Spatiotemporal variation of fire occurrence in the State of Bahia , Brazil , between 2003 and 2019

Wildland fires are responsible for impacts in the fauna and flora in fire sensitive ecosystems and release into the atmosphere greenhouse gases responsible for global warming. This study aims to analyze the spatiotemporal variation of hot spots detected by the AQUA satellite from 2003 to 2019 in the state of Bahia, and to determine the main factors affecting the fire frequency of incidence. A significant downtrend was observed in the number of hot spots recorded over the years and about 65% of them were detected in the months of September and October. The Southcenter and the Northwest state’s regions were the ones with highest hot spot incidence, while the Northeast and East Central regions presented lower incidence. Presidente Jânio Quadros was the municipality with the highest hot spot incidence per area. Besides this one, there were another 351 municipalities (accounting together with almost 75% of the state area) that were classified in the “Extreme”, “Very High” and “High” classes of hot spot incidence. The independent variables that presented a significant correlation with the municipality hot spot density were, in decreasing order of significance: percentage of land covered by savanna; percentage of land covered by agriculture and pastures; percentage of land covered by forests; mean annual temperature; mean annual rainfall; and demographic density. The results obtained in this study should be used by municipal and state environmental agencies seeking to reduce the wildland fire occurrence and thus ensuring biodiversity conservation and reduction in the release of green house gases.


Introduction
Wildland fires are any non-structure fire that occurs in vegetation. They can be classified as controlled burns or wildfires depending on whether the fire is being used in a prescribed manner or if it is out of control (NWCG, 2020). Both, wildfires and controlled burn, are responsible for releasing into the atmosphere large amount of greenhouse gases, thus contributing to global warming. Wildfires in firesensitive ecosystems represent a major threat to the preservation of biodiversity, causing impact on the fauna and flora. Moreover, the smoke often causes respiratory complications and represents, in some locations, a public health issue (Arbex et al., 2004;Soares and Batista, 2007;Shlisky et al., 2009).
In several Brazilian states, the use of fire in controlled burns is cultural and difficult to change. Due to the lack of surveillance by the government and knowledge of the law by the population, usually the burns are held in an irregular manner, without the authorization from the responsible environmental agency, thus against the law. The lack of inspection results in a larger number of control burns that turned into wildfires difficult to extinguish that consume large areas of vegetation.
The detection of wildland fires via satellite began in the 1980s (Wang et al., 2012). Until a few years ago, the Advanced Very High-Resolution Radiometer (AVHRR) and Moderate Resolution Imaging Spectroradiometer (MODIS) were the sensors most effective abroad the satellites for this purpose. Currently, the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor, equipped abroad the Suomi NPP and NOAA-20 satellites, has a greater detection capacity (VIIRS sensors detect about 10 times more hot spots than MODIS sensors, for example). Nevertheless, since both satellites equipped with the VIIRS sensor were put into operation only a few years ago, data from other satellites are still important to analyze variations in fire occurrence through time and space INPE, 2020).
The images generated by the thermal and infrared sensors installed in the satellites are sent to a control center (on land), where they are processed through a detection algorithm (Batista, 2004;Wang et al., 2012). It is essential the use of an efficient algorithm, since a hot spot does not necessarily mean that the vegetation is burning, but that certain area presents, at the moment that the satellite passes, high temperature (Batista, 2004). Only after processing the images received by the satellite, it will be possible to identify which hot spots represent fire in vegetation. Therefore, the data processed by the detection algorithm represents the occurrence of wildland fire. The current algorithm used by the National Aeronautics and Space Administration (NASA) and by the INPE for images generated by MODIS sensors, such as the one abroad the AQUA satellite, is called "Collection 6". The use of the "Collection 6", replacing the "Collection 5", is part of a major reprocessing of the entire MODIS data archive to generate standardized data results with calibration enhancements and algorithm refinements (NASA, 2020a).
Tre Brazilian Weather and Climate Studies Research Center (CPTEC), from the INPE, generates and provides information on the occurrence of fires based on satellite data. Although receiving images from nine satellites that have optical sensors operating in the thermal-average range of 4µm (NOAA-18, NOAA-19, METOP-B, NASA, TERRA, AQUA, NPP-Suomi, NOAA-20, GOES-13 and MSG-3), the images generated by the AQUA satellite processed through the "Collection 6" algorithm has been used as "reference" by the INPE since 2002 to compose comparable time data over the years and thus enable trend analysis for the same periods in regions of interest (White, 2018a;INPE, 2020). Due to the similarities between control burns from wildfires, satellite imagery is unable to discern one from the other. Therefore, a hot spot represents a place where the vegetation is burning, whether controlled or uncontrolled.
This study has the objective to determine the spatiotemporal variation of hot spots detected by the AQUA satellite, processed by the "Collection 6" algorithm, in the state of Bahia, Brazil from 2003 to 2019. Also, to understand why some municipalities burn more than others. The results obtained from this study can be used by municipal and state environmental agencies to assist the development of public policy focused on wildfire prevention and nature conservancy.

