Characterization of Fuel Types for the Canadian Region Using MODIS MCD12Q1 Data

Abstract

The characterization and mapping of fuel types is one of the most important factors to consider in the development of accurate fire behavior models. This study introduces a new methodology for generating a fuel map that can be easily updated on an annual basis. The method involves identifying associations between the Moderate Resolution Imaging Spectroradiometer (MODIS) land cover MCD12Q1 classes and the fuel-type classes categorized by the Canadian Fire Behavior Prediction System (FBP). For this purpose, MCD12Q1 Land Cover Type 1 data (MODIS LCM) were collected for the Canadian region. Concurrently, the Canadian fuel-type map implemented in the Fire Behavior Prediction System (FBP FTM) served as the reference dataset. Both MODIS LCM and FBP FTM were reclassified into a new Canadian FTM (NC-FTM) based on seven fuel-type classes. The method involves three key steps: (1) adapting MODIS LCM and FBP FTM for the classification of the Canadian region, (2) removing ambiguity, and (3) characterizing and assessing the accuracy of the new fuel-type classification using a confusion matrix classification algorithm. The achieved accuracy for the new classification exceeds 85%, highlighting the effectiveness of the approach. The use of MODIS LCM offers a cost-effective method for the annual characterization and mapping of fuel types, providing a practical improvement to the FBP model for Canada. Furthermore, with the proposed methodology, a fuel-type map can be generated for other specific areas of interest in the boreal region.

1. Introduction

Wildland fires are a significant environmental concern across terrestrial ecosystems [1] and present a substantial threat to the environment and society when not properly managed [2,3]. From 2001 to 2019, there was a documented global trend of increasing forest loss caused by fires [4]. Fire factors associated with temperature (e.g., vapor-pressure deficit, length of the growing season) have risen over the past four decades revealing a near-exponential relationship between these factors and annual burned area [5]. Global wildfire activity adheres to seasonal trends, usually experiencing fires from May to October. The highest activity is typically observed in July and August, corresponding to the hottest and driest months of the year (https://s3wfa.esa.int/dashboard, accessed on 3 June 2024). The year 2023 recorded the highest number of fires in the last six years (Figure 1).

While the majority of wildfires worldwide have been recorded in tropical regions, pan-arctic and extra-tropical areas have seen a rising frequency of fires in recent decades [6]. Accordingly, in recent years, the Copernicus Atmosphere Monitoring Service (CAMS) tracked an increase in the number of fires and related emissions in the boreal and arctic regions (Figure 2).

Fire behavior models are defined as mathematical relationships that describe the potential characteristics of fire [7]. They provide methods for conceptualizing forest fuel and its relationship with fire dynamics [8].

Or et al. [9] gave a comprehensive overview of the widely used fire behavior models developed by various frameworks in the United States, Australia, Canada, and Europe and categorized them into three broad types of models for predicting wildfire characteristics: statistical–empirical, semi-physical and physical–mechanistic. In all such models, fuel type is a critical component, and fuel-type mapping is crucial for characterizing wildfire risk [10]. In the last fifty years, various fuel-type classification systems have been proposed globally to summarize their main physical characteristics [11]. In this context, the Fire Behavior Prediction (FBP) system emerges as a key framework [11]. The FBP system was developed within the Canadian Forest Fire Danger Rating System (CFFDRS; [12]) and continues to be widely used in operational fire management, representing an essential part of wildland fire knowledge in Canada [13]. For example, the Canadian Wildland Fire Growth Model (Prometheus; [14]) uses the fuel-type classification based on the FBP system [9].

The FBP system categorizes the predominant fuel types found in Canada and defines their fire behavior characteristics under different burning conditions [15]. It is derived from a database containing more than 400 observations from experimental, wild, and prescribed fires [16].

Considering the primary role of fuel-type mapping in wildfire modeling and its impact on fire risk assessment, maintaining up-to-date, high-quality fuel maps is essential [17,18,19]. However, due to the dynamic nature of vegetation, ongoing climate change, and extensive land use, fuel types can change over time, and fuel-type maps should be updated frequently [18].

Historically, fuel-type mapping has been carried out with field surveys and aerial photointerpretation [7]. Field surveys have the great advantage of limited error, usually used to create field reference datasets at a local scale [20,21]. However, they are challenging and expensive in time and cost. As a result, remote sensing (RS) technologies have progressively replaced them, revolutionizing fuel-type mapping in the last fifty years [11]. Indeed, RS monitoring offers several benefits over traditional field mapping, including real-time data acquisition, extensive time series analysis, and full spatial coverage [22].

Among various RS options, Pettinari and Chuvieco [23] noted that multispectral sensors have been the main source of information for mapping fuel types across different scales. The ability of multispectral imagery to capture a broad range of spectral information from the Earth’s surface makes it an effective tool for enhancing fuel mapping, enabling discrimination between different fuel types according to their unique spectral signatures [24]. In this sense, the Moderate Resolution Imaging Spectroradiometer (MODIS) Land Cover Type 1 (hereafter referred to as MCD12Q1) has considerable importance in Earth surface research and has been extensively used for various applications [25]. This product provides global land cover types, and it has an annual updating cycle. Consequently, it supplies high-quality and timely land cover information, which is crucial to guarantee the updating of the fuel maps [26].

