A Critical Review of Wildfire Risk Prediction Models in Data-Scarce Mediterranean Environments

Abstract

Wildfires are a growing threat in Mediterranean regions where climate variability and land-use practices increase vulnerability to fire risk. Developing effective prediction models is essential for robust wildfire management, particularly in such data-scarce environments. Focusing on data-scarce Mediterranean environments, with reference to environmental conditions observed in Morocco, this review presents prediction models across three methodological categories: spatial risk mapping, temporal forecasting, and fire spread simulation, alongside the satellite data products that support their deployment. Each category is assessed in terms of predictive performance, data requirements, and adaptability to low-resource environments. XGBoost showed strong applicability in data-scarce Mediterranean contexts, while ARIMA was validated for forecasting fire-relevant time series under limited data resources. Freely accessible MODIS-derived products represent a significant asset to the region. Based on this synthesis, a hybrid XGBoost-ARIMA framework incorporating MODIS-derived inputs and SHAP-based interpretability is proposed as a promising candidate architecture to be validated after further investigation. The findings aim to support researchers, land managers, and policymakers in strengthening local wildfire prevention and mitigation efforts by aligning model capabilities with regional data and environmental constraints.

1. Introduction

Wildfires are among the most destructive natural hazards, causing severe ecological losses, damaging local economies, and presenting significant threats to human safety. In the Mediterranean basin alone, 50,000 fires burn approximately 600,000 to 900,000 ha annually, representing 1.3% to 1.7% of total forests [1]. Fire occurrences in Mediterranean countries have increased since the 1970s, driven by both climate change and human activity [1,2], and are expected to intensify further in the coming decades [1]. This trend highlights the pressing need for reliable wildfire prediction and risk assessment tools that serve as early warning and informed mitigation tools.

Morocco is considered a moderate-to-high fire risk country in the Mediterranean region [1,2], yet it is rarely mentioned in the predictive modeling literature, making it a compelling and underrepresented motivating case for this review, such that its findings are intended to serve data-scarce Mediterranean environments more broadly. Within Morocco, the Ifrane province is particularly vulnerable to fire ignition due to its dense cedar forest cover, as illustrated by the extreme risk classification assigned to the region in July 2024 fire season shown in Figure 1.

This paper reviews current prediction approaches across three methodological categories, spatial risk mapping, temporal forecasting, and fire spread simulation, while evaluating the satellite data products that enable their deployment.

2. Materials and Methods

2.1. Search Strategy and Sources

This review uses studies from the Scopus database, which served as an exclusive data source because of its rigorous peer-review standards and extensive indexing of high-impact journals in environmental science and geosciences.

The following are the Boolean search strings that were applied to titles, abstracts, and keywords:

• Fire, wildfire risk assessment, forest fire prediction;
• Machine learning, XGBoost, Random Forest, LSTM, ARIMA;
• Satellite data, MODIS, NDVI, vegetation indices, drought, land use, rainfall;
• Fire spread, simulation, Cell2Fire, fire behavior models;
• Mediterranean regions, North Africa, Morocco.

The search was limited to peer-reviewed journal articles published in English between 2014 and 2026. The year 2014 was selected as the lower threshold for two main reasons. First, it corresponds to the earliest study included in this review [1]. Second, by 2014, MODIS satellite products had accumulated over a decade of consistent global observations, providing sufficient temporal data for the forecasting and risk modeling approaches reviewed in this paper. As for 2026, it reflects the date of the present review and ensures the inclusion of recent methodological developments in wildfire prediction research.

2.2. Inclusion and Exclusion Criteria

The first step of the search was to establish the specific criteria under which our search would fall, to ensure objectivity and relevance to wildfire prediction in data-scarce Mediterranean environments.

Inclusion criteria:

• Peer-reviewed journal articles indexed in Scopus;
• Publications written in English;
• Studies published between 2014 and 2026;
• Research focusing on wildfire risk assessment, early detection, machine learning for prediction, or simulation-based prediction methods.

Exclusion criteria:

• Book chapters, editorials, and non-peer-reviewed sources;
• Studies targeting post-fire ecological impacts without predictive modeling;
• Articles short of methodological transparency or performance evaluation;
• Studies with highly confined hypotheses restricting suitability to other regions.

2.3. Assessment Framework

Each reviewed model was assessed based on three criteria: predictive performance, data requirements, and adaptability to low-resource environments. The ratings ranged from Low to Moderate to High, based on evidence presented within each reviewed study.

Predictive performance was evaluated by comparing models within each study. High was assigned when a model ranked first or showed statistically equivalent performance to the top-performing alternative; Moderate when it performed competitively but was consistently outperformed by at least one alternative; and Low when it considerably underperformed against simpler baselines.

Data requirements were rated based on the volume, specialization, and accessibility of required inputs. Low indicates operation on a single publicly available data source without field measures; Moderate indicates multiple accessible sources requiring preprocessing or integration; and High indicates dependence on reanalysis products, multi-decade archives, or dedicated ground infrastructure.

Adaptability to low-resource environments was evaluated based on data availability, infrastructure needs, and historical records. Models were rated High when their data and infrastructure requirements were met by publicly available sources; Moderate when deployment barriers existed but could be resolved through a targeted data collection; and Low when operational preconditions were assessed as currently unmet resource-limited Mediterranean environments.

These ratings are intended to support evidence-based comparisons rather than to establish definitive rankings.

2.4. Screening and Selection Process

The study selection process followed PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, as illustrated in Figure 2. The initial database search returned 70 records, of which six duplicate records were removed prior to screening. An additional eight records were excluded for the following reasons: topic clearly outside the scope of fire-related prediction (n = 5) and non-English language publications (n = 3). Among the non-English studies, French-language publications identified during the search did not satisfy the inclusion criteria defined in Section 2.2, while one Chinese-language study that met the inclusion criteria was excluded because the language could not be assessed by the authors. This left 56 records for title and abstract screening.

At the title and abstract screening stage, 32 records were excluded in accordance with the inclusion and exclusion criteria defined in Section 2.2. The primary reasons for exclusion were: topic unrelated to wildfire prediction (n = 14), focus on post-fire ecological impacts without predictive modeling (n = 9), and absence of quantitative methodology (n = 9), leaving 24 records for full-text review.

During full-text assessment, three additional records were eliminated due to the lack of quantitative performance metrics (n = 1), incompatible data requirements for transferability to data-scarce environments (n = 1), and scope outside wildfire prediction despite relevant keywords (n = 1). The final set included 20 references in total.

