Figure 2‑1 shows a visual comparison between the spatial patterns of active fire pixel count and FRP contained within the Level 3 Monthly Summary AF & FRP Night-time Products retrieved from one year’s Sentinel-3A and -3B data and from Terra MODIS. The products cover the period March 2020 to February 2021, and very similar spatial patterns are seen, indicating a broad degree of agreement despite the MODIS data including all night-time observations and not only those collected near-simultaneously with SLSTR.
The AF pixel count data show that the SLSTR products include many more AF pixel detections than MODIS, but the grid-cell FRP totals shown in Figure 2‑32 (Section 2.5) are similar between the two records. This is because the additional AF pixels that SLSTR detects in many of the grid cells are dominated by low FRP values. These findings mirror those reported in Xu et al. (2020) – in that the recorded FRP is similar between SLSTR and MODIS but that SLSTR can detect smaller fires due to the AF detection algorithm sensitivity and the smaller mean SLSTR pixel size across the swath compared to MODIS.
Figure 2‑1: Total active fire (AF) pixel count and total FRP of fires detected within 0.25° grid cells from March 2020 to February 2021 using all Sentinel-3A, -3B SLSTR and Terra MODIS data. Note the Terra MODIS active fire pixel count has a scale that is five times smaller than that of Sentinel-3A and -S3B due to the lower number of AF pixels the former sensor typically detects.
The AF detection process is inherently a trade-off between attempting to detect the lowest FRP fires (which are typically the most common type in most geographic areas), whilst also minimising false alarms (i.e., the triggering of the AF detection algorithm by a non-fire phenomena). Raising the minimum FRP detection limit in the contextual AF detection algorithm of Xu et al. (2020) very likely reduces the number of false alarms, but will also prevent some lower FRP fires from being identified. Comparison to global data from Terra MODIS taken within a scan angle limit of 30° and within ±6 minutes of a Sentinel-3 acquisition (Xu et al., 2020) shows that 90% of MODIS-identified active fire pixels (MOD14 Collection 6 product) had a matching Sentinel-3 AF pixel detection, representing an apparent SLSTR product omission error compared to MODIS of 10%. Conversely, of the Sentinel-3 AF pixel detections present in the same dataset, only 56% matched MODIS AF pixel detection, and 79% of the additional detections made by SLSTR had an FRP of less than 5 MW. The additional SLSTR AF detections are mostly below the minimum MODIS FRP detection limit, but SLSTR has a slightly lower minimum FRP detection limit than MODIS due to the smaller mean SLSTR pixel area and the intricacies of the AF pixel detection algorithm used to generate the Sentinel-3 AF products (Xu et al., 2020). Work is ongoing with the Sentinel-3 NTC AF Detection and FRP products to understand how many of these additional SLSTR AF detections are ‘true’ active fire pixels, but it is expected that the vast majority are. Using one month’s C3S Level-2 Summary Product file (August 2020) from Sentinel-3B, we found that the C3S level-2 summary product shows that 95% of MODIS-identified active fire pixels (MOD14 Collection 6 product) had a matching Sentinel-3 AF pixel detection, representing a SLSTR product omission error of 5%. Conversely, of the Sentinel-3 AF pixel detections present in the same dataset, only 63% had a matching MODIS AF pixel detection. In both cases, the SLSTR and MODIS data were limited to 30 view zenith angle and 6 mins time difference from the matching SLSTR-MODIS observations. These detection performances further indicate once again that in general SLSTR is slightly more sensitive to small (i.e., low FRP) fires than is MODIS, as detailed in Xu et al. (2020).
Figure 2‑2a directly compares the FRP of fire clusters (each comprising a set of spatially discrete AF pixels) imaged near simultaneously (within 6 minutes) by SLSTR on Sentinel-3B and by Terra MODIS in January 2019. The MODIS and SLSTR data have a view zenith angle maximum of 30°. The comparison indicates a strong degree of agreement (r2=0.91), particularly considering that Freeborn et al. (2014a) demonstrated a one per-pixel MODIS FRP uncertainty of ±27%, based only on variability in the sub-pixel location of the fire itself. The slope of the ordinary least squares (OLS) linear best fit exceeds 1.0 (at 1.08) primarily because SLSTR quite often detects some low FRP AF pixels at the edge of an active fire cluster that MODIS does not detect. Thus, the former sensor can often provide a slightly higher FRP measurement for the same AF cluster. Similarly, Figure 2‑2b shows the comparison from matching regions observed near-simultaneously by SLSTR on Sentinel-3B and Terra MODIS. Each measurement shown represents the total FRP of the region in August 2020. Again, Figure 2‑33b (Section 2.5) indicates a strong degree of agreement (r2=0.95) and a slope of 1.16 due to SLSTR detecting some low FRP AF pixels that MODIS does not detect.
