The Portuguese Large Wildfire Spread Database (PT-FireSprd)

. Wildfire behaviour depends on complex interactions between fuels, topography and weather, over a wide range of 17 scales, being important for fire research and management applications. To allow for a significant progress towards better fire 18 management, the operational and research communities require detailed open data on observed wildfire behaviour. Here, we 19 present the Portuguese Large Wildfire Spread Database (PT-FireSprd) that includes the reconstruction of the spread of 80 large 20 wildfires that occurred in Portugal between 2015 and 2021. It includes a detailed set of fire behaviour descriptors, such as rate- 21 of-spread


107
Considering that all data sources have limitations and provide information for very limited time frames, combining different 108 sources is key to capture the spread and behaviour variability of wildfires. The example provided in Figure 1 highlights the 109 potential of combining different data sources to overcome inherent acquisition gaps, particularly in the afternoon, when both 110 field and airborne data overcome the satellite gap, and during dawn, when ground-collected and satellite data complement each 111 other. Note that observation frequencies of ground and airborne data strongly depend on daily fire activity patterns.  Storey et al., 2020Storey et al., , 2021. Additional data on constantly 119 evolving wildfires, accompanied by robust replicable methods, is needed, namely in southern Europe where a substantial data 120 gap is manifest (Fernandes et al., 2018). 121 122 Here, we present the Portuguese Large Wildfire Spread Database (PT-FireSprd) that combines data from multiple sources, 123 using a "convergence of evidence" approach to characterise in detail the progression and behaviour of large wildfires in 124 Portugal. Fire behaviour is described in sensu stricto, thus analysis of its drivers and effects is beyond the scope of the current 125 work. The work results from a joint co-creation effort between researchers and fire personnel, integrating data collected from 126 airborne and ground operational resources. regarding the probable location of the fire start, active flaming zones, and specially wildfire progression. It is noteworthy to 207 mention that airborne footage is not synoptic, as different parts of the wildfire (e.g. left flank vs. right flank) are captured at 208 different moments. These, depending on the fire extent and operational priorities can be characterised by significant time lags. 209

Ground data 210
The FEBMON system is linked to user-friendly portable tools that allow collection of georeferenced ground data during 211 wildfires. These tools are typically installed in mobile phones and tablets and are used by fire personnel from several 212 organisations (e.g., fire fighters, forest service). Ground-collected data consists of three main types: i) photos and videos; ii) 213 points that identify active flaming combustion, inactive flaming or smouldering or locations requiring mop-up activities; iii) 214 polygons that delineate an area burned until the time of acquisition (i.e. fire progression). 215

Reports 234
We also used ignition and fire progression data published in reports on the dynamics of the very large wildfires of we used only very limited information regarding ignition location\time and general fire spread patterns, mostly to complement 249 data provided by Guerreiro et al., (2017Guerreiro et al., ( , 2018. 250 251 We chose to include these fire progressions in our database, because they represent the most extreme wildfires that occurred 252 in mainland Portugal, under persistent cloud cover conditions that limited the acquisition of satellite data, and for that reason 253 they constitute relevant case studies, which otherwise would not be represented. 254

Wildfire Progression (L1) 255
Wildfire progression characterises the spatial and temporal evolution of the area burned in a specific fire event. It also contains 256 information regarding the ignition time and location, as well as, flaming zones that correspond to active areas during the 257 wildfire. These include spot fires and reactivation/rekindling areas. In Portugal, a rekindle is a reactivation of the wildfire after 258 its official conclusion and is considered a new incident. For simplicity, we will consider rekindles as reactivations throughout 259 the rest of the manuscript. 260