References [1,2,3,4,5,6,7] provide contextual background on wildfire risk and environmental conditions relevant to Mediterranean and Moroccan landscapes, the selected studies for the final analysis, cited as references [8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23], are summarized in

Table 1. Classification of reviewed references by study region and analytical purpose.

Refs.	Title	Study Region	Purpose
[1]	Climatological evaluation of Haines forest fire weather index over the Mediterranean Basin	Mediterranean Basin	Contextual Background
[2]	Application of remote sensing and machine learning algorithms for forest fire mapping in a Mediterranean area	Northern Morocco	Contextual Background
[3]	Development of a hydrometeorological drought severity composite index based on the integration of multisource characteristics and an explainable artificial intelligence model	The Oum Er Rbia watershed, Morocco	Contextual Background
[4]	Land Use Dynamics and Soil Conversation Strategies In The El Kssiba Region, Atlas Mountains Of Morocco	El Kssiba region of the Middle Atlas Mountains, Morocco	Contextual Background
[5]	The Impact of Climate Change and Land Use on Soil Erosion Using the RUSLE Model in the Tigrigra Watershed (Azrou Region, Middle Atlas, Morocco)	The Tigrigra watershed, the mountainous region of the Middle Atlas, Morocco	Contextual Background
[6]	Litter and biomass traits of some dominant Moroccan understorey fuels in five fire-prone forest regions	The Central Plateau, the Middle Atlas (Western and Eastern), the Western Rif and the Pre-Rif regions	Contextual Background
[7]	Assessment of Flammability of Moroccan Forest Fuels: New Approach to Estimate the Flammability Index	Larache, Ahl Srif, Souk L’Qolla, Dardara, Bellota	Contextual Background
[8]	Comparative study on landslide susceptibility mapping based on different ratios of training samples and testing samples by using RF and FR-RF models	China	Review: Spatial Risk Mapping
[9]	Assessment of Landslide Susceptibility Using Statistical- and Artificial Intelligence-Based FR–RF Integrated Model and Multiresolution DEMs	Iran	Review: Spatial Risk Mapping
[10]	Wildfire Susceptibility Mapping in Karabuk Province, Türkiye Using Machine Learning Algorithms	Turkey	Review: Spatial Risk Mapping
[11]	Wildfire Susceptibility Mapping in Greece Using Ensemble Machine Learning	Greece	Review: Spatial Risk Mapping
[12]	Explainable Artificial Intelligence to Unveil Intrinsic Characteristics of Conditioning Factors Governing Forest Fire Susceptibility	Turkey	Review: Spatial Risk Mapping
[13]	Rainfall forecasting in arid regions in response to climate change using ARIMA and remote sensing	Syria	Review: Temporal Forecasting
[14]	Temporal patterns of active fire density and its relationship with a satellite fuel greenness index by vegetation type and region in Mexico during 2003–2014	Mexico	Review: Temporal Forecasting
[15]	Long short-term memory	N/A (foundational algorithm paper)	Review: Temporal Forecasting
[16]	Wind power generation prediction using LSTM model optimized by sparrow search algorithm and firefly algorithm	N/A	Review: Temporal Forecasting
[17]	Remote sensing diagnosis of Forest fire risk based on state-trend characteristics using machine learning models	China	Review: Temporal Forecasting
[18]	AttentionFire_v1.0: interpretable machine learning fire model for burned-area predictions over tropics	Africa and South America	Review: Temporal Forecasting
[19]	HiQ-FPAR: A High-Quality and Value-added MODIS Global FPAR Product from 2000 to 2023	Global	Review: Satellite Data Products
[20]	A global 1 km resolution daily surface longwave radiation product from MODIS satellite data from 2000–2023	Global	Review: Satellite Data Products
[21]	Cell2Fire: A cell-based forest fire growth model to support strategic landscape management planning	Canada	Review: Fire Spread Simulation
[22]	Sensitivity of ELMFIRE to real-world input datasets for WUI fire modeling	United States	Review: Fire Spread Simulation
[23]	Dynamical linkages between planetary boundary layer schemes and wildfire spread processes: A case study using WRF-Fire version 4.6	China	Review: Fire Spread Simulation
[24]	Improving Groundwater Forecasting Accuracy with a Hybrid ARIMA-XGBoost Approach.	Italy	Methodological Reference

N/A = not applicable (i.e., the study does not have a specific geographic focus or is a methodological paper without a case study region).

. while reference [24] is cited as a supporting methodological source in Section 4.3. Given the limited work conducted in Mediterranean environments, the inclusion criteria were left non-restrictive with respect to study location. As a result, the studies included in the final analysis are geographically diverse, but if transferable, could benefit Morocco and all other similar Mediterranean regions.

3. Literature Review

3.1. Ecological and Data Context in Mediterranean Environments

The environmental conditions documented in Moroccan forests provide a useful context for understanding the types of climatic, ecological, and data constraints this review addresses. Precipitation patterns vary considerably with elevation, ranging from nearly 200 mm in lower areas to over 1100 mm in mountainous zones. Most rainfall occurs between October and May (70–80%), leaving the summer season particularly dry and vulnerable to fire activity. These dry conditions are further intensified by the hot southwesterly chergui winds originating from the Sahara [3]. Mean annual temperature is approximately 18 °C, although extremes can range from −3.5 °C to over 40 °C [3]. Starting in 2019, drought frequency has increased considerably, largely due to increasing temperatures and higher evapotranspiration rates [3]. In parallel, the region has experienced substantial land-use and land-cover changes. Between 1953 and 2024, bare land area has more than doubled, whereas agricultural land has decreased significantly from 36.79% to 14.85% of the total area [4]. Although reforestation initiatives have contributed to forest expansion in certain areas, long-term projections suggest that forest cover may decline by 2050 under the combined effects of population growth, overgrazing, deforestation, and urban expansion [4,5]. Human-related activities further amplify ignition risk, with fuelwood collection, grazing practices, cannabis cultivation, and proximity to roads all being identified as major contributors to fire occurrence in Moroccan forests [2].

The impact of these environmental factors is reflected in a study by Tatli and Türkeş [1], who found that Morocco falls within a zone of moderate-to-high fire risk using the Haines Forest Fire Weather Index (FFWI) and reanalysis data during the 1980–2010 period. FFWI recognized Morocco, Tunisia, and multiple southern European countries, including Greece, Italy, and Spain, as moderate-to-high fire risk zones, primarily due to a dry summer climate and elevated atmospheric instability. The study linked Morocco’s classification to its regional patterns, where elevated temperatures and low humidity systematically increase fire ignition probability.

