Integrating explainable AI and multi-objective optimization for wildfire resource allocation under fixed staffing constraints: a decision-oriented framework

Abstract

Emergency response agencies operating under fixed staffing constraints face a structural challenge: translating spatially heterogeneous wildfire risk into operationally feasible allocation decisions. This study presents a decision-support framework integrating explainable artificial intelligence (XAI) and multi-objective optimization for wildfire resource allocation. Using 544 wildfire events from Taiwan (2011–2024), an Extreme Gradient Boosting (XGBoost) classifier predicts fire severity risk, achieving a test-set ROC-AUC of 0.810 and F1-score of 0.691. Shapley Additive Explanations (SHAP) analysis identifies spatial factors (fire station, township, coordinates) and temporal factors (month, year) as dominant predictors, with forest type and meteorological variables contributing secondary roles. The predicted risk distribution is embedded into a tri-objective model minimizing a response-time proxy (a distance-weighted resource concentration index capturing nonlinear response efficiency), workload imbalance, and a burned-area proxy (an empirically calibrated power-law function linking response delay to nonlinear fire spread escalation), solved via the Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) to generate Pareto-optimal solutions under fixed staffing constraints. Balanced allocation—deploying 60–63% of resources to high-risk zones—achieves a response-time proxy of 1.88 (27.7% reduction vs. uniform deployment), while maintaining workload imbalance at 0.18. Efficiency-oriented strategies reduce response time to 1.83 but increase workload concentration by 28%. The Pareto frontier makes efficiency–equity trade-offs explicit and supports reproducible pre-positioning decisions. Although parameterized using Taiwan wildfire data, the prediction–interpretability–optimization chain imposes no region-specific assumptions and can be recalibrated to any context with comparable institutional constraints.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-50836-0.

Introduction

Climate change has intensified the frequency and severity of extreme heat and drought events, thereby amplifying wildfire risk worldwide. The United Nations Office for Disaster Risk Reduction has identified wildfire-related disasters as one of the fastest-growing categories of climate-driven hazards, associated with substantial economic losses and long-term ecological damage¹. Recent wildfire events across Europe, the Americas, and parts of Asia increasingly exhibit characteristics of high intensity, prolonged duration, and wide spatial spread, indicating a shift from isolated incidents toward systemic and cascading hazard dynamics. These trends frame wildfire not only as an environmental phenomenon but also as a critical challenge for disaster risk management and emergency response systems.

Taiwan represents a particularly demanding wildfire context. Located in a subtropical monsoon zone, more than 60% of its land area is forested, with steep mountainous terrain, seasonal dry winds, and low relative humidity during critical periods. These conditions accelerate fire spread and complicate suppression operations. At the same time, operational constraints further compound the challenge: staffing levels at fire stations are largely fixed, equipment mobility is limited by terrain, and cross-station coordination often requires administrative approval. During dry and windy seasons, individual stations may face multiple concurrent incidents, leading to resource overload, delayed containment, and uneven workload distribution. Addressing this mismatch requires both accurate risk assessment and optimization-based resource planning, yet these two components remain largely disconnected in current practice.

Wildfire research has rapidly advanced in data-driven risk prediction with the growth of meteorological, topographic, and remote-sensing data. Recent work increasingly relies on machine-learning models to capture nonlinear interactions and produce susceptibility or risk maps, with comprehensive reviews documenting the maturation of this research stream and the recurring importance of weather, terrain, vegetation, and human-activity drivers². In parallel, the operations-research literature has developed a well-established body of emergency logistics and response optimization models, including facility location and resource allocation under scarcity and uncertainty^3,4. On the methodological side, evolutionary multi-objective optimization has become a standard approach for generating interpretable trade-off sets in complex, nonconvex decision spaces^5,6. Despite established wildfire suppression optimization efforts, these models often remain separated from data-driven, locally calibrated risk estimates⁷. Explainable artificial intelligence methods—most notably SHAP—have further strengthened the interpretability of complex predictors⁸ and are increasingly used in fire-risk modeling to diagnose feature contributions and spatial heterogeneity. However, prediction, explanation, and allocation are still commonly treated as loosely connected components rather than as an integrated decision-support chain.

Three critical gaps emerge from this disconnection. First, existing wildfire prediction models typically output spatial risk scores or occurrence probabilities, whereas resource allocation models often rely on stylized, uniform, or historical-average risk assumptions. In operationally constrained settings such as Taiwan, this disconnect can lead to high-risk areas being under-allocated while lower-risk areas remain over-resourced. No systematic mechanism currently exists to align prediction outputs with allocation inputs. Second, many dispatch and allocation models implicitly assume unrestricted resource mobility, overlooking the operational reality of fixed staffing, station-based deployment, and the need for cross-station and cross-period coordination. As a result, mathematically efficient solutions may prove infeasible in practice. Third, although black-box predictive models remain prevalent in wildfire risk assessment, public safety governance requires decisions to be transparent, auditable, and justifiable. Limited interpretability constrains the adoption of AI-based tools in policy and operational contexts where accountability is essential. To address these gaps, this study develops a Decision Support Framework for Firefighting Resource Allocation (DSF-FRA) as a three-stage decision chain: (i) a data-driven severity prediction module generating spatially heterogeneous, locally calibrated risk estimates from observable pre-event conditions; (ii) a SHAP-based interpretability module that renders the probability-based risk stratification auditable at the feature level, supporting governance communication; and (iii) a multi-objective optimization module embedding the stratified risk inputs into a tri-objective allocation model solved via MOEA/D, generating Pareto-optimal deployment strategies under fixed staffing constraints. Although instantiated with Taiwan wildfire data, the framework imposes no region-specific assumptions and is designed for recalibration to other emergency response contexts.

Empirically, the study uses 544 Taiwan wildfire events to examine whether AI-derived, locally calibrated risk estimates alter the geometry of efficiency–equity trade-offs relative to conventional uniform-risk assumptions, and whether SHAP-based stratification can serve as an auditable translation layer between probabilistic prediction and constrained allocation. The remainder of this paper is organized as follows: “Literature review” section reviews relevant literature; “Research methodology” section presents the methodology; “Results and analysis” section reports empirical results; “Discussion and Conclusions” sections discuss implications and conclude.

