Quantifying Wildfire Dynamics Through Spatio-Temporal Clustering and Remote Sensing Metrics: The 2023 Quebec Case Study

Abstract

Wildfires have become increasingly frequent and destructive environmental hazards, especially in boreal ecosystems facing prolonged droughts and temperature extremes. This study presents an integrated spatio-temporal framework that combines Spatio-Temporal Density-Based Spatial Clustering of Applications with Noise (ST-DBSCAN), Fire Radiative Power (FRP), and the differenced Normalized Burn Ratio (ΔNBR) to characterize the dynamics and ecological impacts of large-scale wildfires, using the extreme 2023 Quebec fire season as a case study. The analysis of 80,228 VIIRS fire detections resulted in 19 distinct clusters across four fire zones. Validation against the National Burned Area Composite (NBAC) showed high spatial agreement in densely burned areas, with Intersection over Union (IoU) scores reaching 62.6%. Gaussian Process Regression (GPR) revealed significant non-linear relationships between FRP and key fire behavior metrics. Higher mean FRP was associated with both longer durations and greater burn severity. While FRP was also linked to faster spread rates, this relationship varied by zone. Notably, Fire Zone 2 exhibited the most severe ecological impact, with 83.8% of the area classified as high-severity burn. These findings demonstrate the value of integrating spatial clustering, radiative intensity, and post-fire vegetation damage into a unified analytical framework. Unlike traditional methods, this approach enables scalable, hypothesis-driven assessment of fire behavior, supporting improved fire management, ecosystem recovery planning, and climate resilience efforts in fire-prone regions.

1. Introduction

Forest fires represent an increasingly urgent threat to global ecological stability, climate regulation, and natural resource management. These extreme events disrupt biodiversity, release substantial greenhouse gases, and alter ecosystem dynamics, transforming many fire-adapted regions from carbon sinks into carbon sources [1,2,3,4,5,6]. In recent decades, the scale, frequency, and intensity of wildfires have escalated significantly—particularly in the Northern Hemisphere—driven by prolonged droughts, higher temperatures, and anthropogenic pressures that extend fire seasons and elevate landscape flammability [7,8].

Given these growing risks, effective wildfire management necessitates advanced monitoring frameworks that can both detect fires in near-real time and assess their longer-term ecological consequences. Satellite-based remote sensing tools have proven indispensable in this context. High-temporal-resolution sensors, such as MODIS and VIIRS, allow for active fire detection and thermal characterization using Fire Radiative Power (FRP), while post-fire impacts on vegetation are typically assessed using indices like the Normalized Burn Ratio (NBR) and differenced NBR (ΔNBR), derived from higher-resolution multispectral imagery [9,10,11].

Despite their complementary strengths, these data sources are often analyzed in isolation, limiting their utility in understanding the full lifecycle and behavioral complexity of wildfires [12,13]. Prior frameworks frequently overlook the spatial–temporal structure of fire events and their intricate linkages to fire energy and landscape severity. This analytical disconnect restricts our ability to draw robust inferences about how wildfires behave, spread, and impact ecosystems.

To bridge this gap, clustering algorithms, such as Spatio-Temporal Density-Based Spatial Clustering of Applications with Noise (ST-DBSCAN), have gained traction in wildfire studies due to their ability to extract coherent fire event boundaries from dense satellite detections [14,15]. While useful in isolating fire clusters, most implementations stop short of integrating these clusters with quantitative measures of intensity and severity. Additionally, statistical models traditionally used to investigate fire behavior often rely on linear assumptions, which fail to capture the complex, non-linear interactions between fire dynamics and ecological outcomes [16,17].

In response, this study introduces a unified remote sensing framework that holistically analyzes the wildfire lifecycle. First, we use ST-DBSCAN to determine where and when a fire event occurs. We then enrich these spatially and temporally defined events by integrating Fire Radiative Power (FRP) to assess how energetically the fire burns and the differenced Normalized Burn Ratio (ΔNBR) to evaluate what ecological impact it leaves behind. To model the intricate, non-linear relationships between these metrics, we employ Gaussian Process Regression (GPR), a method particularly suited to this task as it also provides robust uncertainty estimates for its predictions [16,17]. The unprecedented 2023 wildfire season in Quebec, Canada, serves as an ideal case study for testing this integrative methodology [18,19,20].

This research is guided by the following hypotheses:

• H1: Fire clusters with longer durations and larger spatial extents exhibit higher mean FRP values, reflecting sustained, high-energy fire behavior.
• H2: Higher FRP levels are associated with greater burn severity (ΔNBR), indicating a positive relationship between radiative intensity and ecological impact.
• H3: The rate of fire spread is positively influenced by both FRP and fire duration, though this relationship may vary depending on local factors, such as wind or fuel continuity. Although external variables, such as wind or fuel continuity, are not directly modeled, their influence is implicitly captured through the residual structure of the GPR model, which accounts for unexplained variance in the data.

By systematically testing these hypotheses, this study contributes a scalable and reproducible analytical pipeline capable of quantifying the spatial structure, thermal behavior, and ecological impact of extreme wildfire events. The results provide critical insights for fire management strategies while also aligning with emerging literature that emphasizes the need for integrated, non-linear, and spatially explicit models of fire dynamics [16,21,22,23,24,25].

2. Study Area and Data Requirements

This section provides details on the geographical context of this study, the satellite datasets utilized for the analysis, and the preprocessing steps undertaken to prepare the data for subsequent spatio-temporal analysis and modeling.

2.1. Study Area

This study focuses on the province of Quebec, Canada (Figure 1), which experienced an exceptionally severe and large-scale wildfire season during June 2023.

