The Pyrogeography of the Gran Chaco’s Dry Forest: A Comparison of Clustering Algorithms and the Scale of Analysis

Abstract

(1) Background: Changes in the spatial, temporal, and magnitude-related patterns of fires caused by humans are expected to exacerbate with climate change, significantly impacting ecosystems and societies worldwide. However, our understanding of fire regimes in many regions remains limited, largely due to the inherent complexity of fire as an ecological process. Pyrogeography, combined with unsupervised learning methods and the availability of long-term satellite data, offers a robust framework for approaching this problem. The purpose of the study is to identify the pyroregions of the Argentine Gran Chaco, the world’s largest continuous tropical dry forest region. (2) Methods: Using globally available fire occurrence datasets, we computed five fire metrics, related to the extent, frequency, intensity, size, and seasonality of fires at three spatial scales (5, 10, and 25 km). In addition, we tested two widely used cluster algorithms, the K-means algorithm and the Gaussian Mixture Model (GMM). (3) Results and Discussion: The identification of pyroregions was dependent on the clustering algorithm and scale of analysis. The GMM algorithm at a 25 km scale ultimately demonstrated more coherent ecological and spatial distributions. GMM identified six pyroregions, which were labeled based on three metrics in the following order: annual burned area (categorized in low, regular or high), interannual variability of fire (rare, occasional, frequent), and fire intensity (low, moderate, intense). The values were as follows: LRM (22% of study area), ROI (19%), ROM (14%), LOM (10%), ROL (9%), and HFL (4%). (4) Conclusions: Our study provides the most comprehensive delineation of the Argentine Gran Chaco’s Dry Forest pyroregions to date, and highlights both the importance of determining the optimal scale of analysis and the critical role of clustering algorithms in efforts to accurately characterize the diverse attributes of fire regimes. Furthermore, it emphasizes the importance of integrating fire ecology principles and fire management perspectives into pyrogeographic studies to ensure a more comprehensive and meaningful characterization of fire regimes.

1. Introduction

Fire is a complex ecological process that has shaped terrestrial landscapes since the establishment of plants on Earth’s surface [1]. Currently, fire regimes are heavily influenced by human pressure [2,3]. This, in conjunction with climate change, is anticipated to exacerbate the ongoing changes in fire patterns [4,5], further altering the frequency, severity, and extent of wildfires in various ecosystems worldwide [6,7,8]. In this context, the ecosystems that have evolved with fire, and require periodic disturbance to maintain their structure and function [9], are affected when fire occurs more frequently or intensely than in the past, as this alters their stability [3,10,11]. In this scenario, disentangling fire regimes is essential for understanding past and future ecological patterns [9,12].

A fire regime describes the repeated patterns of fires in a given location at a given time. Therefore, the characterization of fire regimes typically involves the synthesis of spatial (burned area and fire size), temporal (seasonality, frequency, return interval), and magnitude (intensity and severity) attributes related to wildfires [13]. These attributes are usually summarized into a single representative statistical measure (mean, median, percentile) for a specific area and within a defined time period [2,14,15,16,17,18].

Pyrogeography, the discipline used to analyze the spatial distribution of fire over time [7], provides a suitable framework for understanding the heterogeneity of fire regimes in the Earth’s system at multiple scales. It combines biogeography and fire ecology theory, supported by the availability of global satellite datasets on fire occurrence [19]. Pyrogeographic analysis usually focuses solely on spatial, temporal, and magnitude-related attributes of fires. Once these characteristics of different fire regimes emerge, their relationships with vegetation, climate, and human components can be investigated. Globally, Archibald et al. [2] introduced the concept of pyromes, fire units analogous to biomes (global vegetation units). At a regional and national level, the areas characterized by a specific fire pattern (i.e., fire regime) determine the pyroregions.

Cluster analysis, an unsupervised learning method used for grouping data based on shared attributes, is a powerful tool in pyrogeography. By employing clustering techniques, regions with similar fire attribute values (e.g., burned area, frequency, intensity, and seasonality) can be classified into distinct “pyroregions”, thereby revealing patterns that might otherwise remain hidden. In particular, partitioning algorithms such as K-means and Gaussian Mixture Model (GMM) [20] are effective for analyzing large-scale datasets derived from satellite imagery [2,15,18,21].

K-means, one of the most popular and simplest algorithms [22], groups cells into a predefined number of clusters. K-means is advantageous due to its low time complexity and high computational efficiency [23], making it suitable for initial explorations of fire regime patterns. Conversely, GMM enhances the flexibility of clustering by modeling data as a finite mixture of Gaussian distributions. This methodology facilitates the capture of complex cluster shapes and the effective management of overlapping clusters.

