Version of Record: https://www.sciencedirect.

com/science/article/pii/S030147971830968X
Manuscript_d9010dfd985e361b6e3ab4c60d374ed8

Effects of Landscape Pattern and Vegetation Type on the Fire Regime of a Mesic
Savanna in Mali.

Paul Laris • Aurahm Jo • Suzanne P. Wechsler

P. Laris (corresponding author)

Department of Geography,
California State University Long Beach
1250 Bellflower Blvd.
Long Beach, CA 90840
email: Paul.Laris@csulb.edu
Phone: (562) 985-2222

A. Jo
1145 Northumberland Dr.
Sunnyvale, CA 94087

S. P. Wechsler
Department of Geography,
California State University Long Beach
1250 Bellflower Blvd.
Long Beach, CA 90840

Date of the manuscript revision: July 27, 2018

Original research.

© 2018 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
Effects of Landscape Pattern and Vegetation Type on the Fire Regime of a Mesic Savanna in
Mali.

ABSTRACT

Savanna fires are a critical earth-system process that alter vegetation regionally and contribute
to changes in atmospheric composition globally. The fire regime in savannas has shifted over
time resulting in a large reduction in burned area. Savanna fires, which are human caused and
set for a plethora of reasons, produce complex mosaic burned area patterns that tend to result in
lower overall burned area. Mosaic fire regimes are difficult to detect and map accurately using
available satellite data. Imagery-induced low-resolution bias makes it difficult to link fires with
relevant environmental and anthropogenic factors, while higher resolution imagery is
temporally less frequent. We explore how landscape pattern affects the fire regime in a mesic
savanna by quantifying relationships between the spatial patterns of vegetation, which are
shaped by natural and human factors, widely used ecological indices, and the seasonality and
frequency of fires. The study finds that landscape pattern influences the fire regime; fire
seasonality and frequency varied by landscape index at both the vegetation class and landscape
scales. Percent cover, shape index and largest patch landscape ecological indices demonstrated
the most consistency in burn date trends across scales. The study finds that landscape
fragmentation—specifically a reduction in the size of patches and an increase in their number–
results in an earlier fire regime. We conclude that fire intensity and severity will continue to
decline as agriculture expands and landscapes fragment from agriculture in savannas. Our
methods also demonstrate the ability to integrate landscape indices with coarse-resolution fire
data.

Keywords: savanna; fire; landscape indices; patch, mosaic; seasonal-mosiac; MODIS

1
 1
2 1.0 INTRODUCTION
3
4 There is a growing concern that climate change will cause larger and more catastrophic
5 fires around the world. Indeed, several massive or so-called “mega-fire” events have already
6 been attributed to a changing climate (Flannigan et al., 2009; Pyne, 2008; Williams et al., 2011).
7 Wildfires produce carbon emissions equivalent to 26%–31% of those that stem from fossil fuel
8 combustion and industrial activities (Schultz et al., 2008; van der Werf et al., 2006). It is thought
9 that increased fire activity will be one of the major drivers of future vegetation change under a
10 warming climate resulting in feedback loops that generate even more fire (Overpeck and Udall,
11 2010; Williams et al., 2010). However, while certain biomes have clearly experienced recent
12 increases in fire size and intensity due to climate change, a recent study finds that globally,
13 burned area declined by nearly one quarter between 1998 and 2015 (Andela et al., 2017).
14 How can fire events on the earth be simultaneously intensifying and decreasing? The
15 answer is that the global decline in fire has been driven primarily by large decreases in burned
16 area in the tropical savannas. The documented decline in savanna burning largely overrides the
17 impacts of climate change globally. In large part this is because savanna fire regimes are not
18 climate driven—they are a function of human land use and burning practices (Archibald, 2016;
19 Archibald et al., 2013; Laris, 2013) especially in the mesic savannas of Africa which burn the
20 most. In the African savannas, for example, the total area burned has declined by 1.27-1.7% per
21 year over the last seventeen years resulting in a total decline in area burned between 22.9%-
22 30.6% (Andela et al., 2017). The decline in fire has resulted in both fewer and smaller fires,
23 bucking the trend of increasing mega-fires in the more temperate biomes.
24 That a change in savanna burning could have global impacts is not surprising given that
25 savannas are the most frequently burned biomes on Earth averaging 2.8 million sq-km per year
26 (Giglio et al., 2006; Tansey et al., 2008). This burning potentially impacts ecosystem processes,
27 biogeography (Laris, 2011; Staver et al., 2011), the global carbon cycle (Bond et al., 2005) and
28 human livelihoods (Mistry, 2000). Understanding factors that produce changes in savanna fire
29 regimes are critical for scientists seeking to predict changes in savanna tree and grass cover,
30 improve land and fire management, as well as evaluate greenhouse gas emissions from fires.
31 Yet, as the work by Andela and colleagues (2017) clearly points out, all of the commonly used
32 fire models failed to reproduce the pattern and magnitude of observed declines in savanna fires.
33 In summary, we know very little about how fire regimes change in savannas.
34 Fire is a landscape-scale phenomenon and, as such, landscape pattern plays a critical
35 role in determining whether and how a fire burns through a landscape, a classic case of pattern
36 influencing ecological process (McKenzie et al., 2011; Turner, 1989). Savannas, by definition, are
37 composed of scattered trees and a near-contiguous herbaceous layer that desiccates during a
38 prolonged annual dry season (Scholes and Archer, 1997). Savanna fires are surface fires that
39 burn low-lying fuels (predominantly grasses but also shrubs, small trees and leaf litter) and
40 require a continuous and flammable fuel bed to propagate across a landscape. However, despite

1
 1 the near-contiguous fuel-bed, most savannas burn in highly fragmented patterns (Laris, 2011;
2 Mistry and Berardi, 2005; Randerson et al., 2012; Russell-Smith et al., 1997).
3 It is well established that the fire regime—defined as the frequency, intensity, type and
4 regular spatiotemporal pattern of fires in a particular place—is a key determinant of woody
5 vegetation cover and growth rates in a savanna (Higgins et al., 2000). In savannas, fire intensity
6 is thought to increase during a long dry season as vegetation cures (Govender et al., 2006). In
7 the savannas of Africa, where fires burn more widely and frequently than any other region,
8 there is a need to develop a better understanding of how fire regimes, vegetation patterns and
9 the broader environmental and human factors that create them, are interrelated.
10 The majority of savanna fires are set by people and burn in highly fragmented and
11 patchy patterns. The common practice of setting fires to vegetation as it gradually cures,
12 referred to as patch- or seasonal-mosaic burning, creates a burned mosaic landscape that is
13 heterogeneous (Laris, 2011; Parr and Brockett, 1999; Russell-Smith et al., 1997). While natural
14 features such as riparian woodlands, wetlands, streams, or rocky outcrops play a critical role in
15 governing fire propagation, especially in highly humanized landscapes, their role is secondary
16 to more subtle variations in vegetation cover for much of the savanna. For example—different
17 patches of savanna dry at different rates. The majority of savanna fires are set so as to
18 extinguish at the boundaries of herbaceous vegetation that are either too moist to burn or have
19 already burned at the time of the passage of a fire creating patchy burn patterns ( Archibald et
20 al., 2009; Laris, 2011). The pattern of burning and the landscape mosaics it creates have
21 important implications for ecological processes, biodiversity, carbon sequestration and human
22 land uses (Bird et al., 2012; Parr and Brockett, 1999; Russell-Smith et al., 1997).
23 Unlike in semi-arid landscapes which form the basis of the patch-mosaic theory, in
24 mesic savannas (precipitation 750-1500-mm) there is sufficient rainfall for burned vegetation to
25 recover following a fire over the course of a single wet season. As such, the fire regime in a
26 mesic savanna is not driven by ecological succession, because sufficient fuel to carry a fire is
27 generated on an annual basis. Additionally, burning in mesic savannas is seasonal- as opposed
28 to patch-mosaic burning because the same patches tend to burn on a regular annual basis and at
29 specific times during the fire season (Laris, 2011).
30 Three main characteristics of landscape patches—strength of patch formation (the
31 degree to which patches are differentiated from the surrounding area by their distinct, within-
32 patch homogeneity), patch size, and patch repeatability across a landscape—are commonly
33 used to describe patch structure of vegetation (Koerner and Collins 2013). These characteristics
34 can be quantified using landscape indices (Turner, 1989; White and Pickett, 1985). A key principle
35 of landscape ecology is that the relationship between pattern and process is scale dependent
36 (Wu and David, 2002).
37 Why does research find that subtle natural landscape features—specifically vegetation
38 types—correlate with fire regimes while anthropogenic landscape features do not? A major
39 reason appears to be the scale of analysis. Most research studying relationships between
40 anthropogenic variables and fires have used widely accessible imagery-derived fire products
41 generated from coarse-resolution satellite data such as AVHRR (1-km) or MODIS (0.5-km).