That being said, composing fuel datasets is time consuming and demands careful planning and financial support. As a result, these datasets are not updated regularly, which can be problematic when used as input for wildfire behavior modeling and active wildfire management [27].

To address the need for accurate and updated fuel maps as input for wildfire behavior models, our work aims to develop a simple methodology to classify and map fuel types linking MODIS MCD12Q1 land cover classes to FBP fuel-type classes. The National FBP fuel-type map (FBP FTM) was used as a reference dataset to reclassify the MODIS MCD12Q1 land cover classes in the Canadian Region. A new fuel-type classification system was specifically developed to better represent the fuel characteristics of the Canadian region ecosystems. The novelty of this study lies in the development of a cost-effective methodology for generating annually updated fuel-type maps using freely available MODIS data, which can be applied beyond Canada to other regions of the boreal belt.

In this study, we began by detailing the initial classification of the FBP system in comparison to the land cover class of MODIS MDC12Q1. Next, we linked the MODIS MDC12Q1 land cover classes to the FBP fuel-type classes. Following the implementation of the ambiguity removal process, we generated the counting data matrix to link the FBP FTM classes with the MODIS MDC12Q1 land cover map (MODIS LCM) classes. Finally, we used a confusion matrix classification algorithm for the accuracy assessment of the new fuel-type map. Considering that the FBP model continues to be a widely used approach for assessing fire behavior [28,29,30], this procedure represents a practical improvement for the Canadian region with potential applicability to other areas of the boreal region.

2. Materials and Method

2.1. Study Area

This study was conducted for Canada between latitudes 41°45″ and 83°08″ north and between longitudes 52°37″ and 141°00″ west. According to Dastour et al. [31], Canada has been experiencing an increase in both the frequency and extent of wildfires in recent decades, along with substantial shifts in the timing and duration of fire activity due to climate change, human activities, and vegetation conditions.

From a bioclimatic perspective, Canada can be divided into seven primary vegetation zones, reflecting the major climatic influences across the country [32]. The arctic bioclimate gives rise to the arctic tundra vegetation zone, characterized by low-growing shrubs, grasses, and mosses adapted to very cold and dry conditions with mean annual temperatures ranging from −12 to −6 °C [33]. The boreal bioclimate results in extensive boreal forests and woodlands, dominated by coniferous species such as spruce, pine, and fir, which thrive in the cold, subarctic climate with mean annual temperatures in the below-zero range. The Temperate bioclimate is marked by the presence of four distinct vegetation zones: the Pacific Cool Temperate forest, characterized by coastal rainforests with towering conifers and a lush understory growing at moderate temperatures and elevated precipitation; the Cordilleran Cool Temperate forest, found in the mountainous regions with a mix of conifers and broadleaf species; the Grassland and Steppe, consisting of open plains and prairies with grasses and drought-resistant shrubs with insufficient precipitation to support the growth of forests; and the Eastern Cool Temperate Forest, which includes deciduous forests with species like maple, oak, and birch. Additionally, the Alpine zone is formed under the influence of high elevation on zonal vegetation, featuring a unique assemblage of hardy plants adapted to cold and windy conditions. The dominant vegetation zone is the boreal forest and woodlands, while the second largest vegetation zone in the highlighted area remains the arctic tundra.

2.2. Data

2.2.1. Canadian Land Cover Map from MODIS MCD12Q1

MODIS MCD12Q1 is a descriptive land cover product obtained after processing the annual observation data from the Terra and Aqua sensors. The latest product, MCD12Q1 Version 6.1, is available with a 500 m spatial resolution from the year 2001 to the present (https://doi.org/10.5067/MODIS/MCD12Q1.061, accessed on 3 June 2024). MCD12Q1 contains a suite of datasets that map land cover globally and annually using six different classification schemes: the International Geosphere-Biosphere Programme (IGBP), University of Maryland (UMD), Leaf Area Index (LAI), BIOME-Biogeochemical Cycles (BGC), Plant Functional Types (PFTs), and the Food and Agriculture Organization Land Cover Classification System (FAO-LCCS) [34]. To meet our objectives, we utilized the IGBP Type 1 land cover classification, as it includes a broader range of land cover types [35].

MODIS MCD12Q1 product Version 6.1 for 2014 was collected to create a land cover map of the Canadian region (hereafter called MODIS LCM; Figure 3a). The tiles collected to cover the Canadian region were h09v03, h08v03, h09v04, h10v02, h11v03, h11v04, h13v01, h11v02, h10v04, h12v04, h12v02, h13v04, h13v02, h12v03, h14v01, h15v03, h16v00, h14v04, h15v02, h17v00, h15v01, h14v03, h10v03, h13v03, h14v02, h19v00, and h16v01.