Literature review

This section reviews AI-driven emergency management research with a focus on wildfire risk prediction, emergency resource allocation, and explainable artificial intelligence (XAI). Rather than providing a descriptive overview alone, the review emphasizes the conceptual fragmentation across prediction, allocation, and interpretability, thereby clarifying the theoretical motivation for the proposed integrated decision-support framework (DSF-FRA).

Wildfire risk prediction and data-driven models

Wildfire risk prediction has accelerated with the operational availability of satellite fire products, gridded meteorological datasets, and geospatial covariates describing terrain, land cover, and human presence. Prediction tasks in this literature are typically operationalized into three output types: ignition or occurrence probability surfaces, burned area or spread-related outcomes, and severity proxies that approximate suppression difficulty or potential loss. This task differentiation matters for decision support because each output type implies different label definitions, temporal immediacy, and pathways for translating predictions into actionable deployment inputs. Nonlinear learners are attractive for operational risk mapping because they capture threshold effects and interaction structures among weather, fuel, terrain, and human factors that linear specifications cannot represent.

A major stream focuses on spatial susceptibility or ignition probability mapping. Here, models are trained on historical fire occurrence records and spatial predictors to produce continuous risk surfaces used for prevention planning and resource pre-positioning. Recent work illustrates the use of ensemble-based machine-learning for ignition probability mapping and sensitivity assessment, emphasizing that model interpretation should be coupled with mapping because planning decisions depend not only on relative risk ranks but also on understanding which covariates drive risk heterogeneity across space⁹. However, this stream also faces recurring methodological challenges that directly affect operational reliability, including class imbalance and rare-event learning, spatial autocorrelation that inflates apparent accuracy under random cross-validation, and temporal transfer issues when models trained on one climatic regime are applied to anomalous seasons.

A second stream emphasizes spatiotemporal prediction by integrating multi-source observations. In this setting, satellite fire products serve as both training labels and near-real-time situational signals. For burned area characterization, the MODIS Collection 6 burned area product (MCD64A1) has been widely used as a standard reference dataset because it provides consistent long-term global burned area mapping suitable for retrospective modeling and cross-regional benchmarking¹⁰. For active fire detection, the VIIRS 375 m active fire product has been shown to substantially improve detection of smaller fires relative to coarser-resolution products, supporting applications that require higher spatial granularity and more timely identification of fire activity. This distinction is consequential for modeling: burned area products often align better with season-scale risk analyses, whereas active fire detections are more suitable for short-horizon monitoring and dynamic updating of risk states¹¹.

Fire-weather indices provide a complementary pathway for linking meteorology to fuel moisture and spread potential through physically grounded yet operationally compact representations. The Canadian Forest Fire Weather Index (FWI) system, originally developed as an operational index suite, remains one of the most widely adopted representations in fire-risk studies and early warning practice because it transforms routine meteorological observations into interpretable indices that proxy ignition potential, spread conditions, and fuel dryness¹². From a modeling standpoint, such indices function as engineered features that reduce dimensionality and embed domain knowledge, thereby stabilizing learning when direct fuel measurements are unavailable or when remote-sensing fuel proxies are noisy.

Despite these advances, the dominant endpoint of prediction studies remains a risk score, probability surface, or susceptibility rank. In many papers, this predictive output is treated as the final deliverable, while the translation from probabilistic risk estimates to resource requirements is left implicit. Specifically, prediction outputs rarely specify how a given risk level should map to station-level readiness, cross-station support thresholds, or workload implications under fixed staffing. This is not merely a missing implementation detail: the absence of an explicit translation layer makes it difficult to audit why a particular deployment pattern follows from a model and limits the portability of predictive models into constrained firefighting organizations where decisions must be justified, documented, and repeated under evolving risk conditions.

Emergency resource allocation and multi-objective optimization in disaster operations

Emergency resource allocation is inherently multi-objective, as response systems must simultaneously manage response time, spatial coverage, operational workload, and risk exposure under fixed and often scarce resources. Within disaster operations research, a broad class of optimization frameworks has been developed to address logistics and response planning problems, including facility location, dispatching, and pre-positioning under uncertain demand and time-varying hazard conditions^3,4,13. These models formalize the fundamental tension between efficiency-oriented objectives, such as minimizing response delay, and system-level constraints imposed by limited resources and institutional structures.

In emergency medical services (EMS), location and deployment models have long served as a methodological foundation for resource allocation under uncertainty. Canonical formulations such as the maximal covering location problem (MCLP) explicitly conceptualize coverage as a constrained planning objective rather than a guaranteed service outcome¹⁴. Subsequent extensions incorporated stochastic demand, vehicle relocation, and dynamic redeployment policies, highlighting how operational workload and service equity emerge as endogenous consequences of allocation decisions^15,16. These developments established a station-based planning logic that remains directly relevant for other emergency response systems operating under fixed staffing.

Wildfire suppression and initial attack planning adopt related allocation principles but face additional operational and environmental complexities. In this domain, allocation models have examined the deployment of suppression resources under scenario uncertainty, limited availability, and temporal escalation of fire behavior, while emphasizing that feasible solutions are conditioned by organizational realities such as fixed station staffing, jurisdictional boundaries, and mutual aid agreements^7,17. Compared with stylized EMS settings, wildfire response systems operate under highly heterogeneous spatial risk patterns and pronounced temporal variability driven by weather, fuel conditions, and terrain, which complicates the direct transfer of classical coverage-based formulations¹⁸. Recent studies further demonstrate that effective wildfire resource allocation must account for simulation-based uncertainty propagation, cascading fire behavior, and dynamic routing considerations, particularly when suppression resources are limited and response priorities shift rapidly^19–21.

From a methodological perspective, evolutionary multi-objective optimization techniques, including NSGA-II and MOEA/D, have become standard tools for exploring trade-offs in nonconvex and combinatorial allocation problems^5,6. These approaches are particularly suitable when objectives such as response efficiency, workload balance, and risk mitigation cannot be meaningfully aggregated into a single scalar objective. In the context of emergency and wildfire response, multi-objective formulations have been used to explicitly examine efficiency–equity tensions, spatial imbalance, and robustness considerations under uncertainty^22–24.