Characterization of the study area
Bahia is the fourth largest Brazilian state. With 564,760.43 km 2 , it has an area bigger than several countries. The largest municipality is Formosa do Rio Preto (15,901.74 km 2 ) and the smallest Madre de Deus (32.20 km 2 ). The total population in the state is 14,016,906, that accounts for a mean demographic density of 24.82 inhab./km 2 , varying from 3,859.44 inhab./km 2 at the capital (Salvador) to 0.94 inhab./km 2 in Jaborandi (IBGE, 2010;IBGE, 2020).
The predominant climates in most of the state, according the updated Köppen-Geiger classification (Kottek et al., 2006), are: equatorial with dry winter (Aw), hot arid step (BSh) and equatorial fully humid (Af). Other climates classification that occurs in smaller areas in the state of Bahia are: equatorial monsoonal (Am), warm temperature fully humid warm summer (Cfb) and warm temperature fully humid hot summer (Cfa).
According to satellite imagery from the year 2018, most of the land was covered with natural vegetation (including forests, savanna, steppe and mangrove), occupying a total area of 262.786,12 km 2 , followed by agriculture/livestock areas, with 251,721.61 km 2 . This two land use make up for approximately 90% of the state area (MapBiomas, 2020). Three of six Brazilian biomes are found in Bahia: Atlantic Forest, Cerrado and Caatinga. The protected areas cover about 7% of the state, distributed in 110 Conservation Units (CU) (CBPM, 2020).
According to data from the INPE reference satellite (INPE, 2020), Bahia is the Brazilian state with the fifth highest number of hot spots detected, only behind the states of Pará, Mato Grosso, Rondônia and Tocantins. It is the state from the Northeast Region with the highest wildland fire occurrence. Deforestation data from MapBiomas (2020) accounts for a sum of 87,643 km 2 of deforested areas between 1990 and 2017, which represents a mean of 3,130 km 2 deforested annually.

Datasets
The records of hot spots in the state of Bahia for the period 01/01/2003 to 12/31/2019 were obtained from the INPE Satellite Monitoring Burning Program website, based on data from the AQUA satellite processed through the "Collection 6" algorithm. The values were quantified for the entire state and grouped by month of occurrence and the municipality in which they were detected.
The following variables that may have a significant influence on the number of hot spots were quantified for each municipality: mean annual temperature; mean annual rainfall; population density; municipality area; forest areas (including natural and planted forests, and mangroves); savanna formations (including Cerrado restricted sense, Parque Cerrado, Palmeiral and Vereda); non-forest natural formation areas (including wetlands, grasslands, apicum and other non-forest formation); non-vegetated areas; and agriculture/pasture areas. These variables were chosen due to availability of historical data and because they have been shown to influence the wildland fire occurrence (e.g. Ajin et al., 2016;Suryabhagavan et al., 2016;White, 2018b).
The mean annual temperature and the mean annual rainfall were obtained from Climate-Data (2020) based on climate models and data measured between 1982 and 2012. Data from population density were obtained from the last population census (IBGE, 2010). The municipality areas were obtained from IBGE (2020). And the data from the area of forest, savanna, non-forest natural formations, nonvegetated and agriculture/pasture fields were obtained from MapBiomas (2020) from 2003 to 2018 (the last year with available data).