2
 1 These products contain a well-known “low-resolution bias” which creates a fire bias in which
2 larger fires are detected and mapped while smaller ones are missed (Boschetti et al., 2004; Laris,
3 2005). The error contributed by this bias can be substantial because algorithms are designed to
4 “miss” small fires and fragmented burned area patterns that are common in many savannas
5 (Caillault et al., 2015; Laris, 2005; Randerson et al., 2012). Bias also occurs in the form of the
6 “mixed pixel” problem when mapping key anthropogenic features such as agricultural fields
7 that are often only a few hectares in size (Laris, 2011).
8 The purpose of this study was to determine relationships between patterns of savanna
9 vegetation types and key parameters of the fire regime—seasonality and frequency of fire—as
10 well as their variability, for an area of West African mesic savanna in Mali. We explore the
11 relationship between fire and landscape pattern at multiple scales, recognizing and addressing
12 the aforementioned low-resolution and mixed-pixel biases. In this research we also explore
13 whether landscape ecological indices can be used to scale-up data to develop better linkages
14 between land cover patterns and fire regimes. The methods developed for this research enable
15 statistical comparison between a variety of landscape ecological indices and spatial data to
16 evaluate fire regimes at three different spatiotemporal scales. We address four key research
17 questions:
18
19 1. How does landscape pattern affect the fire regime in a savanna?
20 2. Do spatial patterns of particular types of vegetation have unique effects on specific aspects
21 of the fire regime?
22 3. Which landscape ecological indices at which scale have the most potential for linking
23 vegetation cover and fire regime?
24 4. Can landscape indices be used to resolve the issue of low-resolution bias when using coarse-
25 resolution fire and land-cover data?
26
27 2.0 MATERIALS AND METHODS
28
29 2.1 Study Area
30
31 The study area is located in southern Mali and northern Guinea, entirely within the southern
32 edge of the Sudanian savanna belt (Figure 1). The spatial extent corresponds to the Landsat
33 scene WRS-199/52 (Lon: -7° 17'- -9° 2', Lat: 10° 41'-12° 28'). The study area (32,927 km2) is
34 contained within this scene boundary and is smaller than the boundary because of scene
35 shifting over the 11-year study period.
36 There are three distinct seasons: 1) a hot and dry season from February to June; 2) a
37 warm and rainy season from June to October; and 3) a relatively cooler dry season from October
38 to February. Although the rainy season extends over a five-month period, precipitation events
39 are sporadic, often intense and concentrated in a period of just 70–80 days. Annual precipitation

3
 1 varies from 900-mm in Bamako in the north to 1050-mm in Bougouni in the south (Henry, 2011).
2 The fire season typically begins in October just after the rainy season and ends after the first
3 rains fall in early June.
4 Vegetation is classified as southern Sudanian savanna (Nasi and Sabatier, 1988);
5 however, there is a great variety of vegetation types within the savanna formation. Landscape
6 heterogeneity is largely a function of underlying soil and hydrology as well as agricultural use,
7 the combination of which produces patterns of vegetation cover (Duvall, 2011). Ferricrete
8 outcrops on hard pan cover considerable areas creating xeric conditions. Unproductive soils
9 such as these support sparse vegetation referred to here as short grass savanna which is
10 dominated by short, annual grasses (principally Loudetia togoensis but also Andropogon
11 pseudapricus). Short grass savannas are not cultivated and usually have few widely scattered
12 trees. They comprise up to 25% of the study area, but their distribution is highly uneven (Nasi
13 and Sabatier, 1988). Except for the intensively cultivated areas, a near-continuous layer of tall
14 (over one meter in height) perennial grasses—principally Andropogon gayanus, Hyparrhenia
15 dissolute, Cymbopogon giganteus and Schizachyrium pulchellum—covers the more fertile soils,
16 although there are pockets of dense woodlands where the tree canopy is closed and there is
17 little grass cover. Riparian woodlands line most streams.
18 The natural heterogeneity forms a base that is modified through human use. The pattern
19 of natural vegetation is critical because farmers select only the more fertile lands for agriculture
20 and leave the less productive areas (such as short grass savanna) unfarmed. Vegetation cover in
21 settled areas has been significantly modified. Perennial grasses are less common (except on
22 long-fallow plots of over 15 years), and large portions of the landscape are covered by annual
23 grasses, particularly Andropogon pseudapricus and Pennisetum pedicellatum with scattered trees.
24 The farming system—whether long-term rotational agriculture or a form of more intensive
25 cropping—has important landscape effects. Long fallow cycles can increase landscape
26 heterogeneity by creating a patchwork of fallow plots in different phases of succession that are
27 dominated by different types of grasses as well as tree and shrub cover. A shorter fallow cycle
28 (less than five years for example) tends to produce a more homogeneous pattern composed of
29 predominantly annual grasses mixed with frequently farmed fields.
30
31 [Figure 1 here]
32
33 Figure 1. The savanna mosaic in southern Mali and study area. The study area is defined by the
34 spatial extent of a Landsat scene (199/52) that borders Guinea and Mali.
35
36
37 2.2 Data
38
39 To address the limitations of coarse-resolution land cover data for heterogeneous savannas we
40 use a vegetation cover map generated from 30-m Landsat data as our base. We use the term
41 vegetation cover as opposed to land cover because the former implies subtle differences in species

