Cheng et al., 2022a; McCarthy et al., 2021). At the same time, the use of geostationary platforms such as Himawari-8 indicates the incorporation of high-frequency observations for tracking rapidly evolving fire events (Hong et al., 2023; Zheng et al., 2024). The concurrent use of higher-resolution sensors, including Landsat-8 and Sentinel-2, further suggests that spatially detailed data are being incorporated to complement thermal-based fire detection (Papakis et al., 2025; Zhao & Ban, 2025a).

In contrast, the post-fire stage is characterized by the dominant use of high-resolution optical sensors, particularly Sentinel-2 and Landsat-8. This pattern reflects the critical role of fine spatial resolution and SWIR bands in accurately mapping burned areas and assessing burn severity. The observed use of Sentinel-1 synthetic aperture radar (SAR) further indicates efforts to overcome the limitations of optical imagery, especially under cloud-contaminated conditions, highlighting its complementary role in post-fire analysis (De Luca et al., 2022; Hu et al., 2023b; Zhang et al., 2021b).

The use of auxiliary data also exhibits clear stage-dependent variation (Fig. 6a). The pre-fire stage shows the highest overall use of auxiliary data across all categories. In this stage, the frequent use of climatic and topographic data underscores the importance of environmental drivers, particularly meteorological conditions and terrain characteristics, in wildfire prediction. In the active-fire stage, these variables remain central, while ecological and hydrological datasets are incorporated as complementary sources to better represent fire behavior and its interactions with surrounding environments. In the post-fire stage, although topographic data continues to play a key role, a broader range of auxiliary datasets, including ecological and anthropogenic information, is used to support damage assessment and impact analysis. This pattern reflects the increasing complexity of post-event evaluation.

Fig. 6. Stage-wise distribution of (a) auxiliary data and (b) reference data used in AI-based wildfire monitoring studies. (a) shows the frequency of auxiliary data types (climatic, topographic, ecologic, hydrologic, and anthropogenic) across pre-fire, active-fire, and post-fire stages. (b) presents the number of publications using different reference data sources (satellite, in-situ, aerial, and simulation) for each stage.

The distribution of reference data further reinforces these stage-specific differences (Fig. 6b). Across all stages, satellite-based data serve as the primary reference source, highlighting their central role in wildfire monitoring. However, the relative use of in situ and aerial data increases in the active-fire and post-fire stages, suggesting greater emphasis on diverse reference sources when analyzing fire dynamics and post-fire impacts. This tendency is particularly evident in the post-fire stage, where multiple reference sources are used to address the complexity of burn severity assessment and ecological recovery analysis (Dixon et al., 2023; Farasin et al., 2020; Hu et al., 2023a).

Overall, these findings demonstrate that wildfire monitoring studies adopt distinct data strategies depending on the stage of analysis. Pre-fire research prioritizes long-term, large-scale datasets for risk prediction. Active-fire studies are primarily based on TIR observations from polar-orbiting satellites for fire detection, while also incorporating frequent observations from geostationary satellites to capture rapidly evolving fire events. Post-fire analyses rely mainly on high-resolution imagery for detailed damage assessment. This stage-specific differentiation indicates that effective wildfire monitoring requires the integration of diverse datasets tailored to the evolving characteristics of wildfire processes rather than reliance on a single data source or methodology.

3.4. Trends by AI Algorithms

Input representation in AI-based satellite wildfire monitoring studies has undergone a clear transition over time, reflecting broader changes in data availability and modeling approaches (Fig. 7a). In earlier studies, pixel-based representations were predominant, consistent with the use of traditional machine learning models that rely on structured inputs. Since approximately 2021, however, image-based representations have increased rapidly and now dominate recent studies. This shift indicates a growing preference for approaches that can directly exploit the spatial structure of satellite imagery. In particular, the adoption of image-based inputs is closely associated with the rise of deep learning models, especially CNNs, which are well-suited to extracting spatial features from high-dimensional data (Kang et al., 2022; Pinto et al., 2021). In contrast, object-based approaches have remained relatively limited throughout the study period, suggesting more constrained applicability in wildfire-related tasks.

Fig. 7. Temporal trends in (a) input representation and (b) model architecture used in artificial intelligence-based wildfire monitoring studies from 2016 to 2025. (a) shows the annual number of publications employing image-based, pixel-based, and object-based inputs, while (b) presents the usage trends of neural network-, tree-, and kernel-based models.

A similar evolution is observed in model architectures (Fig. 7b). Neural network-based models have increased sharply since 2020 and have emerged as the dominant approach in recent years. This pattern reflects the increasing reliance on deep learning techniques capable of capturing complex nonlinear relationships in large and heterogeneous datasets (de Almeida Pereira et al., 2021; Kang et al., 2023). In contrast, tree-based models have shown a steady but more gradual increase and have maintained consistent use over time. Kernel-based approaches remain relatively underutilized, suggesting limitations in scalability or effectiveness when applied to high-dimensional remote sensing data.