This period was selected due to the unprecedented nature of the fire activity, which established new historical benchmarks for the region. According to official reports, the 2023 wildfires in Quebec burned over 4.5 million hectares, an area exceeding the cumulative total burned area over the preceding two decades [26]. Quebec’s provincial fire protection agency, SOPFEU, characterized the 2023 season as “historic,” underscoring the extreme conditions and widespread impact, particularly across the northern boreal zones of the province [27].

Given the intensity and scale of these events, four major wildfire zones within Quebec were specifically selected for a detailed, comparative investigation (Figure 2). These zones were chosen as they represent distinct, spatially separate wildfire events, allowing for the evaluation of our analytical framework across varying conditions. The selection was based on their significant geographic extent and the availability of high-quality satellite data necessary for robust analysis.

2.2. Datasets

Three primary types of data were utilized in this study: active fire data for spatio-temporal clustering, multispectral imagery for burn severity assessment, and official government data for validation.

• Active Fire and FRP Data: The core dataset for analyzing spatio-temporal fire dynamics consists of active fire detections sourced from the NASA FIRMS platform. Data from the VIIRS instruments aboard the Suomi-NPP and NOAA-20 satellites were acquired for June 2023. VIIRS provides active fire detections at a spatial resolution of ~375 m and includes essential attributes for each hotspot, namely, geographic coordinates, acquisition date and time, and FRP (MW), which serves as a proxy for fire intensity [28]. A full description of the VIIRS data attributes is provided in

  Table 1. Description of attributes in the VIIRS active fire dataset.

  Attribute Name	Description
  Latitude	Geographic latitude of the fire event.
  Longitude	Geographic longitude of the fire event.
  Brightness	Brightness value of the detected hot spot.
  Scan	Satellite scan width.
  Track	Satellite scan length.
  Acq Date	Date the fire was detected.
  Acq Time	Time the fire was detected.
  Satellite	Satellite from which the data were obtained (S-NPP or NOAA-20).
  Confidence	Confidence level of fire detection.
  Version	Data processing version used.
  Bright T31	Radiance temperature (Kelvin) obtained from the thermal channel.
  FRP (Fire Radiative Power)	Radiant power emitted by the fire (MW).
  Day/Night	Whether the fire was detected during daytime or nighttime.

  .
• Burn Severity Data: To evaluate post-fire effects, the NBR was calculated using multispectral imagery from the Landsat 8 OLI/TIRS sensor. In accordance with established methodologies [29,30], Band 5 (Near Infrared, NIR) and Band 7 (Shortwave Infrared, SWIR 2) were used to assess burn severity, as these bands are highly sensitive to changes in vegetation moisture and structure caused by fire.
• Reference Data for Validation: To quantitatively validate the perimeters derived from the ST-DBSCAN algorithm, the National Burned Area Composite (NBAC) was used as a reference dataset [31]. Produced by Natural Resources Canada, the NBAC provides authoritative, polygon-based perimeters of large fires (>200 ha) and serves as the ground-truth benchmark for our spatial accuracy assessment.

2.3. Data Preprocessing

All datasets underwent preprocessing in QGIS (Quantum GIS) to ensure consistency and suitability for analysis:

• VIIRS Data Integration and Preparation: The separate VIIRS S-NPP and NOAA-20 active fire datasets were merged into a single vector layer. A unified timestamp attribute was generated from the original acquisition date and time fields. While all attributes were retained for descriptive purposes, the core inputs for the clustering algorithm were the geographic coordinates, the new timestamp, and the FRP values.

  Table 2. Sample rows from the processed VIIRS dataset used in ST-DBSCAN analysis (June 2023) [28].

  Fid	Latitude	Longitude	Brightness	Scan	Track	Acq_Date	Acq_Time	Satellite	Instrument	Confidence	Version	Bright_T31	FRP	Day Night
  28883	53.16951	−75.23611	339.63	0.62	0.54	1.06.2023	1643	1	VIIRS	n	2.0NRT	298.55	14.28	D
  28889	53.05344	−75.2285	338.33	0.62	0.54	1.06.2023	1643	1	VIIRS	n	2.0NRT	293.92	5.86	D
  28890	53.05182	−75.31041	325.56	0.63	0.54	1.06.2023	1643	1	VIIRS	l	2.0NRT	298.88	8.55	D
  30491	53.04161	−75.2345	339.82	0.51	0.49	1.06.2023	1824	1	VIIRS	n	2.0NRT	301.41	43.6	D

  presents sample rows from this final preprocessed dataset.
• Landsat Data Preparation: Pre- and post-fire Landsat 8 OLI/TIRS Collection 2 Level-2 Surface Reflectance products were used. These data, already atmospherically corrected, were spatially aligned and clipped to the extent of each fire zone of interest prior to NBR calculation [30].
• Reference Data Preparation: The NBAC vector polygons were reprojected to this study’s coordinate system (UTM) and clipped to the relevant analysis regions to ensure direct spatial alignment with the ST-DBSCAN outputs for validation.

These preprocessing steps resulted in analysis-ready datasets for the subsequent stages of this study. The comprehensive methodology, from data acquisition to final analysis, is illustrated in the workflow diagram in Figure 3.

3. Methods

This section outlines the core analytical steps used to extract and validate spatio-temporal wildfire clusters and quantify fire-related characteristics. All analyses were conducted using QGIS and other open-source tools.

3.1. Spatio-Temporal Clustering with ST-DBSCAN

The foundational step of our analytical framework involved identifying distinct wildfire events from VIIRS active fire detections using the ST-DBSCAN algorithm. This method is critical to our study, as it extends conventional DBSCAN by incorporating a temporal dimension, enabling the detection of wildfire clusters that are cohesive in both space and time.