4
 1 while the latter is concerned with vegetation forms. (For example, the type of grass—perennial
2 or annual—is a critical factor in savanna fire ecology, but both cover types of are often classified
3 as “grassland” in land cover typology). Vegetation classes for this study include short grass
4 savanna, agriculture/short fallow (hereafter short fallow), savanna/long fallow (hereafter
5 savanna), and woodlands (Laris, 2002; 2011).
6 To examine the effects of spatial and temporal scale, we compare vegetation patterns
7 with three fire datasets, each with unique spatial and temporal advantages and disadvantages
8 (Table 1). The MODIS 500-m burned area dataset is generated from an algorithm that
9 determines the change in reflectance between unburned and burned savanna (Giglio et al.,
10 2006). Advantages of the MODIS burned area product are its high temporal frequency and
11 accessibility (daily datasets available online for an 11-year period). Disadvantages are that the
12 data and fire detection algorithm suffer from low spatial resolution and therefore miss many
13 small and often early-dry season fires.
14 The second fire dataset is the 1-km MODIS active fire which uses a thermal channel
15 algorithm to detect “hot spots” or actively burning fires. Advantages are its high temporal
16 frequency (thrice daily), ability to detect and map even small fires, wide availability and
17 relatively long-term (over 10 years). Disadvantages are the scale of the data—even very small
18 fires are mapped due to the thermal energy they release and thus it is not possible to use these
19 data for accurate estimates of area effected by fire. These data do provide an excellent source of
20 the overall spatiotemporal pattern of fires.
21 Finally, we use a 30-m database of burned area generated from Landsat. Advantages of
22 the Landsat database are the high spatial resolution, which effectively maps all surface fires.
23 Disadvantages are the temporal availability of data. Burning regimes have been shown to be
24 consistent between years for the study area (Laris, 2011; Laris et al., 2016). Therefore we use a
25 sample of three years of burned area data (2002-2003, 2006-2007 and 2013-2014) containing high
26 temporal frequency to represent the longer (10 year) study period. We justify this based on
27 previous research that finds
28
29
30 Table 1. Summary of data types and derived variables
31
Variables Data Spatial Description Derived Variables
Resolution
Seasonality of Landsat 30-m Date of average Continuous variable:
Fire pixel burned Julian days modified
so that October 1 = 1
Frequency of MODIS 500-m Number of Count Variable:
Fire Burned times a pixel frequency of fires in
Area burned a pixel as a count

5
 Variance of MODIS 1-km Variance of the Continuous Variable:
average burn Active Fire average burn variance of the
date date average burn date
Spatial Landsat 30-m landscape Continuous variable:
fragmentation indices landscape indices for
calculated for each land cover type
each land cover
type in a
MODIS pixel
1
2
3 2.3 Data and Transformation
4
5 Fire seasonality is a critical determinant of the impact of a fire on young trees. Fire intensity and
6 severity generally increase as the dry season progresses and vegetation cures (Laris et al., 2016).
7 As such, accurate data on the seasonality of fires is critically important. Fire data were obtained
8 to cover the West African fire season, which occurs between October and May (Laris, 2005). We
9 used the MODIS burned area and MODIS active fire point datasets for the years 2003-2014. We
10 generated burned area maps for three fires seasons using Landsat imagery (2002-2003, 2006-
11 2007, and 2013-2014) and used these maps to estimate average burn dates and fire frequency for
12 the study area.
13 We used these three fire data sets to generate three characteristic variables representing
14 the fire regime: (i) fire seasonality (average burn date); (ii) fire frequency; and (iii) variance of
15 average burn date. To link vegetation data generated from Landsat with MODIS fire products,
16 we first rescale the Landsat vegetation data from 30-m to 50-m pixel resolution. We then convert
17 the 50-m data to one based on landscape ecological indices using 500-m pixels to allow
18 comparison with the 500 and 1-km fire datasets. Landscape indices derived from higher
19 resolution data therefore retain the information related to spatial pattern for different vegetation
20 types, while “degrading” the pixel resolution allowing us to scale up the data (Figure 2).
21
22 [Figure 2 here]
23
24
25 Figure 2. Data processing and analysis methods
26
27
28 2.4 Landscape Scale and Indices
29
30 We use three landscape scales - landscape, class, and patch (McGarigal, 2014; McGarigal, et al.,
31 2012). For the purposes of this study, the term “landscape” refers to the area contained in a 500-
32 m MODIS pixel. The term “class” refers to the vegetation type contained in the Landsat

6
 1 vegetation-cover dataset. Patch refers to a group of connected pixels that belong to a single class
2 (or vegetation type).
3
4 We tested two different levels of spatial pattern metrics: Class-level indices (CLI) and
5 Landscape-level indices (LLI) to assess which indices were most useful for linking vegetation
6 pattern with fire regime. The CLI represents the spatial distribution and pattern of a single
7 vegetation type contained in a landscape (MODIS pixel area). The LLI calculates a numeric
8 value for the spatial pattern created by all the vegetation types present within a MODIS pixel
9 (Table 2). We used the uniform tiles method in FRAGSTATS (McGarigal et al., 2012) to calculate
10 all landscape indices. We selected four indices that we expect to influence the spread of a fire
11 across a landscape (Turner and Romme, 1994). In the CLI these included (1) proportion of area,
12 (2) shape index, (3) largest patch index, and (4) patch density. In the LLI these included: (1)
13 largest patch index, (2) patch richness, (3) splitting index, and (4) patch density (McGarigal,
14 2014).
15
16 Table 2. Descriptions of Class-Level Indices and Landscape-Level Indices
17
Indices Explanation
CLI (short grass savanna, short fallow, savanna, woodlands)
Proportion of area Proportion of landscape area for each vegetation type, a simple
measure of amount of area covered by each vegetation type
Shape index Shape index measures shape complexity of a patch compared to a
square shape of the same size
Largest patch index Largest patch index is the percentage of the total landscape area
occupied by the largest patch of a class
Patch density Patch density is the number of patches in a land-cover type per
100-hectares.
LLI (all vegetation types)
Largest patch index Largest patch index is the size of largest patch in the landscape
regardless of vegetation type
Patch richness Patch richness is the number of land-cover types present in the
landscape.
Splitting index Splitting index is equal to 1 when the landscape consists of a single
patch. Splitting index increases as the landscape is subdivided
into smaller patches and achieves its maximum value when the
landscape is maximally subdivided; that is, when every cell is a
separate patch.
Patch density Patch density is the number of patches in a landscape.