Model-level comparisons further reinforce these trends (Fig. 8). RF, as a representative tree-based model, remains the most frequently used model across the entire study period, highlighting its robustness, relatively low preprocessing requirements, and comparatively strong interpretability. At the same time, deep learning models such as U-Net and CNN-based architectures account for a substantial portion of recent studies. This trend reflects the increasing use of neural network-based approaches, particularly in tasks involving image-based wildfire detection and damage assessment (Cho et al., 2022; Hu et al., 2023a). The continued presence of SVM, a representative kernel-based method, among frequently used models indicates that traditional machine learning methods still play a meaningful role, especially in scenarios with limited data or simpler feature representations (Pourghasemi et al., 2020; Rubí et al., 2023). In addition, a diverse set of models, including XGBoost, Long Short-Term Memory (LSTM), Artificial Neural Networks (ANNs), Multi-Layer Perceptrons (MLPs), Logistic Regression (LR), and DeepLabV3+, has been applied in more specialized contexts, although their overall use remains relatively limited.

Fig. 8. Comparison of the most frequently used model types in artificial intelligence-based wildfire monitoring studies. The number of publications employing each model is shown for the entire study period.

These patterns indicate a clear methodological shift from traditional machine learning toward deep learning-based approaches in wildfire monitoring. This transition is closely linked to changes in input representation, because the increasing use of image-based data naturally favors neural network architectures designed to process spatial information. At the same time, the sustained use of tree-based models such as RF indicates that model selection remains dependent on data characteristics, task requirements, and interpretability considerations. Rather than converging toward a single dominant framework, wildfire monitoring research appears to be evolving toward a diverse and complementary set of modeling approaches tailored to specific analytical objectives.

4. Discussion

4.1. Pre-Fire Stage

4.1.1. Multimodal Data Integration and the Expanding Role of Remote Sensing

In wildfire risk and occurrence forecasting, meteorological variables such as temperature, humidity, wind speed, and precipitation remain among the most important predictors. Recent studies, however, have increasingly combined these variables with topography, land cover, vegetation, infrastructure, and population data to represent spatial heterogeneity and human ignition potential that cannot be sufficiently explained by meteorological data alone (Abdollahi & Pradhan, 2023; He et al., 2024; Zhang et al., 2025a; Brown et al., 2025; Pourghasemi et al., 2020; Ahajjam et al., 2025). For example, He et al. (2024) emphasized the roleof population distribution, infrastructure, and high-resolution vegetation imagery in representing human-related and landscape-level fire drivers, whereas Brown et al. (2025) focused on fuel load, canopy structure, and dead fuel moisture to estimate fire intensity potential. Pourghasemi et al. (2020) and Ahajjam et al. (2025) further showed that geomorphological, hydrological, and other geospatial variables can provide additional explanatory information beyond conventional meteorological predictors.

The expanding use of remote sensing data is one of the most important recent developments in wildfire risk modeling. Recent studies have moved beyond using satellite imagery as background information and have instead used specific band combinations, spectral indices, and biophysical variable retrievals to directly estimate fuel amount and moisture status, vegetation structure, and long-term drying conditions. Michael et al. (2021) used MODIS-derived vegetation and water-related indices to represent woody fuel availability and long-term drying trends, while Fernández-García et al. (2023) combined Sentinel-2 biophysical variables and Sentinel-1 radar features to characterize live fuel amount, canopy water status, and scattering properties. Brown et al. (2025) used VIIRS FRP as a satellite-based response variable for fire intensity. In this scene, the diversification of input variables reflects an effort to move AI-based wildfire risk modeling beyond purely statistical prediction toward a more physically grounded representation of the environmental and anthropogenic conditions that shape wildfire occurrence.

4.1.2. Diversification of Prediction Targets

Another important characteristic of AI applications in wildfire research is the increasing diversity of prediction targets. This diversification matters because the selected target determines not only what the model learns, but also how labels are constructed, how class imbalance arises, and which validation metrics are meaningful. Earlier work mainly focused on occurrence-oriented targets, such as classifying fire versus non-fire conditions or estimating wildfire occurrence itself (Bjånes et al., 2021; Alizadeh et al., 2025; Hernandez & Hoskins, 2024; Ferreira et al., 2023). Bjånes et al. (2021) formulated wildfire susceptibility mapping as a binary classification problem in which the positive class was defined using official wildfire ignition records provided by Chile’s National Forest Corporation, whereas the negative class was constructed from randomly extracted no-fire samples drawn from adjacent areas without recorded fire pixels. Similarly, Hernandez & Hoskins (2024) treated wildfire prediction in California’s Central Valley as a fire-versus non-fire classification task, using regional fire history records from the California Department of Forestry and Fire Protection as the basis for wildfire occurrence labels. Because wildfire occurrences are much rarer than non-fire conditions, this binary formulation is inherently affected by class imbalance, which can lead models to favor the dominant non-fire class and make high overall accuracy less informative for evaluating actual fire-detection capability.