3.1.1. Core Principles and Formulation

ST-DBSCAN detects clusters based on the density of points in a spatio-temporal context. The algorithm is governed by three key parameters—spatial epsilon (${ϵ}_s$), temporal epsilon (${ϵ}_t$), and the minimum number of points (MinPts)—along with formal spatial and temporal distance definitions.

For any two data points $p$ and $q$, the spatial distance is computed using the Euclidean distance:

$$d_s\left(p,q\right)=\sqrt{{\left(x_p-x_q\right)}^2+{\left(y_p-y_q\right)}^2}$$

(1)

The temporal distance is defined as the absolute difference in acquisition time:

$$d_t\left(p,q\right)=\left|t_p-t_q\right|$$

(2)

A point q is considered a neighbor of p if and only if both spatial and temporal thresholds are satisfied:

$${\mathrm{d}}_{\mathrm{s}}\left(\mathrm{p},\mathrm{q}\right)\le {\mathsf{ϵ}}_{\mathrm{s}} \mathrm{and} {\mathrm{d}}_{\mathrm{t}}\left(\mathrm{p},\mathrm{q}\right)\le {\mathsf{ϵ}}_{\mathrm{t}}$$

(3)

Based on this neighborhood definition, each point is classified as follows:

• Core point: A point with at least MinPts neighbors within its spatio-temporal neighborhood.
• Border point: A point that is not a core point but lies within the neighborhood of one.
• Noise point: A point that is neither a core nor a border point.

A cluster is defined as a maximal set of density-connected points, representing a single contiguous wildfire event [15,32].

3.1.2. Parameter Selection and Optimization

The ST-DBSCAN algorithm was implemented in QGIS (Version 3.34). Parameter tuning was conducted through an iterative process that combined domain expertise on fire behavior, sensor resolution characteristics of VIIRS, and visual inspection of the clustering results to ensure accurate and interpretable outcomes.

The final parameters were set as follows:

• Spatial Epsilon (${ϵ}_s$ = 2700 m): This threshold enables the linkage of fire detections from adjacent or nearby pixels while avoiding the erroneous merging of separate fire events. It reflects the spatial resolution of VIIRS (~375 m) and accounts for geolocation uncertainties.
• Temporal Epsilon (${ϵ}_t$ = 72 h): A three-day window was chosen to accommodate observation gaps caused by cloud cover or orbital limitations while capturing the full evolution of single fire episodes.
• Minimum Points (MinPts = 5): This value effectively filters out isolated or spurious detections while preserving smaller, yet coherent, fire clusters.

Execution of the algorithm yielded a new vector layer in which each original fire detection point was annotated with attributes identifying its cluster membership.

Table 3. Sample rows from the clustered VIIRS fire detection output layer, showing assigned CLUSTER_ID and CLUSTER_SIZE attributes.

Fid	Latitude	Longitude	Brightness	Scan	Track	Satellite	Confidence	Version	Bright_T31	Frp	Daynight	Acq_Date_T	Cluster_ID	Cluster_SI
28883	53.16951	−75.23611	339.63	0.62	0.54	1	n	2.0NRT	298.55	14.28	D	2023/06/01 16:43:00.000	3896	76,573
28889	53.05344	−75.2285	338.33	0.62	0.54	1	n	2.0NRT	293.92	5.86	D	2023/06/01 16:43:00.000	3896	76,573
28890	53.05182	−75.31041	325.56	0.63	0.54	1	l	2.0NRT	298.88	8.55	D	2023/06/01 16:43:00.000	3896	76,573
30491	53.04161	−75.2345	339.82	0.51	0.49	1	n	2.0NRT	301.41	43.6	D	2023/06/01 18:24:00.000	3896	76,573

provides a sample from this output, displaying assigned CLUSTER_ID values along with corresponding CLUSTER_SIZE statistics.

Subsequently, these annotated points were aggregated by their CLUSTER_ID to derive higher-level summaries for each distinct fire event. This aggregation facilitated the calculation of cluster-level attributes such as DATE_START (earliest timestamp within the cluster) and DATE_END (latest timestamp), which were essential for analyzing the temporal duration and spread dynamics of each fire.

These structured outputs provided the necessary basis for mapping the identified wildfire events and performing further quantitative analyses on their spatial and temporal properties as presented in the Section 4.

3.2. Cluster Validation with IoU Accuracy Assessment

To assess the spatial accuracy of the wildfire clusters generated by the ST-DBSCAN algorithm, we performed a validation analysis using the Intersection over Union (IoU) metric—also known as the Jaccard Similarity Index. This metric quantifies the spatial overlap between algorithm-generated cluster polygons and authoritative fire perimeter data obtained from the National Burned Area Composite (NBAC) database.

Methodology

For each fire zone, the clustered fire points produced by ST-DBSCAN were first converted into polygonal representations. To ensure a consistent and reproducible geometric boundary for all clusters, a convex hull was generated for each set of clustered points. These polygons were then spatially intersected with the corresponding NBAC polygons for the same time period and region.

The IoU is computed as follows:

$$\mathit{IoU}={\displaystyle \frac{\mathit{Area}\left(\mathit{Cluster}\cap \mathit{NBAC}\right)}{\mathit{Area}\left(\mathit{Cluster}\cup \mathit{NBAC}\right)}}$$

(4)

where

• Cluster ∩ NBAC is the intersection area between the ST-DBSCAN cluster and the NBAC fire perimeter.
• Cluster ∪ NBAC is the union of their spatial extents.