However, in many applied allocation studies, the risk component is represented using static risk surfaces, historical averages, or simplified proxies that are treated as exogenous to the optimization process. Such representations may be insufficient when risk is spatially heterogeneous, temporally dynamic, and unevenly distributed across stations and jurisdictions. When risk inputs are weakly coupled to local operational realities, optimization results may appear efficient in aggregate but induce excessive workload concentration, spatial inequity, or coordination pressure at the station level²⁵.

This limitation motivates allocation frameworks in which risk inputs are derived from data-driven models that are locally calibrated and explicitly heterogeneous, and in which the implications of risk variation for station workload, cross-station coordination, and equity outcomes are systematically examined under fixed staffing constraints. Rather than treating risk as a background parameter, this perspective positions risk estimation as an integral component of the allocation problem itself, aligning prediction outputs with allocation-relevant decision structures.

Explainable AI and auditable decision support in public safety

In public safety and emergency management, predictive accuracy alone is rarely sufficient to justify operational or policy adoption. Deployment decisions are embedded in administrative, legal, and political contexts that require explanations to be documented, communicated, and audited. As a result, decision-support systems must satisfy not only performance criteria but also requirements of transparency, traceability, and institutional accountability^26,27.

Explainable artificial intelligence (XAI) methods have emerged in response to these governance constraints. Among them, SHAP provides a theoretically grounded framework for attributing marginal feature contributions to individual predictions, enabling additive and locally consistent explanations for complex machine learning models⁸. Complementary approaches such as LIME approximate local decision boundaries through interpretable surrogate models, offering case-specific explanations without requiring full access to model internals²⁸. Survey studies further synthesize these techniques and evaluate their strengths and limitations in terms of fidelity, stability, and human interpretability^29–31.

In wildfire risk assessment and other hazard-related applications, explainability is particularly salient because model outputs directly influence spatial prioritization and resource allocation decisions that affect communities and jurisdictions. Risk models in this domain typically integrate heterogeneous predictors, including meteorological variables, land cover characteristics, topography, infrastructure proximity, and historical incident patterns. Explanation tools enable analysts to decompose these complex inputs into interpretable components, facilitating scrutiny of why certain locations or periods are classified as higher risk and supporting communication with operational stakeholders^2,32. Recent reviews in disaster risk management further document the expanding role of XAI in hazard-related decision processes, while emphasizing unresolved challenges in translating explanation outputs into actionable and auditable planning decisions³³. Applications in adjacent fire safety domains further illustrate how SHAP-based interpretability can translate into policy-actionable insights. Liu et al. (2025) analyze over 100,000 UK residential fire records using Bayesian-optimized XGBoost with multi-level SHAP analysis, finding that emergency response attributes—particularly alarm delays and detection times—account for 51% of predictive performance in fire spread risk, with direct implications for resource allocation and early warning system design³⁷. Although urban building fires and wildland fires differ structurally in ignition mechanisms and suppression logistics, the underlying governance logic is analogous: SHAP attributions provide the evidentiary basis for connecting model outputs to specific, auditable policy levers rather than treating predictions as uninterpretable risk scores.

Despite growing adoption, most existing applications treat explainability as a post hoc diagnostic or visualization layer applied after model development. Explanations are generated to interpret predictions, but they are rarely integrated into downstream decision mechanisms. In particular, explanation outputs are seldom used to inform how predicted risks are translated into allocation priorities, workload expectations, or coordination requirements across response units. As noted in critiques of XAI in high-stakes domains, explanations that remain detached from decision logic risk becoming justificatory artifacts rather than instruments of accountable decision-making^34,35.

This separation limits the practical role of XAI in operational settings. When explanation outputs remain disconnected from allocation logic, they provide no guidance on how efficiency–equity trade-offs are produced or justified under fixed resource constraints—and for public safety agencies subject to staffing rigidity and external accountability, this gap directly undermines the defensibility of deployment decisions. The present framework addresses this gap by treating explainability as a functional requirement of the end-to-end decision chain rather than a post hoc visualization layer.

Theoretical gaps and research positioning

Despite substantial progress in disaster operations research, emergency resource allocation, and explainable artificial intelligence, several structural gaps remain insufficiently addressed when these strands are considered jointly.

First, a persistent disconnect exists between risk prediction and allocation modeling. In wildfire and disaster risk assessment, predictive studies typically output probabilities, scores, or susceptibility rankings, whereas allocation and optimization models often rely on stylized, static, or uniform representations of risk. The mechanisms by which heterogeneous, data-driven risk estimates should be transformed into allocation-relevant inputs—such as station-level workload implications, spatial prioritization, or coordination pressures—are rarely specified in a systematic manner. As a result, prediction outputs and allocation inputs are frequently aligned only implicitly, limiting both interpretability and operational relevance. This gap has also been noted in disaster operations research, where stochastic or risk-aware models are often developed independently of the decision mechanisms that ultimately consume their outputs²³.

Second, many allocation formulations insufficiently represent organizational constraints that are central to real-world firefighting systems. Fixed staffing levels, station-based deployment structures, and cross-jurisdictional mutual aid arrangements impose rigid boundaries on feasible response configurations. While these constraints are well recognized in practice, they are frequently simplified or omitted in optimization models, particularly when risk is treated as an exogenous surface rather than an endogenous driver of workload imbalance and coordination demand. Consequently, mathematically efficient solutions may be difficult to operationalize, reducing the practical value of otherwise sophisticated multi-objective formulations.

Third, although explainable artificial intelligence methods provide tools for interpreting complex risk models, their role in decision support remains largely confined to post hoc explanation. Existing studies typically use explanation techniques to justify or communicate predictions, but do not integrate explanatory outputs into the logic of allocation or trade-off analysis. As a result, explainability rarely contributes to auditable reasoning about how efficiency, equity, and risk exposure are balanced under fixed resource constraints. Recent critiques further argue that post hoc explanations alone are insufficient in high-stakes decision contexts, as they do not ensure that explanations meaningfully constrain or inform the underlying decision logic³⁴.