Several studies have sought to empirically identify pyromes and pyroregions worldwide (e.g., [14,16,18,21,24,25,26,27]). These studies addressed this issue by applying different approaches in terms of the scale at which the fire metrics were calculated and the clustering algorithms used. From a scale perspective, most authors worked with square grid cells to compute fire metrics, where cell sizes were variable, ranging from 2 × 2 to 100 × 100 km. Most of these studies rely on coarse spatial resolution imagery (250 m to 10 km), largely due to the product’s availability and capability to offer a higher temporal resolution, a crucial aspect in satellite fire regime studies. In terms of clustering algorithms, works have focused on using K-means analysis [14,15,21,28], Affinity Propagation [16], Hierarchical methods [24,25,27], and GMMs [2,18].

The dry forests of the South American Gran Chaco are a good example with which to study variability in fire regimes from a pyrogeographic perspective. The Gran Chaco is the world’s largest continuous tropical dry forest region [29]. Its diverse landscapes, climatic conditions, topography, and human activities [30,31,32] create conditions for varying fire occurrence patterns, suggesting the coexistence of multiple pyroregions within the ecoregion [33,34,35,36,37]. However, a comprehensive understanding of fire regimes in the region is still lacking. The Gran Chaco has a long history of anthropogenic fire, primarily driven by land use changes for forestry and agriculture [38,39,40]. This process has intensified over recent decades, which has further been exacerbated by climate change, leading to profound alterations in its fire dynamics. As a result, the Gran Chaco is considered a global deforestation hotspot [41]. This ecoregion harbors forests of global conservation significance. These forests provide essential ecosystem goods and services and support unique cultural and biological diversity [30,42]. Therefore, a thorough understanding of the existing pyroregions and their characteristics is crucial for conservation efforts and effective management.

Focusing on the Argentine Gran Chaco’s Dry Forest (60% of the Gran Chaco), our goal was to identify landscape units with similar fire regimes, i.e., pyroregions. Specifically, we compared two clustering algorithms, the K-means and GMM, and three scales of analysis by calculating fire metrics using grid cells with 5, 10, and 25 km sides. Our study provides the most comprehensive delineation of the Argentine Gran Chaco pyroregions to date, and highlights both the importance of determining the optimal scale of analysis and the critical role of clustering algorithms in efforts to accurately characterize the diverse attributes of fire regimes. Furthermore, it emphasizes the importance of integrating fire ecology principles and fire management perspectives into pyrogeographic studies to ensure a more comprehensive and meaningful characterization of fire regimes.

2. Materials and Methods

2.1. Study Area

The Argentine Gran Chaco’s Dry Forest (22–34° S; 675,000 km², 25–2900 m.a.s.l.; Figure 1) has a predominantly continental climate, with dry and mild winters and wet hot summers. The mean annual temperature ranges from 16 °C in the north to 26 °C in the south, while the average annual precipitation ranges from 1800 mm in the east to 300 mm in the west [43]. The region is predominantly flat, with a slight slope towards the east and an area of mountains to the west.

The temperature and rainfall gradients define four major subregions: Humid Chaco to the east, Semiarid Chaco in the central north zone, Arid Chaco to the southwest and Serrano Chaco in mountain areas [43] (Figure 1). Within these subregions, plant communities vary in composition and physiognomy, shaped by climatic conditions, geological deposits, disturbance patterns, and use history. This diversity gives rise to a variety of vegetation units, including forests, shrublands, palm groves, savannas, wetlands, and grasslands in varying states of conservation [31,32].

Fire is a recurrent disturbance in the Chaco region, with the majority of events being linked to human activity [38]. Fire activity is most frequent between July and November [34,44]. However, in more humid areas, bimodal seasonality is observed, with notable peaks in late summer and early autumn [35].

2.2. Data Collection

Datasets used in this study were derived from globally available remote sensing products related to burned areas and active fire, based on observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensors on board NASA’s Terra and Aqua satellites. These products were selected due to their high temporal resolution (coverage from November 2000 to the present), which is essential for analyzing fire regimes [2]. Among the currently available products for the region, they represent the most robust options in terms of temporal and spatial coverage. Specifically, we employed the MCD64A1 burned area [38], the FIRED fire event perimeter [39], and the MCD14ML hotspot products [40] (Figure 2).

The MCD64A1 Version 6 Burned Area product is a monthly gridded dataset with a 500 m resolution that provides a per-pixel burned area. Its algorithm combines surface reflectance imagery, active fire detections, and a burn-sensitive vegetation index to determine the burn date for each pixel. The product includes multiple data layers, such as burn date, burn date uncertainty, quality assurance, and the first and last day of reliable change detection within the year. Specific values are assigned to unburned land, missing data, and water pixels [45]. The data is available for processing and download at the Google Earth Engine Platform [46].