IoU values range from 0 to 1, where higher values indicate better spatial agreement. A value above 0.5 is generally considered acceptable in geospatial validation, while values closer to 1 suggest near-perfect overlap [33,34,35].

3.3. Fire Behavior and Impact Metrics

Once the wildfire clusters were identified and validated, we calculated a set of quantitative metrics for each cluster to characterize the behavior, severity, and potential ecological impact of the fire events. These metrics integrate radiative energy release, vegetation consumption, and spatio-temporal dynamics.

3.3.1. Burn Severity: ΔNBR Calculation

Burn severity was assessed using the ΔNBR, a widely accepted remote sensing index for post-fire impact assessment. The NBR is calculated as:

$$\mathrm{NBR}={\displaystyle \frac{\mathit{NIR}-{\mathit{SWIR}}_2}{\mathit{NIR}+{\mathit{SWIR}}_2}}$$

(5)

where NIR (Near Infrared) and SWIR (Shortwave Infrared) refer to reflectance values from Landsat 8 bands 5 and 7, respectively. The ΔNBR is then derived as:

$$\Delta \mathrm{NBR}={\mathit{NBR}}_{\mathit{prefire}}-{\mathit{NBR}}_{\mathit{postfire}}$$

(6)

To assess the ecological impact of the fire events, a burn severity classification was performed using the ΔNBR [36]. Higher ΔNBR values indicate more severe vegetation loss. The ΔNBR values were derived from pre- and post-fire satellite images and classified into severity classes using the proposed unscaled thresholds (

Table 4. Classification of burn severity levels based on unscaled ΔNBR thresholds [30].

Severity Class	ΔNBR Range (Unscaled)
Enhanced Regrowth	<−0.10
Unburned	−0.10 to +0.10
Low Severity	+0.10 to +0.27
Moderate Severity	+0.27 to +0.66
High Severity	>+0.66

) [30].

3.3.2. Fire Intensity: FRP Aggregation

Fire intensity was represented using FRP, reported in megawatts (MWs), as provided by the VIIRS active fire product. For each cluster, we computed the mean FRP by averaging all detection-level FRP values. This aggregation enables comparison across different clusters and provides a standardized measure of energy release [37].

$${\mathrm{FRP}}_{\mathrm{mean}}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\mathit{FRP}}_i$$

(7)

where $n$ is the number of fire points within a given cluster.

3.3.3. Cluster-Derived Spatio-Temporal Metrics

In addition to FRP and the ΔNBR, we computed the following cluster-level metrics:

• Duration (days): The time difference between the first and last detection within the cluster (DATE_END − DATE_START).
• Area (km²): This metric represents the geographic footprint of each fire cluster, calculated directly from the polygon geometry of ST-DBSCAN outputs. After clustering VIIRS fire detections, each cluster was spatially delineated as a polygon using GIS tools, and the area was computed using the projected coordinate system (UTM). This value reflects the total extent of detected fire activity, independent of vegetation damage or burn severity.
• Average Rate of Spread (km²/day): Calculated as the total area divided by the total duration.

$${\mathit{Rate}}_{\mathit{spread}}={\displaystyle \frac{\mathit{Duration}}{\mathit{Area}}}$$

(8)

These metrics collectively provide a multi-dimensional profile of each wildfire event, integrating severity, intensity, and progression.

3.4. Integrative Analysis and Hypothesis Testing

To formally test this study’s hypotheses and quantitatively explore the complex interrelationships between fire behavior metrics, a sophisticated statistical modeling approach was adopted. This moves beyond simple correlation to a more robust analysis of fire dynamics.

3.4.1. Rationale and Methodology: Gaussian Process Regression

GPR was employed to model the potentially non-linear relationships between variables. Traditional linear regression often fails to capture the complex dynamics of fire behavior. GPR, as a non-parametric Bayesian method, is ideal for this task because it models a distribution over functions rather than assuming a fixed functional form [38]. A GPR model is defined by a mean function $m\left(x\right)$ and a kernel (covariance) function k(x, x’):

$$y\left(x\right)\sim \mathcal{G}\mathcal{P}\left(m\left(x\right),\hspace{0.17em}k\left(x,x^{\prime }\right)\right)$$

(9)

The kernel function defines the covariance between any two data points, encoding the assumption about the function’s smoothness. For this study, we used the common squared exponential (RBF) kernel:

$$k\left(x,x^{\prime }\right)={\sigma }_f^2\mathit{exp}\left(-{\displaystyle \frac{{\left(x-x^{\prime }\right)}^2}{2l^2}}\right)$$

(10)

where the hyperparameters ${\sigma }_f$ (signal variance) and $l$ (characteristic length-scale) are learned from the data. A key advantage of GPR is its ability to provide principled uncertainty estimates (95% confidence intervals) around its predictions, which helps in assessing the robustness of the observed relationships.

3.4.2. Hypothesis Testing Framework

The GPR models were implemented in Python using the scikit-learn library and were specifically structured to test the three hypotheses outlined in the Introduction:

• To test H1 (Duration vs. Intensity): We modeled the relationship between fire duration (days) and mean FRP (MW).
• To test H2 (Intensity vs. Severity): We modeled the relationship between mean FRP (MW) and burn severity (ΔNBR).
• To test H3 (Spread vs. Intensity/Severity): We modeled the rate of fire spread (km²/day) as a function of mean FRP (MW) and duration.

The results of these models, detailed in the Section 4, provide empirical evidence for supporting or rejecting this study’s hypotheses. This integrative analysis explicitly links the spatio-temporal patterns of wildfires with their radiative intensity (FRP) and landscape impacts (ΔNBR), thus enhancing our understanding of large-scale wildfire dynamics.