261
To robustly reconstruct wildfire progression, we combined the maximum available data from the different sources mentioned 262 above, with the aim of obtaining convergence of evidence. This allowed reducing the limitations and uncertainties of each 263 individual data source and building higher confidence in the derived wildfire progression. Combining all the available data , we manually delimited the extent and time of the ignition, fire progression and active flaming 266 zones of each wildfire. The reconstruction was always made chronologically, i.e. starting from ignition and ending with the 267 progression prior to wildfire containment. Sentinel-2 and Landsat 8/9 pre-fire images were used to identify areas burned shortly 268 before the wildfire, and post-fire images were used to correct each progression polygon. As an example, Figure 3  Ignition was defined as an area, instead of a point, to account for uncertainties in its location and to have a common data 275 typology for the entire database, in this case, vector polygons. We used mostly official ignition data and initial attack airborne 276 photos to define its location. This was complemented with expert knowledge and information from fire personnel to better 277 define ignition location. For a small set of wildfires (mostly nighttime ignitions), we also used satellite imagery and active-fire 278 data to identify the ignition area. All ignitions were compared with later fire spread patterns and with the final burned area to 279 reduce errors and guarantee consistency (e.g. ignition was contained in the final burned area). Regarding ignition time, the 280 official time of alert was compared with high frequency MSG-SEVIRI FRP detections, to confirm the alert time or, in a very 281 few cases, to anticipate if energy was released before the official ignition time. In addition, MSG-SEVIRI FRP were also useful 282 to identify (or confirm) the timing of reactivation. A clear example is shown in Figure 3, where the significant release of energy 283 around 11:30, combined with ground data, allowed identifying the location and time of the reactivation zone. 284 285 Active flaming zones were mostly derived from ground and/or high spatial resolution satellite imagery. Alternatively, they 286 were defined based on visual interpretation of multiple moderate resolution satellite imagery and often combined with active 287 fire data (mostly VIIRS due to its spatial resolution). Inconclusive visual interpretations were discarded, as well as active zones 288 that did not lead to any relevant subsequent fire spread. The ignition zone and all active flaming zones were always contained 289 within the subsequent fire spread polygon. 290 291 Wildfire progression was represented by a series of consecutive polygons delineating the temporal evolution of the area burned 292 by the wildfire. The number of polygons depended on fire size and data availability. The progression polygons were built using 293 as many data sources as possible, complementing each other in both space and time (see Figure 1). As an example: a common 294 feature found in the data was a pronounced fire spread during daytime, followed by very limited nighttime progression. In 295 these cases, first, the nighttime fire progression was delineated using active fire data (mostly VIIRS) and complemented with 296 ground data, when available. Second, satellite and/or airborne imagery acquired during the following morning were used to 297 perform any necessary adjustments in the nighttime spread polygon(s). Satellite-derived FRE estimates based on SEVIRI/MSG 298 were also used to identify if any substantial fire activity occurred between VIIRS/MODIS nighttime overpass and daytime imagery (satellite and/or airborne). We assumed that fire activity decreased significantly when the wildfire released less than 300 0.5 TJ per 30' period, and anticipated the date/hour of the fire spread polygon accordingly. In smaller wildfires (<500 ha) this 301 threshold was set to 0.1 TJ. These thresholds were defined empirically (see Discussion section). The entire procedure reduced 302 the uncertainties associated with the delineation of the nighttime spread polygons. It should be noted that the fire behaviour 303 within the time span of each progression polygon was unknown and, therefore, was assumed to be free burning in a 304 homogeneous way (Storey et al., 2021). When data were insufficient to determine when a given area burned, the spread 305 polygon was flagged as "uncertain". 306 307 Ignitions/active flaming zones were linked to the resultant spread polygon(s), by assigning a numeric label to a field called 308 "zp_link", providing an explicit connection between both, and allowing to track the source of a given burned progression 309 polygon. When information was insufficient, for example, the start of the progression polygon was unknown, zp_link was 310 defined as "0". After all ignition(s), fire progressions and active flaming zones were defined, each wildfire was divided into 311 burning periods. We assumed that each burning period contained relatively homogeneous fire runs that: 312 313 i) were ignited by the same set of ignitions or active flaming zones; 314 ii) did not exhibit large fire spread direction shifts (less than 45° of variation); 315 iii) were not impeded by barriers (e.g. previously burned area) and; 316 iv) did not exhibit significant changes in fire behaviour (e.g. large ROS variation). 317 318 Regarding the latter criterion, for example daytime and nighttime runs were usually separated in different burning periods even 319 if criteria (i)-(iii) were fulfilled. By definition, a new active flaming zone always marked the beginning of a new burning 320 period; however, not all burning periods started with an ignition or active flaming zone, since this depended on data availability. 321 322 When direct evidence of fire spotting was available (i.e. exact location/timing of the spot fire(s), typically from ground and/or 323 airborne data), if the fire front(s) rapidly (under 1 hour) coalesced with the original fire front, fire progression was merged 324 into a single polygon. In the remaining cases, typically associated with medium distance spotting and/or slow burning fire 325 fronts, the spotting location was defined as a new active flaming zone setting, defining a new burning period. When the exact 326 location/timing of the spot fire was not available, evidence of spotting consisted of observations of non-contiguous burned 327 areas that resulted from the same wildfire. These were typically separated by rivers, lakes and settlements. In these cases, due 328 to lack of data, the polygons separated from the major fire run were defined with zp_link=0 if the distance was larger than 200 329 m. No fire behaviour descriptors were calculated for these burned areas. 330