Positioned at the intersection of these gaps, this study develops an integrated decision-support framework that explicitly links data-driven wildfire risk estimation, explainable modeling, and multi-objective resource allocation. Rather than treating prediction, explanation, and optimization as separable modules, the framework aligns them within a single decision chain designed to support operational feasibility and governance accountability. By examining how heterogeneous risk estimates propagate through allocation decisions under fixed staffing constraints, the study provides a structured approach for diagnosing efficiency–equity trade-offs in wildfire response planning.

Research methodology

This study develops a decision-support framework for firefighting resource allocation by integrating AI-based fire severity prediction, explainable model interpretation, and multi-objective optimization. Rather than optimizing response outcomes under assumed certainty, the framework explicitly examines efficiency–equity trade-offs under fixed staffing constraints. The methodological design consists of four stages: (i) data integration and preprocessing, (ii) fire severity prediction, (iii) model interpretability analysis, and (iv) multi-objective decision simulation.

Data sources and feature selection

Wildfire event data were obtained from the Forest Fire Dashboard maintained by the Forestry and Nature Conservation Agency, Ministry of Agriculture, Taiwan (https://mapportal.forest.gov.tw/portal/apps/opsdashboard/index.html#/f912faeb33614c3f85ba304d3566faca). The system integrates official Forest Fire Incident Records from the Taiwan Forestry Bureau covering the period 2011–2024 and monthly meteorological observations provided by the Central Weather Administration. After temporal and spatial alignment, the merged dataset comprised 544 complete wildfire events.

The wildfire records include county, fire location, burned area, reporting and extinguishing times, and responsible operational units. Meteorological variables include average temperature, relative humidity, precipitation, and wind speed. To ensure both physical interpretability and statistical representativeness, three categories of explanatory variables were selected: (i) climatic factors (average temperature, relative humidity, mean wind speed), (ii) topographic factors (elevation, slope, land cover type), and (iii) temporal and administrative factors (year, month, forest compartment, operational division).

Fire severity was operationally classified into three levels by jointly considering burned area and suppression duration to better reflect operational impact. Events were labeled as high severity if burned area exceeded 20 hectares or suppression duration exceeded 12 h; medium severity if burned area was between 5 and 20 hectares or suppression duration between 4 and 12 h; and low severity if burned area was below 5 hectares and suppression duration below 4 h. Where the two criteria yielded conflicting classifications for a single event, the higher-severity category was assigned to ensure conservative risk stratification. Of the 544 events analyzed, 156 (29%) were classified as high severity, 248 (46%) as medium severity, and 140 (26%) as low severity.

All continuous variables were standardized using Z-score normalization, and categorical variables were encoded using label encoding. The dataset was randomly divided into 80% for training and 20% for testing to support external validation and model generalization. To assess potential temporal non-stationarity, Mann-Kendall trend tests were applied to annual event frequency and annual high-severity-event prevalence rate over the 2011–2024 period. While event frequency exhibits a statistically significant upward trend (Kendall τ = 0.500, p = 0.013), consistent with climate-driven intensification, the high-severity-event prevalence rate shows no significant temporal trend (τ = −0.331, p = 0.100), supporting the stability of the severity structure used for model training and risk stratification.

Fire severity prediction model

Fire severity was predicted using Extreme Gradient Boosting (XGBoost), an ensemble learning algorithm that aggregates multiple decision trees within a gradient boosting framework to reduce both bias and variance. XGBoost is well suited for heterogeneous and nonlinear data structures commonly observed in wildfire datasets.

Key hyperparameters were specified as follows: learning rate = 0.1, maximum tree depth = 6, number of trees = 500, subsample ratio = 0.8, column sampling ratio = 0.8, and L2 regularization coefficient = 1. To address class imbalance across severity levels, imbalance handling was incorporated within a one-vs-rest training strategy. Model training employed five-fold cross-validation to mitigate overfitting.

On the testing dataset, the model achieved a ROC-AUC of 0.810, precision of 0.622, recall of 0.778, and an F1-score of 0.691, demonstrating stable discriminative capability under heterogeneous and nonlinear conditions. The present study addresses class imbalance through cost-sensitive learning via the scale_pos_weight parameter, which adjusts the gradient contribution of minority-class samples during tree construction without altering the underlying data distribution. Alternative strategies include resampling techniques such as SMOTE and random undersampling, which operate on the training sample composition. Liu and Zhuang (2024) demonstrate that combining random undersampling with SMOTE can substantially improve F1 performance in building fire loss prediction under severe class imbalance, elevating average F1 scores by 14–26% points relative to unbalanced baselines³⁶. In the present wildfire context, the moderate imbalance ratio (1:2.49) and the operational requirement for calibrated probability outputs to drive the downstream risk stratification and optimization module favor cost-sensitive training over resampling, as synthetic oversampling can distort the predicted probability distribution in ways that affect the reliability of tercile-based stratum boundaries. Although formal multicollinearity diagnostics were not applied as a preprocessing step, XGBoost’s tree-based architecture is structurally insensitive to linear dependencies among predictors. TreeExplainer computes exact Shapley values derived from the ensemble structure rather than linear approximations, and the stability of the SHAP feature ranking across cross-validation folds suggests that collinearity-induced attribution instability is unlikely in the present setting.

Model interpretability analysis

To enhance transparency and avoid black-box decision-making, Shapley Additive Explanations (SHAP) were employed to quantify feature contributions. SHAP is grounded in cooperative game theory and decomposes each prediction into marginal contributions of individual features.