3.5. Software and Tools

All geospatial data processing, analysis, and visualization steps were performed using open-source software and programming languages to ensure reproducibility. The primary analyses, including data preparation, ST-DBSCAN clustering, IoU validation, and spatial metrics computation, were executed using QGIS (version 3.34). The GPR models and related statistical analyses were implemented using Python programming language (version 3.10), along with libraries such as scikit-learn (version 1.3.0) [39], NumPy (version 1.24.3) [40], and Pandas (version 2.0.3). Data visualization and the final plots were produced using the Matplotlib (version 3.7.1) library [41].

4. Results

4.1. Cluster Statistics and Spatial Patterns

The application of the ST-DBSCAN algorithm across four major wildfire zones revealed distinct spatio-temporal clustering patterns, each reflecting varying fire behaviors during the extreme 2023 Quebec wildfire season. In total, 80,228 active fire points were analyzed, resulting in the identification of numerous high-density clusters and low-density noise points. The total number of clustered points amounted to 78,819, with 1409 points (1.76%) classified as noise due to insufficient spatial or temporal density. Summary statistics for each zone are presented in

Table 5. Summary of ST-DBSCAN clustering results.

Zone	Total Points	Clustered Points	Noise Points	Noise (%)
1	76,573	75,532	1041	1.40%
2	1435	1186	249	17.40%
3	1812	1727	85	4.70%
4	408	374	34	8.30%

.

The spatial distribution of the clusters is visualized in Figure 4. Each map uses unique cluster IDs and colors to reveal the spatial structure and density of fire activity within the respective zones.

In Fire Zone 1 (Figure 4a), the dominant cluster is #3896, covering a vast, contiguous area across the northern extent of the study region. This cluster alone accounts for the majority of the detections, with additional small clusters, such as #40845, #47480, and #54360, occupying marginal spatial niches. The morphology of Cluster #3896 suggests a large-scale, sustained fire event with minimal interruption, consistent with the low noise rate of only 1.4%.

Fire Zone 2 (Figure 4b) contains two dominant clusters, namely, #5542 and #40927, forming dense linear structures that follow likely spread corridors. Smaller clusters, including #45689, #45690, and #47484, appear more fragmented and localized. Despite the compact distribution, the zone shows a higher noise percentage, pointing to discontinuities in ignition events or temporal gaps.

Fire Zone 3 (Figure 4c) is characterized by a large, expansive cluster #7471, which dominates the eastern and central portion of the region. A secondary cluster, #42842, appears in the western edge of the zone. Additional minor clusters, such as #29395, #29396, and #21040, are more scattered. The relatively low noise ratio (4.7%) suggests coherent fire activity across multiple spatial pockets.

Finally, Fire Zone 4 (Figure 4d) exhibits a more fragmented and dispersed pattern. While #7615 represents the primary cluster in terms of both area and point density, #15129, #25395, #25644, and #29391 highlight localized flare-ups. The increased fragmentation in this zone corresponds with a moderate noise rate (8.3%), suggesting short-duration or lower-density events.

These results highlight the spatial variability and structural complexity of wildfire activity across Quebec. While Fire Zones 1 and 3 are characterized by large and spatially continuous clusters, suggesting extended and spatially extensive burn patterns, Fire Zone 4 exhibits more dispersed and fragmented cluster formations. The distinct spatial morphologies observed across zones form a foundational basis for the subsequent analyses of fire radiative intensity and spread metrics (Section 4.2), burn severity mapping (Section 4.3), and fire perimeter validation (Section 4.4).

4.2. Fire Intensity and Spread Dynamics (FRP and Spread Metrics)

To characterize the thermal energy release and spatial progression of the wildfires, FRP, burned area, fire duration, and spread rate were analyzed across all four zones. These metrics are summarized in

Table 6. Fire dynamics metrics across all fire zones.

Zone	Area (km2)	Mean FRP (MW)	Spread Rate (km2/day)	Duration (Days)
1	11,300.7	20.03	376.69	30
2	2673.5	33.78	132.66	12.05
3	2102.9	37.24	31.72	5.55
4	730.3	14.68	131.12	8.54

, providing a comparative view of fire behavior.

Zone 1 demonstrated the largest burned area and longest fire duration (30 days), yet its mean FRP (20.03 MW) was moderate compared to other zones. This indicates a long-lasting but thermally moderate fire regime, where ignition was persistent across space but not characterized by extreme radiative energy.

In contrast, Zone 2 recorded the highest mean FRP (33.78 MW), highlighting a more intense combustion process despite its smaller area. This pattern is visually supported in Figure 5b, where the fire originated near the center and expanded rapidly outward, especially between June 10 and 20. The spread pattern aligns with the observed FRP surge and shorter total duration (12.05 days), suggesting a high-intensity, fast-evolving event.

Zone 3, despite having a smaller area (2102.9 km²) than Zone 2, exhibited the highest mean FRP (37.24 MW) and a relatively short duration (5.55 days). This points to a short-lived but extremely intense ignition episode, potentially associated with dry fuel and rapid combustion. As shown in Figure 5c, fire propagation was multi-directional and spatially concentrated, amplifying localized radiative energy.

Zone 4, on the other hand, displayed a fast spread rate (131.12 km²/day) but a lower mean FRP (14.68 MW). This implies that the fire spread rapidly across the landscape without releasing high thermal energy per unit area. As depicted in Figure 5d, the ignition began near the center and dispersed into fragmented patches, consistent with a low-intensity, wind-driven spread pattern.