331
The definition of the burning period was always dependent on data availability and, in some cases, was subjective. For the 332 progressions derived using only satellite data, the length of the burning period was mostly determined by the timing of the satellite overpass(es) and the FRE temporal evolution. For the progressions derived from more detailed data, the above-334 mentioned criteria were easier to fulfil. In a few cases, uncertainties in fire progressions led to slightly overlapping periods.  The spread direction was calculated using trigonometric rules considering the two above-mentioned vertices between two 394 polygons. The spread direction was calculated both for ROSp and ROSi, where the difference lies only on the origin polygon. 395 FGR was calculated dividing the burned area by each polygon/node (Aj) by the time elapsed between polygons (Δtij) and was 396 expressed in ha/h. An example of the calculation of these fire behaviour descriptors is shown in Figure 5. 397 In addition to the standard fire behaviour descriptors, we also estimated the FRE for each progression polygon. This procedure 401 raised additional challenges. First, MSG-SEVIRI is affected by clouds and smoke, which can hinder the estimation of FRE for 402 some periods of the wildfires, or for their entire duration. Second, due to the coarse resolution of MSG-SEVIRI it was not 403 possible to calculate the FRE for each polygon directly. To circumvent this, FRE was calculated for each 30' bin from ignition 404 until the date/hour of the last wildfire spread polygon. In parallel, we estimated the area burned in each spread polygon every 405 30', using its start/end dates and assuming a constant FGR. Then, for each 30' bin, the total FRE was divided by weighting its 406 value by the proportion of area burned in each spread polygon. Finally, for each spread polygon the 30' FRE estimates were 407 summed only if they covered more than 70% of its duration (Δtij), to ensure that the total FRE was representative. 408

409
We also estimated the FRE flux rate (GJ ha -1 h -1 ) for each spread polygon by dividing the estimated FRE by the corresponding 410 burned area extent and its duration (Δtij). As FRE is highly dependent on the extent burning at a given time window, the FRE 411 flux can provide estimates closer to "instantaneous" values required for other applications. 412

Simplified Wildfire behaviour (L3) 413
We calculated simplified metrics representing a mean fire behaviour across each burning period. This enables higher-level 414 analysis of the data, but at the cost of losing detail and making simplifications to the calculation of the fire behaviour metrics. 415

416
The simplified ROS corresponded to the ROSi estimated for the last spread polygon of a given burning period i.e. the average 417 ROS between the start and the end of each burning period. FGR was defined as the sum of the area burned in the period divided 418 by its duration. The total FRE was calculated considering all energy released by the polygons burning within the burning 419 period, if FRE estimates covered more than 70% of the area burned. 420