Fire occurrence data from northern Morocco documents roughly 1185 ha of forests alone being burned annually, accounting for 43% of the total burned area nationally [2]. A study conducted in the Tangier-Tétouan-Al Hoceima region revealed that hybrid FR-ML models uniformly outperform individual algorithms in mapping fire susceptibility in Moroccan terrain, with FR-RF (AUC = 0.989) on the validation dataset, FR-SVM (AUC = 0.959), and FR-MLP (AUC = 0.858) [2]. However, earlier prediction efforts that relied heavily on expert judgement and did not incorporate field-quantified fuel properties achieved relatively low prediction accuracy (58%) [6]. This highlights the limited availability of data-driven approaches that integrate local environmental inputs.

Complementary field studies in Moroccan Mediterranean forests by [6,7] present the characteristics of local fuel data that can support predictive modeling. The first study investigates the flammability of Moroccan forest fuels across five sites in the Western Rif Mountains. Using a ceramic epiradiator heated at 600 °C, the study measures ignition time, flame height, and combustion duration under varying moisture conditions for 600 g samples of different species [7]. Based on these experiments, the authors developed a Flammability Index (FI) classification to differentiate moderately flammable (FI 4.5–7.5), flammable (FI 7.5–10.5), highly flammable (FI 10.5–13.5), and extremely flammable (FI > 13.5) forest fuels. This classification provides a practical tool for forest managers to assess ignition risk through fuel moisture. However, the index does not cover slope or wind, factors that can highly affect fire behavior and reduce predictive accuracy in complex terrain. The second study examined the physical characteristics of understory fuels in five high-risk Moroccan forest zones: the Western Rif, Pre-Rif, Western and Eastern Middle Atlas, and Central Plateau [6]. Sampling 333 shrub specimens across 35 sites between 2014 and 2016, the authors derived linear regression models to calculate fine fuel biomass from shrub volume, achieving a coefficient of determination ranging from R² = 0.60 to 0.99. Bulk density values ranged from 0.35 to 4.64 mg/cm³, and fuel depths from 1.1 to 7.5 cm. The study concluded that fuel biomass is a species-specific rather than geographically determined variable.

These studies present available ecological and fuel-related datasets that could support future predictive wildfire modeling efforts in data-scarce Mediterranean environments.

3.2. Overview of Prediction Models

3.2.1. Spatial Risk Mapping Models

Spatial risk mapping models compute precise estimates of fire susceptibility across landscapes. They process static or periodically updated datasets and output either continuous susceptibility surfaces or binary hazard classifications. These models serve as planning tools, informing land-use decisions and resource allocation rather than operating in real time.

3.2.2. Temporal and Sequential Forecasting Models

Temporal forecasting models predict the future state of a variable based on its past behavior over time. This category includes models of different complexity levels, ranging from statistical approaches suitable for low-resource settings to computationally intensive neural architectures designed for large-scale fire dynamics.

3.2.3. Satellite Data Products for Fire Monitoring

Satellite-derived data products are not predictive models themselves, but rather globally consistent measurements of diverse variables. These products serve as both input variables for prediction models and as validation benchmarks for model outputs.

3.2.4. Fire Spread Simulation

Fire spread simulators model the physical propagation of an active fire across a landscape following ignition. Unlike risk mapping and forecasting models, these tools process individual fire events, predicting how a fire initiated at a given ignition point will propagate through present conditions over time. Their outputs are usually used for mitigation planning and fuel treatment evaluation.

3.3. Review of Predictive Models by Functional Category

3.3.1. Spatial Risk Mapping Models

FR-RF

The Frequency Ratio–Random Forest (FR-RF) model combines the interpretability and spatial weighting transparency of the Frequency Ratio (FR) with the predictive accuracy and feature-importance assessment of the Random Forest (RF) ML algorithm. They are combined because each method compensates for the other’s limitations. While RF can estimate the overall importance of each conditioning factor, it cannot calculate the spatial correlation between hazard occurrence and the individual classes of each conditioning factor. FR is able to measure these class-level spatial correlations but is unable to classify features based on their contribution level to the model. Therefore, the hybrid FR-RF model resolves both limitations concurrently [8,9]. The FR component converts each class of each conditioning factor into a continuous numerical FR value, standardizing all factors before model training [9]. The FR value determines whether a class is positively or negatively associated with hazard occurrence by comparing it to 1. Subsequently, the RF component processes the FR-weighted dataset by training a large ensemble of decision trees on subsamples of that dataset. The final prediction is determined through a majority vote across all trees [8,9]. The studies reviewed in this section address landslide rather than wildfire susceptibility. However, both hazards cover spatial mapping problems, making them methodologically relevant to the wildfire susceptibility application.

A study covering Fugu County in northern Shaanxi Province, China [8], implemented FR-RF to monitor landslide susceptibility. The dataset consisted of 1339 landslide sites with conditioning factors including slope, aspect, curvature, land utilization, lithology, and NDVI. The authors evaluated the standalone RF model against the integrated FR-RF model across four training/testing splits: 80/20, 70/30, 60/40, and 50/50. The RF model outperformed the FR-RF across all four splits. At the optimal 70/30 split, RF reached an AUC of 0.939 compared to 0.931 for FR-RF [8]. The authors concluded that the integration of FR introduced data redundancy that uniformly reduced predictive performance across all training and testing splits [8].

In contrast, a study conducted in the Sangtarashan watershed of Mazandaran Province, northern Iran [9], reported a noticeably different outcome using a smaller dataset of 129 landslide events with a 70/30 training/testing split. The study included 15 conditioning factors such as elevation, slope, land use/land cover (LU/LC), and NDVI, among others. The results showed that FR-RF outperformed both individual FR and RF models, with AUC values of 0.917, 0.865, and 0.840, respectively [9]. Variable importance analysis identified LU/LC (importance score: 16.95), NDVI (importance score: 16.44), and proximity to roads (importance score: 15.32) as the primary landslide susceptibility predictors [9]. An additional contribution of this study was the evaluation of Digital Elevation Model (DEM) resolution as a cause of variability in model accuracy. Comparing three DEM resolutions, the authors found that the FR-RF model achieved AUC = 0.917 with PALSAR (12.5 m), compared to 0.865 with ASTER (30 m) and 0.863 with SRTM (90 m), respectively, highlighting the influence of input data resolution on susceptibility mapping accuracy [9].

These results indicate that FR-RF performance may vary depending on dataset size, feature structure, and input resolution, suggesting sensitivity to underlying data characteristics.