SHAP analysis reveals that spatial and administrative identifiers—including fire station, township, forest compartment, and geographic coordinates—exert the strongest marginal influence on predicted fire severity. Temporal indicators, particularly month and year, contribute at a secondary level, followed by vegetation proxies such as forest type and land cover. Meteorological variables, including wind speed, temperature, and relative humidity, contribute at a tertiary level but remain non-negligible. High wind speed and high temperature are positively associated with severe fire outcomes, while relative humidity exhibits a consistently negative relationship, consistent with thermodynamic mechanisms of wildfire spread. These patterns are consistent with wildfire science evidence that ignition likelihood and severity escalation are strongly shaped by location-specific factors, including accessibility, land-use patterns, fuel continuity, and management boundaries, in addition to meteorological drivers. The dominance of spatial predictors over meteorological variables reflects the importance of local structural and administrative context in determining fire outcomes in this dataset. The consistency between statistical attribution and physical interpretation aligns with the interpretability principles articulated by Lundberg and Lee⁸. A methodological caveat warrants acknowledgment: spatial and administrative identifiers may partially function as proxies for unobserved confounders, including land management practices, terrain accessibility, and differential reporting intensity across jurisdictions. While the consistency between SHAP attributions and established wildfire science evidence supports the domain validity of these patterns, causal interpretation of individual feature contributions should be approached with caution. The SHAP module is employed here as an auditing instrument for risk stratification rather than as a causal inference mechanism. Within the DSF-FRA, SHAP analysis serves a governance function that extends beyond post hoc model explanation. Rather than generating feature importance visualizations as a terminal output, the SHAP module provides the interpretive bridge between the prediction stage and the allocation stage. Concretely, the predicted severity probabilities from “Fire severity prediction model” section are stratified into three risk tiers (high, medium, low) based on terciles of the predicted probability distribution; these tiers become the demand structure that drives the optimization model in “Multi-objective optimization framework” section. SHAP analysis operationalizes this bridge by making the risk stratification auditable: it reveals which feature combinations drive risk elevation in each spatial unit, allowing decision-makers to verify that the tiers reflect genuine hazard heterogeneity rather than opaque model artifacts. This auditability is essential in public safety governance contexts where allocation decisions must be documented and defensible. By ensuring that the predicted risk structure is interpretable before it enters the optimization process, the SHAP layer ensures that the pathway from observed pre-event conditions to Pareto-optimal deployment configurations is transparent, traceable, and reproducible—properties that standalone prediction models and conventional dispatch heuristics cannot jointly provide.

Multi-objective optimization framework

Following fire risk estimation, a multi-objective optimization model was developed to evaluate firefighting resource allocation under fixed staffing constraints. For strategic planning purposes, potential fire locations are aggregated into three risk strata (high, medium, low) based on terciles of predicted severity probabilities from the AI model. The decision variables () represent the proportions of total available firefighting capacity allocated to each stratum, subject to = 1, and ≥ 0.05 for all k ∈ {H, M, L}. This formulation reflects pre-positioning decisions at the planning level and ensures consistency with the optimization results reported in “Multi-objective optimization and Pareto front analysis” section.

1

where n = 3 denotes the number of risk strata, and α and β are shape parameters calibrated analytically from historical dispatch records via a two-point anchoring procedure (α = 0.313, β = 0.354; full calibration procedure in Table S3, Supplementary Material). This proxy captures the nonlinear relationship between resource concentration and response efficiency using a minimal parametric structure.

Minimizing workload imbalance ()

This objective measures distributional equity across response units, where denotes the mean allocation proportion across all strata:

2

This measure captures organizational workload pressure arising from over-concentration or under-utilization of resources, and serves as an operational proxy for allocation equity.

Minimizing expected burned area

Expected burned area increases nonlinearly with response delay. The coefficients α′ and β′ are distinct from those in Eq. (1) and were estimated from historical wildfire records (2011–2024) using a response-time-binned OLS regression in log-log space that helps mitigate the reverse-causality bias present in incident-level regression (α′ = 0.052, β′ = 3.587, R² = 0.76; Table S3, Supplementary Material). The exponent β′ > 1 confirms superlinear escalation of fire spread with response delay, consistent with established fire behavior models:

3

The tri-objective problem defined by Eqs. (1)–(3) was solved using the Multi-Objective Evolutionary Algorithm based on Decomposition⁶, with a Tchebycheff decomposition scheme, population size of 120, 250 iterations, and neighborhood size of 20. These settings are consistent with standard configurations for tri-objective problems of comparable dimensionality. The algorithm generates a set of Pareto-optimal allocations that collectively characterize the efficiency–equity–control trade-off surface under fixed staffing constraints.

Results and analysis

Model performance and risk prediction structure

This study analyzed 544 forest fire events in Taiwan recorded between 2011 and 2024. Events were spatially concentrated in central mountainous regions and exhibited strong seasonal patterns, with higher incidence during the dry season from October to April. Of the 544 events, 156 (29%) were classified as high severity, 248 (46%) as medium severity, and 140 (26%) as low severity, reflecting a moderately imbalanced but operationally realistic severity structure that provides a meaningful basis for risk-oriented decision analysis.

To ensure causal consistency and prevent information leakage, the preprocessing procedure deliberately excluded all post-event outcomes and response-related variables, including realized burned area, suppression timestamps, and indicators of personnel or equipment deployment such as numbers of firefighters, vehicles, or aircraft sorties. Only variables observable prior to each fire event (ex ante conditions) were retained as predictors. These included spatial identifiers, environmental conditions, topographic attributes, vegetation proxies, and administrative features. This design ensures that the predictive model captures latent wildfire risk conditions rather than endogenous effects of response intensity or institutional intervention.

The final classification model incorporated 25 explanatory variables (see Table S1 in the Supplementary Material for the complete feature list with descriptions) and employed an 80/20 split for training and testing. An Extreme Gradient Boosting (XGBoost) classifier was used to predict wildfire severity levels across the three defined categories (high, medium, and low). For performance evaluation, a one-vs-rest formulation was adopted to assess the model’s ability to discriminate high-severity events from the remaining categories, yielding a test-set ROC-AUC of 0.810 (Table 1). Using a decision threshold of 0.20, identified through five-fold cross-validation to maximize the F1 score on held-out training folds prior to final testing, the model attained a precision of 0.622, recall of 0.778, and F1-score of 0.691. Rather than emphasizing absolute predictive superiority, these results demonstrate stable discriminative capability under heterogeneous and nonlinear conditions, supporting the use of predicted risk probabilities as structured inputs for downstream allocation and decision analysis.

Table 1.

Performance of the main XGBoost wildfire classification model.