7
1

2 2.5 Fire regime – seasonality and frequency of fires

3
4 Landsat data were used to create the vegetation and fine resolution burned area maps. Selection
5 of appropriate vegetation types that determine the fire regime is based on a conceptualization of
6 the drivers of fire in southern Mali. Drivers include agricultural fields and vegetation type
7 which affect spatial extent and seasonality of fires (Laris, 2011; Laris, 2005).
8
9 2.6 Vegetation Map
10
11 To characterize and classify the vegetation types in the study area we created a vegetation cover
12 map using three images from the middle of the study period (2000, 2002, and 2006). We selected
13 four vegetation-cover classes for the analysis: short fallow, savanna, short-grass savanna and
14 woodlands. Two classes, settlements and water bodies were not included as these are non-
15 flammable.
16
17 2.7 Fine temporal resolution burned area maps
18
19 Landsat images covering three fire seasons – 2002-2003, 2006-2007, and 2013-2014 – were
20 chosen because sufficient imagery were available to cover each of three fire seasons. Data for the
21 first two fire seasons were obtained from Landsat 4 Thematic Mapper (TM) and Landsat 7
22 Enhanced Thematic Mapper Plus (ETM+) imagery. For the latter fire season, we used images
23 from TM and Landsat 8 Operational Land Imager and Thermal Infrared Sensor (OLI and TIRS).
24 Burned area maps from Landsat data were generated using standard methods (Laris
25 2011). First, we used ENVI (Exelis Visual Information Solutions) version 5.0 to perform an
26 unsupervised clustering algorithm (ISODATA) using six bands for the Landsat images (bands
27 2-7 for OLI; 1-5 and 7 for TM and ETM+). The clustering algorithm grouped pixels into 25
28 spectral classes. Second, we assessed each class visually as either burned or unburned by
29 comparing to a false-color image (5-4-3 band combination). We then removed pixels burned in
30 prior months (areas burn once per fire season, although burn scars may remain visible and
31 detectable for several months). All water bodies were masked as these can be confused with
32 burned areas. Finally, we combined individual burned area maps to produce seasonal fire
33 maps.
34 The Landsat burned area maps were aggregated to the 500-m scale while retaining
35 critical information on the fire regime (the average burn date, frequency, and standard
36 deviation). Burn frequency maps were generated by resampling the burned area maps to 50-m
37 pixels. The 50-m resolution images were then aggregated to 500-m using block statistics
38 whereby each 500-m pixel was attributed with the sum of the burned areas in the corresponding
39 100 pixels from the 50-m input image. Finally, we combined the 500-m maps for all fire seasons
40 to determine frequency for burning for each pixel to yield a frequency image at 500-m
41 resolution.
8
1 The 50-m burned area data were similarly aggregated to derive a 500-m burn date
2 image. Each 500-m pixel was given a value for the sum of the burned dates in the corresponding
3 100 pixels from the 50-m input image. We summed the burn date images and divided by the
4 burn frequency image to derive an average burn-date image (each pixel value represented an
5 average burn date at 500-m scale).
6 Finally, the standard deviation of burned dates for each 500-m aggregate pixel was
7 calculated using equation (1):
8
( )
9 ∑ equation (1)
10 where:
11 Xburned = per pixel burned date preprocessed to convert “0” values to “no data”
12 Xavg = average burned date for 500-m pixel resampled to 50-m for calculation purposes
13 N = frequency for 500-m pixel resampled to 50-m for calculation purposes (maximum
14 corresponding to the 100 50-m cells for three seasons)
15
16 Monthly burned area maps are available at 500-m resolution for the African continent
17 (Giglio et al., 2006). We calculated average burn date, burn frequency, and standard deviation of
18 the burn date for each pixel for the 11-year study period (2003-2014). MODIS active fire data are
19 also provided as point shapefiles where each point represents the date and location of the fire at
20 1 km resolution. We converted point files to 1-km raster files and then resampled to 500-m. We
21 then calculated the average burn date, burn frequency, and standard deviation of the burn date
22 for each 500-m pixel for the 11-year study period (2003-2014).
23
24 2.8 Statistical Analysis
25
26 Median values of fire regime measures (average frequency, variability of fires) were
27 compared between vegetation types using the Mann-Whitney test (Watt, 1947; Wilcoxon, 1945)
28 due to non-normal (skewed) distribution as determined by the Skewness-Kurtosis test
29 (D'agostino et al., 1990; Royston, 1992).
30
31 3.0 RESULTS
32
33 Research question 1: How does landscape pattern affect the fire regime in a savanna?
34 To explore the impact of landscape pattern on fire regime we developed class level indices for
35 the various vegetation types. Average burned date and fire frequency were derived from each
36 of the data sources and evaluated in the context of the landscape patterns.
37 3.1 Class-Level Indices by Vegetation Type
38 Savanna is the dominant vegetation type with a percent cover of 42%, followed by short grass
39 (25%), short fallow (24%), and woodlands (10%) (Table 3). Savanna patches were larger, on
9
1 average, with a median patch index of 34%, than patches of other vegetation types, while
2 woodlands patches were smaller (median patch index of 3%). In terms of patch density, there
3 were more short fallow patches than other vegetation types per landscape unit with woodlands
4 being less common. Woodlands had the smallest shape index (more square-shaped patches),
5 while savanna/long fallow had the largest shape index (more non-square-shapes).
6

7 Table 3: Median values for vegetation Class-Level Indices

Short grass Short
Savanna Woodlands
savanna fallow
Class-level indices N=100,970 N=98,052 N=108,469 N=69,293
Proportion of area (%) 18 21 42 5
Largest patch index (%) 10 11 34 3
Patch density 12.2 16.3 12.2 8.2
Shape index 1.4 1.5 1.9 1.1
8
9
10 3.2 Fire Regime by Data Source and Vegetation Type
11
12 Average burned date remained relatively consistent among products (Figure 1). The earliest
13 average burned dates occurred in the Landsat data (median Jan 1), followed by MODIS active
14 fire (Jan 8) and MODIS burned area (Jan 18).
15
16 [Figure 3 here]
17
18
19 Figure 3. Variation in burn dates by dataset.
20
21 Landsat data had the earliest average burned dates across all vegetation types (median
22 Jan 1) followed by MODIS active fire (median Jan 8) and MODIS burned area (median Jan 18).
23 As expected, the seasonality of fire varies by vegetation cover. In each of these datasets short
24 grass savanna burned first, with woodland burning last (with the exception of the MODIS
25 burned area data, in which case short fallow burned, on average, one day later) (Table 4).
26
27
28
29 Table 4. Comparison of average* burn date by vegetation type and fire dataset
30
Vegetation Landsat burned MODIS burned area MODIS active fire
area average** average average

10
 Short grass savanna Dec 30 Jan 18 Jan 7
(N=100,966) (N=555,13) (N=13,470)
Short fallow Dec 31 Jan 20 Jan 8
(N=98,043) (N=53,030) (N=12,593)
Savanna Dec 31 Jan 19 Jan 8
(N=108,463) (N=61,274) (N=14,835)
Woodlands Jan 3 Jan 19 Jan 9
(N=69,285) (N=39,649) (N=9,868)

*“Average” refers to average burned dates across pixels with the corresponding vegetation
type. N is the number of pixels used to calculate the average value.
**Dates for the Landsat data correspond to the date of the image for a burn scar. The actual
burn date is 0-14 days earlier because Landsat has a frequency of 15 days. Using a “mean burn
date” would result in a shift in Landsat values to 7 days earlier (Laris, 2011) .