XGBoost

XGBoost (Extreme Gradient Boosting) is a scalable and efficient implementation of the gradient-boosted decision tree (GBDT) algorithm, widely applied to structured datasets. This model is widely used because it captures complex, nonlinear patterns.

The first study took place in Karabük Province in northern Türkiye, to support forest management and emergency preparedness at the province scale [10]. The authors assessed four boosting-based ML algorithms, XGBoost, CatBoost, LightGBM, and AdaBoost, through a set of 13 conditioning factors sorted as topographic (elevation, slope, aspect), vegetation (land use/land cover, NDVI, fuel type), climatic (precipitation, temperature, humidity, wind speed), and anthropogenic (distance to roads, distance to settlements, distance to rivers/water bodies, and population density) [10]. All variables were standardized to 30-m spatial resolution and normalized using min-max scaling. Models were evaluated on both training and testing splits based on overall accuracy, AUC and the Kappa index. XGBoost achieved the highest performance across all metrics on the testing split with an accuracy of 94.5%, AUC of 0.939, and Kappa index of 0.890. CatBoost performed identically in accuracy and AUC, while LightGBM achieved the highest training accuracy (99.8%) but showed a slight tendency to overfit on the testing split. GBM and AdaBoost underperformed with test accuracies of 89.1% and 81.0%, respectively [10]. In the XGBoost features analysis, land cover was recognized as the most impactful predictor, followed by NDVI and temperature, slope, wind speed, and finally distance to residual areas [10]. The authors concluded that XGBoost demonstrated the highest consistency among the boosting-based ML algorithms in their study, combining robust generalization with spatially interpretable patterns [10].

The second study was carried out in Greece, a country that recorded an average of 52 wildfire events per year between 2000 and 2024 [11]. The study evaluates four ensemble regressors: XGBoost, GBM, LightGBM, and CatBoost, trained on burned area data from the Copernicus Emergency Management Service (EFFIS), including 12 conditioning factors: elevation, slope, aspect, roughness, Topographic Wetness Index, distance to roads, rivers, settlements, land use class, grassland cover, dominant leaf type, and a Fire Weather Index (FWI) value [11]. All variables were standardized to a 500-m spatial resolution. In contrast to the previous case, this one deployed the models as regressors rather than classifiers, modeling the continuous likelihood of wildfire rather than a binary prediction [11]. Model evaluation used the standard deviation of Normalized Root Mean Square Error (std-NRMSE), Ensemble Mean to average predictions from all four models, and Ensemble Max to select the maximum predicted value. While XGBoost achieved the second-best performance in this study (std-NRMSE = 0.8451, where lower values indicate better performance), the Ensemble Max model using XGBoost predictions successfully identified 83% of all wildfires recorded between 2000 and 2024 in high-risk zones, and 4% of fire events in low-risk zones [11]. Furthermore, the XGBoost feature importance analysis was dominated by topographic features (54.1%), contradicting the conclusions made in the first study claiming that the vegetation indices are the most influential [11].

The third case study covered the Manavgat district of Antalya Province in the Western region of Türkiye, over a 2283 km² surface [12]. Unlike previous studies, this one goes beyond model comparison by directly addressing the lack of interpretability that limits the use of machine learning models in disaster management situations. The study employed three ensemble algorithms, Random Forest, XGBoost, and Natural Gradient Boosting (NGBoost), and 11 conditioning factors, namely elevation, aspect, curvature, slope, valley depth, Topographic Wetness Index, distance to rivers, distance to roads, temperature, precipitation, and wind speed [12]. The dataset was processed using a 70/30 training/testing split. To address interpretability, two explainable AI methods were adopted: Shapley Additive Explanation (SHAP) for general feature-importance analysis and Local Interpretable Model-agnostic Explanations (LIME) to reveal how specific feature values either pushed individual predictions toward or away from fire classification [12]. NGBoost outperformed the other models with an 81.42% accuracy level; however, McNemar’s test confirmed that the difference between XGBoost and NGBoost was statistically insignificant. This suggests that while NGBoost is superior, XGBoost remains a competitive and statistically comparable alternative [12]. SHAP global analysis identified the two most influential conditioning factors for NGBoost, while LIME local analysis provided additional granularity: for fire-classified samples at 103 m, and for non-fire samples at 1275 m. While the authors agree that NGBoost outperforms XGBoost, they also agree that SHAP and LIME are applicable to any ensemble method, as a tool to enhance transparency and decision support [12].

These findings indicate that XGBoost provides consistent predictive performance across different datasets and configurations, with demonstrated compatibility with interpretability approaches such as SHAP and LIME.

Table 2. Performance summary of spatial risk mapping models.

Model	Reported Metrics	Predictive Performance	Data Requirements	Adaptability to Low-Resource Environments
FR-RF	AUC = 0.931 [8]; AUC = 0.917 vs. RF AUC = 0.840 [9]	Moderate	Moderate	Moderate
XGBoost	Accuracy = 94.5%, AUC = 0.939, Kappa = 0.890 [10]; std-NRMSE = 0.8451 [11]; Accuracy = 81.42% [12]	High	Moderate	High

summarizes the reported metrics for the spatial risk mapping models reviewed in this section.

The comparison of the two spatial modeling approaches identifies notable differences across our established evaluation criteria. For predictive performance, FR-RF produces varying results across studies, with outcomes ranging from having a slightly lower performance than RF to moderate improvements depending on dataset size and structure. In contrast, XGBoost demonstrates an overall strong performance across different geographic contexts, constantly ranking among the best-performing models in the reviewed literature. As for data requirements, both models use conditioning factor datasets derived from publicly available sources, including topographic, vegetation, climatic, and anthropogenic variables, making them comparable in terms of input data. However, FR-RF requires additional data preprocessing steps, such as feature transformation and input resolution adjustment. These differences may influence their adaptability to low-resource environments. The variability observed in FR-RF results introduces doubt about its behavior under incomplete data. In contrast, XGBoost shows stable performance across variations in dataset characteristics and supports integration with interpretability frameworks such as SHAP to facilitate the translation of model outputs into actionable insights.

3.3.2. Temporal Forecasting Models

ARIMA

The Autoregressive Integrated Moving Average (ARIMA) model is a widely used statistical framework for time-series forecasting. It transforms non-stationary data into stationary form through differencing to then model those resulting series using three parameters: p (autoregressive order), d (differencing order), and q (moving average order). Stationarity is typically assessed using the Augmented Dickey–Fuller (ADF) test, while parameter selection is commonly guided by the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC).