Dataset	Metric	Value
Test set	ROC-AUC	0.810
	Precision	0.622
	Recall	0.778
	F1-score	0.691
	Decision threshold	0.20

All metrics were computed on the held-out test set using the threshold that maximized the F1 score on held-out cross-validation folds. A complete list of the 25 explanatory variables used in the model is provided in Table S1 (Supplementary Material).

The predicted probability distribution exhibited pronounced skewness, with elevated risk concentrated in specific spatial units and seasonal contexts. The reported ROC-AUC of 0.810 is consistent with benchmarks from comparable wildfire severity classification studies using administrative and spatial predictors in data-constrained settings, where AUC values typically range from 0.75 to 0.85 depending on label definition and spatial resolution. The AUC–F1 gap reflects the moderate class imbalance (28.7% positive prevalence) combined with the deliberate use of a low decision threshold (0.20), selected via five-fold cross-validation to prioritize recall for high-severity events—operationally appropriate given the asymmetric costs of false negatives in emergency resource planning. Such clustering implies that uniform allocation of firefighting resources may be inefficient, as it fails to account for localized escalation of wildfire risk. Instead, region-specific and risk-informed assessments can provide quantitative guidance for patrol prioritization, early warning deployment, and pre-positioning strategies under constrained resources. For subsequent allocation analysis, predicted wildfire risk probabilities were stratified into three relative risk tiers—high, medium, and low—based on terciles of the predicted probability distribution. This distribution-based stratification ensures balanced representation across risk tiers and facilitates systematic comparison of allocation strategies under heterogeneous risk conditions.

To enhance interpretability, SHAP values were employed to quantify the marginal contribution of each feature to the model output. The SHAP summary plot (Fig. 1) shows that spatial and administrative identifiers—such as station, township, compartment, and geographic coordinates—exert the strongest influence on predicted fire severity, followed by temporal indicators such as month and forest type. Although the model incorporates meteorological, topographic, vegetation, and administrative variables, SHAP analysis reveals that spatial context dominates risk differentiation in this setting. Figure 2 further corroborates this pattern by ranking predictors according to their mean absolute SHAP values, confirming that spatial and administrative variables consistently dominate model explanations, while meteorological and vegetation-related factors play a secondary but non-negligible role. From a wildfire science perspective, this pattern is consistent with evidence that ignition likelihood and escalation potential are strongly shaped by location-specific factors, including accessibility, land-use patterns, fuel continuity, and management boundaries. Importantly, because only ex ante variables were included, the SHAP explanations reflect latent hazard structures rather than operational reactions, aligning with interpretability and causal consistency principles emphasized in explainable AI research⁸. Together, these results provide a transparent and domain-consistent representation of wildfire risk suitable for informing resource allocation decisions.

Fig. 1.

SHAP summary plot illustrating the feature-level contributions to wildfire severity classification.

Fig. 2.

Mean absolute SHAP values showing the overall importance of each predictor in the main classification model.

Note

Each dot represents the SHAP value for one observation; the color gradient (low to high) indicates the normalized value of the corresponding feature.

Multi-objective optimization and pareto front analysis

After obtaining the spatial risk distribution from the AI model, the next step was to translate these predictions into resource allocation decisions. A tri-objective optimization framework was therefore formulated to evaluate alternative pre-positioning strategies under limited response capacity. The three objectives represent key dimensions of emergency response performance: minimizing the response-time proxy (efficiency), minimizing workload imbalance (equity), and minimizing a burned-area proxy ​ (control effectiveness).

The burned-area proxy is derived from regression analysis of 2011–2024 historical wildfire records as an empirical power-law function of the response-time proxy (R² = 0.76), consistent with Eq. (3) in Sect.  3.4. This specification reflects the nonlinear escalation of fire extent with response delay: early suppression achieves disproportionately large reductions in final burned area, while delayed response produces rapidly compounding losses.

Decision variables represent the proportions of total available firefighting capacity allocated to the three risk strata (). The allocation proportions satisfy , with a minimum allocation constraint of 0.05 per stratum to prevent complete neglect. The optimization was solved using the Multi-Objective Evolutionary Algorithm based on Decomposition⁶, which employs a Tchebycheff decomposition scheme to maintain solution diversity while promoting Pareto convergence. A population size of 120, 250 iterations, and a neighborhood size of 20 were adopted, consistent with standard configurations for tri-objective problems of comparable dimensionality. A total of 116 non-dominated solutions were obtained.

The Pareto front reveals clear and interpretable geometric trade-offs. In the ​–​ plane (Fig. 3), an inverse relationship emerges, indicating an efficiency–equity tension under fixed staffing. In the ​–​ plane (Fig. 4), diminishing returns are observed, where further concentration of resources yields progressively smaller reductions in expected fire impact. In the –​ plane (Fig. 5), a U-shaped relationship appears, suggesting that both excessive concentration and excessive dispersion undermine control effectiveness.

Fig. 3.

Pareto front between f₁ (response time proxy) and f₂ (workload imbalance).

Fig. 4.

Pareto front between f₁ (response time proxy) and f₃ (burned area proxy).

Fig. 5.

Pareto front between f₂ (workload imbalance) and f₃ (burned area proxy).

These patterns indicate that resource allocation forms a continuous decision spectrum rather than discrete strategy groups. The operational and policy implications of this spectrum structure are examined in Sect.  4.3.

Allocation spectrum and policy implications

Building on the Pareto analysis in Sect.  4.2, the 116 non-dominated solutions do not form distinct strategic clusters. Instead, they span a continuous allocation spectrum ranging from highly concentrated deployment in high-risk zones to increasingly uniform distribution across risk strata.

At one end of the spectrum, solutions allocating more than 65% of capacity to high-risk strata achieve the smallest response-time proxy (≈1.80–1.84) and lowest burned-area values (≈0.02–0.03). These configurations represent a surge-style operational logic that prioritizes rapid containment but produces substantial workload concentration (≈0.22). This end of the spectrum reflects short-horizon control at the expense of workforce sustainability.

As allocations gradually shift toward more even distribution, workload imbalance decreases systematically, with ​declining from approximately 0.22 in concentrated solutions to below 0.10 in dispersed solutions. However, both the response-time proxy and the burned-area proxy increase substantially:rises by approximately 20–45% and increases by a factor of 3–15 relative to concentrated solutions. Near-uniform allocations therefore achieve minimal workload imbalance but correspond to delayed control and sharply increased expected fire impact.