Quality Control and Quality Assurance (QC/QA) 421
All L1 to L2, and L2 to L3 processing was done using Matlab scripts complemented with quality controls checks to identify 422 errors in the original L1 data. These included simple checks to incorrect field names, incoherent data format (e.g., date/hour), 423 and consistency on the fire spread structure defined by the di-graphs, as for example: i) time elapsed between node was always 424 positive;and ii) every spread polygon with a positive zp_link was always associated with a predecessor valid node (either of 425 "z" or "p" type), among others. 426 427 During the processing of L1 data to L2, we did frequent quality checks to identify potential errors, for example, null values of 428 ROS or FGR associated with valid fire spread polygons, fire progression polygons that did not have a known start/end date, or made independent calculations of the ROS and FGR and compared them with the ones estimated using the developed Matlab 431 code. All these quality control steps assured that the data produced were reliable and of the best possible quality. The process 432 was iterative, requiring frequent corrections to the L1 data and the re-run of the quality check. 433 434 Finally, for each wildfire we defined a confidence flag that provides an overall information of how reliable the fire progression 435 data were. Although directly related to L1, ultimately it should also provide the user an estimate of the confidence associated 436 with L2 and L3. This was defined empirically based on the uncertainties that arose in the process of building the fire progression 437 polygons and was graded into a 5-level system where 1 refers to the lower quality and 5 to the highest quality (Table A1). database spans a wide fire behaviour variability both between (e.g. Figure 6A,B,F) as well as within each wildfire (e.g. Figure  444 6C,E,D). The total burned area extent of the wildfires contained in the database was around 460,000 ha, which represents about 445 half of the area burned in the 2015-2021 period. On average, progression was reconstructed for 93% of the area burned by the 446 80 wildfires, leaving 7% deemed "uncertain". Wildfire behaviour descriptors were estimated for 88% of the burned area extent 447 (ca. 400,000 ha). The time elapsed between two consecutive fire progression polygons ranged between 30' and 14h30 with an 448 average value of 3h15. The mean duration of the burning periods was around 8h00, with a standard deviation of 4h50. researchers and fire personnel. The percentiles were translated into empirical classes, ranging from "very low" to "extreme" 489 fire behaviour. In general, as ROS increases so does the FGR. However, the relationship between ROS and FGR depends on 490 the morphology of the fire perimeter: elongated fast-spreading wildfires had relatively higher ROS and lower FGR (e.g. Figure  491 6B, C) and more complex burned area perimeters had relatively lower ROS and higher FGR (e.g. a flank run with an extensive 492 active fireline; see Figure 6A and the last polygons of Figures 6E and 6F) in an agricultural area around 19h30. In this second burning period, fire behaviour was significantly different from the first. 521 The mean ROS was ca. 1500 m/h, reaching a maximum value of 3720m/h between 16:30 and 17:30. On average, the fire grew 522 at a rate of 455 ha/h, however, significant variability was observed with values reaching 1236 ha/h coinciding with the ROS 523 peak. Framing the fire behaviour descriptors with the empirical classes represented in Figure 8, the behaviour in the second 524 burning period was often framed in the "Very High" class, i.e. between percentiles 90 and 97.5. As a consequence of the 525 behaviour exacerbation, the wildfire released around 38 TJ, with peaks of about 9 and 12 TJ observed during the afternoon. 526 The energy flux rate was highest between 16:00 and 16:30, coinciding with an abrupt increase in ROS (Figure 10d). For several reasons, it is easier to collect information for larger wildfires than for smaller ones. The wide range in fire size 559 present in the PT-FireSprd database suggests that it is representative of wildfires burning under a broad range of conditions. However, smaller wildfires (between 100 and 500 ha) are slightly under-represented in the database creating a potential bias. 561 This can be particularly relevant if one considers the proportion of smaller wildfires that occur every year. Thus, fire behaviour 562 descriptors may also be biased towards larger values which may have an implication, for example, on the fire behaviour classes 563 defined in Figure 8. Note that for typical fuel loads, say 15-20 t ha-1 (Fernandes et al., 2016), the third class in Fig. 8 already  564 corresponds to fires very difficult to control directly (Hirsch and Martell 1996). Nevertheless, these classes should be 565 considered as a first exploratory approach with the aim of creating a simple and clear communication baseline between 566 researchers and fire personnel based on quantitative fire behaviour data. Ultimately, the database will allow framing the 567 behaviour of new wildfires according to historical patterns. Adding smaller wildfires to the PT-FireSprd database will certainly 568 help to better represent a wider range of fire behaviour. 569 570 Confidence in the wildfires of 2015-2016 was lower than for the most recent ones due to relevant advances in operational fire 571 monitoring resulting in better quality and higher quantity of fire data. Since 2018, the FEBMON system has improved and 572 grown, providing larger quantity and higher quality data, thus leading to more reliable and detailed fire progression 573 reconstructions. The distribution of the duration of the spread polygons between 2015 and 2021 ( Figure B3) shows 574 heterogeneity of the database across time, but also the evolution introduced by the implementation of the FEBMON system. 575 Results suggest that estimates of ROS and FGR might be underpredicted in wildfires with lower confidence, most probably 576 due to the lack of data to thoroughly cover the afternoon, but especially the early night period (i.e. between VIIRS/MODIS 577 day and nighttime overpasses, Figure 1). This issue is further discussed in section 5.