A 2024 study implemented ARIMA using IBM’s Statistical Package for the Social Sciences (SPSS) to forecast monthly rainfall across six stations in the Tartous Governorate of Syria, a coastal Mediterranean region with climatic similarities to the broader MENA zone [13]. The study combines 38 years (1983–2020) of satellite-derived precipitation data from the PERSIANN-CDR dataset, extracted using ArcMap GIS 10.8, with ground-based observations from 71 stations covering 1991–2009. To address discrepancies between the two data sources, the authors applied station-specific linear correction equations, and only stations with a coefficient of determination (R²) above 0.75 were included in the modeling process. ARIMA was subsequently used to forecast one-year-ahead monthly precipitation, which were then evaluated against satellite observations for the October 2020 to September 2021 period. The reported results were consistent across all six stations, with an evident match between ARIMA-predicted values and corrected satellite observations.

A second study applied seasonal ARIMA models within an operational fire risk mapping framework in Mexico [14], using MODIS satellite data from 2003 to 2014. The target variable of this study was the Dead Ratio (DR), a satellite-derived fuel greenness index representing the proportion of non-living vegetation and associated combustible material. The DR was derived from monthly MODIS NDVI composites using the relative greenness and maximum live ratio equations proposed by [14]. The modeling process covered 28 vegetation type combinations across four macro-regions of Mexico (Northwest, Northeast, Center, and South). Model selection was based on AIC, and residual diagnostics were determined using Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) plots, with the retained models having an absent residual autocorrelation at all time lags. Overall, 14 of the 28 vegetation–region combinations achieved R² values above 0.80, while nine fell within the 0.70–0.80 range.

The dependence of ARIMA on linear relationships and stationary input data limits its ability to capture complex interactions between climatic, vegetation, and anthropogenic drivers. As a result, ARIMA is more suitable for forecasting individual fire-relevant variables than modeling fire occurrence probability as a function of multiple interacting factors.

LSTM

Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) networks, are suitable architecture designed to process and effectively analyze sequential time series data. LSTM networks are distinguished for their mitigation of the vanishing gradient problem, a traditional limitation of RNNs, through a gating mechanism capable of capturing long-term dependencies, enhancing model accuracy in time-series applications [15].

One study addressed the randomness and nonlinearity in wind power forecasting using a hybrid model combining Bidirectional LSTM (BiLSTM), Random Forest (RF), the Sparrow Search Algorithm (SSA), and the Firefly Algorithm (FA) for its evaluation [16]. This FA-SSA-BiLSTM-RF architecture relies on BiLSTM for bidirectional time series processing and the enhancement of long-term dependency modeling, while SSA and FA are used for hyperparameter optimization and convergence improvement. The dataset comprised 40,000 records of wind speed, wind direction, temperature, humidity, pressure, and hourly power output, obtained from the Kaggle Wind Power Generation archive covering the period 2015–2020, and a 5:1 training-to-validation split ratio. Model performance was evaluated using accuracy, mean absolute error (MAE), mean squared error (MSE), F1-Score, and prediction duration. The FA-SSA-BiLSTM-RF model achieved a prediction accuracy of 98.5%, converging within 12 to 15 iterations, and maintaining a stable performance under varying load conditions [16]. However, the model’s performance declined under extreme weather, highlighting the abrupt and severe environmental variability of the model. Additionally, the model is computationally intensive, requiring roughly 100 MB of memory, potentially restricting deployment in resource-constrained monitoring environments [16].

The second study focused on developing a remote sensing diagnostic framework for forest risk in China’s Greater Khingan Range, combining static “state indicators” and dynamic “trend indicators” within a comparative LSTM and Random Forest (RF) framework [17]. A multi-source dataset ranging from 2010 to 2024 was compiled from TerraClimate meteorological reanalysis, SRTM topographic data, VIIRS nighttime light imagery, OpenStreetMap road and settlement layers, and NASA MODIS remote sensing products, providing 18 indicators, including meteorological, topographic, vegetation, and human disturbance. The LSTM model was configured with the TensorFlow framework and a sigmoid output layer for binary fire probability estimation, trained using the Adam optimizer with early stopping, and validated through a sliding time window. LSTM results analysis showed an AUC of only 0.63 and an accuracy of 0.60, underperforming compared to the RF model, which achieved an AUC of 0.94 and an accuracy of 0.89 on the same dataset [17]. The authors attribute LSTM’s underperformance to the study’s small sample size and the non-sequential nature of the feature vectors.

These findings indicate that while LSTM models are effective for capturing temporal dependencies, their performance is highly dependent on dataset size, data structure, and computational resources.

AttentionFire_v1.0

AttentionFire_v1.0 was initially designed to capture complex interactions between climate variables, human activity, and burned area factors, with a special focus on the tropical regions of Africa and South America (ASA), which together account for over 70% of global burned areas [18]. AttentionFire_v1.0 relies on a Long Short-Term Memory (LSTM) network supplemented with an attention mechanism, allowing it to detect the temporal and spatial importance of diverse climatic, fuel-related, and anthropogenic components of fire prediction [18]. In contrast to conventional black-box ML models, AttentionFire_v1.0 produces transparent insights into the dominant drivers and their impacts on fire dynamics [18].

One application of this model used data from the Global Fire Emissions Database (GFED4), alongside NCEP-DOE reanalysis, LUH2 land cover data, and population and road density metrics, with a focus on Northern Hemispheric Africa (NHAF), Southern Hemispheric Africa (SHAF), and Southern Hemispheric South America (SHSA) [15]. The model architecture included Temporal Attention for the identification of critical time intervals and Variable Attention for the detection of key drivers [18]. The model was benchmarked against Artificial Neural Networks (ANNs), Decision Trees (DT), Random Forests (RF), Gradient-Boosting Decision Trees (GBDT), and standard LSTM models, using a 90/5/5 train-validation-test split on data ranging from 1997 to 2015. Among all tested models, AttentionFire had the lowest Mean Absolute Error (MAE) and attributed 66–80% of the variability in burned areas to climate wetness indicators [18]. Furthermore, short-term VPD variations were identified as the dominant drivers of fire ignition in NHAF during the wet-to-dry season and the onset of dry season (September to December), while long-term precipitation patterns were found to be the primary influence on fire occurrences in SHAF and SHSA during the wet and wet-to-dry seasons (December to March in SHAF, and November to April in SHSA) [18].

While the model demonstrates clear interpretability and scalability, AttentionFire_v1.0 is computationally demanding, necessitating 22 times longer training time and 141% more memory than simpler baseline models [18].