Table 1. Code and description for land cover MODIS MDC12Q1 classes [35].

Code	Description
1	Evergreen Needleleaf Forests: dominated by evergreen conifer trees (canopy > 2 m). Tree cover > 60%. Almost all trees remain green all year. Canopy is never without green foliage.
2	Evergreen Broadleaf Forests: dominated by evergreen broadleaf and palmate trees (canopy > 2 m). Tree cover > 60%. Almost all trees and shrubs remain green year round. Canopy is never without green foliage.
3	Deciduous Needleleaf Forests: dominated by deciduous needleleaf (larch) trees (canopy > 2 m). Tree cover > 60%. Consists of seasonal needleleaf tree communities with an annual cycle of leaf-on and leaf-off periods.
4	Deciduous Broadleaf Forests: dominated by deciduous broadleaf trees (canopy > 2 m). Tree cover > 60%. Consists of broadleaf tree communities with an annual cycle of leaf-on and leaf-off periods.
5	Mixed Forests: dominated by neither deciduous nor evergreen (40–60% of each) tree types (canopy > 2 m). Tree cover > 60%. Consists of tree communities with interspersed mixtures or mosaics of the other four forest types. None of the forest types exceeds 60% of landscape.
6	Closed Shrublands: dominated by woody perennials (1–2 m height), >60% cover. The shrub foliage can be either evergreen or deciduous.
7	Open Shrublands: dominated by woody perennials (1–2 m height), 10–60% cover. The shrub foliage can be either evergreen or deciduous.
8	Woody Savannas: tree cover, 30–60% (canopy > 2 m).
9	Savannas: tree cover, 10–30% (canopy > 2 m).
10	Grasslands: dominated by herbaceous annuals (<2 m). Tree and shrub cover is less than 10%.
11	Permanent Wetlands: permanently inundated lands with 30–60% water cover and >10% vegetated cover.
12	Croplands: at least 60% of area is cultivated cropland.
13	Urban and Built-up Lands: at least 30% impervious surface area including building materials, asphalt, and vehicles.
14	Cropland/Natural Vegetation Mosaics: mosaics of small-scale cultivation 40–60% with natural tree, shrub, or herbaceous vegetation.
15	Permanent Snow and Ice: at least 60% of area is covered by snow and ice for at least 10 months of the year.
16	Barren: at least 60% of area is non-vegetated barren (sand, rock, soil) areas with less than 10% vegetation.
17	Water Bodies: at least 60% of area is covered by permanent water bodies.
255	Has not received a map label because of missing inputs

provides a list of 17 land cover MODIS MDC12Q1 classes with descriptions [35].

2.2.2. Canadian Fuel-Type Map from CWFIS

The National FBP fuel-type map, hereafter called the FBP FTM, was developed for the Canadian Wildland Fire Information System (CWFIS). This was derived primarily from forest attribute data for a resolution of 250 m as described in detail by Beaudoin et al. [36].

The purpose of this dataset is to provide FBP fuel types that facilitate the production of daily FBP grids and enhance situational awareness regarding national fire potential.

The FBP FTM (version 2014b) is available for download according to the Canadian Forest Service (https://cwfis.cfs.nrcan.gc.ca/downloads/fuels/archive/, accessed on 3 February 2024). While a more recent version (2019) is accessible, it is currently in the development phase. Hence, for the scope of this paper, we chose to utilize the 2014b version.

The FBP FTM provides a classification system consisting of 17 fuel types [37,38] and 5 non-fuel types, as specified in the accompanying metadata of the FBP FTM (https://cwfis.cfs.nrcan.gc.ca/downloads/fuels/archive/National_FBP_Fueltypes_version2014b.zip). The 17 fuel types are categorized into five major groups (coniferous, deciduous, mixedwood, slash, and open) and represent the prevalent fuel types found in Canada [15] (

Table 2. Overview of the FBP FTM classification (https://cwfis.cfs.nrcan.gc.ca/downloads/fuels/archive/National_FBP_Fueltypes_version2014b.zip). According to Taylor et al. [37], fuel types can be organized into five major groups (coniferous, deciduous, mixedwood, slash, and open), with a total of 17 discrete fuel types and 5 no-fuel types. C stands for coniferous, D stands for deciduous, M stands mixedwood, S stands for slash, and O stands for open.

Code	FBP Fuel Type	Description
101	C1	Spruce–Lichen Woodland
102	C2	Boreal Spruce
103	C3	Mature Jack or Lodgepole Pine
104	C4	Immature Jack or Lodgepole Pine
105	C5	Red and White Pine
106	C6	Conifer Plantation
107	C7	Ponderosa Pine–Douglas Fir
108	D1	Leafless Aspen
109	M1	Boreal Mixedwood–Leafless
110	M2	Boreal Mixedwood–Green
111	M3	Dead Balsam Fir Mixedwood–Leafless
112	M4	Dead Balsam Fir Mixedwood–Green
113	S1	Jack or Lodgepole Pine Slash
114	S2	White Spruce–Balsam Slash
115	S3	Coastal Cedar–Hemlock–Douglas
116	O1a	Matted Grass
117	O1b	Standing Grass
118	Water	Water
119	Non-Fuel	Non-Fuel
120	Unknown	Wetland (FBP fuel-type unknown)
121	Urban or Built-Up Area	Urban or Built-Up Area
122	Vegetated Non-Fuel	Vegetated Non-Fuel

; Figure 3b). In the FBP FTM used in our study, no pixels were counted in 9 out of the 22 classes. Specifically, classes 104, 106, 110–115, and 117 were found to be empty.

