Human‐Driven Fire Regime Change in the Seasonal Tropical Forests of Central Vietnam

To better understand fire regimes and their relation to climate in the seasonal tropical forests of continental Southeast Asia, we developed the first multi‐century tree‐ring based fire history chronology for the region. The chronology included 776 fire scars collected at Bidoup NuiBa National Park (BNNP) in the Central Highlands of Vietnam and spans the period 1636–2020. Fires were recorded in 116 years, representing 47% of the years covered by the 249‐year period between the first fire scar (1772) and the last (2020). While only 9% of years within the sampled BNNP forests experienced fires before 1905, 70% recorded fires between 1906 and 1963 and 90% showed evidence of fire after 1963. Fire occurrence was highly correlated with climate indices (wet season Nino 3.4 and dry season regional Palmer Drought Severity Index) during the period 1906–1963, but showed no significant correlation after 1963. Our fire reconstruction from BNNP suggests that the fire regime has shifted from one driven primarily by climate to one in which human activities dominate the occurrence of fire within these seasonal tropical landscapes.

Understanding how fire regimes influence biodiversity and forested ecosystems is a priority for natural resource management. Fire regimes are primarily characterized by their frequency, seasonality, intensity, and fire extent. In tropical forests, the fire frequency or return interval may range from years to centuries and the intensity from low to high (i.e., flame lengths from centimeters to meters) (Cochrane, 2009). Variability in frequency and intensity are both driven by vegetation and climatic conditions, but in many regions are also influenced by humans. Fuel flammability is particularly sensitive to moisture, which is controlled by climate. Periodic droughts associated with El Niño events and other sources of supra-annual climatic variability may also increase the likelihood of fires. When a drought is severe enough, forest fires can spread into forests where fire is relatively rare (Baker & Bunyavejchewin, 2017). In some regions, the length of the fire season has also increased (Westerling et al., 2006). Model predictions suggest that warming and drying climatic conditions are likely to lead to more frequent, and potentially more intense, fires (Bergeron et al., 2004;Dale et al., 2001;Flannigan et al., 2009;Meyn et al., 2010). However, while fires require suitable climatic conditions and sufficient fuel loads, they also need an ignition source. Dry lightning has long been an important ignition source across many biomes, but so too have humans, particularly in recent centuries (Farjon et al., 1999;Gowlett, 2016;Thompson et al., 2021). In many regions with a long history of settlement, humans have replaced other ignition sources and are now the primary source of ignition (Baker & Bunyavejchewin, 2009;McLauchlan et al., 2020). For example, over 85% of the 1.5 million forest fires in the eastern and coastal areas of the United States were caused by humans during the period 1992-2012 (Balch et al., 2017). Humans have also altered fire regimes through fire suppression. For example, in the western US active fire suppression initiated in the early 1900s has dramatically altered forest structure and composition (Hagmann et al., 2021). This has made forests that were once dominated by frequent, low-intensity prone to devastating high-intensity fires.
Understanding fire regimes and how they have changed over time is critical for understanding and predicting future fire regimes. Disentangling the interactions between climate, fuels, and humans requires better data on the timing and intensity of past fires and their drivers. This would also provide important insights into the historical development of forests within landscapes and across regions and would benefit fire management across multiple scales (Vadrevu et al., 2019). There is now a large and growing body of research on this subject; however, it has been almost exclusively limited to temperate forests because it is heavily dependent on dendroecological techniques that require trees that form annual growth rings (Amoroso et al., 2017).