More recent studies have expanded the scope of prediction to include a wider range of targets, such as hotspots, burned area, fire duration, burn severity, and fire intensity (Zhang et al., 2022; Ahajjam et al., 2025; Han et al., 2024; Brown et al., 2025). Zhang et al. (2022) used a CNN2D-LSTM architecture to predict global monthly burned area from the Global Fire Emissions Database version 4, shifting the problem from binary occurrence to continuous spatiotemporal estimation of fire-affected extent. Ahajjam et al. (2025) broadened this perspective by predicting wildfire occurrence, burned area, and fire duration as separate targets. They used the MCD14ML product, an active fire detection dataset derived from MODIS observations and distributed through the NASA Fire Information for Resource Management System (FIRMS). Brown et al. (2025) shifted the focus toward strictly continuous physical quantities by forecasting FRP in megawatts rather than predicting discrete occurrence labels. Together, these studies show that AI-based pre-fire research is moving from a narrow focus on occurrence prediction toward a broader set of management-relevant targets, including fire extent, persistence, severity, and intensity. This trend indicates that AI can support multiple stages of wildfire management, but it also means that different studies often address fundamentally different prediction problems. Direct comparison of model performance based solely on reported accuracy is therefore increasingly problematic.

The diversification of prediction targets also creates ambiguity in model validation, because a model that performs well for occurrence classification may not necessarily be suitable for estimating burned area, fire duration, burn severity, or fire intensity. Models that forecast spatial fire extent based on accumulated hotspot or burned area products are often assessed using overlap-based metrics such as intersection over union or Dice coefficients. Strictly continuous variables, such as FRP and fire duration, generally require regression metrics such as root mean square error or mean absolute error. Targets such as burn severity, which are categorical but ordered, require yet another evaluation framework. Accordingly, studies addressing different tasks should be evaluated under task-specific validation schemes rather than compared under a single common performance standard.

4.1.3. Limitations of Data-Driven Models and the Need for Physical Interpretability

A fundamental limitation of current AI-based wildfire risk modeling is that most approaches are data-driven models that learn statistical relationships from observations rather than physics-based models that directly describe ignition mechanisms. Although such approaches can provide high predictive performance, they often provide limited explanation of how a particular prediction is generated or why a given learning structure is appropriate for the wildfire risk problem under consideration. This limitation matters because pre-fire risk models are often expected to support preventive planning, resource allocation, and early warning, where decision-makers need to understand not only where risk is high but also which drivers are responsible for that risk. In practice, even if a wildfire risk warning is issued correctly in advance, it cannot effectively support decision-makers unless the underlying wildfire drivers responsible for that warning are clearly identified. Moreover, the lack of transparency in the causal basis of model outputs can reduce confidence in early warnings and make it difficult to distinguish the relative roles of meteorological, fuel-related, topographic, and human ignition factors. In this respect, the central issue in wildfire risk and forecasting research is no longer only how to improve predictive accuracy, but also how to justify the inference structure through which wildfire risk is derived from environmental and anthropogenic variables (Maniatis et al., 2022; Babu et al., 2025; Zerouali et al., 2025).

To reduce this problem, some pre-fire studies have attempted to introduce more explicit inference structures rather than relying exclusively on fully implicit end-to-end models. Maniatis et al. (2022) used SVM to classify land cover from Sentinel-2 imagery and then combined the resulting land-cover information with other fire-related factors through the analytic hierarchy process to construct fire risk maps. In this case, the weight assigned to each factor in the final risk estimate is relatively transparent. Similarly, Babu et al. (2025) separated wildfire risk into a static hazard component and a dynamic risk component to distinguish the effects of long-term landscape susceptibility from those of short-term meteorological and vegetation conditions. These approaches do not fully resolve the problem of physical interpretation, but they show how AI-based wildfire risk models can be structured in a more interpretable way by separating major risk components or making factor contributions more explicit.

However, such interpretive efforts remain limited. Even when the inference structure is made more explicit, it does not directly establish the causal mechanisms of ignition. Current interpretation attempts often remain at the level of showing which combinations of variables or model structures the system depends on, rather than physically justifying the inference structure itself. Zerouali et al. (2025) presented higher-performing composite models using Sentinel-2 Multispectral Instrument data and stacked hybrid machine learning models but also suggested that future studies should incorporate explainable AI techniques such as Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations. In this context, explainable AI should be used not only as a post hoc diagnostic tool, but also as a way to examine whether model predictions are consistent with known wildfire processes. Hybrid physical–AI approaches that incorporate process-relevant information on fuel availability, fuel moisture, terrain effects, and human ignition potential may provide a realistic pathway for making data-driven wildfire risk models more physically grounded. These observations indicate that current pre-fire AI research is moving toward great interpretability, although stronger links between model structure and wildfire processes are still needed.