Between these two extremes, a near-knee region emerges in the Pareto geometry, typically around ≈0.60, ≈0.30, and ≈0.10. Solutions in this region achieve near-optimal performance across all objectives: ​ remains within 5% of the concentrated-end minimum, improves by approximately 20%, and ​ remains comparably low, all without inducing the severe trade-offs observed at either extreme (Table 2). This region represents an institutionally feasible robustness zone, where minimum coverage is maintained in lower-risk areas while sufficient capacity is preserved for high-risk escalation.

Table 2.

Representative allocation ranges and associated objective values across the Pareto spectrum.

Allocation region		​	​		​	​
Concentrated end	0.65–0.69	0.23–0.28	0.06–0.09	1.80–1.84	0.21–0.24	0.020–0.029
Near-knee region	≈ 0.60	≈ 0.30	≈ 0.10	1.82–1.90	0.17–0.21	0.02–0.03
Dispersed end	0.40–0.47	0.28–0.33	0.19–0.32	2.20–2.60	0.04–0.09	0.09–0.30

The near-knee region denotes solutions where marginal improvement in any single objective induces disproportionate deterioration in others, representing a stability zone for multi-objective performance. Lower f₁ indicates shorter response time, lower f₂ indicates more balanced workload, and lower f₃ indicates reduced expected burned area.

Rather than representing discrete strategies, the Pareto solutions therefore illustrate a continuous decision logic for wildfire response under fixed staffing. The near-knee region identifies an operationally robust allocation range that balances efficiency, equity, and control effectiveness. This finding has direct implications for wildfire governance: effective pre-positioning does not rely on selecting among categorical policies but on operating within a stability range defined by the trade-off structure itself.

Discussion

Theoretical implications and comparison with existing research

By integrating AI-based wildfire risk estimation with multi-objective optimization, this study develops a reproducible and interpretable decision framework for firefighting resource allocation under constrained personnel availability. The results confirm that under bounded operational capacity, no single allocation simultaneously optimizes efficiency, equity, and fire control; improvements in one dimension require sacrifices in others. The three-dimensional Pareto fronts exhibit asymmetric, locally nonlinear structures, meaning that marginal gains in one objective can induce disproportionate losses elsewhere. This geometry demonstrates that multi-objective optimization functions not only as a prescriptive tool but also as a diagnostic instrument—revealing the shape, sensitivity, and limits of feasible decision spaces under institutional constraints, and showing that allocation choices span a continuous spectrum with an operationally robust near-knee region, as detailed in Sect.  4.3.

Many wildfire resource allocation models have relied on simplified representations of fire risk. For example, Haight and Fried⁷ and Ntaimo et al.¹⁷ emphasized scenario uncertainty and suppression effectiveness while assuming relatively stylized or historically averaged risk distributions. In contrast, the present framework incorporates spatially explicit, machine-learning-derived risk estimates, enabling allocation decisions that respond to heterogeneous fire behavior patterns across locations rather than uniform or scenario-based risk inputs.

This distinction has two substantive implications that go beyond data modernization. First, heterogeneous, locally calibrated risk inputs fundamentally alter the geometry of the trade-off space. Under uniform or historically averaged risk, the efficiency–equity frontier is relatively shallow: concentrating resources in nominally high-risk areas yields modest efficiency gains because the underlying risk variation is small. When risk is empirically heterogeneous and spatially structured, as in the present framework, the same concentration decision produces sharply asymmetric workload consequences across stations—generating the steep, nonlinear trade-off curves observed in the Pareto fronts reported in Sect. 4.2. The shape of these frontiers is therefore not an artifact of the optimization algorithm but a direct consequence of the AI-derived risk structure embedded in the objective functions. This finding has an immediate governance implication: agencies that base allocation planning on stylized or aggregate risk inputs are likely to underestimate the equity costs of efficiency-oriented deployment strategies. Second, the use of SHAP to make probability-based risk stratification auditable before optimization introduces a form of institutional accountability that neither prediction studies nor conventional optimization models provide independently. Prediction studies can explain why a location is high-risk but cannot specify how that risk should translate to resource requirements. Optimization models can generate efficient allocations but cannot explain why particular deployment configurations are preferred. The DSF-FRA closes this gap by making the risk stratification and the translation to allocation inputs explicit, auditable, and reproducible—properties increasingly required in public safety governance contexts where deployment decisions are subject to administrative review^26,34.

The efficiency–equity trade-off identified here is consistent with observations in broader emergency logistics and healthcare allocation literature^3,4, where balancing service efficiency and fairness has long been recognized as a structural challenge. However, existing studies have rarely examined this tension in wildfire response systems under explicitly fixed staffing and station-based deployment constraints. By demonstrating that these trade-offs persist even when personnel levels and station locations are inflexible, the present study extends insights from general emergency logistics into a wildfire-specific institutional context.

Furthermore, the integration of SHAP-based interpretability responds to growing concerns regarding transparency and accountability in AI-assisted public safety decision support. While machine learning models have been increasingly applied to wildfire prediction, their opacity has constrained operational trust and governance relevance. By aligning feature-level explanations with plausible ex ante hazard mechanisms, the framework demonstrates how explainable AI can support allocation reasoning rather than serving solely as a post hoc interpretive layer.

Policy and managerial implications and implementation considerations

The findings offer several implications for wildfire prevention and firefighting governance. First, AI-based risk estimation enables proactive, risk-informed pre-deployment strategies. The spatial importance patterns revealed through SHAP analysis indicate that personnel and equipment can be positioned in anticipation of elevated risk, reducing reliance on reactive cross-station mobilization and mitigating response delays during critical periods.

Second, the Pareto-optimal solution set supports flexible and tiered dispatch policies. In high-risk scenarios, efficiency-oriented strategies may be selected to prioritize rapid containment, even at the expense of temporary workload imbalance. In moderate- or low-risk contexts, more balanced allocations can be adopted to preserve workforce sustainability and coverage. Compared to conventional dispatch practices that rely on rule-based heuristics or reactive mobilization, the proposed framework enables systematic exploration of allocation alternatives under varying risk scenarios. This represents a methodological advance over static staffing guidelines currently used in many fire agencies.

