6 Comparison with datasets of burnt areas
6.1 Datasets and methods
As mentioned in Section 4.1, visual inspection of wildfires locations and findings from previous studies (Davies and Legg, 2016) have shown that it is unclear whether the location recorded in the IRS is where the fire was reported or where fire-fighting resources were marshalled instead of from the core or ignition point of the fire itself. In addition, our analysis showed extensive differences or disparities between recorded fuel type information from IRS Property Types and fuel types extracted from land use maps at wildfire locations. Moreover, while all wildfires have records of an estimated burnt area class (in m2) a limited number exists of wildfires where the burnt/damage area is also recorded in numeric terms (in ha); hence it is unclear how well the categorical and numeric estimates of burnt areas compare to each other and which of the two is closer to the actual burnt area.
For this purpose, in this section we provide a quantitative assessment of a) how close the recorded wildfire locations are to the actual burnt areas; b) how representative the recorded vegetation/habitats are of the main fuel types burnt and c) how close is the damage/burnt area reported in the IRS to the actual burnt area by respective fire incidents.
We used polygons of burnt areas in Scotland delineated from satellite imagery analysis to conduct the assessment. We obtained access to the following spatial datasets:
- Burnt area polygons (n=23) generated by NatureScot using Sentinel-2 multispectral satellite imagery (SNH, 2019) for wildfires that have occurred in 2018, 2019 and 2020.
- Burnt area polygons (n=101) generated by the European Forest Fire Information System (EFFIS; https://effis.jrc.ec.europa.eu/) using MODIS satellite imagery (areas >30ha) for the 2014-2020 period. Polygons from 2014 to 2019 have been previously obtained for the Scottish Fire Danger Rating System (SFDRS) project and were re-used in this project after gaining permission from EFFIS.
Both datasets were in ESRI shapefile format. Each burnt area polygon contained information about the location of the fire, date the fire occurred and/or detected and the size of the burnt area. Both polygon datasets were loaded into QGIS along with the locations the IRS wildfire incidents. There were a few overlapping burnt area polygons between the two datasets; when this was detected, we removed the EFFIS ones because the NatureScot dataset is of better spatial accuracy. We used a time window of ± 10 days around the IRS wildfire date and used these dates to associate individual IRS wildfire incidents with burnt area polygons. We then selected the wildfire incidents that were closer to the burnt area polygons via visual inspection of the satellite imagery and added the burnt polygon IDs to the attributes of associated wildfire incidents. This process showed that in cases of bigger fires, more than one wildfire incident was associated with a burnt area polygon, often with them having different dates.
We delineated additional burnt areas (n=23) to increase the number of wildfires analysed by mapping the burnt area of associated IRS wildfire incidents following a similar approach used by NatureScot. We chose to delineate new burnt areas that corresponded to wildfire incidents within the Highland LA because these were the wildfires more likely to be bigger, and hence easier to detect from satellite imagery, and were also more likely to have a significant impact on seminatural vegetation. In brief, new burnt areas were digitised using the following steps:
- We initially identified 88 IRS wildfire incidents (different to the ones corresponding to the NatureScot and EFFIS burnt area polygons) with good spatial coverage of the Highland LA that were likely to be big fires on peatland and shrubland, using fuel type information and burnt area estimates.
- We used the locations and recorded dates of the selected wildfire incidents to check whether relatively cloud-free Sentinel-2 satellite images were available from Sentinel Hub Playground (Access: https://apps.sentinel-hub.com/sentinel-playground) for the dates prior and after the wildfire incidents and visually inspected the available images in colour infrared (combination of bands 8, 4 and 3) to detect scars of burnt areas.
- We downloaded individual extracts of Sentinel-2 images from Sentinel Hub Playground in infrared band combination as jpegs, loaded them in QGIS and georeferenced them in OS British Grid using easily-identified control points. We then overlaid the wildfire incident locations and visually detected burnt areas that were in proximity to them, which were then digitised to create polygons of burnt areas (in ESRI shapefile format) (see Figure 6.1).
- We added the following information in the polygon attribute table: Incident Id of the corresponding wildfire incident; Unique polygon ID; Date of satellite imagery; and calculated burnt area (in ha). When more than one wildfire corresponded to the burnt area polygon, we added the Incident Id of the fire incident that was closest to the burnt area perimeter.
- Due to high cloud coverage, it was not always feasible to detect burnt areas corresponding to selected wildfires.
A total of 125 unique wildfire incidents were included in the analysis that corresponded to 71 burnt area polygons (16 from NatureScot, 38 from EFFIS and 17 delineated by us). This number of fire incidents was lower than anticipated considering the number of available polygons, but we found little correspondence mainly between wildfire incidents and the EFFIS burnt areas. The locations of the selected burnt areas are shown in Figure 6.2.