The C3S Level 3 Gridded FRP Products are intended primarily for large scale analysis of fire patterns, seasonality, anomalies and trends. Such characteristics are part of a regional ‘fire regime’, which describes the role of landscape fires within an area over time, and under this broad definition, the fires’ physical attributes such as frequency, size, intensity, type and seasonality. Fire regimes may change with changing climate and with other changes in human impact associated with, for example, land use management and land use change (Moritz 2012; Flannigan et al., 2009; Archibald et al., 2013; Hantson et al, 2015). Characterizing past and current fire regimes has historically been performed by analysing field data such as charcoal records, fire-scar networks and fire occurrence databases. However, the regular and continuous information on landscape fires that can be provided by EO satellites is increasingly being used to determine certain fire regime characteristics (e.g., Chuvieco et al., 2008; Freeborn et al., 2014). Seasonality is one of the most important component characteristics of a region fire regime, and one which products such as the C3S Level 3 gridded FRP and the MODIS Active Fire Products are well suited to determine. To evaluate the C3S Level 3 Gridded FRP Products we therefore used them to derive fire regime seasonality characteristics, and compared these to the same metrics derived from the MOD14 data used to create the similar MODIS MOD14CMQ and MOD14C8Q climate modelling grid (CMG) products (Justice et al., 2002).
Figure 2‑2: Inter-comparison of night-time global FRP records obtained from Sentinel-3B SLSTR and Terra MODIS within ± 6 minutes of each other. (a) Fire cluster-based comparison. Each measurement is for a single fire cluster in January 2019, representing a spatially contiguous set of active fire pixels. Both datasets are atmospherically corrected. (b) Region-based comparison. Each measurement shows total FRP from the regions detected at almost the same time by both SLSTR and MODIS in August 2020.
The fire season metrics derived from the C3S Level 3a Daily Gridded AF & FRP Night-time Product files were compared to those derived from MOD14 data gridded to the same 0.1° spatial resolution grid. The comparisons are made globally as well as within the geographic regions of the Global Fire Emission Database (GFED) (e.g., see Figure 1‑6). Comparisons were made in terms of AF detections but also with FRP (as in Figure 2‑1). The 14 GFED regions are defined based on containing similar biomass burning regimes rather than on country borders (Van der Werf et al., 2017).
Figure 2‑3: Daily active fire count and FRP comparison between SLSTR and MODIS. (a) Total global daily FRP; (b) Cumulative percentage of total global daily FRP; (c) Total global daily active fire count; (d) Cumulative percentage of total global daily active fire count. Notice there are a few days in May and November 2020 for which SLSTR has no input Level-2 FRP data.
As an example at the global scale, Figure 2‑3a shows daily global total FRP data as derived from Sentinel-3A, -3B SLSTR and Terra MODIS daily gridded global products. All three show a very similar temporal development. At the global scale, the fire season peak occurs in September, starts in May and ends in November for each of S3A, S3B and MODIS. The cumulative percentage in Figure 2‑3b shows a similar trend. However, if we define the start of the fire season as 10% of the total FRP and end of the fire season as 90% of the total FRP (as in Freeborn et al. 2014b), the start of the fire season occurs in May 2020 for all three satellites. However, the end of the fire season occurs in November 2020 according to MODIS but December according to SLSTR. The peak, as well as the duration, of the global “fire season” agrees well with results from Giglio et al. (2006).
The AF count data shows a similar seasonal pattern to FRP, but the SLSTR AF count is ~ 5 times higher than the MODIS AF count due to the former’s greater sensitivity to smaller fires (i.e., those with lower FRP) (Figure 2‑3c). Again, if we define the start of the fire season as 10% and end of the fire season as 90% of the total AF pixel count, the start of the fire season differs in MODIS and SLSTR, being in April 2020 for MODIS, and in May 2020 in the case of SLSTR. Similarly, the end of the fire season will be in December 2020 according to MODIS, but for SLSTR the end of the fire season is one month later, in January 2021. Since the MODIS and SLSTR FRP records are more similar than the AF pixel count records, due to the extra AF pixels detected from SLSTR being mostly low FRP, and the FRP patterns are more in agreement with the seasonality patterns seen in Giglio et al. (2006), we focus on the FRP for fire season analysis from here on.