These contrasting dynamics illustrate the importance of jointly evaluating FRP, spread rate, and duration to fully capture the behavior of wildfires. While high FRP typically suggests intense combustion, its ecological impact depends on how quickly and extensively the fire propagates.

To complement the quantitative summaries, FRP histograms were generated for each fire zone to reveal the underlying distribution of thermal energy. As shown in Figure 6, Fire Zones 2 and 3 exhibit positively skewed distributions with high FRP peaks, supporting the presence of intense, short-term combustion events. Conversely, Fire Zone 4 presents a broad, flatter distribution, suggesting widespread but lower-intensity burning. Fire Zone 1, despite its large burned area, maintains a unimodal distribution centered at moderate FRP values, indicative of stable and sustained fire activity over time.

4.3. Burn Severity Mapping (NBR Classification)

To evaluate the ecological impacts of wildfire activity across the study area, burn severity classification was performed for all ST-DBSCAN-derived fire clusters using unscaled ∆NBR values. Following the classification thresholds defined in [30] (Table 4), the burned areas were categorized into five severity classes: Enhanced Regrowth, Unburned, Low, Moderate, and High/Severe. This classification framework provides essential information on vegetation damage, combustion intensity, and post-fire regeneration potential, supporting spatially explicit assessments of ecological impact. Proportional distributions for each fire zone were computed and are presented in

Table 7. Percentage distribution of burn severity classes by fire zone.

Fire Zone	% Enhanced Regrowth	% Unburned	% Low Severity	% Moderate Severity	% High/Severe
1	6.73	14.54	15.27	62.16	1.30
2	0.06	0.01	0.14	15.96	83.84
3	1.76	4.59	5.91	18.84	68.90
4	5.76	8.59	13.80	19.70	52.15

, while individual cluster-level values are summarized in

Table 8. Burn severity classification by cluster.

Fire Zone	Cluster ID	% Enhanced Regrowth	% Unburned	% Low Severity	% Moderate Severity	% High/Severe
1	3896	4.16	12.07	24.82	55.62	3.33
1	40845	2.78	6.10	16.25	73.01	1.86
1	47480	20.00	40.00	0.00	40.00	0.00
1	54360	0.00	0.00	20.00	80.00	0.00
2	5542	0.01	0.01	0.17	20.41	79.40
2	40927	0.01	0.02	0.53	15.72	83.72
2	45689	0.00	0.00	0.00	15.79	84.21
2	45690	0.27	0.00	0.00	17.34	82.39
2	47484	0.00	0.00	0.00	10.53	89.47
3	7471	8.52	22.93	29.01	34.84	4.70
3	21040	20.00	20.00	40.00	20.00	0.00
3	29395	11.49	24.14	9.20	49.43	5.75
3	29396	5.26	10.53	15.79	63.16	5.26
3	42842	5.75	12.94	20.31	52.40	8.61
4	7615	9.58	20.77	24.60	40.58	4.47
4	15129	6.13	18.06	23.23	46.77	5.81
4	25395	16.3	17.75	20.00	42.22	3.70
4	25644	5.00	25.00	7.50	50.00	12.50
4	29391	6.67	14.67	13.33	54.67	10.67

, which shows burn severity classification by cluster. The spatial patterns are visualized in Figure 7.

Fire Zone 1 exhibited the most heterogeneous severity pattern, owing to its extensive area (11,300.7 km²) and prolonged burn duration (30 days). As shown in Figure 7a, the dominant class was Moderate severity (62.16%), followed by Low severity (15.27%) and Unburned (14.54%). High/Severe impact was relatively limited (1.30%), suggesting a fire regime that was widespread but not exceptionally intense. This distribution reflects a landscape-wide fire affecting large vegetation cover with relatively controlled combustion conditions. The Enhanced Regrowth rate was 6.73%, indicating some areas began ecological recovery shortly after the burn.

In contrast, Fire Zone 2 demonstrated a much more intense and destructive burn pattern. As visualized in Figure 7b, High/Severe severity dominated the region, accounting for a striking 83.84% of the total affected area, while Moderate and Low severity levels were almost negligible (15.96% and 0.14%, respectively). Several clusters, such as Cluster 47484, exhibited High/Severe values exceeding 89%. The percentage of Unburned and Enhanced Regrowth areas was nearly zero, indicating extensive and uniform combustion with minimal resistance or natural firebreaks.

Fire Zone 3 presented a moderately severe but more spatially varied fire impact, with 68.90% of the burned area classified as High/Severe, complemented by Moderate- (18.84%) and Low-severity (5.91%) zones. As shown in Figure 7c, Cluster 7471, the most dominant in this zone, exhibited a more balanced structure with notable Unburned (22.93%) and Enhanced Regrowth (8.52%) zones, suggesting potential fire suppression or topographic barriers that mitigated spread in some regions.

Fire Zone 4, the smallest in spatial extent (730.3 km²), displayed a patchy yet impactful burn profile. According to Figure 7d, the zone was dominated by Moderate (19.70%) and High/Severe (52.15%) severity. Notably, Enhanced Regrowth accounted for 5.76%, and Unburned areas made up 8.59% of the total footprint. This indicates a fragmented ignition pattern and topographically induced variability in fire spread and intensity. Smaller clusters such as 25644 and 29391 exhibited over 10% High/Severe severity.

These findings underscore that higher FRP values or longer durations do not necessarily imply uniformly severe burn impacts. Local factors, like vegetation type, fuel load continuity, slope, and wind exposure, can significantly alter burn severity outcomes. The integration of spatio-temporal clustering with ∆NBR-based severity analysis enables a multi-dimensional understanding of fire behavior, supporting more nuanced post-fire ecological assessments and management strategies.