FIRED is a global collection of country-level fire event perimeter products at a spatial resolution of 500 m. It is generated using the open-source Python package FIREDpy (Fire Event Delineation for Python), which automates the download and processing of source files from the MODIS burned area product for a specified area of interest. FIREDpy applies a spatiotemporal flooding algorithm that consolidates hundreds of MODIS grid cells into single fire perimeter shapefiles. Each fire event is accompanied by ancillary information, including ignition date, event duration, burned area, and other relevant attributes [47]. The data is available for download at the Earth Lab Data collection at CU Scholar [48].

MCD14ML is a monthly hotspot location product with a 1 km spatial resolution. The algorithm uses brightness temperature imagery to identify hotspots (active fires or thermal anomalies). The product includes each hotspot’s geographic coordinates, detection date, and type (e.g., vegetation fire or other sources such as volcanic activity), as well as additional information. In this study, we used the standard science-quality subset (consistent and well-calibrated), which was pre-processed by the MODIS Fire Team at the University of Maryland’s Science Computing Facility. These data are available in shapefile format, with a 2–3-month delay, through the Archive Download Tool provided by NASA’s Fire Information for Resource Management System (FIRMS).

2.3. Data Preparation

Data processing was performed using the Google Earth Engine (GEE) platform [46]. GEE is a cloud-based platform designed for large-scale temporal and spatial data visualization and analysis. It provides access to a wide range of public remote sensing products, such as the MCD64A1, and also allows users to import third-party products like FIRED and MCD14ML. Its massive computational capacity and integrated analytical tools make it particularly suitable for the fast and efficient processing of large geospatial data. Users can write programs using the platform libraries (currently available for the Python and JavaScript languages) to process and analyze data. This enabled the generation of harmonized data layers for subsequent, more specific statistical analyses, such as the application of different clustering algorithms in R version 4.4.1 programming software [45], facilitating a robust assessment of regional fire regimes.

The study area was analyzed at different scales by partitioning its surface into grids of three distinct cell sizes—5 × 5 km (25,846 cells), 10 × 10 km (6741 cells), and 25 × 25 km (1132 cells)—and computing fire metrics for each cell. Given the wide range of spatial resolutions employed in previous studies to determine pyroregions and pyromes—ranging from 10 to 50 km at the regional scale [14,15,18,25] and 25 to 100 km at global scale [2,26,28], with overlapping values between both scales—we chose to compare all of these grid cell sizes to identify the most suitable resolution to accurately represent fire regimes in the Argentine Gran Chaco’s Dry Forest. This approach seeks to achieve a balance between capturing the local variability in fire regimes while ensuring a sufficient amount of data is obtained for robust statistical analysis. To characterize the multidimensional space that defines fire regimes over the 2001–2021 period, five key metrics were calculated for each cell: (1) annual burned area, (2) interannual variability of fire, (3) fire intensity, (4) fire size, and (5) Fire Seasonal Concentration Index [2,49] (Figure 2).

Burned area is the most commonly investigated aspect of fire regimes [2,15,17,18,21,24,25,28,50]. It is widely used by climate and vegetation modelers, as well as regional land managers, to estimate carbon fluxes and assess fire risk [51]. Moreover, understanding patterns in burned area is essential for identifying the factors that control fire activity [52] and for detecting changes in fire regimes over time [53,54]. Annual burned area was computed using the MCD64A1 burned area product [45]. First, the burned area for each year from 2001 to 2021 was obtained by summing the monthly burned area layers, and then the mean was computed for the entire period.

The interannual variability of fire, i.e., the year-to-year fluctuations in burned area, is largely driven by climatic factors [55,56]. Given that future projections indicate warmer and/or drier conditions, understanding this variability is particularly relevant. The coefficient of variation of the annual burned area (CVABA) was used as a proxy for IVF, given its robust correlation with this measure. Areas that burn frequently have lower CVABA values than areas that burn infrequently, as the interannual variability in burned areas is higher when fires occur less often [2].