In tropical forests, where few tree species form annual growth rings, our knowledge of long-term forest dynamics and the role of disturbances is poor. Despite a long history of human occupation in the seasonal tropics, contemporary understanding of the influences of human settlement and climate on fire regimes remains rudimentary in these regions. Large areas of continental southeast Asia experience a strongly seasonal climate and recent studies have identified dozens of tree species with annual growth rings (e.g., Baker & Bunyavejchewin, 2006;Buckley et al., 2010;Sano et al., 2008;Zuidema et al., 2011) that might help to better understand the history of fire in these landscapes. In some parts of the region, extensive areas of fire-prone pine savannas are dominated by Pinus kesiya or Pinus merkusii (Ashton, 2014;Hiep et al., 2005). The Central Highlands of Vietnam have more than 100,000 hectares (ha) of P. kesiya forest (Lung, 1983) that routinely experience fire. P. kesiya forms annual growth rings (Pumijumnong & Wanyaphet, 2006) and trees that survive fires often have fire scars that have formed as a result of heat-induced cellular death across parts of the vascular cambium. Tree-ring records from Vietnamese conifers indicate that the Central Highlands of Vietnam have also experienced several periods of extreme drought over the past 700 years . The Da Lat Plateau in the Central Highlands has also been occupied for centuries to millennia by the indigenous Lat people (Farjon et al., 1999).
The Central Highlands region of Vietnam has a strongly seasonal climate with a dry season that lasts for 4-5 months with little to no rain. The high temperatures, lack of rainfall, and low humidity, particularly toward the end of the dry season in April and May, create conditions suitable for fires in these landscapes every year. Regional variability in climate, particularly the intensity of the dry season, is strongly influenced by large-scale ocean circulation dynamics, such as the El Niño-Southern Oscillation (ENSO) . During strong El Niño conditions continental Southeast Asia, including Vietnam, experiences intense drought conditions, which have been associated with larger and more intense fires. Long-term records of fire activity might provide useful insights into the interaction between climate drivers such as ENSO, human activities, and fire across these seasonal tropical landscapes.
In this paper we use tree rings and fire scars from P. kesiya and Keteleeria evelyniana to develop a multi-century record of fire activity in the Central Highlands of Vietnam. We use this record to provide a long-term historical context for 20th Century fire activity within the region.

Study Sites, Tree Species, and Fire Scar Collection
The study sites are located at Bidoup Nui Ba National Park (BNNP) in the southern part of Vietnam's Central Highlands (12°4′60″N, 108°40′0″E). BNNP occupies a total area of 64,800 ha and is one of five recognized biodiversity hotspots in Vietnam. Ninety-nine percent of BNNP is forested and the vegetation of the Park is dominated by montane evergreen forest with small patches of coniferous forests and mixed conifer-broadleaved forests. The climate of the Central Highlands is seasonal and monsoonal with mean annual rainfall of 1,810 mm. The dry and cool season lasts for 4 months (December-March); mean monthly rainfall over this period is 35 mm. Mean monthly rainfall over the warm and wet season (May-November) is >200 mm. The transition from the dry to wet season occurs between late March and early May. Mean annual temperature is 18°C. Climate data are based on observations from 1980 to 2020 at the Dalat meteorological station which is ∼40 km from the study sites.
We used a targeted sampling method to collect fire-scarred samples from 12 sites across the landscape (10 sites for P. kesiya and two sites for K. evelyniana). We qualitatively assessed stands based on tree diameter at breast height and crown architecture to identify stands that supported relatively old trees to ensure the longest possible fire chronology. At each site, we identified all fallen P. kesiya and K. evelyniana within an area of ∼1 ha. We focused on recently fallen and wind-blown trees, which are relatively common within the landscape, to enable cutting of cross-sections for fire scars. However, we also obtained samples from some living trees ( Figure S1 in Supporting Information S1). Fire-scarred trees and stumps with multiple visible fire scars were prioritized to build the longest possible fire chronology. Partial cross sections of living trees and whole cross-sections of stumps and fallen trees that showed evidence of fire scars were collected with a chainsaw at 20 cm above the base of the stem. Where there was evidence of fire scars in living K. evelyniana trees, we used increment cores to collect samples from the fire scars.
Tree cores and fire-scarred cross-sections were transferred to the laboratory where they were air-dried. To ensure a high-quality surface for dendrochronological analyses the samples were sanded using a belt-sanding machine with a series of increasingly fine sandpaper grits. To ensure accurate dating of all fire scars, we used a multistep procedure to crossdate the samples (Stokes & Smiley, 1968). We began by visually crossdating series taken from the same tree using skeleton plots. We then crossdated across trees using distinctly narrow or wide years or sequences of years in the tree-ring series. This allowed us to accurately allocate annual growth rings to the year each was established.