Ultimately, what matters is not simply building a more complex model, but demonstrating why its input variables, learning structure, and final inference process are appropriate for the wildfire risk problem being addressed. From this perspective, future pre-fire research needs to reduce the gap between physics-based and AI-based approaches by linking data-driven learning structures to the actual processes of wildfire occurrence. Clarifying how input variable states influence prediction outcomes, together with developing more interpretable model structures, would increase the reliability of real-world decision-making and contribute to the ultimate goal of pre-fire research.

4.2. Active-Fire Stage

4.2.1. Utilization of Spatial and Temporal Information

One of the most notable trends in active-fire stage research is the increasing use of spatial and temporal information. Although conventional contextual approaches already incorporated simple neighborhood statistics, such as local means or contrasts with surrounding pixels, current AI-based approaches attempt to learn spatial patterns, contextual relationships, and temporal evolution more directly. This shift is particularly important in wildfire monitoring because fire- and smoke-related signals appear as rare and spatially limited anomalies embedded in heterogeneous and temporally fluctuating background conditions. In this sense, the use of spatiotemporal information is not merely a methodological refinement, but a necessary strategy for distinguishing transient fire signals from complex background variability.

In active fire detection, many studies have sought to overcome the low sensitivity and limited generalization performance of fixed-threshold algorithms by using AI models that can learn nonlinear and complex relationships between input variables and target labels. Vector-based machine learning models have used bi-spectral and contextual features to represent local thermal anomalies, whereas image-based deep learning models such as CNNs and U-Nets have been used to capture neighboring spatial patterns more directly (Sung et al., 2024; Maeda & Tonooka, 2022; Thangavel et al., 2023; Singh et al., 2025; Sung et al., 2025; Jang et al., 2019). Image-based deep learning models such as CNNs and U-Nets have also been widely used because they are effective at capturing spatial context and therefore better reflect the core concept of contextual algorithms: comparison with surrounding areas (Zhao et al., 2023; de Almeida Pereira et al., 2021; Hong et al., 2023; Zhang et al., 2025b; Kang et al., 2022). These approaches extend the logic of conventional contextual algorithms by allowing the spatial relationships between fire pixels and surrounding backgrounds to be learned rather than predefined by fixed thresholds. In particular, Kang et al. (2022) compared vector-based RF and image-based CNN models and showed that the image-based model achieved higher accuracy and faster detection by more effectively using contextual information.

Detection approaches that use temporal information have also been explored, particularly for geostationary satellites. Because geostationary observations provide frequent measurements, temporal deviations from normal brightness temperature (BT) patterns can be used to identify abnormal fire-related signals. Some studies used the difference between the temporal baseline normally observed for a target pixel and the observed BT as an input variable for machine learning models, showing that wildfire detection performance can be maximized when both spatial and temporal information are used (Zheng et al., 2024; Kang et al., 2022). In addition, recurrent neural network-based deep learning algorithms such as LSTM have been widely applied to improve detection accuracy by reflecting time-series patterns of BT (Zhao & Ban, 2022; Ji et al., 2023; Zhang et al., 2025b; Hong et al., 2023). In particular, Zheng et al. (2024) predicted BT using a Temporal Convolutional Network and then used the difference between predicted and observed BT as an input variable for RF. These studies indicate that temporal modeling is especially useful for active fire detection because fire signals are defined not only by their absolute radiometric values but also by their sudden departure from normal temporal behavior.

A similar tendency is observed in smoke detection. Image-based deep learning models have been widely applied to smoke detection using high-resolution polar-orbiting satellite imagery or ground-based camera imagery because smoke plumes must be interpreted in relation to surrounding spatial patterns rather than through isolated pixel values alone (Ba et al., 2019a; Bhamra et al., 2023; Zhao et al., 2022; Larsen et al., 2021; Wang et al., 2025; Chaturvedi et al., 2025). Temporal information has also been incorporated into smoke detection. For example, Bhamra et al. (2023) used an LSTM model to extract temporal context from tiled wildfire images and their previous frames, showing that time-series information can improve discrimination of smoke behavior and appearance.