2.3. Pre-Processing

As previously outlined, the MODIS LCM and the FBP FTM represent spatial information on different types (classes) of physical coverage across Canada’s region, each characterized by distinct resolutions.

Given that the original MCD12Q1 product is stored in a hierarchical data format (HDF) and employs a sinusoidal projection, data pre-processing is necessary. This pre-processing includes format conversion, mosaicking, reprojection, and collocation (Figure 4). The Geospatial Data Abstraction Library (GDAL) was utilized for this pre-processing. GDAL 3.8.4 is a software library for reading and writing raster and vector geospatial data formats and is released under the permissive X/MIT style free software license by the Open Source Geospatial Foundation [39]. The MODIS HDF data format was converted into GeoTIFF, while the data projection was changed from sinusoidal to WGS84. Additionally, image mosaicking and subsetting were accomplished using GDAL. The outcome of this pre-processing step for the MODIS MCD12Q1 product was the LCM of the Canadian region (Figure 4, top panel, left side).

Conversely, the acquired FBP FTM has the original Geotiff format. The data projection was processed in GDAL and converted from Canada Lambert Conformal Conic to WGS84.

Additionally, the two maps were collocated to obtain spatially overlapping maps. In particular, the pixel values of MODIS LCM were resampled into the geographical raster of the FBP FTM. The spatial resolution of the MODIS LCM was resampled at 250 m using the nearest neighbor resampling method. This method ensured the retention of the original class values in the resampled image (for instance, each pixel with a resolution of 500 m was split into 2 × 2 pixels with a resolution of 250 m). The pre-processing steps are summarized in Figure 4 (top panel).

2.4. Processing: Ambiguity Removal, Counting Data Matrix, and Confusion Matrix

The pixel-by-pixel comparison of MODIS LCM vs. FBP FTM maps was necessary to find an association between the different class categorizations of the two maps. Unfortunately, this association was ambiguous for two reasons: (1) the effect of the resampling of the MODIS LCM from 500 m to 250 m, which increases the omission error in the pixel classification, and (2) the different number of classes between the MODIS LCM and FBP FTM maps.

To reduce this ambiguity, the two maps were reclassified. First, we outlined the conceptual framework of the reclassification procedure (Figure 4, bottom panel). Then, we provided a visual example to illustrate the methodology (Figure 5).

Initially, the same homogeneous clusters of 3 × 3 pixels common to each map were identified. Theoretically, in a perfect world (without the spherical effect of the Earth), the dimension of the best cluster would be 2 × 2 pixels, while in the real world, a cluster of 3 × 3 pixels is able to consider the effect of non-perfect image co-registration.

Once clusters of 3 × 3 pixels were identified, the next step was to apply the ambiguity removal procedure. The ambiguity removal process identifies all clusters of 3 × 3 pixels that share the same class in both respective maps.

After implementing the ambiguity removal procedure, we generated a counting data matrix (CDM) to associate the FBP FTM classes with the MODIS LCM classes. This involved the counting of pixels from FBP FTM classes within MODIS LCM classes and vice versa. Unfortunately, the number of classes was different between the two maps. Consequently, some classes of the MODIS LCM and the FBP FTM data were merged according to the criteria of ecological and biophysical similarity (further details are provided in Section 3).

To assess the agreement between the two maps, an aggregation of the MODIS LCM and FBP FTM classes was performed, resulting in the generation of a confusion matrix (CM). The CM was a square matrix N×N, where N is the number of classes. For this purpose, the CM was constructed based on the CDM, where similar classes were merged into larger classes to produce a square matrix.

A visual example of the implemented procedure is shown in Figure 5 to provide a clear exposition of the methodology presented above. In particular, we refer to the maps shown in Figure 5a and Figure 5b as the Truth Map (TM) and Predicted Map (PM), respectively.

In the TM, we counted four classes identified with the colors red, green, blue, and orange, while in the PM, we counted three classes identified with the colors cyan, magenta, and purple.

The ambiguity removal procedure consists of two steps:

Step 1: A 3 × 3 sliding window was applied to both the TM and the PM. If the value of all pixels within the 3 × 3 window shared the same color, then the value within the 3 × 3 window was retained; otherwise, all pixels were re-classified with the color white (see Figure 5c,d). The white pixels, which were not common to both maps, were excluded from the ambiguity removal procedure.

