Uncertainty assessment of remote sensing- and ground-based methods to estimate wildfire emissions: a case study in Calabria region (Italy)

In a general framework characterized by ever-increasing evidence of impacts attributable to climate change, the quantitative estimation of wildfire emissions (e.g., black carbon, carbon monoxide, particulate matter) and the evaluation of its uncertainty are crucial for mitigation and adaptation purposes. Global atmospheric emission models use mainly remote sensing fire datasets, which are affected by significant uncertainties. To assess the errors of remote sensing-based inventories, we compared the temporal and spatial behavior of the last version of the satellite-based Global Fire Emissions Database (GFED4s) with a more accurate ground-based wildfire emissions inventory, for the 2008–2016 period. The study area was Calabria (southern Italy), among the Italian regions with the highest contribution to national wildfire emissions. This study highlights a reliable agreement of time evolution of Burned Areas (R2 = 0.87), but an overestimation of their extent by satellite compared to ground observations (approximately + 18%). Nevertheless, satellite data systematically underestimated Dry Matter and emissions by forest and grassland wildfires (ranging between -66% and -97%). Furthermore, detailed information on land cover allowed assessing the vegetation parameters uncertainties on ground-based emission inventory. The Mass Available Fuel values, which are constantly modified by wildfires, and land use changes, and not frequently updated, showed not to affect the emission estimations. Finally, the relationship between ground-based and remote sensing-based inventories for the analyzed period highlighted that the preliminary satellite emissions related to 2017–2019 require careful validation before any applications.


Introduction
Global biomass burning (BB) produced by wildfires is one of the most critical sources of gaseous pollutants and particulate matter (Andreae and Merlet 2001;Akagi et al. 2011). As an example, BB produces black carbon (BC) and carbon monoxide (CO) globally in the atmosphere for 59% and from 33% to 50%, respectively Bond et al. 2013). In the Mediterranean region, wildfires mainly human-caused (Ruffault and Mouillot 2017) represent a troubling hazard that yearly causes Jessica Castagna jessica.castagna@unical.it Extended author information available on the last page of the article. human deaths and significant ecological and economic losses (Galizia et al. 2021) due to a considerable extension of Burned Areas of approximately 400,000 ha per year in the last four decades (Vallejo Calzada et al. 2018). Moreover, according to the future projections, the heatinduced fires are expected to grow significantly (Ruffault et al. 2020) because of the changing climatic conditions, entailing increasing heatwaves and drought events (Turco et al. 2017;Turco et al. 2018;Senatore et al. 2022).
BB emissions estimations are primary for chemical and transport models regarding climate-impact and air-quality studies. Global and local emission inventories are still affected by significant uncertainties, although numerous studies and improvements have been performed in the last years (Van der Werf et al. 2010;Chiriaco et al. 2013;Van Der Werf et al. 2017). Despite the satellite uncertainties, in the last decades, the scientific community preferred remote sensing data to study global biomass burning due to their capability of providing global coverage (Turco et al. 2019a;Pan et al. 2020). Indeed, although a very high spatial resolution characterizes ground-based data, they show a limited spatial coverage and a lack of consistent qualitycontrol procedures (Turco et al. 2019a).
The remote sensing datasets inventories are divided into two main approaches: Burned Areas-based and Fire Radiative Power-based (Pan et al. 2020). For example, Global Fire Emissions Database (GFED) ( Van der Werf et al. 2006;2010;Van Der Werf et al. 2017) and Fire INventory from NCAR (FINN)  are the main global BB emissions dataset based on Burned Areas, while Global Fire Assimilation System (GFAS) (Kaiser et al. 2012) and Quick Fire Emissions Dataset (QFED) (Darmenov and da Silva 2015) are based on the Fire Radiative Power.
One of the most applied remote sensing inventories is the Global Fire Emissions Database -GFED (Giglio et al. 2013;Randerson et al. 2015;Van Der Werf et al. 2017), which is based on the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor. Remote sensing is limited to the satellite passage that photographs the presence of a Burned Area over a pixel. Besides, it is unable to establish the start/end time of the fire, its duration, the fire regime, and the final extent of each event (Mouillot et al. 2014). These uncertainties concerning fire detection propagate on the satellite-based emission inventory. This work aims to inform end-users of a broadly applied remote sensing-based (i.e., the last version of GFED, called GFED4s) and a ground-based emission inventory about the uncertainties of emissions, generally used as input of atmospheric models. In particular, the case study concerns the Calabrian peninsula (Italy), located in the central Mediterranean basin, from 2008 to 2019. The ground-based Burned Area's high spatial resolution and the territory's detailed knowledge allowed a more in-depth analysis of the various sources of uncertainty. Specifically, we assessed Burned Area's temporal and spatial consistency and the agreement in time between the Dry Matter and different emission estimations.