4.4. Validation of Clustered Fire Boundaries (IoU Analysis)

To assess the spatial accuracy of the ST-DBSCAN clustering results, convex hulls derived from each fire zone were compared with the official fire perimeters provided by the National Burned Area Composite (NBAC) dataset, curated by the Canadian Forest Service (CFS) [31]. The comparison employed the Intersection over Union (IoU) metric, which quantifies the degree of spatial overlap between two polygon geometries.

The overlay results are presented in Figure 8, where ST-DBSCAN-derived cluster boundaries (colored polygons) are compared with NBAC official perimeters (blue outlines). Fire Zone 1 achieved the highest overlap with an IoU of 62.6%, followed by Zone 3 (55.9%), Zone 2 (51.1%), and Zone 4 (45.5%). These values reflect varying degrees of spatial alignment between the data-driven cluster outputs and official fire records.

The NBAC data used in this comparison represent authoritative burned area records from Natural Resources Canada and serve as a reference benchmark for spatial fire extent validation.

4.5. Gaussian Process Regression Analysis

To investigate the non-linear relationships between fire radiative energy and key fire behavior indicators, GPR models were developed for each fire zone. These models assess how FRP relates to fire duration, spread rate, and burn severity.

In the FRP vs. fire duration plots (Figure 9), all fire zones exhibit a positive, non-linear relationship, indicating that higher FRP values are generally associated with longer fire durations. The GPR models reveal a steep initial increase followed by a plateau-like behavior, particularly in Zones 2 and 3 (Figure 9b,c), where FRP values exceeding 150 MW correspond to durations greater than 15 days. This asymptotic trend suggests a saturation effect, where further increases in FRP result in diminishing returns in duration. In contrast, Zone 4 (Figure 9d) displays a flatter curve with higher variance, indicating more irregular fire behavior despite elevated FRP levels.

The FRP vs. burn severity analysis (Figure 10) similarly reveals a positive but non-linear relationship, with increasing severity associated with greater FRP. In Zone 2 (Figure 10b), the GPR curve displays a steep rise up to approximately 30 MW, beyond which the curve flattens, indicating a saturation point in burn severity. This pattern suggests that beyond a certain energy threshold, further increases in FRP do not lead to proportionally higher ecological impact. In contrast, Zones 1 and 4 (Figure 10a,d) exhibit flatter or more scattered curves with wider confidence intervals.

To capture the joint impact of FRP and time on fire propagation, a three-dimensional GPR model was constructed and visualized (Figure 11). This model highlights how higher FRP values, combined with temporal progression, contribute to accelerated spread rates. In Zones 2 and 4 (Figure 11b,d), the GPR surfaces exhibit steep gradients during specific time intervals. In Zone 2, the surface shows a sharp incline followed by a leveling off. The surface in Zone 4 displays localized peaks and irregularities. In contrast, Zones 1 and 3 (Figure 11a,c) show more gradual and uniform surfaces.

These GPR-based insights confirm that fire radiative energy serves as a reliable proxy for multiple fire behavior attributes. Especially when combined with time dynamics, FRP emerges as a key driver of fire intensity, severity, and propagation potential.

5. Discussion

This study introduced an integrated remote sensing framework to characterize the spatio-temporal dynamics, intensity, and severity of the extreme 2023 wildfires in Quebec. By coupling ST-DBSCAN clustering with FRP and ΔNBR metrics and testing relationships with GPR models, this research moves beyond simple fire mapping to a more holistic, hypothesis-driven analysis of fire behavior.

The application of ST-DBSCAN yielded meaningful fire cluster boundaries largely consistent with official NBAC perimeters, demonstrated by Intersection over Union (IoU) values reaching up to 62.6%. The high IoU in Fire Zone 1, in particular, confirms the algorithm’s effectiveness in delineating large, spatially cohesive fire complexes, reflecting its capacity to capture long-duration, sustained events. In contrast, the lower IoU scores in more fragmented zones like Fire Zone 4 (45.5%) highlight how cluster accuracy is influenced by fire morphology, a challenge consistent with the literature on automated perimeter delineation [23].

Burn severity analysis revealed substantial differences in landscape impact, directly linking to the fire’s thermal behavior. For example, the extreme intensity of Fire Zone 2 (mean FRP of 33.78 MW) translated directly into the most severe ecological impact, with 83.8% of its area classified as High/Severe burn. This supports the established connection between radiative energy and combustion intensity [11,42]. However, our findings also highlight a critical nuance: high FRP does not uniformly result in high severity, emphasizing the modulating role of local factors, a complexity also observed by Roy et al. [22].

The GPR models provided the deepest insights by quantifying the non-linear relationships between fire intensity, spread dynamics, and severity. A key observation was the saturation behavior observed in Zones 2 and 3 for the FRP–duration relationship (Figure 9b,c), where the duration initially increased sharply with FRP up to ~150 MW, after which the trend flattened. This suggests that fire duration responds strongly to FRP up to a certain energy threshold, beyond which additional energy does not proportionally extend fire persistence—likely due to limits imposed by fuel availability or fire suppression thresholds.

In the FRP–severity analysis, Zone 2 (Figure 10b) exhibited a sigmoidal response, with severity increasing sharply up to ~30 MW and plateauing thereafter. This saturation trend supports the idea that after a certain point, increasing fire intensity no longer translates into more severe vegetation damage—potentially due to combustion completeness or biophysical resistance thresholds. Conversely, Zones 1 and 4 (Figure 10a,d) displayed weaker or inconsistent severity responses, with broader GPR confidence bands. This variability may stem from vegetation heterogeneity, topographic fragmentation, or variable fuel structures, which diminish the predictive power of FRP in these areas.