Step 2: The two maps were compared with a 3 × 3 sliding window. If the value of all pixels inside the 3 × 3 window in TM and PM was not equal to the color white, then the value inside the 3 × 3 window was retained in both the TM and PM. Otherwise, all pixels were re-classified with the color white (see Figure 5e,f).

We counted the distribution of color class pixel by pixel in both the TM and the PM. Specifically, we set a color class in the TM and counted the distribution of pixels that fell within the respective color class of the PM (see

Table 3. Counting data matrix results of ambiguity removal example for Truth Map (TM) and Predicted Map (PM).

		Truth
		Red	Green	Blue	Orange	Total
Predicted	cyan	0	0	36	18	54
	magenta	9	0	0	0	9
	yellow	0	9	0	0	9
	Total	9	9	36	18	72

). Finally, we compared the TM and the PM using the CM algorithm (see

Table 4. Confusion matrix example for Truth Map (TM) and Predicted Map (PM).

		Truth
		Blue + Orange	Red	Green	Total
Predicted	cyan	54	0	0	54
	magenta	0	9	0	9
	yellow	0	0	9	9
	Total	54	9	9	72

).

2.5. Confusion Matrix and Analysis of Accuracy

A cross-validation analysis was carried out to provide a quantitative assessment of the map accuracy.

The mapping of fuel types was achieved using a classification based on the multiclass CM algorithm developed by Congalton [40]. The CM classifier is considered one of the most widely recognized and well-established image classification methods due to its robustness and ease of use, making it commonly employed in vegetation and land cover mapping.

According to Stehman [41], the CM classifier offers a detailed assessment of the agreement between sampled reference data and the classification results, as well as a description of the misclassifications recorded for each class. Additionally, Congalton [40] presents various statistical indices derived from the confusion matrix that measure accuracy according to different needs. These include Overall Accuracy (OA), the k coefficient, User Accuracy (UA), and Producer Accuracy (PA).

Producer Accuracy is the probability that a value in a given class was classified correctly.

User Accuracy is the probability that a predicted value assigned to a certain class truly belongs to that class. The probability is based on the fraction of correctly predicted values to the total number of values predicted to be in a class.

The Overall Accuracy is calculated by summing the number of correctly classified values and dividing by the total number of values. The correctly classified values are located along the diagonal of the confusion matrix.

The k coefficient measures the agreement between classification and truth values. A kappa value of 1 represents perfect agreement, while a value of 0 represents no agreement. The kappa coefficient is computed as follows:

$$k={\displaystyle \frac{N{\sum }_{i=1}^n{m}_{ii}-{\sum }_{i=1}^n{T}_i{P}_i}{N^2-{\sum }_{i=1}^n{T}_i{P}_i}}$$

(1)

where

i is the class number;

N is the total number of classified values compared to truth values;

m_ii is the number of values belonging to the truth class i that have also been classified as class i (i.e., values found along the diagonal of the confusion matrix);

P_i is the total number of predicted values belonging to class i;

T_i is the total number of truth values belonging to class i.

Moreover, according to Landis and Koch [37], Kappa (k) values can be interpreted as follows: ≤0 indicates no agreement, 0.01–0.20 as none to slight agreement, 0.21–0.40 as fair, 0.41–0.60 as moderate, 0.61–0.80 as substantial, and 0.81–1.00 as almost perfect agreement.

3. Results and Discussion

Table 5. Counting data matrix comparing MODIS LCM and FBP FTM classes after the ambiguity removal procedure. MODIS LCM classes are presented on the y-axis (see Table 1), while FBP FTM classes are presented on the x-axis (see Table 2).

	101	102	103	105	107	108	109	116	118	119	120	121	122
1	10242	1418904	14715	344223	54	1179	17055	126	405	0	45	9	0
2	0	0	0	270	0	9	0	0	0	0	0	0	0
3	0	9	0	0	0	81	126	9	0	0	0	0	0
4	0	18	0	9	0	48456	7776	0	0	0	9	0	108
5	9	25128	0	819	0	40914	485262	9	171	0	18	0	9
6	0	0	0	0	0	315	0	0	0	0	0	0	0
7	11556	3303	0	0	0	35928	252	4104	1917	405	63	0	3083535
8	51624	562959	32418	909	126	325638	118899	40500	936	0	405	0	1359
9	471879	109674	0	9	0	326097	44154	55170	2817	216	630	9	108873
10	2799	909	18	45	0	16236	297	154665	10053	2318391	1161	90	5417388
11	10251	48420	0	18	0	684	1620	81	9207	1395	5319	0	9504
12	0	0	0	0	0	612	0	8550	846	9	81	162	1965087
13	0	0	0	0	0	9	0	0	18	45	0	21132	180
14	0	0	0	0	0	99	0	0	0	0	0	0	936
15	0	0	0	0	0	0	0	0	6192	3439593	0	0	3213
16	0	0	0	0	0	0	0	0	8928	2932209	0	0	16893
17	0	36	0	0	0	9	0	27	2815065	35433	36	0	1035