Limitations and future improvements 609
The generic limitations of the input data have been thoroughly described in Section 1. In particular for Portugal some 610 limitations of the data must be pointed out. Fire progression perimeters and fire points collected in the ground by fire personnel 611 have relevant spatio-temporal uncertainties. For example, there is often a lag between the date/hour a polygon is drawn in the 612 ground and the actual date/hour it burned completely. Another relevant issue is that of data acquisition / reporting errors done 613 by fire personnel, which may be reduced by improved training and experience. The number of users of the FEBMON system 614 has been growing in recent years and, with adequate training, it is expected that the quality and quantity of ground data will 615 increase in upcoming years. In fact, over 27,000 aerial and 2,500 ground photos were taken in the year 2022 which represents 616 a relevant increase compared to previous years. 617 618 Regarding airborne data, the discussion can be separated into two components. First, initial attack photos, which can be 619 extremely useful to draw initial fire progression and infer probable ignition areas, are not collected for every wildfire to which 620 a helicopter is dispatched, and sometimes are of poor quality. Additional training and increasing the awareness of fire personnel 621 for the relevance of the data they collect is necessary. Second, aeroplane data are acquired at relatively low altitude, precluding 622 a synoptic view of the wildfire. Time lags between data acquisition for different parts of the wildfire (e.g. left vs. right flanks) 623 may be large and introduce relevant spatio-temporal uncertainties in the delineation of the fire progression. In addition, 624 perimeters are drawn manually and depend on the training and experience of the fire expert. In upcoming years, the integration 625 of new airborne sensors, specially with multispectral capability, the ability to perform high-altitude scans and the use of airborne fire observations. With this new capacity, it will be possible to integrate deep learning processes in the data analysis, 628 increasing both the quantity and quality of the available fire data. This integration will also allow a well-organised structure in 629 data collection, management and analysis, improving decision-support systems. Finally, the use of UAVs during nighttime 630 (pioneered in 2022 in Portugal) will complement aeroplane/helicopter data during periods of low data availability. 631 632 Regarding official fire data, errors in the delineation of final burned area perimeters and in the ignition location, often located 633 outside of the fire perimeter, need to be corrected to increase the quality of the PT-FireSprd database. Regarding satellite data, 634 implementing (semi-) automatic algorithms to delimit fire perimeters (e.g., Chen et al., 2022) will increase the availability of 635 fire perimeters and reduce the uncertainties associated with manual perimeter delimitation. Improvements in the spatial 636 resolution geostationary satellites, such as the recently launched Meteosat Third Generation (MTG), will certainly improve 637 fire behaviour estimates, as already observed in HIMAWARI-8 and last generation GOES satellites. 638