Third, transparency enhances managerial communication and institutional legitimacy. SHAP-based explanations translate complex model behavior into interpretable feature contributions, enabling non-technical stakeholders to understand why specific areas are classified as high risk and how those classifications influence allocation outcomes. This approach parallels the evidence-based governance orientation demonstrated in building fire literature, where SHAP attributions have been used to connect model outputs to specific policy levers such as alarm system optimization and resource pre-positioning³⁷. These explanations can be integrated into spatial visualization tools such as GIS-based dashboards or web-based decision-support platforms, supporting training, scenario analysis, and post-event review.

Implementing the proposed framework requires several institutional and technical prerequisites. Fire agencies must establish systematic data collection protocols for meteorological, topographic, and incident records, along with GIS infrastructure capable of spatial risk visualization and scenario comparison. Personnel training programs are needed to familiarize dispatchers and managers with interpreting SHAP-based explanations and navigating Pareto-optimal solution sets. Finally, periodic model retraining and validation are essential to maintain predictive reliability as environmental conditions and fire behavior patterns evolve.

Model Assumptions, Limitations, and Future Research Directions

The framework rests on several simplifying assumptions. Fire stations are modeled as independent decision units, abstracting away inter-station coordination mechanisms that may exist in practice. Risk estimates are treated as deterministic inputs to the allocation model, without explicitly propagating prediction uncertainty into optimization outcomes. The temporal structure assumes discrete planning periods, whereas wildfire dynamics are continuous and may require dynamic reallocation. Additionally, the model employs 25 explanatory variables across 544 observations, yielding a sample-to-feature ratio of approximately 21.8:1. While this falls within acceptable bounds for gradient boosting methods, external validation on larger or geographically diverse datasets would further establish generalizability.

Several limitations follow from these assumptions. The framework relies on historical and aggregated datasets, and variations in reporting practices, spatial resolution, or institutional data availability may influence model performance. A further limitation concerns the geographic transferability of the nonlinear shape parameters calibrated from Taiwan wildfire records (α = 0.313, β = 0.354 for f₁; α′ = 0.052, β′ = 3.587 for f₃). These parameters encode region-specific relationships between resource concentration and response efficiency, and between response delay and fire spread escalation. In regions with denser road networks or shorter inter-station distances, the response-time proxy may exhibit a flatter escalation curve; in high-fuel-load environments such as Mediterranean or boreal forests, the burned-area exponent may substantially exceed 3.587, amplifying the efficiency–equity trade-off. Practitioners applying the DSF-FRA to other regions should treat the calibration procedure described in Supplementary Table S3 as a mandatory recalibration step rather than a transferable parameter set, and should validate Pareto front geometry against locally observed response-time and spread data before operational deployment.

Second, the analysis does not incorporate real-time operational factors such as traffic conditions, terrain accessibility, dynamic weather updates, or communication delays. Extending the framework toward adaptive, event-driven deployment would require coupling the optimization module with live data streams—GIS-based routing engines, nowcasting meteorological feeds, or IoT sensor networks—and computationally efficient re-optimization methods such as warm-starting evolutionary algorithms or surrogate-assisted multi-objective optimization. The risk stratification employs tercile boundaries derived from the empirical probability distribution, which ensures balanced stratum representation and adapts implicitly to the aggregate risk level of each planning period. However, the sensitivity of the MOEA/D Pareto front geometry to these specific boundary choices has not been formally assessed. Shifting stratum boundaries—for instance, by adopting seasonally recalibrated thresholds that reflect the bimodal distribution of fire activity between dry and wet seasons—could alter both the demand structure entering the optimization and the location of the near-knee region. Assessing this sensitivity systematically, and developing dynamic stratification schemes updated on a monthly or seasonal basis, represents an important direction for future work.

Third, the objective formulation does not explicitly incorporate social vulnerability, public risk perception, or long-term community resilience—a deliberate scope decision that prioritizes operational feasibility and data availability. Future work may extend the framework to multi-hazard environments, such as concurrent wildfire and flood events or cascading disasters involving infrastructure failure, as well as to dynamic allocation settings where resources are continuously reallocated in response to evolving conditions. Despite these limitations, the proposed framework provides a reproducible template for integrating AI-based risk assessment with multi-objective resource allocation, offering a pathway toward more transparent and adaptive wildfire response governance.

Conclusions

This study demonstrates that integrating artificial intelligence with multi-objective optimization provides a robust analytical foundation for wildfire emergency management. Rather than producing opaque recommendations, the proposed framework transforms resource allocation into an interpretable decision space in which efficiency, equity, and control effectiveness can be explicitly examined and balanced. Under limited resources, optimal firefighting strategies are not singular solutions but families of near-optimal alternatives along a Pareto frontier. This perspective allows decision-makers to align deployment choices with policy values and operational constraints, enhancing both rationality and transparency in emergency response planning.

Future research may extend this framework in several directions. First, cross-regional and multi-agency scenarios could be examined to assess robustness under cooperative command structures. Second, integration with real-time sensing, remote sensing, and GIS platforms could enable interactive decision-support systems with simulation capabilities. Third, incorporating social costs, ethical considerations, and long-term resilience indicators into the objective functions would further strengthen the role of AI as a tool for legitimate and accountable disaster governance.

Overall, this study offers a verifiable and theoretically grounded approach to intelligent firefighting resource allocation and provides a solid foundation for future research in AI-assisted disaster risk management.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

B.-W.H. conceived and designed the study, performed the analysis, prepared the figures, and wrote the manuscript.

Funding

The author declares that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The dataset supporting the findings of this study is publicly available at Mendeley Data: Hsu, B.-W. (2026). Code and data for: Integrating Explainable AI and Multi-Objective Optimization for Wildfire Resource Allocation. Mendeley Data, V1. 10.17632/f8gknmkhvf.1.

Declarations

Competing interests

The authors declare no competing interests.

Ethics Approval

This study did not involve human participants or animals and therefore did not require ethics approval.

Consent to participate

Not applicable.

Consent for publication

Not applicable.