Spatio-temporal patterns of wildfires in the Niassa Reserve –Mozambique, using remote sensing data

Wildfires are among the biggest factors of ecosystem change. Knowledge of fire regime (fire frequency, severity, intensity, seasonality, and distribution pattern) is an important factor in wildfire management. This paper aims to analyze the spatiotemporal patterns of fires and burned areas in the Niassa Reserve between 2002-2015 using MODIS data, active fire product (MCD14ML) and burned area product (MCD64A1). For this, the annual and monthly frequencies, the trend of fires and the frequency by types of forest cover were statistically analyzed. For the analysis of the spatial dynamics of forest fires we used the Kernel density (Fixed Method). The results show a total of 20.449 forest fires and 171.067 km2 of burned areas in the period 2002-2015. Fire incidents were highest in 2015, while the largest burned areas were recorded in 2007. The relationship between increased fires and burned areas is not linear. There was a tendency for fires to increase, while for burnt areas there was stabilization. Forest fires start in May and end in December. August-October are the most frequent period, peaking in September. Fires occur predominantly in deciduous forests and mountain forests because of the type of vegetation and the amount of dry biomass. There is a monthly spatial dynamics of wildfires from east to west in the reserve. This behavior is dependent on vegetation cover type, fuel availability, and senescence.

Fire detection is performed using a contextual algorithm that exploits the strong emission of 106 medium infrared fire radiation and is based on the brightness temperature derived from the 4 and 11 107 µm channels. [26]. The location of the fire corresponds to the center of a 1x1 km pixel, signaled by the 108 algorithm as containing one or more fires within the pixel. To avoid false alarms (commission errors), 109 only highly reliable fire pixels (> 80% reliability) were considered.

111 Product of burned area MODIS MCD64A1 112
The burned area data were obtained from the MODIS burned area sensor, monthly product 113 MCD64A1 Version 6, available from the MODIS Active fire and Burned Area products website 114 (http://modis-fire.umd.edu/ ). MCD64 is the latest product from the MODIS Burned Area product. 115 This is a 500 m product, global grid level 3. It is based on an automated hybrid approach that exploits 116 the 1 km MODIS active fire potential and 500 m surface reflectance input data [27]. 117 The hybrid algorithm applies dynamic boundaries to composite images generated from a burn 118 sensitive vegetation index, which in turn are derived from shortwave infrared channels, MODIS band 119 5 and 7, and a measure of temporal texture. Data layers include recording date, recording data 120 uncertainty, quality assurance and the first and last day of reliable change detection of the year 121 Overall the MCD64A1 has improved detection of burned areas over past collections. But 122 specifically in significantly better detection of small fires and adaptability to different regional 123 conditions in multiple ecosystems.

153
Where: n is the number of points observed; h is the bandwidth; K is the Kernel function; x is the 154 coordinate vector representing the location of the estimated point; and Xi is the vector of the ith 155 coordinate that represents each point observed in relation to the estimated.

157
One of the key steps in estimating kernel density, in addition to choosing the kernel function, is 158 setting parameter smoothing, such as bandwidth size ('search radius' in ArcGIS 10.2). Often 159 determined on the basis of subjective choices, specialized knowledge or possibly supported by 160 empirical decisions. 161 According to [32], the choice of bandwidth depends on the purpose of density estimation. If the 162 goal is to explore the data and suggest models and hypotheses about it, it is sufficient to choose the 163 smoothing parameter subjectively by visual inspection. However, it is very difficult to define this 164 subjective value, which can generate ambiguous results, because the values depend on the scale 165 adopted and the specific characteristics of the studied area  Since there is spatial variability of data in the reserve, and the total fire distribution is regular or 172 homogeneous with respect to the total reserve area, both methods could be adopted. However, for 173 reasons of comparing the monthly density, we opted for fixed bandwidth. For this we used the mean 174 distance method (RDmean), which can be analyzed by local or global approach. Defined by [31] as: 175 Where: A is the total size of the study area; N is the total number of Heat Focuses.

179
The parameters used and the bandwidth result are presented in Table 1.  The total monthly distribution chart of forest fires and burned areas comprises the historical 197 series with accumulated monthly totals from 2002 to 2015. (Fig. 3). 198

181
The occurrence of fires and burned areas was recorded between May and December, the  It is noteworthy that there is a spatial-monthly dynamics of outbreaks and burnt areas in the 218 reserve. From May to June (early burning season), fires start east of the reserve (Fig. 5), with the 219 highest occurrence of fires and burned areas in semi-deciduous open forests (Fig 4), and migrate 220 gradually to the center of the Reserve.