Table 3. Performance summary of temporal forecasting models.

Model	Reported Metrics	Predictive Performance	Data Requirements	Adaptability to Low-Resource Environments
ARIMA	R2 > 0.75 across 6 stations [13]; R2 > 0.80 for 14/28 vegetation combinations [14]	Moderate	Low	High
LSTM	Accuracy = 98.5% on large wind dataset [16]; AUC = 0.63, Accuracy = 0.60 on smaller forest fire dataset [17]	Moderate	High	Low
AttentionFire_v1.0	Lowest MAE among all tested models (ANN, DT, RF, GBDT, LSTM); explains 66–80% of burned area variability [18]	High	High	Low

presents a performance summary of the temporal forecasting models evaluated in this subsection.

The three temporal forecasting approaches have shown great differences in terms of model complexity and operational requirements. AttentionFire_v1.0 achieved the strongest predictive performance among the evaluated models, with the lowest MAE and the ability to interpret fire drivers. LSTM had a varying performance across datasets, producing high prediction accuracy in large, structured datasets while underperforming in smaller or less sequential inputs. ARIMA produced consistent results in univariate forecasting tasks, with reported R² values generally above 0.70 and often exceeding 0.80. Similarly, the models differ in their data requirements. AttentionFire_v1.0 is configured on multi-source datasets. LSTM processes large volumes of sequential data to leverage its architecture. ARIMA is applied to single-variable time series. Consequently, these differences influence the level of adaptability of each model to low-resource environments. Both AttentionFire_v1.0 and LSTM require high computational capacity and extensive datasets, restricting their functionality in low-resource contexts. In contrast, ARIMA has less demanding input requirements and a transparent modeling process, making it a better fit for resource-limited settings.

3.3.3. Satellite Data Products for Fire Monitoring

MODIS Data Products

The Moderate Resolution Imaging Spectroradiometer (MODIS), onboard NASA’s Terra and Aqua satellites, has provided continuous observations of the Earth’s surface since 2000. Its range of derived data products includes indicators of vegetation condition, land surface temperature, atmospheric moisture, and radiation balance, all of which are relevant to fuel conditions and fire ignition risks. Two recent studies have adopted different MODIS products to address common persistent limitations in data quality and coverage, producing outputs that can support wildfire prediction applications.

The first study introduced HiQ-FPAR, a reprocessed global dataset of the Fraction of Absorbed Photosynthetically Active Radiation (FPAR) covering the period 2000–2023, developed using the Spatio-Temporal Information Composition Algorithm (STICA) [19]. FPAR quantifies the proportion of incoming solar radiation within the 380–710 nm wavelength range, a direct indicator of vegetation greenness, biomass accumulation, and fuel load availability. The standard MODIS FPAR product (MOD15A2H) can be limited due to the spatiotemporal noise caused by cloud cover, aerosol contamination, sensor anomalies, and its 3D radiative transfer algorithm. STICA mitigates these issues through a two-step enhancement process. In the first stage, a Multiple Quality Assessment (MQA) procedure assesses each pixel based on three indicators: Algorithm Path (AP), FPAR Standard Deviation (STD-FPAR), and Relative Time-Series Stability (RETSS), to differentiate between high-quality and less reliable observations. In the second stage, spatial smoothing is performed through the Inverse Distance Weighting (IDW) method across 9 × 9 pixel windows, producing a spatially enhanced component (FPAR_S). Temporal smoothing is then applied using Simple Exponential Smoothing (SES) over adjacent image composites, producing a temporally stabilized component (FPAR_T). The final HiQ-FPAR value for each pixel is computed as a weighted combination of FPAR_S, FPAR_T, and the raw FPAR to identify temporally unstable sources. Results show that HiQ-FPAR outperformed both MODIS FPAR and Sensor-Independent FPAR (SI-FPAR) with an RMSE of 0.130 and an R² of 0.722, compared to 0.154/0.630 and 0.146/0.717 for MODIS and SI-FPAR, respectively. Nevertheless, HiQ-FPAR inherits some limitations from its MODIS foundation. Sensor anomalies in 2001, 2016, and 2022 introduce data gaps that can be directly traced to the reliance of the algorithm on MODIS data. Additionally, validation against biome type B5, the densest and most prone vegetation category, revealed weaker results compared to other biome types because of persistent cloud cover and elevated aerosol concentrations.

The second study used the ELITE-MODIS Surface Longwave Radiation (SLWR) product, providing daily global estimation at 1 km spatial resolution from 2000 to 2023 [20]. The product includes two radiation components: Surface Longwave Upwelling Radiation (SLUR) to reflect thermal energy emitted by the land surface, and Surface Longwave Downwelling Radiation (SLDR) to capture thermal energy returned from the atmosphere to the surface. The methodology includes five MODIS level-1B and level-2 swath products (MOD/MYD02, 03, 05, 06, and 35) from the Terra and Aqua database. Clear-sky SLUR is retrieved through a hybrid regression model that relates MODIS TOA radiances from thermal channels 29, 31, and 32 to surface-emitted radiation; however, under cloudy conditions, SLUR is estimated using the ELITE Broadband Emissivity (BBE) product and (MOD06). Clear-sky SLDR is estimated using the derived SLUR, Column Water Vapor (CWV) from MOD05, and TOA channel 29 radiance, while cloud-sky SLDR is computed through a single-layer cloud model (SLCM) driven by cloud base temperature (CBT). Finally, daily averages are produced by integrating the instantaneous estimates through linear sine interpolation, synchronized with hourly ERA5 reanalysis data. The product showed promising results, based on the validation across 369 in situ flux sites from nine global networks: daily SLDR achieved an RMSE of 25.31 W/m² and R² of 0.86, while daily SLUR achieved an RMSE of 17.77 W/m² and R² of 0.93. However, SLDR performed less uniformly in specific climate types, introducing biases. The authors concluded the main limitations to be the propagation of errors from underlying MODIS products and cloudy-sky retrieval uncertainty.

Collectively, these two studies establish MODIS as a scientifically mature and accessible satellite platform for data retrieval.