This spatiotemporal trend is also evident in fire spread modeling. Many studies have adopted image-to-image architecture such as U-Net to predict fire spread (Khennou & Akhloufi, 2023; McCarthy et al., 2021; Zhang et al., 2021b). Cheng et al. (2022b) used an LSTM model to extract temporal information from input time-series imagery. More broadly, many studies have attempted to exploit both spatial and temporal information to achieve higher predictive performance. For example, Dong et al. (2024) predicted FRP using an autoencoder architecture that integrates a spatial block for extracting spatial information from static variables and a spatiotemporal block for capturing spatiotemporal patterns from dynamic variables. Compared with detection tasks, fire spread modeling requires a stronger representation of temporal evolution because the target is not only the presence of fire but also its changing location, intensity, and direction over time.

4.2.2. Increasing Modality of Input Variables

Another important trend in active-fire stage research is the increasing diversity of input data modalities. This trend reflects the fact that wildfire-related signals observed in satellite imagery are not determined by fire activity alone, but are also strongly influenced by background environmental and observational conditions. Meteorological, ecological, topographic, hydrological, and anthropogenic factors can affect the appearance, intensity, transport, and detectability of fire- and smoke-related signals. Accordingly, recent studies have increasingly combined satellite imagery with numerical model outputs and GIS-derived variables to represent the environmental context in which active fire, smoke, and fire spread signals are observed.

In active fire detection, environmental variables can alter thermal characteristics and thereby affect BT values observed from satellites. Several studies have jointly used numerical model outputs and GIS datasets representing weather, vegetation, topography, and observation conditions (Kang et al., 2023; Sung et al., 2024; Ochoa et al., 2024; Papakis et al., 2025). In particular, Sung et al. (2024) quantified the effects of vegetation and topographic factors on BT and improved wildfire detection performance by applying preprocessing steps that corrected for the influence of these environmental variables. These studies suggest that satellite-observed active fire signals should be interpreted not as purely radiometric phenomena, but as thermal anomalies shaped by environmental context.

A similar logic applies to smoke detection and fire spread modeling. Because smoke formation, transport, and dispersion are influenced by atmospheric and terrain conditions, some studies have incorporated meteorological and topographic variables that may affect smoke behavior and transport direction (Bhamra et al., 2023; Ding et al., 2024; Luo et al., 2025). In fire spread modeling, environmental variables such as topography, weather, vegetation, hydrology, and human-related factors have been more extensively used because they directly influence spread direction, intensity, and propagation conditions (Zou et al., 2023; McCarthy et al., 2021; Dong et al., 2024; Zakari et al., 2025). In particular, Rubí et al. (2023) showed that adding anthropogenic variables, such as distance to roads and urban areas, in addition to commonly used environmental variables, can better represent fire behavior in populated regions. Overall, the increasing modality of input variables indicates a shift from image-only detection toward context-aware modeling of active-fire processes.

4.2.3. Fusion Use of Multi-Satellite Imagery

A third major direction in active-fire stage research is the fusion of multi-satellite imagery. Because no single satellite system can simultaneously provide high spatial resolution, high temporal frequency, and stable observation under all atmospheric conditions, recent studies have increasingly attempted to combine different sensors in a complementary manner. This strategy is important because active fire monitoring requires both timely detection and reliable spatial representation, while individual sensors inevitably involve trade-offs among spatial resolutions, revisit frequency, and atmospheric sensitivity.

In active fire detection, many studies have attempted to integrate polar-orbiting and geostationary satellite observations to overcome the spatiotemporal resolution limitations of single-sensor observations. Some studies used polar-orbiting fire products as reference data to enhance geostationary satellite-based detection, whereas others combined imagery from different satellite platforms to improve temporal continuity and spatial detail (Zheng et al., 2024; Kang et al., 2023; Zhao & Ban, 2022; Gao et al., 2025; Eleutério et al., 2025; Badhan et al., 2024). These efforts show that multi-satellite fusion is not simply a data expansion strategy, but a way to compensate for the observational limitations of individual satellite systems.

Similar fusion-oriented approaches have also appeared in smoke detection and fire spread modeling. In smoke detection, some studies have attempted to integrate satellite observations with ground-based camera imagery. For example, Bhamra et al. (2023) improved smoke-detection performance and reduced latency by combining geostationary satellite-based active-fire detection products with ground-based camera imagery and meteorological observations. Such attempts may also help reduce confusion between smoke and clouds by complementing satellite observations with ground-level visual information and local atmospheric information. These studies suggest that future active fire monitoring will increasingly rely on sensor fusion frameworks that integrate spatial detail, temporal frequency, and all-weather observation capability.

In fire spread modeling, there have been attempts not only to integrate polar-orbiting and geostationary satellite observations but also to incorporate SAR imagery. While most studies have used polar-orbiting optical satellites (Farguell et al., 2021; Khennou & Akhloufi, 2023; Cheng et al., 2022b), some have added temporal information by combining geostationary and polar-orbiting imagery (McCarthy et al., 2021; Dong et al., 2024). Other studies, such as Zhang et al. (2021b), have incorporated SAR imagery to overcome the timeliness limitations of optical satellites under cloud-covered conditions. These efforts may improve the temporal resolution of model outputs and enhance the capacity for timely wildfire response.