1 Research question 2: Do spatial patterns of particular types of vegetation have unique effects on specific
2 aspects of the fire regime?
3
4 We further this analysis to evaluate whether spatial patterns of particular types of
5 vegetation have unique effects on specific aspects of the fire regime. The effect of fire
6 seasonality on vegetation types was evaluated by comparing fire seasonality (burned date and
7 frequency) within the lower quartile (less percent cover) of a particular vegetation type and
8 upper quartile (higher percent cover) of a particular vegetation type (Figures 4 and 5) (maps
9 illustrating the spatial patterns for each cover type for early are found in the supplementary
10 data).
11 Fires tended to occur earlier in areas with more short grass savannah (Figure 4A). Fires
12 occurred later in areas with higher percentage of savanna and woodlands. Only in short fallow
13 was there no significant difference in average burned dates between areas with lower versus
14 higher percent cover (p=0.11, Figure 4B).
15
16
17 [Figure 4 here]
18
19 Figure 4. Average burned date by vegetation cover type. 1st quartile indicates lower percent
20 cover; 4th quartile indicates higher percent cover. * Indicates statistically significant differences;
21 Mann-Whitney test (α=0.5).
22
23 Burning was more frequent (more fire events) in areas with higher percentage of short
24 grass savanna (p<0.001; Figure 5A). For example, burning occurred 234 times in areas with
25 higher percentage of short grass savanna, compared to 157 times in areas with lower
11
 1 percentage. Areas with high percentage of savanna burned more frequently (p<0.001). In
2 woodlands, there was little observed difference in frequency by percent cover, although
3 statistically significantly different (Figure 5D).
4
5 [Figure 5 here]
6
7
8 FIGURE 5. Fire frequency (events) by vegetation cover type. 1st quartile indicates lower percent
9 cover; 4th quartile indicates higher percent cover. * Indicates statistically significant differences;
10 Mann-Whitney test (α=0.5).
11
12 These results demonstrate that the spatial patterns of vegetation influence both the
13 seasonality and frequency of the fire regime. Short grass savanna burned more consistently and
14 frequently earlier in the season, short fallow burned less consistently and frequently, with little
15 difference in average burn date, and savanna and woodlands more often burned later.
16
17 3.3 Landscape Ecological Indices
18
19 Research question 3: Which landscape ecological indices at which scale have the most potential for linking
20 vegetation cover and fire regime?
21 Patch scale indices were evaluated by vegetation type using the upper and lower
22 quartiles of percent cover within each vegetation type. The overall trend in fire seasonality was
23 calculated using the landscape ecological indices to quantify trends in average burned dates
24 derived from the three data sources (Table 5).
25
26 Table 5. Average burn date based on Landsat, MODIS active fire and MODIS burned area by vegetation
27 cover type and patch-scale index.
Landscape Landsat Burned
MODIS burned area MODIS active fire
Indices Area
By Vegetation
Lower Upper Trend Lower Upper Trend Lower Upper Trend
Type
Percent Cover
Short Grass
102 79 E 114 104 E 104 93 E
Savanah
Short Fallow 92 92 -- 106 116 L 97 101 L
Savannah 83 101 L 112 106 E 98 99 --
Woodlands 91 102 L 110 114 L 99 106 L
Shape Index
Short Grass
101 83 E 114 106 E 103 96 E
Savanah
12
 Short Fallow 92 92 -- 107 115 L 98 101 L
Savannah 84 97 L 112 109 E 99 99 --
Woodlands 92 101 L 110 113 L 99 104 L
Largest Patch
Short Grass
101 79 E 114 105 E 103 93 E
Savanah
Short Fallow 92 92 -- 107 115 L 97 101 L
Savannah 83 101 L 112 106 E 98 99 --
Woodlands 91 102 L 109 114 L 100 106 L
Patch Density
Short Grass
90 93 L 111 107 E 99 98 --
Savanah
Short Fallow 94 89 E 109 117 L 99 101 L
Savannah 96 88 E 108 113 L 99 100 --
Woodlands 92 99 L 111 110 -- 101 101 --
E (trending earlier), L (trending later), -- (no discernable trend). Oct 1=1.

1 The trend in fire seasonality varies by vegetation type. Using the percent cover index, for
2 example, as the percent cover of short grass savanna increases, fires occur earlier, while as the
3 percent cover of woodlands increase, fires tend to occur later in the season. The shape index and
4 largest patch indices were relatively consistent among cover types. Short grass savanna tended
5 to burn earlier across scales of data, while short fallow and woodlands tended to burn later. The
6 seasonality of savanna burning was inconsistent across scales of data for both of these indices.
7 The patch density index was inconsistent in determining trends of fire seasonality across scales.
8 Short grass savanna showed the most consistency across scales and landscape indices
9 (specifically percent cover, shape index and size of largest patch), with an earlier burn date.
10 Patch density (the increasing number of total patches) trends in the opposite direction. That is,
11 as the number of patches increase, the burn date shifts later.
12 The percent cover, shape index and largest patch landscape ecological indices
13 demonstrated the most consistency in burn date trends across scales. These indices therefore
14 demonstrate the greatest potential for linking vegetation cover and fire regime.
15 To further assess the relationship between landscape indices and fire regime variables,
16 we used the Landsat fire dataset to compare pixel values in the upper quartile range of each
17 landscape index with those in the lower quartile. We reclassified burn dates as either “early”
18 (October to December) or “late” (January to May) based on the commonly used convention for
19 the December 31st cut-off date for early fires (Laris et al., 2016) for comparison purposes, and to
20 illustrate patterns. Table 6 (Figures 1-4 in the supplement) depicts the results for the four
21 vegetation types in terms of percent cover, shape index, largest patch index and patch density
22 based on lower and upper quartiles for timing and frequency. Fire frequency was high for all

13
 1 cover types and vegetation patterns, generally greater than once every other year over (>1.5/3
2 years).
3 Short grass savannas tended to burn earlier when represented using percent cover,
4 shape index, largest patch index, and decreased patch density. Areas with a high proportion of
5 short grass savanna (such as the western side of the imagery) were characterized by fewer
6 (lower patch density), larger (largest patch index) and more irregularly shaped (shape index)
7 patches, with fires occurring earlier in the fire season and more frequently. Areas with more
8 irregularly shaped patches had an average burned date that was 18-days earlier than more
9 square-shaped patches. Larger patches tended to burn 24 days earlier than smaller patches.
10 Short grass savanna areas with smaller patch density or fewer patches burned 3 days earlier
11 than areas with many patches.
12
13 Table 6: Comparison of average burn dates by lower and upper quartile for vegetation classes by patch
14 index

Percent Cover Shape Index Largest Patch Patch Density

Lower Upper Trend Lower Upper Trend Lower Upper Trend Lower Upper Trend

Short grass 101 77 E 100 82 E 101 77 E 89 92 L

Short fallow 91 91 -- 91 91 -- 91 91 -- 93 87 E

Savanna 79 100 L 81 96 L 80 100 L 96 85 E

Woodlands 89 101 L 90 100 L 90 101 L 91 98 L

15 E (trending earlier), L (trending later), -- (no discernable trend). Oct 1=1.

16
17 Areas with a high proportion of short fallow (Figure 2 in Supplement) were likely to
18 have many patches (higher patch density), with fires occurring earlier in the season. Areas
19 burning earlier also burned more frequently. There is little variation in terms of average burn
20 dates for the short fallow class. The exception are areas with many patches (high patch density
21 value) which burned 6 days earlier than areas with few patches.
22 Areas with a high proportion of savanna vegetation were likely to have fewer, larger,
23 and more irregularly shaped patches, with fires occurring more often later in the season (Figure
24 3 in Supplement). The southeast portion of the imagery, which is dominated by areas of
25 contiguous savanna cover, has a high concentration of later fires. Areas with more irregular,
26 non-square shaped patches had a 15-day later average burned date than more regular-shaped
27 patches. Larger patches tended to burn 20 days later than smaller patches. Savanna areas with
28 smaller patch density burned 15 days later than areas with many patches.
29 Areas with a high proportion of woodlands were likely to have many, larger and
30 irregularly shaped patches with fires occurring later in the season and less frequently than in

14
 1 areas with a lower proportion of woodlands (Figure 4 in Supplement). Areas with a larger
2 percentage cover of woodlands such as the far southeast had a 12-day later average burned date
3 than areas with a lower proportion of woodlands. Areas with more irregular, non-square-
4 shaped woodlands patches had a 10-day later average burned date. Larger patches tended to
5 burn 11 days later than smaller woodland patches. Woodland areas with higher patch density
6 burned 7 days later compared to areas with few woodland patches.
7
8 3.4 Landscape Scale Indices
9
10 Table 7 shows the results for the analysis of the LCI by mean burn dates. The results of the
11 Landsat scale analysis demonstrate that larger patch size results in later burn dates, while
12 greater patch density (more patchy or fragmented landscape) results in earlier burn dates
13 regardless of vegetation type. As the data indicate, the Landsat data have the earliest average
14 burn dates and MODIS burned area, the latest. Interestingly, the trends for different indices
15 differ by data set in several cases.
16
17
18

15
1 Table 7: Mean values of burn dates for the Landscape Scale Indices by data set