Wildland fire incidence per municipality
The state municipalities were grouped according to the classification originally proposed by  with the inclusion of a new frequency class: Extreme (Table 1). This updated classification seeks to better describe the frequency of wildland fire based on the number of hot spots per area detected by the AQUA satellite during a period of one year.
Table 1 -Frequency of incidence of hot spots detected by the AQUA satellite over one year. The classification follows the originally proposed by  with the inclusion of the "Extreme" frequency class.

Frequency class Number of hot spots detected per year
Very Low None or one hot spot for an area > 600 km 2 (Equivalent to < 0.0017 hot spots/km 2 ).

Low
One hot spot for an area > 300 and ≤ 600 km 2 (Equivalent to > 0.0033 and ≤ 0.0017 hot spots/km 2 ).

Average
One hot spot for an area > 150 and ≤ 300 km 2 (Equivalent to > 0.0067 and ≤ 0.0033 hot spots/km 2 ).

High
One hot spot for an area > 75 and ≤ 150 km 2 (Equivalent to > 0.0133 and ≤ 0.0067 hot spots/km 2 ).
Very High* One hot spot for an area > 25 and ≤ 75 km 2 (Equivalent to > 0.04 and ≤ 0.0133 hot spots/km 2 ) Extreme** One hot spot for an area ≤ 25 km 2 (Equivalent to ≥ 0.04 hot spots/km 2 ) Note: *Size of the class altered from the original proposed by  for the inclusion of the Extreme class. ** New class proposed in this study.

Statistical analysis
The Student's t test was used to test the existence of significant differences among the number of hot spots recorded in different months of the year. Linear regression was used to evaluate the trend in the number of hot spots detected throughout the time series. A correlation matrix was constructed with the Pearson (r) correlation coefficients for the variables analyzed in the study.

Results
A total of 229,554 hot spots were detected by the AQUA satellite using the Collection 6 algorithm in the state of Bahia between 01/01/2003 and 12/31/2019. The year of 2007 had the highest number of detections (29,468), while 2018 had the lowest number (4,956). The mean during the assessed period was 13,503 hot spots per year with a standard deviation of 6,871. The linear regression analysis indicates a significant downtrend during the time series (r 2 = 0,48; p < 0,01) ( Figure 1).
When classifying the hot spots according to the months of the year in which they were registered and using the Student's t test to compare the means, it was possible to define 4 different groups ( Figure 2). In the group "A" are the months with the highest number of hot spots detected: September (mean of 4,484 per year, 35.4% of the total) and October (mean of 4,044 per year, 29.9% of the total). In the group "B" are the months of August (mean of 1,254 per year, 9.3% of the total) and November (mean of 1,098 per year, 8.1% of the total). Following, the group "C" with the months of July (mean of 491 per year, 3.6% of the total) and December (mean of 404 per year, 3% of the total). In the group "D" are the months with the lowest detection, that were, in descending order: June, January, March, February, May and April (mean between 290 and 188 hot spots per year, corresponding to 2.2 to 1.4% of the total).
Hot spots were detected in 412 of the 417 municipalities in Bahia during the period assessed ( Table 2). The municipality with the highest incidence was Formosa do Rio Preto (13,526). Proportionally to its area, Presidente Jânio Quadros registered the highest frequency of incidence of hot spots, followed by Cândido Sales and Belo Campo.
The correlation matrix was done using the proportional values for the following variables that was significantly affected by the municipality area: forest area (r = 0.29; p < 0.001), savanna formation area (r= 0.94; p < 0.001); non-forest natural formations (r = 0.79; p < 0.001), non-vegetated area (r = 0.49; p < 0.001), agriculture/pasture fields area (r = 0.78; p < 0.001); as also in the number of hot spots registered (r = 0.81; p < 0.001). So, it doesn't matter if a municipality has 10 or 1000 km 2 of forest area, for example, but how much that area represents of the total municipality size. The same is true for the number of hot spots. A municipality A, for example, which registered 10 times more hot spots than the municipality B, may present a lower hot spot density if its size is more than 10 times larger than the size of municipality B.  Table 2 -List of municipalities of Bahia state (Brazil) and their number of hot spots (HS) detected; mean annual hot spots; area; area divided by mean annual hot spots; hot spot frequency of incidence; forest area; savanna area; non-forest natural formations area; non-vegetated area; agriculture/pasture area; demographic density; mean annual rainfall and mean annual temperature. Municipalities are grouped in increasing order of the area by the mean annual number of hot spots.  ,80  322  Candeias  20  1,18  251,63  213,88  Average  53,42  12,92  3,12  11,00  165,14  321,87  1797  24,60   323  Santa  Teresinha  56  3,29  719,26  218,35  Average  103,37  3420,17  0,33  0,42  614,37  13,64  651  23,00   324  Cipó  12  0,71  155,72  220,60  Average  43,06  16,09  0,11  2,02  110,60  122,78  891  24,90   325  Coração de  Maria  29  1,71  378,42  221,83  Average  28,82  1822,82  0,10  1,51  347,96  64,34  1145  23,10   326  Casa Nova  730  42,94  9647,07  224,66  Average 5154,80  183,48  1563,73  32,32  1934,50  6,73  480  25,40  327 Camacan 44  Based on the analysis of the correlation matrix (Table 3) the most significant correlations obtained and related to the wildland fire occurrence were: the hot spot density increased as the percent of the municipality covered by savanna formations increased and as the percent of the municipality covered by agricultural and pasture fields decrease. Other less significant correlations were the hot spot density increases as the percent of a municipality covered by forest decreases; as the mean annual temperature decreases; as the mean annual rainfall decreases; and as the demographic density decreases. Non-significant correlations were found between the hot spot density and the percentage of non-forest natural formations, as also with the percentage of non-vegetated area.