221
From July to August there is a migration and predominance of fires and burned areas in 222 deciduous forests (Fig. 4A and 4B). Especially in the central east, central and midwest regions (Fig.  223 5). However, in September, fires migrate and predominate in deciduous forests as well as mountain 224 forests ( Figure 4A and 4B), and are located in the center and west of the reserve ( Figure 5). From 225 October to December there is a migration of fires to the west of the Niassa Reserve, with a 226 predominance of Mountain forests.  240 which may be related to meteorological phenomena such as temperature increase and low rainfall over the 241 analyzed period 242 The occurrence of fires and burned areas was recorded between May and December, with the peak 243 recorded in September (Fig 3) It is also observed that the highest occurrence of fire and the highest severity occur between August, 249 September and October, with about 70% of fire and 73.6% of burned area, with maximum peak in September. 250 (Fig 3). What can be explained by a combination of factors, September is the month with the lowest rainfall 251 (Fig. 1b), low relative humidity, and highest dry biomass accumulation, which causes greater ignition 252 occurrence and ease of propagation fire, especially in deciduous and mountain vegetation. Similar results 253 were found in studies conducted on similar phytophysiognomies in southern Africa [38], [41], [42]. 254 The end of the fires occurs between November and December, when rainfall falls on average exceed 150 255 mm (Fig. 3). This result was not expected, as the hydrological year begins in October. However, according to 256 [13] the increase in precipitation does not necessarily correspond to the immediate reduction in the occurrence 257 of fires. As the soil and combustible material are still low in moisture, precipitation will be absorbed to the 258 point where the combustible material will no longer ignite. However, October is the month with the highest 259 amplitude or variability of fire occurrence in the reserve, which may be related to the beginning of the 260 hydrological year in the second half of this month. 261 262 Spatial pattern

263
The spatial pattern of fires was influenced by the type of forest cover. The fires occurred predominantly 264 in areas of deciduous vegetation and mountain forest. The burned areas followed the same trend with greater 265 severity in deciduous vegetation and mountain forest. (Fig. 2). 266 These results were expected, as approximately 60% of the reserve is occupied by deciduous forests, and the 267 phytophysiognomy presents very low values of vegetation vigor in the dry period. However, Mountain forests 268 were not expected to have high numbers of forest fires and burned areas, which may be related to the greater 269 accumulation of dry biomass, especially from September to October. 270 It is noteworthy that there is a spatial-monthly dynamics of outbreaks and burnt areas in the reserve.  The main factors that determine the beginning of the fire season in this region and in this 274 phytophysiognomy are the high temperature and the low precipitation. According to [25] this is the first 275 phytophysiognomy that enters the senescence process in the Niassa Reserve and consequently higher 276 probability of fire in this period. 277 According to [43] fires are limited by the availability of fine fuel, especially in semi-deciduous open 278 forests, which in turn are dependent on soil moisture and nutrient availability. Thus, the low amount of dry 279 biomass in Semi-deciduous Open Forests (Eastern Reserve Region) determines the gradual (spatial-temporal) 280 migration of fires to the Reserve Center in deciduous forests (Fig 5). 281 From June to August there is a migration and predominance of fires and burned areas in the deciduous forests 282 (Figs 4 and 5), especially in the central east, central and midwest regions. In this phytophysiognomy, in this 283 period, there is the beginning of senescence, and consequent reduction of vegetation vigor [25]. 284 With the end of dry biomass in deciduous forests and the onset of senescence in mountain forests, 285 starting in September, the same process of migration of fires to mountain forests west of the reserve occurs. 286 This region covered by wetter forests, with a long phenological cycle, requires more time to provide 287 flammable conditions for forest fires. The trend of increased occurrence of fire outbreaks and burned areas in 288 this phytophysiognomy continues until December.

289
According to [44], fires migrate when there is no longer enough fuel to sustain them, when weather 290 conditions are not prone to burning or when they encounter topographic or anthropogenic barriers or 291 previously burned areas. It is evident in the reserve area that the lack of sufficient fuel for its sustainability is 292 fundamental to the dynamics of fires. But, as already mentioned, the beginning of the burnings and their 293 migration are also strongly dependent on senescence and the availability of dry fuel.