Fire intensity refers to the rate of energy released along the burning front. This attribute is useful for distinguishing between ground fires and crown fire regimes. Ground fires typically burn at lower intensities and spread more slowly due to the limited availability of fuels, whereas crown fires involve higher energy release, tend to spread faster, and burn for longer and more extensively as they consume both surface and canopy vegetation. Therefore, understanding fire intensity is essential for characterizing fire behavior and assessing its ecological and management implications [57]. Fire intensity was estimated using the MCD14ML hotspot product [58], which provides the fire radiative power (FRP), a measure of the rate of radiant energy released by active fires [59]. The product was previously filtered by type to only include active vegetation fires with a minimum confidence level of 60%, recorded over the entire 21-year period. To minimize the influence of extreme values and potential outliers, fire intensity was quantified using the 95th percentile of the FRP distribution, rather than the maximum value [2,49].

Fire size is a key attribute of fire activity, being closely linked to burned areas, fire intensity, seasonality, and landscape topology. Across ecosystems, fire size distributions typically exhibit a strong skew, with a high frequency of small fires and relatively few large events [2]. Fire size was derived from the FIRED event perimeter product [47]. Given the strong skew in fire size distributions toward smaller events, the 80th percentile (5 × 5 km grid) and 95th percentile (10 × 10 km and 25 × 25 km grids) of the size distribution were used as the final values to minimize errors from unusually large fires. We considered a minimum of 10 records to calculate both fire size and fire intensity at all scales, except for fire sizes at 5 km, where a minimum of 5 records was used to avoid discarding too many grid cells due to the lower fire density in certain areas of the study region.

Finally, the Fire Seasonal Concentration Index captures the annual concentration of fire activity. It helps distinguish systems with brief fire periods (due to limited flammable conditions) from those with extended fire seasons, such as those in the dry tropics. Human activity can also alter these natural cycles, often anticipating the fire season and lengthening it beyond what weather alone would suggest [60]. The Fire Seasonal Concentration Index was obtained from the MCD64A1 product. To calculate the Fire Seasonal Concentration Index, mean monthly values of the burned pixel frequency distribution were transformed into vector quantities, with the magnitude determined by the number of burned pixels and the direction by the month of the year, expressed in angular units (e.g., 15° for January, 45° for February, etc.). The Fire Seasonal Concentration Index was then calculated as the ratio of the magnitude of the resultant vector from the 12 monthly vectors to the average annual burned pixels, expressed as a percentage [49]. This index indicates the degree to which fires are concentrated over a 12-month period. It ranges from 0 to 100%, where 0% means that fires are evenly distributed across all months, and 100% means that they are highly concentrated in a single month.

2.4. Data Analysis

Clustering analyses were performed at three spatial scales, using K-means and GMM algorithms to identify groups of cells with similar fire metric values, thus delineating different pyroregions. Prior to clustering, all fire metrics were Box–Cox-transformed, z-score-normalized, and centered. Internal validation analysis was then conducted to evaluate cluster quality and determine the optimal algorithms and spatial scale for the study. The stats and mclust packages in R were used for the K-means and GMM analyses, respectively. The mclust package was chosen because it facilitates the statistical comparison of groups with varying volumes, shapes, and orientations, using the Bayesian Information Criterion (BIC) to ascertain the most suitable data clustering techniques (Figure 2).

K-means algorithm:

Let X = {xi}, i = 1, …, n be the set of n-dimensional points (grid cells in our case) to be grouped into a set of K clusters, C = {ck k = 1, …, K}. K-means analysis can be used to perform a group partition that minimizes the squared error between the empirical mean of a cluster and the points in the cluster. Let μk be the mean of cluster ck. The squared error between μk and the points in cluster ck is defined as follows (1):

$$\mathrm{J}(\mathrm{ck})={\sum }_{x_i \in c_k}{\left|\left| x_i-{\mu }_k \right|\right|}^2$$

(1)

The goal of K-means is to minimize the sum of squared error over all K clusters (2):

$$\mathrm{J}(\mathrm{C})={\sum }_{k=1 }^K{\sum }_{x_i \in c_k}{\left|\left| x_i-{\mu }_k \right|\right|}^2$$

(2)

The process starts by randomly selecting k initial cluster centers (centroids) and assigning each data point to the closest centroid, forming initial clusters. Next, the algorithm calculates the mean distance of all the points and the center for each cluster, updating the centroid position to this new mean. Once updated, the data points are reassigned to the nearest centroids, forming new clusters. This iterative process continues until the centroids stabilize and no longer change, indicating convergence. The aim of K-means is to minimize the within-cluster variance, i.e., the sum of squared distances between data points and their respective centroids. Some limitations of K-means include its sensitivity to outliers, the need to predefine the number of clusters, its assumption of spherical clusters of similar size, and the hard assignment of each data point to only one cluster [22,23].