Rank
Using the updated classification originally proposed by , 39 municipalities from the state of Bahia (together occupying 16.1% of the state area) were classified with "Extreme" frequency of hot spot incidence; 153 municipalities (occupying 46.3% of the state area) with "Very High" frequency of hot spot incidence; 90 (16.3% of the state area) with "High"; 69 (11.3% of the state area) with "Medium"; 26 (4.4% of the state area) with "Low"; and 40 (5.6% of the state area) with "Very Low". Most of the municipalities with the lowest wildland fire incidence were located in the Northeast and in the East Central regions while the ones with the highest fire frequency were located in the Southcenter and in the Northwest regions ( Figure  3).

Discussion
Even though 229,554 hot spots were detected in the state of Bahia during the period analyzed, it is important to state that this number only represents a small part of the real number of wildland fires that actually occurred. The MODIS sensor used by the AQUA satellite is relatively outdated. It is capable of detecting only fires with a front line width greater than 30 meters, inferior to recent launched satellites equipped with VIIRS sensor that can detect wildland fires with smaller front line (NASA, 2020b). Additionally, satellite based fire detection may also be restricted by the presence of clouds, by the canopy of the forest when the fire is consuming only the understory, when fires started and ended during the interval between the satellite passage, or when the fire occurs in a mountainside opposite to the satellite observation path (INPE, 2020). Nevertheless, since the AQUA satellite has been in operation for several years, its data is important to compose comparable time series.
The downtrend recorded in the number of hot spots in the state of Bahia is also verified when analyzing the data from the whole country. According INPE (2020), 2007 was the year with the highest number of hot spots detected by the AQUA satellite in Brazil, with a relative decrease in the following years. The same was registered in this study. The possible explanation to the reduction in the numbers from the country, as a whole, can be the decrease in the deforestation (White, 2018b(White, , 2019INPE, 2020). Unfortunately, data of deforestation from Bahia are unavailable for the years 2018 and 2019. However, data from 2003 to 2017 indicates a significant decline in such period (MapBiomas, 2020). Other factors such as the El Niño and La Nina can also have influence the data in this study, since recent studies proved the relation between El Niño-Southern Oscillation and interannual fire activity (Jiménez-Muñoz et al., 2016, White, 2018b, 2019. Despite the downtrend contradicts recent studies that indicate global warming is increasing wildfire occurrence in the last years (Moritz, 2012), it is important to mention that the 17 year period analyzed in this study is too short to identify patterns related to climate change.
The predominant climate in Bahia is tropical with dry winter and rainy summer, in most of the municipalities the dry season lasts for at least 5 months (Mendonça and Danni-oliveira, 2007). At the end of this period and the beginning of the rainy season (September and October for most of the state) the wildland fires are more frequent. In a smaller region denominated zona da mata, a 300 km extension across all the Bahia coastal zone, the summer is dry and the winter rainy (White and Ribeiro, 2011). Therefore, in this region, the wildland fires are more frequent between December and March. The highest number of hot spots in the months of September and October found in this study is related to the predominant climate in the state that follows the pattern observed in most of South America countries. All South American countries below the equator, with the exception of Chile, present higher fire activity from August to November (White, 2019). In Brazil, the fire season lasts from July to December, with the peak in the months of August and September (Soares and Batista, 2007).