3.3.4. Fire Spread Simulation

Cell2Fire

Cell2Fire is an open-source, cell-based wildfire growth simulator designed to support data-driven landscape management and planning [21]. The model partitions a forest landscape into a rectangular grid of homogeneous square cells, each characterized by fuel type, topographic attributes, weather conditions, and moisture content. Fire propagation is monitored using the Canadian Forest Fire Behavior Prediction (FBP) System, which provides models for estimating the Head Rate of Spread (HROS), Flank Rate of Spread (FROS), and Back Rate of Spread (BROS). These three spread rates determine an elliptical fire growth geometry for each burning cell, following the strongest wind direction and updating dynamically at each time step to illustrate changes in environmental conditions. A fire ignited at a given cell spreads outward along eight axes to adjacent cells. Once the fire reaches the center of a neighboring cell, that cell becomes an independent ignition source and initiates its own elliptical spread calculations. Cell2Fire is able to process each burning cell independently, thanks to its implementation in C++ with OpenMP parallelization, achieving a computational speed of 30 times faster than the vector-based simulator Prometheus. Additionally, the modular architecture of the model enables its use as a standalone tool or an integrated component within broader landscape management systems. The model was benchmarked against Prometheus across three experimental datasets: the historical Dogrib fire in Alberta, two sub-instances derived from that landscape, and 10 hypothetical wildfires across 5 areas of British Columbia. Across the British Columbia instances, Cell2Fire achieved a mean 1-MSE of 91% and a mean SSIM of 68%. For the Dogrib simulation, Cell2Fire maintained an average of structural similarity of 87.91% and an average accuracy of 91.82% relative to Prometheus. The authors concluded that the difference between the two simulators is mainly due to the discrete cellular-automata approach of Cell2Fire versus Prometheus’ continuous wave-propagation model.

Cell2Fire assumes homogeneity within each cell, an approximation whose accuracy is highly dependent on spatial resolution. The model is computationally efficient and compatible with satellite-derived inputs such as MODIS products.

ELMFIRE

The Eulerian Level Set Model of FIRE spread (ELMFIRE) is a wildland fire spread model that uses the Eulerian level set method to solve a partial differential equation where a smooth scalar function spreads through the domain, with the fire front defined as the zero contour of that function. Fire spread rates are calculated by the Rothermel semi-physical model, using fuel type, topography, wind dynamics, and fuel moisture content. The model has been extended to simulate Wildland–Urban Interface (WUI) fires through the addition of an urban fire spread feature based on the Hamada model, allowing the smooth integration of wildland and structure fire propagation within a single functional framework.

A recent study conducted in 2026 applied this WUI model to three historical California fires to simulate real-world input variability using wind data, fuel moisture content (FMC), structural representation, and road firebreaks [22]. The results present wind data and FMC as equally dominant sources of uncertainty, with RTMA spatially distributed wind data simulations outperforming point-station WFAS data simulations by more than 50% in burned area discrepancy. FLAMMAP multi-stage processing introduced sufficient uncertainty to reduce burned area accuracy by up to 40% and structure damage accuracy by up to 70% relative to NFDRS-sourced values. Structural representation had a smaller impact on burned area scope (under 15%) but significantly reduced structure damage prediction accuracy by roughly 40% when structures were presented as non-burnable rather than modeled using the Hamada urban spread component. As for the road firebreak, when misrepresented, it could reduce accuracy by up to 60%. The authors concluded based on the three studied fires that wind direction and fuel moisture are the primary non-compensable determining factors of simulation reliability.

WRF-Fire

WRF-Fire is a coupled atmosphere model that combines a fire spread component with the Weather Research and Forecasting (WRF) mesoscale atmospheric model, enabling two-way dynamic interaction between fire behavior and atmospheric conditions. The fire spread component uses the Rothermel model to estimate the rate of spread, while the atmospheric component simulates wind fields, temperature, humidity, and turbulence using a full planetary boundary layer (PBL) scheme. The output of WRF-Fire is a fire-induced heat release modifying the local wind field in real time, which in turn updates the fire spread trajectory, a feedback mechanism absent in one-way simulators such as Cell2Fire and ELMFIRE.

A study applied WRF-Fire version 4.6 to the Jinyun Mountain wildfire in Chongqing, China, using a triple-nested domain configuration with a 300-m atmospheric resolution and a 30-m fire grid [23]. The model was evaluated through hourly observations from six meteorological stations and MODIS-derived burned area data. The study included five PBL parametrization schemes, MYJ, MYNN2, MYNN3, BouLac, and UW, under unified land surface and surface layer configurations. The results varied across schemes, with burned area agreement relative to MODIS ranging from 47.38% to 92.82%. The MYNN3 scheme showed the highest spatial correspondence with the observed burned areas, the lowest systematic bias in 2 m temperature simulations, and a more consistent physical representation of fire-induced turbulence. In comparison, the MYJ and UW schemes failed to capture local circulation changes produced by fire-generated thermal forcing, consistently underrepresenting turbulence intensity. While the BouLac scheme had a lower RMSE in wind speed, its local adaptability to represent non-uniform thermal circulation under complex terrain was limited.

Although WRF-Fire provides a physically detailed simulation framework, it requires substantial data and computational resources. The model relies on reanalysis datasets such as ERA5 for initialization, high-resolution terrain inputs, and extended runtime on computational infrastructure.

Table 4. Performance summary of fire spread simulation models.

Model	Reported Metrics	Predictive Performance	Data Requirements	Adaptability to Low-Resource Environments
Cell2Fire	Mean accuracy vs. Prometheus = 91.82%; mean 1-MSE = 91%; 30× faster than Prometheus [21]	High	High	Moderate
ELMFIRE	Burned area discrepancy > 50% without spatially distributed wind data [22]	High	High	Low
WRF-Fire	Burned area agreement = 92.82% with MODIS under best PBL scheme (MYNN3) [23]	High	High	Low

provides a comparative summary of the fire spread simulation models evaluated in this subsection.

Comparing the reviewed fire spread simulators across predictive performance, data requirements, and applicability to low-resource environments reveals great considerations that should be addressed. All three models produce sufficient accuracy within their respective prediction performances. Cell2Fire achieved a mean accuracy of 91.82% relative to Prometheus, while ELMFIRE reproduced burned areas and structural damage with reasonable agreement under high-quality inputs. WRF-Fire, under its best-performing configuration, reached 92.82% agreement with MODIS-derived burned area observations. However, the models differ substantially in their data requirements. Cell2Fire relies on gridded data inputs, topographic data, and weather inputs that can be obtained from publicly available sources such as MODIS products and digital elevation models (DEMs). In contrast, ELMFIRE shows sensitivity to input data quality, with its reduction in accuracy exceeding 50% in the absence of spatially distributed wind data and decreasing in structure damage prediction by up to 70% when high-quality fuel moisture inputs are unavailable. WRF-Fire requires detailed atmospheric initialization, nested domain configurations, and significant computational resources. These differences influence their adaptability to low-resource environments. The dependence of ELMFIRE on high-quality meteorological inputs may limit its use in regions where such data are limited. Similarly, the computational and infrastructural requirements of WRF-Fire may restrict its deployment in settings lacking dedicated resources. In contrast, Cell2Fire has lower implementation requirements due to its open-source framework and compatibility with widely available satellite data. However, its application remains dependent on the availability of locally validated fuel classification parameters.