GMM algorithm:

Let x = {x1, x2, …, xi, …, xn} be a sample of n independent, identically distributed observations. The distribution of every observation is specified by a probability density function using a finite mixture model of G components:

$$f(x_i;\Psi )={\sum }_{k=1}^G{\pi }_k{f}_k(x_i;{\theta }_k)$$

(3)

where Ψ = {π1, …, πG − 1, $\theta$1, …, $\theta$G} are the parameters of the mixture model, $f(x_i;\Psi )$ is the kth component density for observation xi with parameter vector $\theta$k, (π1, …, πG − 1) are the probabilities (such that ${\pi }_k>0, {\sum }_{k=1}^G{\pi }_k=1$), and G is the number of mixture components [61]. Assuming that G is fixed, the mixture model’s parameters are usually unknown and must be estimated. The log-likelihood function corresponding to Equation (3) is given by l (Ψ; x1, …, xn,) $={\sum }_{i=1}^n\mathrm{l}\mathrm{o}\mathrm{g}\left(f\right(x_i;\Psi ))$. Given that the direct maximization of the log-likelihood function is complicated, the maximum likelihood estimator (MLE) of a finite mixture model is usually obtained via the expectation maximization algorithm (EM) [62,63].

In GMM, each cluster is represented by a specific Gaussian distribution, defined by mean and variance, and the algorithm iteratively adjusts these parameters to achieve the best fit [23,64]. Data assignment to each model is flexible, so data points have a probability of belonging to different groups. GMM have the advantage of automatically determining the optimal number of clusters. However, their primary limitation lies in their high computational complexity, particularly for large datasets or high-dimensional spaces [23].

The evaluation of the goodness of clustering was achieved through the implementation of internal validation measures, which employ mathematical formulations that rely exclusively on properties intrinsic to the data, avoiding any reliance on an external classification system [65]. Commonly used metrics include Connectivity and the Dunn Index. Connectivity assesses how well a cluster respects local densities and groups elements with their nearest neighbors in the data space. Connectivity values range from 0 to infinite, with lower values indicating higher cluster Connectivity [66]. The Dunn Index (0 to +∞) considers both the compactness and separation of clusters, where higher values indicate more compact and well-separated clusters [65].

3. Results and Discussion

3.1. Gaussian Mixture Model and K-Means Cluster Analysis at Different Scales

The GMM algorithm identified between six and nine clusters (i.e., pyroregions) as optimal at scales of 25 and 5 km, respectively, based on the BIC index (

Table 1. Optimal number of pyroregions (i.e., clusters) identified for the Argentine Gran Chaco by GMM (Gaussian Mixture Model) and K-means clustering algorithms at different spatial scales, based on internal validation measures and the BIC index.

Clustering Algorithm and Grid Scale (km)	Internal Validation Index		Optimal Number of Clusters
GMM	Connectivity	Dunn index	Based on BIC
5	7750.7070	0.0134	9
10	2124.4070	0.0177	7
25	494.5726	0.0181	6
K-means	Connectivity	Dunn index	Based on the Connectivity and Dunn indices
5	1327.0380	0.03246	2 and 10
10	649.8726	0.0312	2 and 9
25	178.0877	0.0644	2 and 10

Values in bold are optimal for each algorithm.

). Among these scales, Connectivity and Dunn indices highlighted the 25 km scale as the most appropriate, ensuring optimal goodness of clustering (i.e., minimum and maximum value for Connectivity and Dunn indices, respectively). For K-means analysis, the optimal number of clusters was determined using Connectivity and Dunn indices, as this algorithm does not inherently provide such information. The two indices revealed significant disparities; the Connectivity index consistently identified two clusters as optimal across all scales, whereas the Dunn index suggested that using between nine and ten clusters was optimal (Table 1). Despite these differences, both indices determined the 25 km scale as the most suitable. The decision to evaluate grid sizes of 5 km, 10 km, and 25 km was based on both preliminary tests (Figure A1 and Appendix A) and on the lack of consensus in the literature, where spatial resolutions vary widely—from 10 to 50 km at the regional scale [14,15,18,25] and from 25 to 100 km at the global scale [2,26]. These overlapping ranges underscored the importance of a scale comparison tailored to the specific fire dynamics of the study area.

As internal validation measures consistently identified the 25 km scale as the most appropriate for both GMM and K-means, the comparison between the two algorithms was conducted solely at this scale. An additional benefit of the 25 km scale lies in its greater area coverage in pyrogeographic analysis. Larger grid cells increase the likelihood of including more fires, allowing for the calculation of fire metrics based on percentiles, for which we established a minimum quantity of data (See Section 2.2). In our study area, the 5 and 10 km scales leave 52% and 40% of the Argentine Gran Chaco uncovered (Figure 3D,E), respectively, while the 25 km scale reduces this gap to only 21% (Figure 3C). This difference has significant implications for fire management in critical areas. For instance, the 5 and 10 km scales exclude the Serrano Chaco forest (Córdoba mountains) from the analysis (Figure 3D,E), a region with high fire activity and where megafires (fires ≥ 10,000 ha) occur regularly [34,67].