shows the counting matrix when comparing MODIS LCM and FBP FTM data after the procedure of ambiguity removal. For the characterization and accuracy assessment of fuel types based on the CM classification algorithm, a rationalization of classes in both maps was necessary. In particular, some classes of the MODIS LCM and the FBP FTM data were merged following the criteria of ecological and biophysical similarity. In our study, the process of merging the classes of both maps was carried out in two steps. First, out of the 17 FBP fuel types and 5 no-fuel types, only 13 classes contained information about our study area, while the remaining 9 classes were empty (see Section 2.2.2). These 13 classes were merged into 7 classes, following the approach employed by Simpson et al. [42]. Then, we reduced the MODIS LCM classes to seven classes in order to have the same number of classes needed for the square matrix (see Section 2.4). More in detail, we began by considering the basic classification of Simpson et al. [42], which takes into account the following classes: Coniferous 1 (C1), Coniferous 2 (C2, C3, C4, C5, C7), Deciduous (D1), Mixedwood (M1), Open cropland and grassland (O1), Non-fuel (NF), and Water. However, in our case, we merged C1 and C2 into the same class (C). We also considered Urban and Built-Up area (UBU). Additionally, we merged O1 and the Vegetated No-fuel (VNF) class into the same class named Open (O). Thus, we obtained seven classes in total, namely Coniferous (C), Deciduous (D), Mixedwood (M), Open (O), No-fuel (NF), Water and Wetland (WW), and Urban and Built-Up area (UBU) (

Table 6. Confusion matrix results for MODIS LCM and FBP FTM classes (codes and descriptions are available in Table 1 and Table 2).

		Coniferous (C)	Deciduous (D)	Mixedwood (M)	Open (O)	Water and Wetland (WW)	No-Fuel (NF)	Urban and Built-Up Area (UBU)
		101, 102, 103, 105, 107	108	109	116, 122	118, 120	119	121	Total	User Accuracy (%)
Coniferous (C)	1, 2, 8, 9, 11	3076695	653607	181728	215613	19764	1611	18	4149036	74.15
Deciduous (D)	4, 6	27	48771	7776	108	9	0	0	56691	86.02
Mixedwood (M)	3, 5	25965	40995	485388	27	189	0	0	552564	87.84
Open (O)	7, 10, 12, 14	18630	52875	549	10634265	14121	2318805	252	13039497	81.55
Water and Wetland (WW)	17	36	9	0	1062	2815101	35433	0	2851641	98.71
No-fuel (NF)	15, 16	0	0	0	20106	15120	6371802	0	6407028	99.45
Urban and Built-Up area (UBU)	13	0	9	0	180	18	45	21132	21384	98.82
	Total	3121353	796266	675441	10871361	2864322	8727696	21402	27077841
	Producer Accuracy (%)	98.57	6.12	71.86	97.82	98.28	73.01	98.74		OA (%) 86.61 k 0.809

).

In the second step of the process, we associated the MODIS LCM classes that matched well from an ecological point of view with the seven classes obtained from FBP FTM (e.g., the Evergreen Needleleaf Forests class from MODIS LCM was associated with the Coniferous class from FBP FTM). For those MODIS LCM classes that had uncertain destination, we applied a multiple-to-one criterion based on the maximum number of pixels. This involved assigning one or more MODIS LCM classes to each FBP FTM class, according to the class taxonomy, using a translation between products, as reported in Table 6. After completing this process, the classes were associated as follows:

• MODIS LCM classes 1, 2, 8, 9, and 11 were associated with FBP FTM classes 101, 102, 103, 105, and 107 in fuel type Coniferous (C);
• MODIS LCM classes 4 and 6 were associated with FBP FTM class 108 in fuel type Deciduous (D);
• MODIS LCM classes 3 and 5 were associated with FBP FTM class 109 in fuel type Mixedwood (M);
• MODIS LCM 7, 10, 12, and 14 were associated with FBP FTM classes 116 and 122 in fuel type Open (O);
• MODIS LCM class 17 was associated with FBP FTM classes 118 and 120 in fuel type Water and Wetland (WW);
• MODIS LCM classes 15 and 16 were associated with FBP FTM class 119 in fuel type No-Fuel (NF);
• MODIS LCM class 13 was associated with FBP FTM classes 116 and 122 in fuel type Urban and Built-Up area (UBU).

The above-mentioned associations between classes were then formalized into a new fuel-type categorization system, which is detailed in

Table 7. New fuel-type classification system and link between MODIS LCM classes and FBP FTM classes after rationalization (codes and descriptions are available in Table 1 and Table 2).

	Code	Rationalization LC Classes	Rationalization FT Classes
Coniferous	C	1, 2, 8, 9, 11	101, 102, 103, 105, 107
Deciduous	D	4, 6	108
Mixedwood	M	3, 5	109
Open	O	7, 10, 12, 14	116, 122
Water and Wetland	WW	17	118, 120
No-fuel	NF	15, 16	119
Urban and Built-Up area	UBU	13	121

.