Radial growth in trees of the monsoonal tropics of continental southeast Asia follows a distinct phenological cycle. Growth begins in April-May with the onset of the monsoon. During this period the earlywood in conifers such as P. kesiya is formed. Growth slows and latewood forms toward the end of the monsoon period (October/ November) and ceases before the onset of the dry season (December) (Pumijumnong & Wanyaphet, 2006). Fire scars typically occur in between the latewood and earlywood because this is a period of growth dormancy associated with the dry season (and the period of the year during which all fires occur). To ensure that the fire records are correctly aligned with the climate data, we assigned fire scars that were formed during the dormant season to the subsequent year. For example, if a fire scar was recorded after the latewood of the 2019 growth ring, it was assumed to represent a fire that occurred during the dry season in 2020 (i.e., after radial growth in 2019 had ceased but before radial growth in 2020 had begun).

Statistical Analyses
We identified years in which fires occurred from the fire scars in the cross-dated wood samples. We used the Fire History Analysis and Exploration System (Brewer et al., 2016) to combine the exact year and season of forest fires into a master chronology for each site and to analyze fire-climate relationships. Mean, median, minimum, maximum, and Weibull median fire intervals were calculated for each site. Within a single site we assumed that years with no fire scars experienced either no fires or fires that were not sufficiently intense to wound the cambium of the trees. In years in which a relatively low percentage of trees were scarred, we assumed that a fire had occurred, but that the fire was either a low-to moderate-intensity fire or that it was spatially heterogeneous in intensity. For those years in which a high percentage of trees had fire scars, we assumed that a moderate-to high-intensity fire had occurred. We note that given the distances amongst the individual sampling plots, we do not assume that the presence of fire at more than one site is indicative of landscape-scale fire. That would require a sampling intensity that was beyond the scope of this study.
To identify synchronous fire events in the Central Highlands of Vietnam, we developed a composite fire chronology for each site. We then combined the site series into a master composite fire chronology for BNNP. To understand variability in historical fire activity within the broader BNNP landscape, we identified six clusters of sites based on proximity and local landscape conditions. For convenience we refer to them based on the name of the nearest local ranger stations. Composite fire chronologies provide an overview of past fires at the landscape level. Since fire return intervals at each site are influenced by the number of samples collected (Horne & Fulé, 2006), we produced three composite fire chronologies. The broadest composite (C01) consisted of all samples that recorded one or more fire scars. The intermediate composite (C10) included fires that scarred a minimum of two trees and at least 10% of the samples. The final composite (C25) comprised fires that scarred a minimum of three trees and at least 25% of all sampled trees. Filtering fire scar data helps minimize potential misinterpretations of site fire histories when comparing sites with different sample sizes and sample area (Falk et al., 2011). The composite of all trees recording at least one fire provides a detailed record of fire events, while the filters eliminate relatively small fires. Fire return intervals were calculated for each site and for the entire study area.
The location of fire scars within the annual growth ring can provide useful information on the season in which the fire occurred (Capprio & Swetnam, 1995). Based on well-described patterns of the seasonal phenology of growth ring formation in P. kesiya (Pumijumnong & Wanyaphet, 2006), we classified the season in which a fire occurred based on the location of the fire scar within the annual ring. Scars in the latewood/earlywood boundary were classified as Dormant season (D) fires; scars formed in the earlywood as Early Monsoon season (E) fires; and scars formed in latewood cells as Late Monsoon season (L) fires. If the fire season could not be determined, we recorded it as unknown (U).