Landscape Landsat MODIS MODIS active

Index burned area fire
Lower Upper Trend Lower Upper Trend Lower Upper Trend
Largest patch 89 95 L 117 104 E 102 98 E

Patch richness 91 94 L 109 113 L 99 101 L

Split index 95 89 E 104 118 L 100 98 E

Patch density 94 90 E 107 115 E 100 101 --

2 E (trending earlier), L (trending later), -- (no discernable trend). Oct 1=1.

3
4
5 Landscape level indices using Landsat data (Figure 6) show burn dates for four indices
6 based on the landscape scale and thus not distinguished by vegetation type. Areas with smaller
7 patches (6A) and more fragmentation (higher split index) (6C) had an 8-day earlier average
8 burned date and burned less frequently. Areas with fewer vegetation types (low richness) (6B)
9 had a slightly earlier (2 day) average burned date and nearly identical frequency. Areas with
10 many patches (high patch density) (6D) had a 6-day earlier average burned date and a lower
11 frequency.
12
13
14 [Figure 6 here]
15
16
17 Figure 6. Landscape Level Indices. Median (SD). * indicates significant difference in average
18 burned date and frequency of fires between the upper and lower quartile; Mann-Whitney test
19 (α=0.5).
20
21 In general, areas that burned earlier tended to be more fragmented and composed of
22 smaller patches. Early fires are associated with smaller largest patch index (small patches),
23 smaller patch richness (less vegetation type diversity), larger splitting index (landscape
24 subdivided into smaller, irregularly-shaped patches), and larger patch density (many patches).
25 Thus, early fires tend to be fragmented, resulting in many, small fires burning in areas with
26 fewer vegetation types than late fires. These early, small fires burned more homogeneous areas
27 (fewer land cover classes) than later fires. At the landscape scale, areas affected by early fires
28 also tended to burn less frequently. Late fires were associated with larger more contiguous
29 patches.

16
1
2 3.5 Low-resolution bias and its effects
3
4 Research question 4: Can landscape indices be used to resolve the issue of low-resolution bias when using
5 coarse-resolution fire and land-cover data?
6 Comparison of the results by data set demonstrates the documented tendency for
7 MODIS burned area data to represent burning occurrences later in the season. Not only is low-
8 resolution bias apparent in its expression of later burned dates (see Figure 3), it is evident for all
9 vegetation types and indices (Table 4). Average dates for MODIS burned area are significantly
10 later (2-3 weeks, Table 4) than those for the Landsat dates and there is greater variance. The
11 MODIS active fire dates fall in between. The burn dates for short fallow areas show no trend for
12 percent cover, shape index or largest patch per the Landsat data burned area data, yet these
13 same indices trend later (4-10 days) when MODIS active or burned area data are used. Savanna
14 burn dates trend later for the first three indices when Landsat data is used and earlier when
15 MODIS burned area data are used. There is no trend with the MODIS active fire data. Finally, in
16 terms of woodlands, the trend is later for all data sets and all indices (except patch density,
17 which is neutral for the MODIS data).
18
19 4.0 DISCUSSION
20
21 This research demonstrates the influence of both landscape pattern and vegetation type
22 on a fire regime in the savanna. At the landscape level, results from the LLI show that smaller
23 and more irregularly shaped patches with less diversity of vegetation types were associated
24 with fires occurring earlier in the season. A landscape composed of a few vegetation types,
25 highly fragmented with small patches of vegetation, burned the earliest. Conversely, areas with
26 little fragmentation and large, homogenous patches of vegetation burned later. We conclude
27 that earlier fires are associated with landscape heterogeneity and fragmentation, but not
28 necessarily with a higher diversity of vegetation types.
29 The study also finds that fire seasonality and frequency varied according to vegetation
30 type. Results from the CLI indicate that fires occurred earliest in short grass savanna, followed
31 by short fallow, savanna, and then woodlands. These findings are consistent with finding from
32 other studies of burning practices in West Africa (Caillault et al., 2015; Devineau et al., 2010;
33 Laris, 2011; Laris et al., 2016). The seasonality of fire generally follows the seasonal changes in
34 fuel moisture levels for these vegetation types. As Laris (2002; 2011) has argued, short, annual
35 grasses have faster drying rates than other vegetation types. People have long set fires first in
36 areas with short annual grasses with the goal of fragmenting the landscape to prevent large and
37 damaging fires. Areas under long fallow cycles, characterized by more perennial grasses and
38 denser woody vegetation, hold moisture longer. This study finds that these areas tend to have
39 non-square shapes and larger-sized patches. Fires in these savanna areas with larger more
40 contiguous patterns occur later in the season.

17
 1 Results indicate that landscape heterogeneity by vegetation type plays a critical role in
2 determining the seasonality of fire. Few, large, and non-square-shaped patches were associated
3 with earlier burned dates in short grass savanna areas; while in savanna/long fallow areas, these
4 same patterns were associated with later burned dates. In woodlands, which are generally
5 composed of many, smaller irregularly shaped patches, fires occurred later. In short fallow
6 areas, an increase in patchiness was associated with earlier burn dates.
7 The latter finding is important for long-term trends in savannas where agriculture is
8 expanding. Our results indicate that as landscape fragmentation increases (more and smaller
9 patches) fires occur earlier regardless of vegetation type. Importantly, the class level results
10 show that fires occur earlier in areas with an increase in patchiness of short fallow agriculture.
11 This result suggests that the gradual fragmentation of the landscape with short fallow farming
12 systems produces an earlier fire regime with somewhat less frequent fire. As others have
13 shown, increasing agriculture area tends to reduce fire once a particular spatial threshold is
14 reached, because fires can no longer propagate across a landscape (McKenzie et al., 2011). Our
15 findings suggest that, prior to the attainment of the landscape threshold preventing fire
16 propagation, an increase in area under rotational agriculture results in a regime of earlier, less
17 frequent fires and smaller fires. This finding is not surprising as previous research has found
18 that farmers tend to burn fallow patches adjacent to their fields as early as possible as a
19 defensive measure to protect crops from fire. In addition, interviews with farmers suggest that
20 people often attempt to prevent fire from burning areas in close proximity to villages, as is the
21 case with much frequently farmed land, which would explain the lower fire frequency in these
22 areas (Laris, 2002) . This process has expanded in area as the number of livestock has increased
23 in recent years and grasses have become increasingly valued as fodder. Finally, as several
24 studies have shown, a regime of patch- or seasonal-mosaic results in a decline in area burned
25 overall and critical to establishing this regime is early and patchy burning (Hudak and Brockett,
26 2004; Laris, 2011; Russell-Smith et al., 1997). The work of Andela et al (2017) corroborates our
27 finding as well. Although their research was conducted at a very different scale, they found that
28 agricultural expansion and intensification were primary drivers of declining fire activity in
29 savannas globally.
30 The above findings have implications for land management agencies seeking to work
31 with local villages and land owners to improve fire management. We find that un-developed
32 savanna landscapes have distinctly different fire regimes than those with significant levels of
33 agriculture. This suggests that if environmental agencies seek to reduce the intensity, frequency
34 or extent of the fires in African savannas, they should develop distinct management plans for
35 different types of landscapes. In undeveloped savannas, the traditional practice of setting fires
36 early according to grass type serves to fragment the landscape and render it more useful and
37 manageable for local inhabitants (Laris 2011). As such, management agencies should work with
38 local people to maintain this burning regime. Contrasting, in areas where agricultural land and
39 short fallows are increasing as a fraction of the total landscape, management agencies should
40 seek to understand the struggles local inhabitants face when attempting to manage the negative
41 effects of fire. For example, the increase in short fallow results in a shift from perennial to