4.2.4. Limitations and Future Directions

The application of satellite remote sensing and AI has substantially advanced studies of the active-fire stage, but several limitations remain. The most fundamental limitation is the scarcity and regional imbalance of reliable reference data. Reference data required for active-fire stage studies are difficult to obtain because, by the nature of wildfires, field data is difficult to collect during the event itself. As a result, most studies are limited to countries such as the United States, Canada, Australia, China, and South Korea, where large wildfires occur frequently or where wildfire information is systematically recorded and distributed (Fig. 3). This regional bias limits the evaluation of model generalization in underrepresented regions such as Central America, Africa, Southeast Asia and North Asia, and makes it difficult to assess omission and commission errors for small or less-documented fires (de Almeida Pereira et al., 2021; Kang et al., 2023; Zheng et al., 2024; Zhao et al., 2023). In addition, many regional studies are conducted using non-public local data, but these datasets are not sufficiently shared (Jang et al., 2019; Kang et al., 2022; Sung et al., 2024; Farguell et al., 2021; Wang et al., 2025). It is therefore necessary to construct shared benchmark datasets, promote reference data sharing, and conduct cross-regional evaluations across underrepresented areas to improve the generalization performance of AI-based active fire models.

Another limitation is that, although environmental variables related to wildfire characteristics have been used in some modeling studies, their relationships with wildfire behavior remain insufficiently analyzed and interpreted. In many cases, meteorological, topographic, and hydrological variables are treated mainly as additional inputs, while their influence on model predictions is evaluated only through a change in performance (Dong et al., 2024; Khennou & Akhloufi, 2023; McCarthy et al., 2021). This matters because context-aware active fire models cannot be fully trusted or transferred across regions unless the roles of these variables in detection, smoke behavior, and fire spread are more clearly understood. More comprehensive analyses are needed to determine how each environmental variable interacts with conventional input variables and influences final model predictions. Explainable AI methods may provide a practical way to evaluate whether learned relationships are consistent with known fire processes and to design more interpretable model structures (Ochoa et al., 2024; Zhang et al., 2025a; Zakari et al., 2025).

A further limitation is the insufficient consideration of individual dataset characteristics when integrating multimodal data. This issue arises not only when satellites with different spatiotemporal resolutions are used simultaneously, but also when satellite data are combined with field-based records. In particular, many studies have used active fire detection products from polar-orbiting satellites as reference data for training geostationary satellite-based algorithms without sufficiently accounting for differences in spatiotemporal resolution between the two systems (Zheng et al., 2024; Zhao & Ban, 2022). Because polar-orbiting and geostationary satellites differ in both observation timing and spatial resolution, additional dataset refinement is needed to ensure reliable learning and validation, including temporally constrained sample selection and manual inspection or filtering (Kang et al., 2023; Zhang et al., 2025b). Similarly, when field-based wildfire occurrence records are integrated with satellite imagery, additional processing is required to reconcile differences in dimensionality and spatiotemporal representation of wildfire information (Sung et al., 2024, 2025). Future studies should therefore place greater emphasis on harmonizing the temporal, spatial, and representational characteristics of different data sources, rather than treating them as directly interchangeable inputs.

4.3. Post-Fire Stage

4.3.1. Optical Remote Sensing as the Dominant Post-Fire Framework

Optical remote sensing data currently provides the strongest foundation for AI-based post-fire wildfire research. This dominance is particularly evident in burned area and burn severity mapping, where multispectral optical imagery provides direct spectral information on vegetation loss, charcoal deposition, and post-fire surface change. Recent work indicates that the field has progressed beyond simply using optical imagery as model input toward the construction of learnable datasets and benchmark frameworks centered on optical data. Sdraka et al. (2024) presented Forest wiLdfire Observations for the Greek Area, which combines Sentinel-2 and MODIS and functions as both a training dataset and a benchmark. Using an open dataset that includes expert-refined burned area polygons, they compare spectral indices, traditional machine learning methods, and deep learning models within a unified framework.

Additional studies, such as Knopp et al. (2020), Hu et al. (2021), and Cho et al. (2022), also reflect the growing use of optical imagery as the basis for trainable frameworks for burned area mapping. These studies show that optical data are not only observational sources but also the main basis for constructing AI-ready inputs, including image patches, pixel-level samples, segmentation masks, and benchmark datasets (Hu et al., 2021; Cho et al., 2022; Sdraka et al., 2024; Ouadou et al., 2025). In this sense, the strength of optical data in current post-fire AI research lies not only in their spectral sensitivity to burned surfaces but also in their role as the primary basis for constructing labeled training data and comparable evaluation frameworks. As a result, a wide range of AI algorithms has been tested most actively in this domain, and performance gains have accumulated most clearly on optical datasets.