We conducted a series of superposed epoch analyses (SEA) using the R package BurnR (Malevich et al., 2018) to evaluate the relationship between our fire history record and various climate predictors. For the SEA we used the C25 fire chronology for all sites combined (i.e., BNNP-wide) that included a minimum of three scarred trees and at least 25% of sampled trees recording a fire. The first fire year was in 1894 and the last fire was in 2015, over this period our sample depth ranges from 28 to 206 samples. We then used SEA to analyze fire events in relation to Nino 3.4 value (Rayner et al., 2003) in July, August, and September (JAS) (i.e., the core months of the rainy season), and Southern Oscillation Index (SOI) (Ropelewski & Jones, 1987), and instrumental Palmer Drought Severity Index (PDSI) value in January, February, and March (JFM) (the core months of the dry season). We extracted instrumental PDSI values for BNNP (12-12.5 N, 108-108.5 E) from the Royal Netherlands Meteorological Institute (https://climexp.knmi.nl) for the period 1901-2020. For all analyses, we included climate data extending 3 years prior to the fire year. Statistical significance was calculated with a 95% bootstrapped confidence interval based on 1,000 simulations.
We used breakpoint analysis to test for structural changes in the cumulative number of fires from 1636 to 2020. To do this, we applied the empirical fluctuation process test (Brown et al., 1975) with the ordinary least squared-based cumulative sums tests following Ploberger and Krämer (1992). The Bayesian Information Criterion was used to determine the number of breakpoints in the time series. We conducted the breakpoint analyses using the strucchange package (Zeileis et al., 2002). Where we identified a significant change in fire activity, we split the fire record at the identified year and conducted separate fire-climate analyses for the different time periods. All analyses were conducted using the R statistical programming software (v. 4.1.0, R_Core_Team, 2021).

Results
We collected a total of 221 cross-sections and 49 cores from 12 sites across the BNNP landscape. However, only 183 partial sections and 23 cored trees were successfully cross-dated. The remaining samples could not be crossdated due primarily to the presence of decayed wood and were excluded from further analysis. The resulting chronology included 776 fire scars covering the period 1636-2020 (Table 1). Over the 249 years between the between the first fire scar (1772) and the last (2020), there were 116 years (47%) in which at least one fire was recorded within the BNNP landscape. We grouped the 12 sites into six clusters based on their locations within BNNP for further analysis (Figure 1 and Tables S1 and S2 in Supporting Information S1). The highest number of fire scars recorded across the clusters was 238 at Bidoup (BD) (mean fire scars per site = 79.3); the lowest number was 26 at Da Long (mean fire scars per site = 26). The number of fire scars per cross-section ranged from 1 to 16, with the highest number of scars in a single section found at BD. The number of fires in the area increased dramatically after 1900 (Table 1).
Breakpoint analysis suggested that significant changes in the cumulative number of fires occurred in 1906 and 1963 (S = 6.796, p < 0.001, OLS-CUSUM test, Figure S6 in Supporting Information S1). From 1772 (the year of the first observed fire scar) to 1905, only 17% of years within the sampled area experienced forest fires; during 1906-1963 71% of all years showed evidence of some fire and after 1963 90% of all years experienced forest fires (Table 1, Figure 2a). Because of the relatively small number of samples dated prior to 1905, we focused on the period 1906-2020. There was a statistically significant difference between the median fire interval prior to 1964 and from 1964 onward across the studied sites (t = 2.55, p = 0.02, Figures 2b and 2c).
The season of fire-scar formation was also recorded for all 776 fire scars. As expected, almost all scar positions were in the dormant season (99%) or late in the growing season (1%). There was no evidence of fires occurring during the rainy season ( Figure S3 in Supporting Information S1).
Over the past 249 years the fire regime at BNNP has been dominated by frequent, low-intensity fires that have had patchy spatial impacts ( Figure S2 in Supporting Information S1). Fires sufficiently intense to wound the cambium of trees occurred somewhere within BNNP almost every year since 1900. While some El Niño years All samples (1636All samples ( -2020All samples ( ) 1636All samples ( -1905All samples ( 1906All samples ( -1963All samples ( 1964All samples ( -2020 Length fire chronology ( are associated with heightened fire activity ( Figure S2 in Supporting Information S1), there is no consistent or significant pattern of association of extensive fire events with El Niño. For example, while the highest number of fire scars (47 at 9 of the 12 sites) occurred during the extreme 1982-1983 El Niño event and high numbers of fire scars were also recorded during the very strong El Niño events of 1994-1995 and 1997-1998, there were also high numbers of fire scars during weak El Niño years (1987)(1988)(2002)(2003) and very low numbers of fire scars during other El Niño years (e.g., 1918-1919, 1965-1966) (Figure S4 and Table S3 in Supporting Information S1).