18
 1 annual grasses on the landscape. While perennial grasses will resprout following a well-timed
2 fire, annuals will not. As such, as pasture lands become increasingly rare, due to a reduction in
3 undeveloped savanna coupled with an increase in herd numbers, local inhabitants seek to
4 prevent burning of annual grasses. This poses a potential problem, however, as traditionally
5 annual grasses are burned early to create a mosaic. Agencies should thus work with clusters of
6 villages to develop new management strategies whereby some strategic burning of annuals is
7 conducted early to fragment the landscape and prevent fires ignited later in the dry season
8 (accidentally or purposefully) from destroying valued pasture. Villages need help rethinking
9 the fire “mosaic” and management agencies can help especially by providing high resolution
10 imagery of burn patterns, grass cover, and other important landscape features. Villagers and
11 management agencies could then work jointly to create new burning strategies aimed at
12 preserving values herbaceous resources, while also continuing a form of the age-old strategic
13 practice of burning a mosaic to protect some areas and render the landscape more useful for
14 inhabitants.
15 Through this study we also developed a method to retain information on vegetation
16 patterns derived from finer resolution Landsat imagery, which enabled analyses of burn
17 seasonality and frequency by vegetation type. While the coarse resolution bias imposed by
18 MODIS data was not overcome, our unique approach contributed to a greater depth of
19 understanding related to the spatial patterns of the fire regime by vegetation type. Specifically
20 we hypothesized that retaining information on vegetation patterns derived from the scale of
21 Landsat imagery in the form of landscape ecological indices at the 500-m scale would improve
22 results. We found no evidence that this was the case. MODIS burned area product had the
23 greatest degree of low-resolution bias when compared with the other two fire datasets; average
24 burn dates associated with the MODIS burned area were significantly later than those for
25 Landsat and MODIS active fire data regardless of vegetation type confirming the findings of
26 others (e.g., Caillault et al., 2015). Laris (2005) showed that mosaic fire regimes in Mali have a
27 distinct temporal component: early fires tended to be smaller and more fragmented than later
28 fires. As such, it is no surprise that the MODIS burned area product (which tends to miss
29 smaller fires) had the latest overall burn date average. Moreover, the trend analysis found
30 disagreement for the shift in the seasonality of fires for areas with more short fallow lands (both
31 MODIS products trended later for increases in short fallow). Short fallow patches are likely the
32 smallest of all areas as agricultural fields are often no greater than a few hectares. Here again, it
33 is no surprise that the MODIS products could not capture this fine-scale landscape pattern and
34 its influence on fire. The mixed pixel (low-resolution) bias issue is depicted in figure 7 where the
35 500-m pixels contain a variety of vegetation types with different landscape patterns.
36
37 [Figure 7 here]
38
39

19
 1 Figure 7. The fragmented savanna landscape. Grids 1-4 represent four MODIS 500-m pixels. The
2 finer grid cells represent the four vegetation classes contained in the classified and resampled
3 50-m Landsat pixels.
4
5
6 5.0 CONCLUSIONS
7
8 This study finds that landscape pattern influences the fire regime in a mesic savanna of
9 West Africa. Specifically, we find that fire seasonality and frequency varied by landscape index
10 at both the vegetation class and landscape scales. Results find that fires occurred earliest in
11 vegetation types that tend to dry the quickest following the end of the rains including short
12 grass savanna and annual grasses on short fallow lands. Savanna and woodland patches tended
13 to burn later. Importantly, this study finds that landscape fragmentation—specifically a
14 reduction in the size of patches and an increase in their number–results in an earlier fire regime.
15 This holds regardless of vegetation type; the landscape-scale analysis finds that early fires are
16 associated with many and smaller patches and a less diverse mixture of vegetation types. The
17 study also finds that as the landscape becomes increasingly fragmented, due to agriculture and
18 accumulation of annual grasses, the fire regime shifts earlier and fires are slightly less frequent.
19 We therefore conclude that fire intensity and severity will decline as agriculture expands in
20 savannas.
21 The methods developed for this research present a viable approach to bridging
22 inconsistencies between the temporal and spatial scales of available data. We find that our use
23 of landscape ecological indices to retain information on landscape pattern while scaling up the
24 data identified certain land cover types where fire frequency trends were consistent. However,
25 these were not consistent among all cover types for each index across scales. We conclude that
26 current coarse resolution fire products are at a scale that is inappropriate for developing models
27 linking fire regimes to changes in land-use and cover analysis in the African savannas; the
28 fragmented landscape patterns—and the fire patterns they produce—are too fine to be captured
29 using 500-m data.
30 In savanna landscapes, where the fire regime is a key determinant of tree cover,
31 understanding the relationships between the fire regime and the natural and anthropogenic
32 factors that create landscape pattern is critical for predicting future changes in vegetation cover
33 as well as emissions from these fires. This study finds that both natural and anthropogenic
34 landscape features influence the fire regime and that analysis of these factors requires data at a
35 scale representative of the landscape pattern. Land and fire managers working in savanna
36 environments should seek to work with local inhabitants to integrate existing burning practices
37 with the shifting goals of land managers in the context of an increasingly fragmented landscape.
38

20
 1 ACKNOWLEDGEMENTS
2 The authors wish to thank Fakuru Camara for his assistance in the field.
3
4 Funding:
5 Funding for this research was provided through National Science Foundation (Grant #1313820).
6
7 Competing Interests Statement
8 The authors declare no conflict of interest. The funding agency had no role in the design of the
9 study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and
10 in the decision to publish the results.
11

21
REFERENCES

Andela, N., Morton, D., Giglio, L., Chen, Y., van der Werf, G., Kasibhatla, P., . . . Kloster, S.
(2017). A human-driven decline in global burned area. Science, 356(6345), 1356-1362.
ArcGIS, E. (2012). 10.1. Redlands, California: ESRI.
Archibald, S. (2016). Managing the human component of fire regimes: lessons from Africa. Phil.
Trans. R. Soc. B, 371(1696), 20150346.
Archibald, S., Lehmann, C. E., Gómez-Dans, J. L., and Bradstock, R. A. (2013). Defining pyromes
and global syndromes of fire regimes. Proceedings of the National Academy of Sciences,
110(16), 6442-6447.
Archibald, S., Roy, D. P., Van Wilgen, B. W., and Scholes, R. J. (2009). What limits fire? An
examination of drivers of burnt area in Southern Africa. Global Change Biology, 15(3), 613-
630. doi:10.1111/j.1365-2486.2008.01754.x
Bird, R. B., Codding, B. F., Kauhanen, P. G., and Bird, D. W. (2012). Aboriginal hunting buffers
climate-driven fire-size variability in Australia’s spinifex grasslands. Proceedings of the
National Academy of Sciences, 109(26), 10287-10292.
Bond, W. J., Woodward, F. I., and Midgley, G. F. (2005). The global distribution of ecosystems in
a world without fire. New phytologist, 165(2), 525-538.
Boschetti, L., Flasse, S. P., and Brivio, P. A. (2004). Analysis of the conflict between omission and
commission in low spatial resolution dichotomic thematic products: The Pareto
Boundary. Remote Sensing of Environment, 91(3), 280-292.
Caillault, S., Ballouche, A., and Delahaye, D. (2015). Where are the ‘bad fires’ in West African
savannas? Rethinking burning management through a space–time analysis in Burkina
Faso. The Geographical Journal, 181(4), 375-387.
D'agostino, R. B., Belanger, A., and D'Agostino Jr, R. B. (1990). A suggestion for using powerful
and informative tests of normality. The American Statistician, 44(4), 316-321.
Devineau, J.-L., Fournier, A., and Nignan, S. (2010). Savanna fire regimes assessment with
MODIS fire data: their relationship to land cover and plant species distribution in
western Burkina Faso (West Africa). Journal of Arid Environments, 74(9), 1092-1101.
Duvall, C. S. (2011). Biocomplexity from the ground up: Vegetation patterns in a West African
savanna landscape. Annals of the Association of American Geographers, 101(3), 497-522.
Ehrlich, D., Lambin, E. F., and Malingreau, J.-P. (1997). Biomass burning and broad-scale land-
cover changes in Western Africa. Remote Sensing of Environment, 61(2), 201-209.
Flannigan, M. D., Krawchuk, M. A., de Groot, W. J., Wotton, B. M., and Gowman, L. M. (2009).
Implications of changing climate for global wildland fire. International Journal of Wildland
Fire, 18(5), 483-507.
Giglio, L., Van der Werf, G., Randerson, J., Collatz, G., and Kasibhatla, P. (2006). Global
estimation of burned area using MODIS active fire observations. Atmospheric Chemistry
and Physics, 6(4), 957-974.