4.3.2. Temporal Design in Post-Fire Mapping: Bi-Temporal and Mono-Temporal Approaches

This optical-centered trend is especially evident in research on burned area and burn severity. A substantial body of earlier work relied on multispectral optical imagery and spectral differencing for burned area delineation (Petropoulos et al., 2011; Roy et al., 2019), and recent deep learning studies have further developed this direction. The contribution of recent AI-based studies lies not in introducing change detection itself, but in reorganizing pre- and post-fire imagery, spectral indices, and spatial patterns into learnable inputs for burned-area and burn-severity mapping. Sdraka et al. (2024) emphasized a change-detection formulation using both pre-fire and post-fire imagery, while Sui et al. (2024), Wu et al. (2024), and Zhang et al. (2024) similarly showed the usefulness of bi-temporal optical inputs for separating burned and unburned areas or assessing burn severity. Thus, current AI-based post-fire mapping builds on the long-standing use of fire-induced spectral change while extending it through paired-image learning, index-guided modeling, attention-based feature extraction, and semantic segmentation (Sui et al., 2024; Wu et al., 2024; Khankeshizadeh et al., 2024; Ouadou et al., 2025).

At the same time, some studies have examined whether acceptable classification performance can be achieved using mono-temporal post-fire imagery alone. Van et al. (2024) conducted wildfire severity mapping using only post-fire Sentinel-2 imagery, combining raw bands with a large set of spectral indices and selecting key variables through SHAP and forward stepwise selection. Similar efforts can be found in uni-temporal or single post-fire optical mapping studies, such as Petropoulos et al. (2011), Hu et al. (2021), and Seydi et al. (2022a). These studies are better understood as practical options for cases in which pre-fire imagery is unavailable, temporally inconsistent, or unsuitable for rapid post-fire assessment, rather than as replacements for bi-temporal approaches. Accordingly, current optical research trends can be characterized as being centered primarily on bi-temporal input, while also exploring the operational potential of mono-temporal approaches in parallel.

4.3.3. Complementary Role of SAR and Optical–SAR Fusion

Although temporal design has mainly developed within optical workflows, it also points to the continuing importance of alternative sensing modalities when optical observations are unavailable or unstable. In this respect, SAR-based approaches have both distinct advantages and significant limitations. SAR can operate day and night and is less affected by clouds or smoke, making it useful for post-fire mapping when optical imagery is obscured (Belenguer-Plomer et al., 2019; Belenguer-Plomer et al., 2021; Zhang et al., 2021b; Zhang et al., 2024). However, recent comparisons indicate that SAR used alone generally provides lower burned-area separability than optical data, and its effectiveness can vary depending on wavelength, polarization, vegetation structure, and the availability of pre-event observations (Zhang et al., 2024).

Several recent studies have therefore focused on enriching SAR inputs rather than using backscatter alone. Zhao and Ban (2025b) proposed a strategy that moved beyond simple log-ratio intensity from RADARSAT Constellation Mission compact-polarimetric C-band SAR by incorporating compact-polarimetric decomposition and the Compact-Pol Radar Vegetation Index as additional features. The authors reported that decomposition features and vegetation indices provided complementary information beyond log-ratio intensity, and that performance improved when all three input types were used together. Xi et al. (2024) also designed a dual-path framework based on the complementarity of Sentinel-1 and Sentinel-2, arguing that optical imagery is effective for capturing spectral differences, whereas SAR contributes geometric and physical information. A similar fusion-oriented logic appears in Seydi et al. (2022b), which combines multitemporal Sentinel-1 coherence with post-event Sentinel-2 imagery, and in Zhang et al. (2021b), which explicitly tests the synergy of Sentinel-1 and Sentinel-2. Taken together, current SAR-based wildfire research suggests not that SAR replaces optical data as a stand-alone alternative, but that it functions most effectively as a complementary source that improves observation reliability under cloud, smoke, or data-gap conditions.

4.3.4. Temporal Uncertainty, Phenology, and the Role of Spectral Indices

This issue connects to a broader methodological challenge increasingly raised by recent studies: how to handle uncertainties related to temporal gaps and changes in vegetation condition, often discussed under the broader problem of phenology. Spectral indices can effectively summarize fire-induced change, but their values are also sensitive to land-cover type, acquisition timing, post-fire regeneration, atmospheric conditions, and confusion with water bodies, shadows, or urban structures (Sdraka et al., 2024; Wu et al., 2024; Fassnacht et al., 2021; Boer et al., 2008; Storey et al., 2016). These sensitivities are particularly important in AI-based post-fire mapping because unstable index values can be learned as inconsistent or region-specific signals rather than as generalizable indicators of burned area or burn severity.