Figure 2‑4: Peak timing of the fire season as determined from Level 3a Daily Gridded FRP Product of S3A and S3B, and from daily Terra MODIS data. Temporal resolution of the data is one day. The abbreviation of the names is as follows: Boreal North America (BONA); Temperate North America (TENA); Central America (CEAM); NH South America (NHSA); SH South America (SHSA); Europe (EURO); Middle East (MIDE); NH Africa (NHAF); SH Africa (SHAF); Boreal Asia (BOAS); Central Asia (CEAS); SE Asia (SEAS); Equatorial Asia (EQAS); Australia and New Zealand (AUST).
Figure 2‑4 shows the peak month of the fire season both globally and in the GFED regions of Figure 1‑6, as derived from SLSTR and from Terra MODIS. Globally, the three satellite datasets (S3A, S3B and Terra MODIS) agree very well: for example, they all peak at the beginning of September 2020. For the GFED regions, 10 out of the 14 regions show a very good agreement, with the three satellites showing a fire season peak that is identical to within a few days. Note that landscape fire is dynamic phenomena having a high daily variance, and further analysis is needed to study the difference for regions showing greater disparity between each of the datasets.
Figure 2‑5 and Figure 2‑6 (image has been divided into two distinct parts to enhance user accessibility and clarity) show the seasonal pattern of daily cumulative FRP for both the globe and for the 14 GFED regions, again as derived from each of the three satellite datasets. The data from SLSTR shows a very strong agreement with that from MODIS in almost all regions, despite these figures being based on all Terra MODIS data and not just that collected near-simultaneously with SLSTR. Globally the coefficient of variation (r2) between the data from S3A and from Terra MODIS is 1.00, and the slope of the OLS linear best fit is 1.07. For S3B and Terra MODIS the values are 1.00 and 1.08 respectively. For the GFED regions, the coefficient of variation (r2) between S3A and MODIS ranges from 0.77 to 1.00 and the slope from 0.79 to 1.15. Between S3B and MODIS the coefficient of variation (r2) ranges from 0.85 to 1.00, and the slope from 0.78 to 1.13.
Figure 2‑5: Seasonal pattern of cumulative monthly FRP (%) at the global and GFED region scale, as derived from the C3S Level 3 Daily FRP Product and Terra MODIS - Part a. The abbreviation of the names is as follows: Boreal North America (BONA); Temperate North America (TENA); SH South America (SHSA); Europe (EURO); Middle East (MIDE); Boreal Asia (BOAS); Central Asia (CEAS); SE Asia (SEAS).
Figure 2‑6: Seasonal pattern of cumulative monthly FRP (%) at the global and GFED region scale, as derived from the C3S Level 3 Daily FRP Product and Terra MODIS - Part b. The abbreviation of the names is as follows: Central America (CEAM); NH South America (NHSA); NH Africa (NHAF); SH Africa (SHAF); Equatorial Asia (EQAS); Australia and New Zealand (AUST).
The C3S Level 3a 27-Day Gridded AF & FRP Night-time Product is simply the accumulation of 27 C3S Level 3a Daily Gridded AF & FRP Night-time Products, so its evaluation will simply focus on verifying the correctness of the lower temporal resolution statistical summary derived from the former data. The 27-Day products were verified thoroughly in this way, with the active fire pixel count, FRP and all other parameters being found to agree with the accumulation of the daily product data. This indicates the correctness of the C3S 27-Day products. An example is shown in Figure 2‑7 where it is shown that the active fire pixel count of the 27-Day product from 29 May to 24 June 2020 perfectly agrees with that derived from the accumulation of the 27 individual daily products.
Figure 2‑7: Example of verification of the Level-3a 27-Day gridded AF & FRP Night-time product. (a) Total active fire pixel count from the C3S 27-Day gridded FRP product covering 29 May to 24 June 2020; (b) Total active fire pixel from the 27 daily C3S daily products covering the same period. The active fire pixel count is identical in all the grid cells between (a) and (b).