4. Discussion

4.1. Comparative Analysis of Reviewed Models

Table 5. Comparative evaluation of reviewed prediction models across three assessment criteria.

Model	Category	Predictive Performance	Data Requirements	Adaptability to Low-Resource Environments
FR-RF	Spatial Risk Mapping	Moderate	Moderate	Moderate
XGBoost	Spatial Risk Mapping	High	Moderate	High
ARIMA	Temporal Forecasting	Moderate	Low	High
LSTM	Temporal Forecasting	Moderate	High	Low
AttentionFire_v1.0	Temporal Forecasting	High	High	Low
HiQ-FPAR (MODIS)	Satellite Data Products	High	Low	High
ELITE-MODIS SLWR	Satellite Data Products	High	Low	High
Cell2Fire	Fire Spread Simulation	High	High	Moderate
ELMFIRE	Fire Spread Simulation	High	High	Low
WRF-Fire	Fire Spread Simulation	High	High	Low

MODIS: Moderate Resolution Imaging Spectroradiometer; FPAR: Fraction of Absorbed Photosynthetically Active Radiation; SLWR: Surface Longwave Radiation; FR-RF: Frequency Ratio-Random Forest; XGBoost: Extreme Gradient Boosting; ARIMA: Autoregressive Integrated Moving Average; LSTM: Long Short-Term Memory; AttentionFire_v1.0: Interpretable machine learning fire model with attention mechanism; Cell2Fire: Cell-based forest fire growth simulator; ELMFIRE: Eulerian Level Set Model of FIRE spread; WRF-Fire: Weather Research and Forecasting model coupled with a fire spread component; HiQ-FPAR: High-quality MODIS Fraction of Absorbed Photosynthetically Active Radiation product.

presents a summary of the within-category benchmarking described in Section 3.3, with each model evaluated according to the three assessment criteria defined in Section 2.3.

4.2. Interpretation of Key Findings

The within-category benchmarking conducted in Section 3.3 shows a consistent pattern that holds across all three modeling categories: as architectural complexity increases, data and infrastructure requirements increase proportionally, and adaptability to low-resource environments decreases. While the identified pattern does not reflect the absolute limitation of the more complex models, it highlights the mismatch between those models’ operational requirements and the data conditions commonly observed in data-scarce Mediterranean regions. Furthermore, this pattern does not imply that simpler models are superior; it simply demonstrates how model selection in data-scarce environments is governed by a different set of criteria than model selection in well-equipped ones.

Another observation concerns the role of interpretability. In the spatial risk mapping category, SHAP analysis significantly increased the applicability rating of XGBoost by resolving the transparency barrier that limits machine learning adoption in operational fire management systems [10,12]. The divergence in SHAP-identified feature importance across the three XGBoost studies, with land cover dominant in one context and topographic variables in another, is itself informative: it confirms that fire driver importance is landscape-specific, which encourages the implementation of SHAP analysis as a critical requirement for any regional deployment rather than an optional enhancement [10,11,12].

The third observation concerns the complementary rather than competing nature of the category-level findings. In the context of wildfire prediction, XGBoost identifies where fire risk is highest across a landscape at a given point in time, while ARIMA detects the evolution of fire-relevant environmental conditions along a temporal trajectory. Cell2Fire simulates fire propagation once ignited based on locally validated fuel data. The reviewed evidence suggests that no single model within any category addresses all three capabilities simultaneously.

Finally, the MODIS data products reviewed in Section 3.3.3 represent a shared enabling infrastructure across all model categories. The availability of over two decades of globally consistent FPAR and surface longwave radiation records [19,20], both freely accessible, suggests that the primary constraint may be less the absence of suitable data products than the absence of structured efforts to integrate these products into locally validated models.

4.3. Candidate Framework for Future Investigation

Drawing on the within-category comparisons summarized in Table 5, a hybrid XGBoost-ARIMA framework including SHAP-based interpretability may be a promising candidate framework for wildfire risk assessment in underrepresented Mediterranean regions. The rationale for this framework integration is derived from category-level findings: ARIMA addresses the temporal dimension of fire risk by tracking the evolution of environmental conditions over time, XGBoost addresses the spatial dimension by mapping susceptibility across a landscape using conditioning factors, and SHAP identifies the leading drivers in a given prediction. The sequential integration of these components, where ARIMA-generated temporal forecasts serve as dynamic inputs for XGBoost alongside static landscape variables, has been explored in a previous study, where such a hybrid framework has outperformed both standalone models [24]. The proposed four-stage framework is illustrated in Figure 3.

The reliance of this configuration on freely accessible MODIS-derived inputs and open-source tools further supports its applicability beyond Morocco to other fire-prone regions across the Mediterranean basin.

It should be noted that this candidate framework has not yet been validated locally, and the reviewed evidence does not eliminate alternative configurations. The framework is presented here as a potential starting point whose individual components have been independently documented in climatically and topographically similar environments [10,11,12,13,14], rather than as an operational recommendation.

5. Conclusions and Future Work

This review assessed different prediction tools across three functional categories alongside the satellite data products that support their deployment. XGBoost and ARIMA emerged as the most consistently documented candidates for data-scarce Mediterranean contexts across the three assessment criteria, with MODIS products providing the accessible data infrastructure to support their deployment, and Cell2Fire representing a promising candidate for future regional integration as locally calibrated fuel data become available. However, several limitations should be acknowledged. First, the systematic search was restricted to Scopus-indexed, English-language publications between 2014 and 2026, which may have excluded relevant studies published in other languages or other periods. Second, the reviewed models were evaluated qualitatively across three assessment criteria rather than through direct quantitative benchmarking on a common dataset; performance comparisons across categories may be subjective. Third, applicability assessments remain prospective pending local field validation. These limitations do not invalidate the findings but clarify that the proposed framework should be treated as a structured starting point for future work rather than a validated operational system.

Future work will focus on testing the proposed XGBoost-ARIMA-SHAP framework within the Moroccan context. Once validated, its adoptability for similar regions within the Mediterranean basin should be assessed.