6.2 Distances to burnt area polygons
Table 6.1 shows the number of wildfire incidents within respective Local Authorities (LAs) and their nearest distances from the locations of wildfire incidents to the perimeters of the corresponding burnt area polygons, calculated using the Distance to Nearest Hub tool in QGIS. 104 of the 125 wildfires included in this analysis occurred within the Highland LA, while another ten (10) wildfires occurred in the Moray LA and just one (1) to three (3) wildfires occurred in the remaining LAs. Hence, we can be confident about the findings of this analysis only for those wildfires within the Highland LA. Median nearest distance to burnt area polygons for wildfires that occurred in the Highland LA was 260 m. There were 27 and 79 Highland wildfire incidents that were located less than 100m and 1,000m to the perimeter of their corresponding burnt areas, respectively, while there were 11 wildfire incidents located at distances more than 2km but only one at a distance more than 4km. Nearest distances to corresponding polygons were relatively small for the remaining wildfire incidents, with median distances ranging from just 4m to 389m, apart from the two wildfires in the Argyll and Bute LA that had a median distance of 1,741m.
|Argyll & Bute||2||239||1,741||1,741||3,244|
|Dumfries & Galloway||3||0||4||583||1,736|
The results of this analysis suggest that the location of wildfire incidents recorded in the IRS is relatively close to the actual corresponding burnt areas, at least for the wildfires occurring in the Highland LA, despite these wildfires generally occurring in very remote areas of low accessibility. However, these fires tend to be bigger, hence it is more likely for the recorded location to be closer to the fire perimeter. In fact, most of the wildfire incidents in the Highland LA (n=72) corresponded to burnt area sizes above 100 ha, while 39 and 14 wildfire incidents corresponded to burnt area sizes above 500 and 1,000 ha, respectively.
6.3 Land cover/fuel type
We intersected the burnt area polygons with the available land use and fuel type maps (EUNIS, LCS88, LCM15 and EFFIS FT) within QGIS and calculated the proportions of areal covers of each respective Broad Habitat (BH), along with the dominant (in terms of overall coverage) BH for each burnt area polygon. Overall, heathlands and shrub were the most frequent fuel type, followed by bogs and peatland. For example, based on the LCS88 classes 78 wildfires were on shrub and 38 on bogs and peatlands, whilst there were also eight (8) grassland fires and four (4) fires on conifer forest. Looking at the BHs of the associated wildfire incidents (classified from Level 3 Property Types), there were 103 wildfires on shrubland, 10 grassland fires and 8 fires on conifer forest.
The dominant BHs of the burnt area polygons were compared to BHs classified from IRS Level 3 Property Types for the associated wildfire incidents, excluding those wildfires that were classified as bogs and peatlands since the IRS property types do not provide this distinction. The results of this comparison are given in Table 6.2 and show very good agreement for wildfires where shrubland was the dominant fuel type, especially when the LCS88 and LCM15 classes were considered (94% and 92% agreement, respectively). There was moderate agreement for conifer forest fires and for grassland fires when LCM15 and LCS88 classes were used, respectively (50% agreement for both land use maps). Overall agreement was high for both the LCS88 and LCM15 maps (82% and 81%, respectively), but this was partially due to the dominance of wildfires on shrubland in the dataset and the exclusion of wildfires on bogs and peatlands from the comparison (most of which were recorded as shrubland fires in the IRS).
|Property Type BH||EUNIS BH||LCS88 BH||LCM15 BH||EFFIS FT BH|
6.4 Burnt area comparisons
We compared the calculated burnt area sizes of fire polygons with the burnt area estimates in the corresponding IRS wildfires recorded as a) damage area classes (available for all 125 wildfire incidents) and b) numeric damage area estimates in ha (available for 49 wildfire incidents).
Regarding damage area classes, the calculated burnt area of the corresponding polygons fell within the same area class for 51 wildfires, but this was not the case for 72 wildfire incidents, while there were two (2) wildfire incidents with no burnt area estimate recorded in IRS. 44 of the 51 wildfires with matching calculated and estimated burnt areas had occurred within the Highland LA.
There was a strong linear relationship between calculated burnt polygon areas and numeric damage areas for the 48 wildfires with both values available (Figure 6.3a, R2=0.70). However, this strong relationship was mainly driven by the big fire of March 2018 in the Highlands (burnt area of 6,275 ha). Removing this wildfire from the analysis resulted in a much weaker relationship (R2=0.26), due to large differences between calculated and estimated burnt areas for the wildfire incidents that had damage areas recorded the IRS mainly between 2 ha to 50 ha and in general less than 500 ha (Figure 6.3b).
Overall, the findings of this analysis suggest that in most cases the burnt area estimates recorded in the IRS fail to reflect accurately the extent of damage caused by the respective wildfires, while usually IRS records tend to underestimate the size of burnt areas and thus potentially underestimate the damage caused to impacted seminatural habitats (Figure 6.3).
There is a problem
Thanks for your feedback