The C3S Level 3 Monthly Summary AF & FRP Night-time Product was compared to the MODIS MOD14 product. As the spatial patterns of absolute AF pixel counts and FRP have been analysed already in Section 2.3.1, we focus here on the analysis of fire season and metrics defined at the global and GFED region scale. Degree of agreement between the C3S and MODIS products are quantified using the coefficient of determination (r2) and the slope of the ordinary least squares (OLS) linear best fit.
Figure 2‑8 and Figure 2‑9 (image has been divided into two distinct parts to enhance user accessibility and clarity) show the seasonal pattern of cumulative monthly FRP at both the global and GFED region scale, from both the C3S FRP products and Terra MODIS. As with the C3S daily product (Figure 2‑5 and Figure 2‑6), the seasonal pattern from SLSTR appears very close to that derived from MODIS. The r2 value of the cumulative FRP formed by the S3A and Terra data is 1.00. The slope of the linear best fit is 1.06, whilst that from S3B and Terra is 1.00 and 1.07 respectively. For the GFED regions, the r2 between the S3A and MODIS values range from 0.76 to 1.00 and slopes 0.60 to 1.14, whilst S3B and MODIS are 0.83 to 1.00 and 0.69 to 1.12.
Figure 2‑8: Seasonal pattern of cumulative monthly FRP (%) at the global and GFED region scale, as derived from the C3S Level 3 Monthly Summary FRP Product and Terra MODIS Part a. The abbreviation of the names is as follows: Boreal North America (BONA); Temperate North America (TENA); SH South America (SHSA); Europe (EURO); Middle East (MIDE); Boreal Asia (BOAS); Central Asia (CEAS); SE Asia (SEAS).
Figure 2‑9: Seasonal pattern of cumulative monthly FRP (%) at the global and GFED region scale, as derived from the C3S Level 3 Monthly Summary AF & FRP Night-time Product and Terra MODIS Part b. The abbreviation of the names is as follows: Central America (CEAM); NH South America (NHSA); NH Africa (NHAF); SH Africa (SHAF); Equatorial Asia (EQAS); Australia and New Zealand (AUST).
Figure 2‑10 shows the start, peak, end and duration of the fire season as derived from the C3S Level 3 Monthly Summary FRP Product and from Terra MODIS, both globally and in the GFED regions. We used 25% of the total FRP to define the beginning of the fire season, and 75% to define the end of the fire season. Overall, the start, peak, end and duration of the fire season agree very well between S3A, S3B and MODIS. Globally, the maximum difference between each of these metrics for each satellite is one month. For all the GFED regions, 50% of the 14 regions have 100% agreement for all the four fire season metrics analysed, 70% have a difference of one month, and almost all the regions have a difference of less than two months. Some of these differences are caused by regions like the Middle East, which do not have a marked fire seasonality. Another reason is that SLSTR is typically more sensitive to smaller fires than MODIS at night, and thus it may record an earlier start to the fire season, as in Europe, compared to MODIS. Compared to the daily product (Figure 2‑4), the monthly temporal resolution of the C3S Level 3 Monthly Summary FRP Product means that the minimum level of agreement (apart from zero) is 1 month.
Figure 2‑10: Fire season metrics as determined from Level 3 Monthly Gridded AF & FRP Night-time Product of S3A and S3B and from daily Terra MODIS data. The temporal resolution of the data is one month and analysis is conducted globally and for the GFED regions. (a) Start of the fire season; (b) Peak of the fire season; (c) End of the fire season; (d) Duration of the fire season. The abbreviation of the names is as follows: Boreal North America (BONA); Temperate North America (TENA); Central America (CEAM); NH South America (NHSA); SH South America (SHSA); Europe (EURO); Middle East (MIDE); NH Africa (NHAF); SH Africa (SHAF); Boreal Asia (BOAS); Central Asia (CEAS); SE Asia (SEAS); Equatorial Asia (EQAS); Australia and New Zealand (AUST).
This document has been produced in the context of the Copernicus Climate Change Service (C3S). The activities leading to these results have been contracted by the European Centre for Medium-Range Weather Forecasts, operator of C3S on behalf of the European Union (Delegation Agreement signed on 11/11/2014 and Contribution Agreement signed on 22/07/2021). All information in this document is provided "as is" and no guarantee or warranty is given that the information is fit for any particular purpose. The users thereof use the information at their sole risk and liability. For the avoidance of all doubt , the European Commission and the European Centre for Medium - Range Weather Forecasts have no liability in respect of this document, which is merely representing the author's view. |