Breakpoint analysis indicated a significant change in fire activity in 1906 and 1963 ( Figure S6 in Supporting Information S1). We limited our analyses of the fire-climate interactions to the period 1906-2020 because (a) the number of fire scar samples from before 1906 was too small for statistical analysis and (b) we had instrumental data for ENSO variability (i.e., Nino 3.4) from 1906. SEA indicated that during the period 1906-1963 fire occurrence was negatively associated with instrumental PDSI JFM and Nino 3.4 JAS 2 years before the fire ( Figure 3). However, there were no correlations between fire occurrence and regional PDSI JFM , SOI JFM , or Nino 3.4 JAS for Vietnam's Central Highland in the period after 1963.

Discussion
Our dendrochronological reconstruction of landscape-scale fire activity at Bidoup NuiBa National Park (BNNP) in the Central Highlands of Vietnam is the first of its kind in seasonal tropical forests in Southeast Asia. Our results indicate three distinct periods in the fire history record that we developed for BNNP: (a) prior to 1905 when there are relatively few tree-ring records, (b) 1905-1963 when there are many tree-ring records and fire activity is frequent at the landscape scale, but infrequent at any single site, and (c) 1964-present when fire activity was frequent at both the landscape and the site scales. Prior to 1900, the fire regime at BNNP was characterized by patchy, low-intensity fires. There were only 32 fire scars recorded during the period 1636-1905. This may reflect either infrequent fires, reduced detectability of fires due to small sample sizes, or a combination of both. However, absent an independent data set on fire activity or historical records from BNNP, it is difficult to determine which is more likely.
In contrast, after 1905 the sample size in our data set is much larger and it is possible to break the 20th Century into two distinct periods with high confidence ( Figure S7 in Supporting Information S1). During the first period  there is considerable fire activity across the many sites, but relatively few sites experience fire in any given year. In addition, fire activity is significantly correlated with regional climatic conditions. Both the wet Orange and red bars indicate statistically significant correlations at the 95% and 99% confidence level, respectively, based on bootstrap sampling (see Methods for details).
season Nino 3.4 and dry season regional drought (PDSI) indices are significantly associated with the number of fires across BNNP. After 1963 fire becomes even more frequent and the number of sites that experience fire in any given year increases. However, during this period, the relationship between the number of fires and regional climate indices disappears. If fire is occurring nearly every year, inter-annual variability in climate is unlikely to be driving this shift. It is worth noting that our analyses focus on fire occurrence across the landscape (i.e., the number of fires). Other features of fire activity, such as total area burned or fire intensity, which may be more sensitive to interannual climate variability, were not explicitly quantified. Whether they are correlated with interannual variability and, if so, whether that relationship has also weakened or disappeared since 1963, remains as yet unknown.