Notably, MODIS LCM classes 1, 2, 8, 9, and 11 were linked with the Coniferous FBP FTM class. Specifically, classes 8, 9, and 11 were assigned to this category based on the higher number of pixels falling in the Coniferous class (Table 5). From an ecological perspective, the rationale for assigning classes 8 and 9 (Woody Savanna and Savanna) of the LCM to the Coniferous class lies in the fact that both classes can be considered open and sparse forests where the predominant vegetation life form is coniferous [43]. In support of this, we found that these classes have already been grouped with needleleaf forests for Canada in previous studies based on semantic definitions [44].

According to Sulla-Menashe and Friedl [35], class 11 (Permanent Wetland) is described as lands that are permanently inundated with water, featuring 30–60% water cover and more than 10% vegetated cover. In this context as well, the vegetated cover appears to pertain to a coniferous life form. This is likely because two-thirds of the Canadian boreal forest is covered by wetlands, which are seasonally or permanently water-saturated [45,46]. Our decision to include class 11 in the Coniferous group stems from the fact that, during drought periods when the water table descends, wetlands can also support burning. Hence, we opted for a precautionary approach that considers class 11 as a potentially flammable category in the event of possible dry periods [47].

Table 5, Table 6 and Table 7 show the results of this rationalization, the CM, and the accuracy coefficients. The OA across all classes is 86.61% higher than the 85% level considered adequate [48]. Moreover, according to Landis and Koch [49], the k coefficient of 0.809 suggests an almost perfect agreement between LCM and FTM. This result shows that the use of LCM data provided a valuable characterization and mapping of fuel types.

The PA and UA are higher than 6.12% and 74.15%, respectively (Table 6). Concerning a set level of 85%, C, O, WW, and UBU exhibit high PA, while D, M, WW, NF, and UBU show high UA. Not surprisingly, WW and UBU areas showed high UA and PA levels. Table 5 presents the confusion matrix generated by the cross-validation analysis, which sheds light on several observed patterns. Table 5 clearly shows that the primary source of low UA and PA, and consequently most of the mapping errors, stems from confusion between a subset of ecologically similar classes. Specifically, classes representing D and M were often confused with C, while the NF class was frequently misinterpreted with O (Table 6). These patterns indicate that classification errors are largely concentrated among classes that encompass ecological and biophysical gradients [45,46], and that are quite similar both functionally and in terms of properties.

Finally, Figure 6 depicts the main outcome of this study, the new Canadian FTM (NC-FTM). It showcases the newly generated fuel map for the Canadian region, derived through our methodology.

Fire scientists need updated fuel maps to implement strategies for fire danger assessment and reduction. The National FBP fuel type map, released in 2014, has undergone significant revisions in both algorithms and datasets over the intervening years. Currently, a new National FBP fuel-type map is in the process of development (https://cwfis.cfs.nrcan.gc.ca/downloads/fuels/development/, accessed on 3 February 2024). This implies a challenge associated with updating information in the FBP FTM.

The same need to revisit the time of fuel maps is essential not only for the CWFIS but also for other agencies with the same scope. For example, the European Forest Fire Information System (EFFIS) fuel map was developed in 2017 with a resolution of 250 m, while the last fuel map for Europe was developed in 2022 with a 1 km resolution for the continental scale [50].

Remarkably, Figure 7 illustrates the pixel distribution for classes for the Canadian region from 2018 to 2022, revealing an evident change in the intra-class pixel distribution across the different land cover types. This further emphasizes the need for an annual update of the land cover map. With the use of the MCD12Q1 dataset, we can overcome this limitation, as this product is updated every year.

Previous studies have provided significant advancements in fuel map development, offering valuable methodologies for this evolving field. For instance, Aragoneses et al. [17] developed a FirEUrisk European fuel map with a spatial resolution of 1 km, using land cover data, biogeographic datasets, and bioclimatic modeling. Their approach employed a hierarchical classifier, and the resulting map was validated against the Land Use/Cover Area Frame Survey (LUCAS; [51]) data for seven fuel classes. This study achieved an OA of 88.4% and a k of 0.78. Sismanis et al. [52] proposed a methodology for local wildland fuel mapping in Mediterranean ecosystems, achieving a resolution of 20 m. The study used a time series of six Sentinel-2 images, combined with landscape and climatic data, to identify nine fuel classes through spectral–spatial machine learning classifiers. Validation with LUCAS data reported an OA of 88.00% and a k of 0.840.

In a more localized context, Smith et al. [53] implemented a methodology to produce highly detailed fire fuel maps for local sites in Alaska at a resolution of 5 m, using AVIRIS-NG imagery. The method relied on a random forest classifier, a supervised algorithm based on decision trees, to identify 17 fuel classes. Related validation results showed an OA of 81.50% and a k of 0.700.

Our study achieved an OA of 86.61% and a k of 0.809, aligning closely with the validation performance reported by Aragoneses et al. [17] and Sismanis et al. [52]. Notably, the methodologies by Sismanis et al. [52] and Smith [53] were designed and validated for local-scale applications in distinct geographical contexts (Greece and Alaska, respectively). In contrast, our approach is tailored for continental scales, offering broader applicability across diverse landscapes.