Change-based features such as dNBR therefore remain central to post-fire analysis, but their effectiveness depends on how temporal gaps, vegetation recovery, and observation conditions are handled (Roy et al., 2019; Wu et al., 2024; Zhang et al., 2024). For AI-based post-fire mapping, the key issue is not simply whether spectral indices are used, but how variability caused by phenology, delayed image acquisition, and early vegetation recovery can be transformed into stable and learnable change information.

4.3.5. Evolving Role of Spectral Indices and Future Directions

This broader concern also leads to a reconsideration of the role of spectral indices themselves. Indices such as NBR, dNBR, RdNBR, and NDVI have long served as core tools for detecting burned area and burn severity (Chuvieco et al., 2002; Key and Benson, 1999, 2006). They continue to be widely used in recent burned area mapping studies as baselines, change indices, or model inputs (Roy et al., 2019; Roteta et al., 2019; Lizundia-Loiola et al., 2020; Collins et al., 2018; Ba et al., 2019b; Liu et al., 2020). For example, in optical-based research, dNBR and related indices that summarize the difference between pre-fire and post-fire imagery are commonly treated as key change features. Wu et al. (2024) used dNBR for burn severity assessment and label generation, and Zhang et al. (2024) likewise regarded optical dNBR as an important component of burned area separation.

However, these indices alone do not guarantee stable segmentation under all conditions. Sdraka et al. (2024) pointed out that threshold-based approaches have limited generalizability and are repeatedly affected by spectral confusion. Earlier burned area mapping studies have also shown that fixed-threshold approaches often require adjustment depending on image conditions and regional characteristics (Alonso-Canas and Chuvieco, 2015; Roteta et al., 2019). In post-fire optical mapping and mono-temporal segmentation research, spectral indices are often used not as final discriminators but as auxiliary features (Knopp et al., 2020; Petropoulos et al., 2011; Brand and Manandhar, 2021; Seydi et al., 2022a; Hu et al., 2021). Thus, in recent wildfire AI research, spectral indices are better understood not as independent final solutions but as intermediate representations that summarize change signals and support labeling, feature construction, or model input.

The direction of SAR research is also expanding beyond the use of single C-band data. Increasing attention is being given to approaches that combine repeat-observation-based change detection with information from different frequency bands and derived polarimetric features. In this sense, recent work on C-band, L-band, and compact-polarization SAR can be read as part of a broader move toward multi-frequency wildfire monitoring rather than as a simple continuation of single-backscatter analysis. The NASA-ISRO Synthetic Aperture Radar Mission is a representative example of this direction. Designed to operate in both L-band and S-band simultaneously and to provide systematic repeat observations, it points toward future wildfire monitoring frameworks in which cross-frequency complementarity and time-series change information can be analyzed together. Although it has not yet been established as an operational wildfire burned area system, it illustrates the direction in which SAR-based wildfire observation may expand in the near future.

5. Conclusions

This study examined recent advances in satellite-based wildfire monitoring using AI from a full-lifecycle perspective, covering the pre-fire, active-fire, and post-fire stages. The results show stage-dependent differences in research focus, with the highest concentration in the post-fire stage, followed by the active-fire stage, and relatively less attention to pre-fire prediction and long-term recovery monitoring. Across these stages, recent studies show a clear shift toward multimodal data integration, image-based and spatiotemporal modeling, and deep learning-based workflows. However, most studies have been conducted at regional scales using heterogeneous datasets, which limits generalization. Several challenges persist, including geographical data imbalance, limited reference data, inconsistencies in spatiotemporal resolution, and a lack of physical interpretability in data-driven models. To address these issues, future research should focus on four actionable directions: (1) constructing and sharing benchmark datasets and reference data across diverse and underrepresented regions, (2) conducting cross-regional evaluation with task-specific validation metrics, (3) improving model interpretability through explainable AI and hybrid physical–AI approaches, and (4) developing robust strategies for harmonizing multi-source data with different spatial, temporal, and observational characteristics. Overall, this study provides a comprehensive overview of current research and highlights key directions for future research in satellite-based wildfire monitoring.

Author Contributions

Conceptualization: Sung TJ; Data curation: Yang SY, Kim WH, Sung TJ; Formal analysis: All authors; Funding acquisition: Im JH; Methodology: Yang SY, Kim WH, Sung TJ; Project administration: Im JH; Supervision: Im JH; Visualization: Yang SY, Kim WH; Writing–original draft: Yang SY, Kim WH, Sung TJ, Lee SH; Writing–review & editing: Yang SY, Kim WH, Sung TJ, Im JH.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported. This article was commissioned by the Editorial Board and underwent independent peer review.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-02310080). T. Sung was supported by the Basic Science Research Program (RS-2024-00409427) through the National Research Foundation of Korea (NRF), funded by the Ministry of Education.

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.