1
Govender, N., Trollope, W. S., and Van Wilgen, B. W. (2006). The effect of fire season, fire
frequency, rainfall and management on fire intensity in savanna vegetation in South
Africa. Journal of Applied Ecology, 43(4), 748-758.
Henry, C. (2011). An integrated approach to estimating groundwater storage variability and
recharge in southern Mali, Africa. Unpublished Masters Thesis, Simon Fraser University,
Burnaby, BC.
Higgins, S. I., Bond, W. J., and Trollope, W. S. (2000). Fire, resprouting and variability: a recipe
for grass–tree coexistence in savanna. Journal of ecology, 88(2), 213-229.
Hudak, A., and Brockett, B. (2004). Mapping fire scars in a southern African savannah using
Landsat imagery. International Journal of Remote Sensing, 25(16), 3231-3243.
Koerner, S. E., and Collins, S. L. (2013). Small-scale patch structure in North American and
South African grasslands responds differently to fire and grazing. Landscape ecology,
28(7), 1293-1306. doi:10.1007/s10980-013-9866-0
Laris, P. (2002). Burning the seasonal mosaic: preventative burning strategies in the wooded
savanna of southern Mali. Human Ecology, 30(2), 155-186.
Laris, P. (2011). Humanizing savanna biogeography: linking human practices with ecological
patterns in a frequently burned savanna of southern Mali. Annals of the Association of
American Geographers, 101(5), 1067-1088.
Laris, P. (2013). Integrating land change science and savanna fire models in West Africa. Land,
2(4), 609-636.
Laris, P., Dadashi, S., Jo, A., and Wechsler, S. (2016). Buffering the savanna: fire regimes and
disequilibrium ecology in West Africa. Plant ecology, 217(5), 583-596.
Laris, P. S. (2005). Spatiotemporal problems with detecting and mapping mosaic fire regimes
with coarse-resolution satellite data in savanna environments. Remote Sensing of
Environment, 99(4), 412-424.
McGarigal, K. (2014). FRAGSTATS help. Documentation for FRAGSTATS, 4.
McGarigal, K., Cushman, S., and Ene, E. (2012). FRAGSTATS v4: spatial pattern analysis
program for categorical and continuous maps. University of Massachusetts, Amherst,
Massachusetts, USA. goo. gl/aAEbMk.
McKenzie, D., Miller, C., and Falk, D. A. (2011). The landscape ecology of fire: Springer Science and
Business Media.
Mistry, J. (2000). World savannas: ecology and human use. World savannas: ecology and human
use.
Mistry, J., and Berardi, A. (2005). Assessing fire potential in a Brazilian savanna nature reserve.
Biotropica, 37(3), 439-451.
Nasi, R., and Sabatier, M. (1988). Projet inventaire des ressources ligneuses au Mali.
Overpeck, J., and Udall, B. (2010). Dry times ahead. Science, 328(5986), 1642-1643.
Parr, C., and Brockett, B. (1999). Patch-mosaic burning: a new paradigm for savanna fire
management in protected areas? Koedoe, 42(2), 117-130.
Pyne, S. J. (2008). Passing the torch. The American Scholar, 77(2), 22.

2
Randerson, J., Chen, Y., Werf, G., Rogers, B., and Morton, D. (2012). Global burned area and
biomass burning emissions from small fires. Journal of Geophysical Research: Biogeosciences,
117(G4).
Royston, P. (1992). Comment on sg3. 4 and an Improved D'Agostino Test. Stata Technical
Bulletin, 1(3).
Russell-Smith, J., Ryan, P. G., and Durieu, R. (1997). A LANDSAT MSS-Derived fire history of
kakadu national park, monsoonal northern australial, 1980-94: Seasonal extent,
frequency and patchiness. Journal of Applied Ecology, 748-766.
Scholes, R., and Archer, S. (1997). Tree-grass interactions in savannas. Annual review of Ecology
and Systematics, 28(1), 517-544.
Schultz, M. G., Heil, A., Hoelzemann, J. J., Spessa, A., Thonicke, K., Goldammer, J. G., . . . van
Het Bolscher, M. (2008). Global wildland fire emissions from 1960 to 2000. Global
Biogeochemical Cycles, 22(2).
Staver, A. C., Archibald, S., and Levin, S. A. (2011). The global extent and determinants of
savanna and forest as alternative biome states. Science, 334(6053), 230-232.
Tansey, K., Grégoire, J. M., Defourny, P., Leigh, R., Pekel, J. F., van Bogaert, E., and Bartholomé,
E. (2008). A new, global, multi-annual (2000–2007) burnt area product at 1 km resolution.
Geophysical Research Letters, 35(1).
Turner, M. G. (1989). Landscape ecology: the effect of pattern on process. Annual review of
Ecology and Systematics, 20(1), 171-197.
van der Werf, G. R., Randerson, J. T., Giglio, L., Collatz, G. J., Kasibhatla, P. S., and Arellano Jr,
A. F. (2006). Interannual variability in global biomass burning emissions from 1997 to
2004. Atmospheric Chemistry and Physics, 6(11), 3423-3441.
Watt, A. S. (1947). Pattern and process in the plant community. Journal of ecology, 35(1/2), 1-22.
White, P. S., and Pickett, S. T. (1985). Natural disturbance and patch dynamics: An introduction.
Unknown Journal, 3-13.
Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics bulletin, 1(6), 80-83.
Williams, A. P., Allen, C. D., Millar, C. I., Swetnam, T. W., Michaelsen, J., Still, C. J., and Leavitt,
S. W. (2010). Forest responses to increasing aridity and warmth in the southwestern
United States. Proceedings of the National Academy of Sciences, 107(50), 21289-21294.
Williams, J., Albright, D., Hoffmann, A. A., Eritsov, A., Moore, P., Mendes de Morais, J., . . . Van
Lierop, P. (2011). Findings and implications from a coarse-scale global assessment of recent
selected mega-fires. Paper presented at the International Wildland Fire Conference (5th,
9–13 May 2011, Sun City, South Africa). FAO.
Wu, J., and David, J. L. (2002). A spatially explicit hierarchical approach to modeling complex
ecological systems: theory and applications. Ecological modelling, 153(1), 7-26.