Furthermore, a key distinction lies in the fire behavior models employed across these studies. While Aragoneses et al. [17] and Sismanis et al. [52] utilized the Fire Behaviour Fuel Types (FBFT; [54]) as a reference for the fuel classification system, our study adopted the Canadian FBP system [37]. Since the FBFT was originally developed in the United States, care must be taken when using the classification crosswalk scheme for Europe [55]. In contrast, the Canadian FBP system is specifically designed for boreal regions, making it particularly well suited for the northernmost bioclimatic conditions. This distinction highlights the relevance and scalability of our methodology for managing wildfires in continental contexts, particularly for boreal and arctic regions.

The methodology used in this study enabled us to establish a functional connection between the MODIS LCM and the FBP FTM classes. This connection presents some drawbacks such as a reduction in resolution (from 250 m to 500 m) and a decrease in the number of fuel types (from 17 to 7 fuel-type classes). These considerations highlight the trade-offs inherent in using a simplified classification system. When our map is used as input for fire behavior models, the limited number of fuel types may impact its effectiveness, especially in highly heterogeneous or fine-scale environments. In this regard, it is important to adopt a worst-case scenario approach when forecasting and assessing fire developments, ensuring that potential risks are not underestimated, even when individual fuel-type characteristics are generalized. Expert operators can use their local knowledge to fine-tune the parameters of fire behavior models for the represented territory, enhancing its accuracy and usefulness.

Among the key advantages and benefits offered by our methodology is the possibility of updating the fuel map annually, offering a fuel-type characterization that remains relevant and responsive to changing conditions. In addition, the proposed method is cost-effective and easy to access, relying on freely available sources.

A more general result of our study is the development and testing of a methodology for generating updated fuel-type mappings that could be potentially applied in other settings. In fact, because the MCD12Q1 dataset encompasses globally distributed land cover types, the approach we proposed for Canada can potentially be expanded to other regions.

However, at the level of the boreal belt, there is considerable heterogeneity in species composition, forest structure, and environmental conditions [56,57]. For example, the North American boreal forest has more coniferous species adapted to colder and often wetter conditions, whereas the Eurasian boreal forest includes large expanses of larch forests, particularly in Siberia, which are adapted to colder and drier environments. Differences in species composition are proposed to explain the varying fire behaviors observed between American and Eurasian forests [58,59].

Because of these differences, countries that host significant portions of the world’s boreal forests, alongside Canada, are developing tailored systems for vegetation fire behavior prediction as, for example, the BEHAVE system in the United States [60].

In Russia, an official state system for predicting forest fire behavior has yet to be established. However, individual research groups are actively working on developing a classification of vegetation fuels [61,62]. Forest inventories, which often serve as the foundation for such efforts, are updated infrequently. Our approach, initially developed and tested using Canadian fuel maps, demonstrates potential applicability to other regions and fuel classifications.

Regarding the application of our results across the Arctic, Figure 8 provides a practical example of how the new fuel-type classes generated through this study (Table 7) could be applied in other boreal regions, such as Finland. The map, which is valid for 2023, is derived from MDC12Q1 dataset of 2022 (Figure 8). Regarding a practical implementation of our research, the annual fuel-type maps generated through our methodology has been integrated as an informational layer in INFRA (INtegrated Fire Risk mAnagement), a web-based wildfire risk management service developed under the European project Arctic PASSION (https://arcticpassion.eu/blog/infraservice, accessed on 13 December 2024). The integration of our methodology into INFRA highlights its applicability, providing a valuable resource for advancing wildfire monitoring and management strategies.

Finally, it is worth emphasizing that the classification used in this study can be relatively easily updated, customized, and refined for the specific territory by an expert operator, especially when focusing on local or regional scales. The significant knowledge about fire held by Indigenous communities, deeply tied to their specific landscapes, can bolster the assignment process of MODIS LCM classes to FBP FTM classes [63,64] and widely enhance operational wildfire management.

4. Conclusions

Fuel-type maps are essential for fire behavior modeling, firefighting resource allocation, and the monitoring of vegetation recovery after fire. This study aimed to provide a methodology for fuel-type mapping in the Canadian region, taking advantage of MODIS MCD12Q1 land cover data. The Canadian FBP fuel-type map of 2014 served as a reference dataset to assess the results obtained for the considered test area. The analysis results demonstrated that the use of land cover based on remotely sensed data provided a valuable characterization of fuel types, achieving a classification accuracy exceeding 85%. Results obtained from these investigations can be broadly extended to all the boreal region ecosystems. The key advantage of this approach is that the use of MCD12Q1 data enables annual updating, thus addressing the general limitation related to the revisit time of fuel maps. The approach proposed in this study can be fruitfully applied to meet the frequent requirements of national and international environmental protection agencies, as well as forestry and environmental managers.