639
Regarding methodological uncertainties, the major challenge was to assign the correct date/hour to a specific burned area. For 640 example, when raw data sources indicated that an area burned but active areas were absent or small, there were always 641 uncertainties as to when it actually burned completely, which could lead to a relevant ROS/growth rate underestimation. These 642 uncertainties were larger between dusk until VIIRS overpass(es) and between the later and dawn. One approach to reduce 643 these uncertainties was to use FRE data to monitor the daily cycle of fire activity and help to better define the start/end date of 644 a progression polygon. The method was empirical and future work is needed to better define the thresholds for setting the 645 ignition or reactivation times, as well as the end of a fire progression. Exploratory analysis done in a few wildfires of the PT-646 FireSprd database suggest that FRE has a significant drop after the head of the fire stops, which may take several minutes/hours 647 until reaching the FRE thresholds used. This moment is commonly accompanied by a flank growth that burns slower and 648 releases lower amounts of energy. These fire dynamics probably explain why ROS was likely underestimated in low 649 confidence wildfires and why FGR was less affected by data confidence. Improvements can be achieved in the future, through 650 the use of more sophisticated methods (e.g. change point detection), more ground observations during the head to flank run 651 transition, and higher spatial resolution data from geostationary satellites. Part of these improvements can be used to partially 652 update the 2015-2021 wildfires of the PT-FireSprd database. 653

654
In terms of characterising uncertainties and its effects, future work should also adopt a metrological approach to propagate 655 uncertainties to the descriptors, providing useful information to users. By providing an uncertainty assessment, the PT-FireSprd 656 database would be on the pathway of Fiducial Reference Measurement (FRM) compliance. 657

658
The continuous update of the PT-FireSprd database will require a joint effort by researchers and fire personnel. The automation 659 of data collection procedures (discussed above), as well as dedicated training to fire personnel, are key factors to guarantee 660 both the quality as well as a sustainable update of the database. In the upcoming years, other fire behaviour descriptors could be included such as type of spread (surface vs. crown fire), fireline intensity, flame size, spotting (including maximum distance) 662 and/or PyroCb occurrence. Finally, methods described in the current work can be, at least partially, applied to many other fire-663 prone areas of the globe and contribute to the much-needed data on observed wildfire behaviour. 664

Data Availability 665
The dataset contains generic metadata file with relevant information for each wildfire (Table A2), such as the fire ID, official 666 incident ID (ANEPC, 13 digit number), fire name, municipality, civil parish, start date, duration (hours), extent (ha), among 667 others. The fire name was defined as Municipality_DDMMYYYY, where DD is day, MM month and YYYY the year. In 668 case more than one wildfire occurred in the same municipality on the same day, we added an additional string at the end of the 669 fire name (e.g. "_2"). 670

671
The dataset is then divided in 3 Levels, with three corresponding folders: 672 • Fire Spread (L1): Each year has a separate folder that contains one folder per wildfire labeled with the fire name. It 673 contains a polygon shapefile with the attributes listed in Table A3. 674 • Fire behaviour (L2): A single polygon shapefile that contains all wildfires and estimated fire behaviour metrics for 675 each individual fire spread polygon. The attributes are listed and explained in Table A4. 676 • Fire behaviour (L3): A single polygons shapefile that contains the simplified fire behaviour metrics calculated for 677 each burning period. The attributes are described in Table A5. 678

679
The generic metadata is connected to L1 data through the fire name field, and to L2 and L3 through the fire "ID" field. 680

681
The data are freely available at https://doi.org/10.5281/zenodo.7495506 (last access: 30th December 2022; Benali et al. 2022). 682 We intend to update the database annually with wildfires from the current fire season and implement continuous improvements 683 to the procedure. Also, if additional information from past wildfires becomes available, we will update the database either by 684 changing existing fire spread polygons or by adding new wildfires. Updates for future years depend on the availability of input 685 data and associated funding.  (Table A1, Table A2, Table A3, Table A4 and Table A5 near Table A1. Confidence flag value, class and interpretation. The flag is defined for each wildfire.

981
Flag value Flag Class Interpretation 1 Very Low The major fire progressions were observed only with satellite data, with important associated uncertainties.