Across large areas of continental Southeast Asia common and widespread forest types such as the pine savannas and deciduous dipterocarp forests support a significant ground layer dominated by C 4 grasses (Bunyavejchewin et al., 2011;Nguyen & Baker, 2016;Ratnam et al., 2016). During the annual dry season these grasses cure and become flammable. Growth of annual grasses during the wet season and the accumulation of pine needle litter in P. kesiya forests at BNNP can generate as much as 20 tonnes per hectare of flammable material each year (Huong, 2007). Several studies have reported that the peak season for fire occurrence in the Central Highlands of Vietnam is during the late dry season (late February to early April) (Huong, 2007;Le et al., 2014). This creates conditions conducive to fire occurrence where ignition sources are present. While fires can occur in these forests in most years due to the availability of fuel and appropriate climatic conditions during the dry season, they appear to have been historically ignition limited. The strong association between La Niña (i.e., wetter) conditions during the rainy season 2 years prior to fire occurrence likely reflects an increase in production of fine fuels such as grasses and pine needles and the subsequent time required for them to dry and cure. The significant negative correlation with PDSI JFM suggests that the more intense dry season conditions more effectively reduce fuel moisture levels. The combination of a wetter rainy season followed by a drier dry season would increase total fuel loads and then increase flammability and connectivity of the ground story fuel matrix. Dry lightning is relatively rare across the region with most lightning accompanied by rains during the monsoon. And, while the Da Lat Plateau has been occupied for millennia by the indigenous Lat people who practised slash and burn agriculture, they lived at relatively low densities and are not known to have burnt large areas of forest (Farjon et al., 1999). As such, extensive fires most likely occurred following years in which a wetter wet season followed by a drier dry season produced more abundant and more flammable fuel loads and one or a few ignitions could lead to fires burning over atypically large areas.
The SEA showed statistical associations between the fire scar chronology and both the instrumental PDSI for the dry season and Nino 3.4 for the wet season during the period 1905-1963, but showed no significant correlation for either after 1963. This suggests a fundamental shift in the fire regime at BNNP from one in which fire occurred frequently within the landscape, but was spatially patchy, to one in which almost all sites burn in most years. The timing of this shift is associated with several government policies designed to reduce the density of rapidly growing urban areas by resettling people in rural settings. Resettlement programs in the 1960s and the establishment of the New Economic Zone program in the 1980s and 1990s (Dien et al., 2013) encouraged transmigration of people from high-density urban areas to low-density regional areas ( Figure S5 in Supporting Information S1). The program offered people housing and permits to clear forest land for agricultural activities. The rapidly increasing number of people within these landscapes and the widespread use of fire to clear land for crops and to prepare agricultural sites during the dry season for sowing during the early wet season meant that ignition was no longer a limiting factor (Le et al., 2014). In effect, the landscape became saturated with potential ignition sources.
The strong intra-annual seasonality of precipitation associated with the Asian monsoon also influences leaf phenology across the region. Many tree species lose some or all of their leaves during the dry season. This increases insolation and reduces relative humidity on the forest floor, drying the fine fuels and increasing their flammability. The combination of a ubiquitous ignition source, the annual dry season, and abundant and flammable fine fuels means that widespread fires now occur in most parts of these landscapes nearly every year.

Conclusions
Long-term variability in fire regimes in tropical landscapes is poorly understood because of the absence of historical and proxy records of fire. We used fire scars from two conifer species with reliably annual growth rings to develop an annually resolved, multi-century fire history for a forested landscape in the Central Highlands of Vietnam. The BNNP fire history shows that fire has been present within this landscape for hundreds of years, but that since the early 1960s the fire regime has categorically changed. A large influx of people associated with resettlement programs led to higher population densities and a dramatic increase in ignition sources. Since then, fire in these landscapes is effectively no longer ignition limited. In strongly seasonal forests such as these, the ubiquitous presence of ignition sources has led to nearly annual fires across the landscape.

Data Availability Statement
• Data is publicly uploaded to Zenodo (https://doi.org/10.5281/zenodo.8020559). • The Nino 3.4 data used for analyzing the association between fire and climate factors in the study are from Rayner et al. (2003). • The SOI data used for analyzing the association between fire and climate factors in the study are available at Ropelewski and Jones (1987) and https://crudata.uea.ac.uk/cru/data/soi/. • The instrumental PDSI data used for analyzing the association between fire and climate factors in the study are available at the following source: https://climexp.knmi.nl. • Version 2.0.2 of the FHAES software used to combine the exact year and season of forest fires into a master chronology available via and developed openly at https://www.frames.gov/fhaes/home. • The R package BurnR used for evaluating the relationship between our fire history record and various climate predictors available via and developed openly at Malevich et al. (2018) and https://cran.r-project.org/web/ packages/burnr/vignettes/introduction.html.