Through the analysis of the correlation matrix it was possible to conclude that the municipality with lower mean annual temperature and lower mean annual rainfall presented higher hot spot density. The correlation with the rainfall was expected, since vegetation moisture content increases with rainfall, making it more difficult to burn (Soares and Batista 2007;White and Ribeiro 2011;White, 2018a). The temperature also influences the moisture of the vegetation and, consequently, the facility which it burns. Despite higher temperatures favor the drying of the vegetation (Soares and Batista 2007;White, 2018b), in tropical regions, higher temperature is associated with higher rainfall amount. This occurs in the Bahia municipalities, since the correlation between both variables was positive and significant. Therefore, even though higher temperatures can contribute with a higher wildland fire occurrence, its effect was possibly suppressed by the rainfall.
Although meteorological parameters play a key role in the wildland fire occurrence, it is essential to analyze human activities in order to understand better fire occurrence since, in Brazil, 99% of the ignitions sources are associated with human activities and only 1% are originated from a natural source, that is lightning (Soares and Batista, 2007). A good way to understand the main factors associated with the wildland fire occurrence is analyzing the land use variable White, 2018b). In this study, the municipalities with a greater percentage of its total area covered by savanna formations presented a greater hot spot density. Also, the municipalities with higher percentage of its total area covered by forests presented lower hot spot density. Those results are in accordance with the literature, since tropical forests, including the Atlantic Forest biome, has a low wildfire risk of occurrence (Ajin et al., 2016;Suryabhagavan et al., 2016;White, 2018a), while savanna formations have a high wildfire risk of occurrence (Nolasco and Sanhueza, 2011;White, 2019). Also, the percentage of land covered by pastures and agricultural fields presented a negative significant correlation with the hot spot density. Such correlation was unexpected, since fire is a commonly used tool for clearing areas in farming activities and has already been shown by several authors as presenting a positive significative correlation with wildland fire occurrence (Ajin et al., 2016;Suryabhagavan et al., 2016;White, 2018b). A possible explanation for such finding could be that, as presented in Table 3, the municipalities with a greater percentage of land use covered by savanna formations presented less percentage of land used to farming activities. Therefore, if the savanna formations burn more frequent than pastures and agricultural fields, could end up mistakenly reflecting in the statistical analysis that farming activities diminish wildland fire activity.
The demographic density also presented a negative significant correlation with the hot spot density. Some authors affirm that more densely populated regions would be more prone to burn, since most of the wildland fires are initiated by humans (Soares and Batista, 2007). Nevertheless, larger populations are usually concentrated in urbanized areas, with lower vegetated formations and consequently with lower wildland fire occurrence.
The map indicating the wildland fire incidence in the municipalities of the state of Bahia is an important visual tool to assess the future risk of wildland fire occurrence, since data from past fires is an important variable used to predict areas with higher risk of future fires .

Conclusions
During the period analyzed, the number of hot spots in Bahia presented a significative downtrend over the years. The months of September and October were the ones with higher fire occurrence. The percentage of land covered by savanna formations was the variable that had the greatest influence on the municipal hot spot density. The others variables that also had significative influence in the hot spot density were, in decreasing order of significance: percentage of land covered by agriculture and pasture fields; percentage of land covered by forests; mean annual temperature; mean annual rainfall; and demographic density.
Almost 75% of the state area was classified with "High", "Very High" or "Extreme" frequency of incidence of hot spots. The municipalities with the highest hot spot incidence were located in the Southcenter and in the Northwest regions, while the ones with the lowest hot spot incidence were in the Northeast and East Central regions.
The data obtained thought this study can and should be used in the development of public policies focused in the reduction of environmental harmful wildland fire events, as well as, to subsidize programs of biodiversity management and conservation.