Low
The major fire progressions were observed only with satellite data with moderate uncertainties 3

Moderate
The major fire progressions were observed with satellite data with low/moderate uncertainties and complemented with other sources.   Numerical link between a ignition or active flaming zone ("z") polygon and a wildfire progression ("p") polygon 1,2,3... -the link between types "p" and "z" with known dates and hours; 0 -used for type "a" or when progression in "uncertain" or when the link between "p" and "z" is unknown burn_period Burning period 1,2,3,..; 0 for the same cases as "zp_link".  Source of the data fserv -forest service ; sat -satellite data ; airb -airborne data; fops -fire personnel; ek -expert knowledge; rep -external reports zp_link Numerical link between a ignition or active flaming zone ("z") polygon and a wildfire progression ("p") polygon 1,2,3... -the link between types "p" and "z" with known dates and hours; 0 -used for type "a" or when progression in "uncertain" or when the link between "p" and "z" is unknown burn_period Burning period 1,2,3,..; 0 for the same cases as "zp_link".
area Burned area extent (ha) > 0 for progression polygons, -1 for ignition or active flaming zones.
growth_rate Fire growth rate (ha/h) >0 for progression polygons with zp_link value >0; -1 for previously burned areas or uncertain progression polygons ros_i Average rate-of-spread (m/h) calculated since ignition\active flaming areas or a progression marking the start of the burning period >0 for progression polygons with zp_link value >0; -1 for previously burned areas or uncertain progression polygons ros_p Parcial rate-of-spread (m/h) calculated between consecutive ignition\active flaming areas and progression polygon, or between two consecutive progression polygons >0 for progression polygons with zp_link value >0; -1 for previously burned areas or uncertain progression polygons spdir_i Spread direction associated with "ros_i" ( ° from North) 0 to 359.99; -1 for the same cases in "ros_i" spdir_p Spread direction associated with "ros_p" ( ° from North) 0 to 359.99; -1 for the same cases in "ros_p" duration_i Duration (hours) associated with the "ros_i" metric >0 known progression polygons; -1 for ignition\active flaming zones, previously burned áreas or uncertain progression polygons >0 for known progressions with at least 70% of FRE observations between "sdate" and "edate"; -1 for the remaining polygons FRE_flux Fire Radiative Energy flux (TJ ha -1 h -1 ) >0 for known progressions with at least 70% of FRE observations between "sdate" and "edate"; -1 for the remaining polygons FRE_perc Percentage of FRE observations between "sdate" and "edate" Between 0 and 100 for known progression polygons; -1 for the remaining.
* values will change when the database will be updated with new wildfires.  year Year 2015-2021* sdate Start date and hour of the burning period yyyy-mm-dd hh:mm; "na" for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) edate End date and hour of the burning period yyyy-mm-dd hh:mm; "na" for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) inidoy Start day-of-year of the burning period (hours in decimal values) 1 to 366; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) enddoy End day-of-year of the burning period (hours in decimal values) 1 to 366; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4)   qc  Confidence flag for each wildfire  See table A1 area Burned area extent (ha) >0 growth_rate Average fire growth rate (ha/h) >0; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) ros Average rate-of-spread (m/h) >0; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) max_ros Maximum rate-of-spread (m/h) observed in the burning period >0; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) spdir Spread direction associated with "ros_i" ( ° from North) 0 to 359.99; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) duration Duration (hours) of the burning period >0; -1 for burning periods which only have progression polygons with unknown "zp_link" (see Table A4) FRE Fire Radiative Energy (TJ) >0 for known progressions with at least 70% of the area burned during the burning period covered with FRE estimates; -1 for the remaining polygons FRE_flux Fire Radiative Energy flux (TJ ha -1 h -1 ) >0 for known progressions with at least 70% of the area burned during the burning period covered with FRE estimates; -1 for the remaining polygons FRE_perc Percentage of FRE observations between "sdate" and "edate" Between 0 and 100 * values will change when the database will be updated with new wildfires.