The three-dimensional GPR surface models integrating FRP and time (Figure 11) further revealed distinct zone-level dynamics. In Zone 2 (Figure 11b), the surface was steep but stabilized after a peak, indicating a burst of intense propagation within a short window. In Zone 4 (Figure 11d), the GPR surface was irregular and contained local peaks, possibly reflecting fragmented fire patches and complex topography—consistent with the low IoU values and spatial disjointedness of fire clusters in this zone. By contrast, Zones 1 and 3 (Figure 11a,c) displayed smoother and more uniform surface gradients, suggesting steady but less volatile spread behavior, aligned with their longer durations and moderate spread rates.

These results validate and nuance the hypothesis-driven structure of this study. H1 was supported through the observed positive and saturating relationship between FRP and fire duration in Zones 2 and 3. H2 was confirmed in Zone 2, though variable patterns in Zones 1 and 4 suggest that this relationship is modulated by local heterogeneity. H3 received partial support: while spread rate increased with both FRP and time in some zones (notably Zones 2 and 4), this relationship was less stable in others, likely due to topographic, meteorological, or fuel continuity factors.

Overall, the GPR-based modeling approach proved highly effective in capturing zone-specific, non-linear fire dynamics. It complements previous work based on linear models (e.g., Freeborn et al. [17]) and aligns with recent machine learning studies that emphasize data-driven fire behavior prediction (e.g., Kumar et al. [10]).

Importantly, this study’s primary contribution is its multi-faceted integrative framework. Unlike prior works that often treat fire characteristics in isolation (e.g., Archibald et al., 2013 [43]), our approach demonstrates the value of linking spatial patterns, energy release, and ecological outcomes to test specific hypotheses. While the current analysis focused on satellite-derived metrics, future extensions should incorporate dynamic abiotic variables (e.g., weather, topography) and biotic factors, such as vegetation type, to improve predictive capacity. Scaling this framework across additional geographies and timeframes may further support the development of generalized fire behavior models under changing climatic conditions.

6. Conclusions

This study presents an integrated remote sensing and clustering-based framework for characterizing the spatio-temporal dynamics and ecological impacts of large-scale wildfire events. By leveraging ST-DBSCAN clustering, FRP, the ΔNBR, and GPR, we analyzed the structure, intensity, severity, and spread dynamics of the extreme wildfires that occurred in Quebec, Canada, during the 2023 fire season.

∆∆The results underscore several key findings:

• Spatio-temporal clustering using ST-DBSCAN effectively delineated coherent fire regions, with validation against NBAC perimeters showing moderate to high spatial agreement (IoU up to 62.6%), supporting the reliability of the approach.
• FRP-based intensity analysis revealed that higher radiative energy outputs were robustly associated with longer fire durations. However, the relationship with the rate of spread was found to be zone-dependent, indicating that fire intensity is not the sole driver of its propagation speed.
• Burn severity classification using the ΔNBR highlighted substantial spatial variability across fire zones. Zones with high FRP values often, but not always, coincided with severe burn areas—indicating the influence of local factors.
• GPR models confirmed the non-linear yet interpretable relationships between FRP, fire duration, spread rate, and burn severity, proving highly useful for quantifying complex fire dynamics.

Collectively, the proposed framework offers a modular and scalable method for integrating satellite-derived fire data to support decision-making in wildfire management. By quantifying relationships between fire behavior metrics, it enables proactive risk assessment, resource prioritization, and ecological recovery planning.

This study presents an integrated remote sensing and clustering-based framework for characterizing the spatio-temporal dynamics and ecological impacts of large-scale wildfire events. By leveraging ST-DBSCAN clustering, Fire Radiative Power (FRP), the differenced Normalized Burn Ratio (dNBR), and Gaussian Process Regression (GPR), we analyzed the structure, intensity, severity, and spread dynamics of the extreme wildfires that occurred in Quebec, Canada, during the 2023 fire season.

The results underscore several key findings:

• Spatio-temporal clustering using ST-DBSCAN effectively delineated coherent fire regions, with validation against NBAC perimeters showing moderate to high spatial agreement (IoU up to 62.6%), supporting the reliability and spatial accuracy of the approach.
• FRP-based intensity analysis revealed that higher radiative energy outputs were robustly associated with longer fire durations, confirming Hypothesis 1 (H1). However, the relationship between FRP and spread rate was zone-dependent, suggesting that fire intensity is not the sole driver of propagation speed, thereby offering partial support for Hypothesis 3 (H3).
• Burn severity classification using ΔNBR highlighted substantial spatial variability across fire zones. Zones with high FRP values often—but not always—coincided with severe burn areas, supporting Hypothesis 2 (H2) in regions with stable vegetation characteristics while highlighting the role of local factors in modulating fire impact.
• GPR models confirmed the presence of non-linear, zone-specific relationships between FRP, fire duration, spread rate, and burn severity. These models proved highly useful for quantifying complex fire dynamics and identifying saturation effects and threshold behaviors, which linear models would not capture.

Overall, the proposed framework offers a modular and scalable method for integrating satellite-derived fire data to support decision-making in wildfire management. By quantifying the relationships among fire behavior metrics, it enables proactive risk assessment, resource prioritization, and ecological recovery planning in data-rich, large-scale fire regimes.

Future research should aim to incorporate additional environmental variables—such as meteorological conditions, vegetation types, and topography—into the analytical pipeline to improve predictive performance. Moreover, extending this framework across multiple fire seasons and diverse geographies can support the development of generalized fire behavior models under changing climatic conditions.