According to validation indices, K-means outperformed GMMs (Table 1). Nevertheless, the considerable disparity in the optimal number of clusters for K-means indicates that the most effective algorithm should necessitate an examination of results in light of fire ecology principles and fire management perspectives. The K-means analysis that resulted in two clusters was represented by groups of similar sizes (412 and 483 cells), exhibiting the strong aggregation of grid units into two broad clusters, with minimal isolated cells (Figure 3A). Cluster 2 had higher Annual Burned Area and fire frequency values, but smaller and less marked seasonality than Cluster 1. Cluster 2 coincides with the fire concentration area identified by Cavallero et al. [37] in a study of burned areas of Argentina using MODIS data. However, their subsequent assessment revealed significant variations in total burned area and fire frequency within this region [37], which remained undetected when using only two clusters. Additionally, Cluster 2 spans the entire precipitation gradient, whereas recent studies observed unimodal fire seasonality in drier areas, and bimodal seasonality in humid areas, also showing higher fire frequency and burned area [35]. The differences found in previous studies, along with well-documented climate gradients and vegetation heterogeneity across the area (Figure 1 and Figure 3A) [31,32], suggest that classifying the Argentine Gran Chaco into only two pyroregions fails to accurately capture its fire regime diversity. This oversimplification could lead to ineffective fire risk management strategies.

In contrast, the K-means result with 10 clusters identified small, fragmented, and spatially dispersed pyroregions (each comprising less than 12% of grid cells; Figure 3B). This pattern contradicts the expected spatial autocorrelation of fire activity [37,68]. Additionally, post-fire plant communities often exhibit increased flammability, reinforcing fire recurrence [69,70]. From a management perspective, designing and implementing differentiated fire risk strategies for multiple coexisting pyroregions within a confined area presents significant challenges. Ultimately, the GMM approach identified six pyroregions with a more spatially coherent distribution, demonstrating stronger and more consistent aggregation of grid cells into clusters (Figure 3C).

3.2. Pyroregions at 25 Km Scale

The six pyroregions identified using the GMM are characterized by a unique combination of fire metric distributions. They were labeled based on the three metrics that exhibited the most pronounced differences among them, in the following order: annual burned area (categorized in low, regular and high), interannual variability of fire (categorized in rare, occasional and frequent) and fire intensity (categorized in low, moderate, intense) (Figure 4).

The LRM pyroregion is defined by a low annual burned area and rare, moderate-intensity fires. It is distinguished by its large fire events (>1000 ha) and pronounced seasonality (Figure 3C and Figure 4). Uniquely, this is the only pyroregion in the study area to experience megafires—fires exceeding 10,000 hectares [71]. LRM is the most extensive pyroregion (22% of study area) and is predominantly within the Semiarid Chaco. On a global scale, LRM aligns with the RIL pyrome identified by Archibald et al. [2], which is similarly characterized by large fire sizes and the occurrence of megafires. Nevertheless, they differ in temporal and magnitude aspects, with the RIL pyrome exhibiting more than twice the frequency and intensity of the LRM pyroregion. Regionally, the southern portion of LRM overlaps with mountain areas dominated by natural woody vegetation, nearly half of which has experienced fires classified as moderate in frequency (1–3 events over 20 years; [37,72]. Livestock management in these areas involves the use of fire for the removal of woody vegetation and the renewal of forage biomass during the dry season [73]. Under favorable weather and fuel moisture conditions, these fires can spread rapidly, resulting in large burnt areas [67,74].

The ROI pyroregion is characterized by regularly burned areas, with occasional intense fires that are typically large and seasonally concentrated. This region accounts for 19% of the study area and spans the Semiarid Chaco and the southwest portion of the Humid Chaco (Figure 3C and Figure 4). Globally, this pyroregion is included within human-derived pyrome (ICS), which is associated with deforestation and agricultural activities [2]. At a regional scale, Cavallero et al. [37] highlighted high-intensity land use changes in this area between 2000 and 2019, driven by the conversion of xerophytic and subhumid forests into croplands and pastures. Burned areas have played a significant role in this transformation, with almost 70% of the land converted to agricultural use, having been burned with moderate frequency over the past two decades.

The ROM pyroregion has a regular burned area, with occasional and moderate-intensity fires. These fires are medium-sized and seasonality is marked. ROM spans 14% of the region and is distributed across the Humid Chaco and towards more humid areas of the Semiarid Chaco (Figure 3C and Figure 4). Although it is part of the human-derived pyrome, the ROM pyroregion only shares similarities in terms of fire size. Regionally, Cavallero et al. [37] observed a similar moderate fire frequency in natural woody areas located in the north-eastern part of our study area, belonging to the ROM pyroregion (Figure 3C). In this region, controlled and prescribed fires are used for vegetation management in silvopastoral systems [75].

The LOM pyroregion has a low-burning area, with occasional and moderate-intensity fires. Fire events are medium–small-sized and seasonally concentrated. The LOM region occupies 10% of the area and is predominantly located in the Semiarid Chaco (Figure 3C and Figure 4). This pyroregion also belongs to the human-driven pyrome [2], but is characterized by fires that are less intense and half the size. At the regional level, the pyroregion is situated outside the areas of concentrated fire activity identified by Cavallero et al. [37], aligning with the low-burning area observed in our study.

The ROL pyroregion is characterized by a regular burned area, featuring occasional and low-intensity fires that are typically large. Unlike the other pyroregions, where fires are highly seasonal (Fire Seasonal Concentration Index: 56%–71%, Figure 4), fires in ROL are more evenly distributed throughout the year, making it a notable exception. ROL covers 9% of the area and is primarily located within the Humid Chaco. Recent studies determined the bimodal seasonality of fires in Humid Chaco [35], which may explain the less pronounced seasonal fire patterns detected for ROL. This pyroregion occurs mainly in natural forest and mixed areas, located in the northeastern part of our study area (Figure 3C). According to Cavallero et al. [37], these areas experience moderate burn area and fire frequency, which is consistent with our results (Figure 1 and Figure 3C). In these more humid areas, there is a significant biomass accumulation and fuel desiccation plays a critical role in fire occurrence. Fires typically only occur under exceptional environmental conditions, such as heat waves or prolonged dry spells, and often impact valuable native and cultivated forests. Notably, this region was heavily affected by extensive and prolonged wildfires in 2020 and 2022, particularly in the province of Corrientes (east of the region) [76,77].

In addition, we defined a seventh pyroregion, consisting of cells excluded from the clustering analysis due to insufficient data for calculating the fire metrics (as we established a minimum number of data to calculate distributions for fire size and fire intensity metrics), despite still representing a relevant fire regime within the study area. The absence of sufficient data may be attributed to low fire activity during the 20-year period analyzed or to the inability of MODIS sensors to detect smaller fires due to their coarser spatial resolution. This pyroregion occupies 21% of Argentine Gran Chaco, and is mostly located in the Arid and Semiarid subregions (Figure 1 and Figure 3C). Although this region is globally associated with three distinct pyromes (RIL, RCS, ICS; [2]), at the regional scale, it lies within areas identified as having minimal concentrations of burned area during the study period [2]. In these Arid and Semiarid regions, limited fuel accumulation [37] and/or low human presence likely constrain fire occurrence.

Our study on pyroregion identification in the Argentine Gran Chaco region contributes to the development of specific fire risk management strategies according to fire regimes. For example, the LRM pyroregion is notable for its frequent megafires. These extreme fires are characterized by erratic spread, high intensity, and extensive fire fronts that overwhelm civil protection and firefighting capacities [78]. In this context, fire management actions may include planned and prescribed fuel treatments to reduce the risk of wildfires and adaptive management to create landscape mosaics [79]. The occurrence of megafires is increasing in other South American ecosystems as a result of climate change and current land use practices. This requires proactive landscape planning and investment in preventive measures. Sustainable land use practices such as agroforestry, silvopastoral systems, cultural burning and green firebreaks can play a critical role in reducing fire risk [80]. Another feature of the fire regime that may result in different management strategies is the length and timing of the fire season, as these define the timeframe for different actions. Longer fire seasons, as observed in the ROL pyroregion, prolong conducive burning conditions [81]. This increases the probability of escaped fires related to deforestation or forage regrowth, and reduces the time window for planned fire use in order to manage fuels [82].

3.3. Limitations and Future Research

The study is not without limitations, and several areas for improvement remain. For example, incorporating additional fire metrics, such as fire severity, could provide a more comprehensive representation of the multidimensional space that characterizes fire regimes [83]. However, the region still lacks a comprehensive database on fire severity, with data mainly limited to small-scale, local [84], and event-specific studies [85]. Additionally, the spatial and temporal resolution of some datasets may restrict the ability to detect smaller fire events and capture long-term variability in pyroregions. As satellite sensors accumulate more data over time, the time span of pyrogeographic analyses can be extended, leading to a more accurate characterization of pyroregions. Similarly, as the availability of high-resolution satellite data increases, a more precise and detailed delineation of fire patterns can be achieved.

Although global-scale fire regime characterizations (i.e., pyromes) already exist, regional and national studies, especially in South America (which harbors some of the most threatened forests in the world; [10]), should not be overlooked. This is evident in our findings. While most of the pyroregions identified in this study spatially align with the global-scale pyromes, our regional analysis reveals significant differences and heterogeneity in fire patterns that are not accounted for at the global level. This highlights the importance of carefully selecting the appropriate scale of analysis to accurately capture such patterns.

Looking forward, this methodological framework could be applied to other complex ecosystems worldwide, known to have a wide diversity of fire regimes, but this kind of pyrogeographic approach is still poorly developed. Moreover, understanding the interactions between pyrogeography and factors such as vegetation, biophysical, climate, and human impacts [26,86] can enhance the predictions of alterations in future fire regimes. This is particularly critical given the pressing environmental challenges societies face in the context of global change. These advancements can improve fire risk management, as well as mitigation and adaptation strategies, ultimately promoting sustainable coexistence with fire.

4. Conclusions

Our study, drawing on two decades of satellite fire occurrence data, provides the most comprehensive delineation of the pyroregions of the Argentine Gran Chaco’s Dry Forest to date. The regionalization of pyroregions was based on data availability, the scale of analysis, cluster algorithms, internal validation measures, and fire ecology and management criteria. This approach identified six distinct pyroregions within the ecoregion, as well as an additional indeterminate fire regime, offering valuable insights into the complexity of fire dynamics in the region.

The selection of an optimal scale for analysis played a crucial role in maximizing spatial coverage, as limited fire data can hinder the accurate calculation of fire metrics. Comparing multiple spatial resolutions was essential to identify the most appropriate scale for capturing regional fire regime patterns. Our findings support the use of the 25 km grid as the most suitable resolution, balancing detail with data sufficiency. This highlights the importance of a multi-scale approach in fire ecology research [87], particularly in regions like the Gran Chaco.

Internal validation measures identified a disparate number of optimal clusters for the K-means algorithm, ranging from two to ten. Instead, GMM identified an optimal number of clusters (6 clusters) based on BIC. Given the disparity in the number of optimal clusters identified by statistical criteria, we considered that the final delineation of pyroregions necessitated the consideration of fire ecology principles and fire management perspectives. The GMM algorithm demonstrated a more intuitive ecological and spatial distribution, coupled with stronger grid cell aggregation, which is an important factor for fire management.

The GMM algorithm at a 25 km scale provided the most suitable pyroregion delineation based on the above-mentioned criteria and data availability. Six different pyroregions were assessed for the Argentine Gran Chaco’s Dry Forest, plus an additional indeterminate pyroregion, for which fire metrics could not be calculated due to insufficient data. (i.e., fire activity was low in the 20-year period analyzed and/or fires were undetected at the spatial resolution of MODIS sensors). This indeterminate fire regime occurs in arid and semiarid areas with low human presence, where fuel and/or human ignitions may be limiting fire activity.

Pyroregions in the Semiarid Chaco recorded mainly low-burning areas and moderate-intensity fires, LOM and LRM, with the former having occasional and medium-sized fires, whereas the latter shows rare but large fires, including megafires. Pyroregions including sub-humid and humid areas have higher burned areas and more frequent fires than those including Semiarid Chaco areas. The main differences among humid and semiarid areas are based on fire intensity (ROM, ROL, ROI). In general, the majority of pyroregions have a marked seasonality of fire occurrence, diminishing towards more humid areas (i.e., ROM and ROI). Finally, a pyroregion characterized by a highly burned area, large and frequent but low-intensity fires (HFL), was identified in the south of Humid Chaco. While most of the pyroregions identified in this study align with the human-derived pyrome at the global scale, a regional perspective uncovers significant heterogeneity that remained unaccounted for at the global level.

This work highlights the effectiveness of applying pyrogeography concepts and machine learning methods, supported by satellite products, to uncover patterns in fire regimes. It emphasizes the importance of determining the optimal scale of analysis to accurately characterize the diverse attributes of the pyroregions, and underscores the critical importance of selecting an appropriate clustering algorithm to adequately detect patterns in heterogeneous landscapes like the Gran Chaco Dry Forest. Furthermore, it points out that pyroregions cannot be delineated solely on the basis of scale and cluster analysis. It is equally important to integrate fire ecology principles and fire management perspectives to ensure the more comprehensive and meaningful characterization of pyroregions.