Relevance of wildfire emissions in the study area
The investigated area is a region of southern Italy (Calabria) (Fig. 1c), which has an extension of 15,080 km 2 with rugged topography consisting of 9.0% lowland, 49.2% upland, and 41.8% mountains. The region displays marked disparities in environmental conditions concerning climate regimes, landscape, vegetation, soils, and cropping systems. The climatic extremization trend affecting the Mediterranean region is particularly evident in Calabria (Mendicino and Versace 2007;Senatore et al. 2020). Longterm meteorological observations evidenced drier summers, primarily due to temperature rise and precipitation reduction (see Fig. 1a, b from http://interactive-atlas.ipcc.ch/, last access on 29th of November 2022), which led to a higher incidence of fire activities (Bedia et al. 2014;Bencardino et al. 2019;Castagna J et al. 2021a;2021b) with a local and transboundary impact of BB emissions (Castagna et al. 2019).
To assess the wildfires issue in Calabria, we considered regional emission amounts, provided by the Italian emission inventory at the local scale and estimated by the Italian Institute for Environmental Protection and Research (ISPRA -http://emissioni.sina.isprambiente.it/ serie-storiche-emissioni/, last access on 29th of November 2022), from Forest Land (FL), Other Wooded Land (OWL), which is largely diffused in this territory, and Grass-Land without OWL (GL*), for 2010, 2015, and 2019, as years included in our investigation period (2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019). This inventory highlights that, in the investigated period, Calabria contributed heavily to the total national amount of pollutant emissions due to FL, GL*, and OWL subcategories (see Figs. 8 and 9 in the Appendix). The Calabrian percentage contribution to the national amount of pollutant emissions linked to FL, OWL, and GL* sectors varied from 10%, 7% and 7%, respectively, in 2010, to greater values observed in 2015 (28% for FL, 20% for OWL, and 13% for GL*) and in 2019 (22% for FL, 27% for OWL, and 20% for GL*). Differences observed among regional contributions can be attributed to different investments in fire management, extinguishing plans, as well as agricultural abandonment, reforestation, and reduction of forest area (Turco et al. 2019b). The latent interplay between socio-economic development and environmental conditions has been frequently investigated in multifaceted research fields dealing with climate change, land use, soil degradation, and other issues involving global, regional, and local spatial scales (Salvati and Zitti 2005;Harte 2007;Zambon et al. 2018;Maletta and Mendicino 2020). In southern Italian areas, for example, where agriculture is a principal economic activity, and the land is highly exploited, the fire causes are strongly related also to socio-economic factors (Masala et al. 2012). The poor development of industrial and socio-economic activities in Calabria may explain the relevant emissions contribution produced by FL, OWL, and GL* compared to all contributing macro sectors for this region.
For this study, we chose air pollutants with the largest contribution from wildfires compared to the total production from the remaining sources. Figure 2 reports the total emission amounts and the main air pollutants, highlighting (E-OBS) daily gridded dataset, which is a European land-only, highresolution gridded observational dataset produced using the blended daily meteorological station data from the European Climate Assessment and Dataset (ECA&D) (available at http://interactive-atlas.ipcc. ch/). c) Map of the Calabrian topography highlighting areas with revised biomass volume (Scrinzi et al. 2017) the emission contribution attributed to only FL, OWL, and GL* subcategories (FL+OWL+GL*) and from All Sources Activities without FL+OWL+GL* (ASA*). The percentage contributions from FL+OWL+GL* greater than 10% were evidenced in this graph, showing a relevant incidence especially for carbon monoxide (CO) and black carbon (BC), with values equal to 36% and 48%, respectively, during 2015, while they accounted for 29% and 41%, respectively, in 2019. The subsequent most essential contributions regard sulfur dioxide (SO 2 ) and fine particulate matter (PM 2.5 ), with percentage values both equal to 21%, in 2015, while in 2019, PM 2.5 had a slightly higher value of 17%, compared to the value of 15% registered for SO 2 . Among these pollutants with a relevant contribution, we selected those with an important impact on climate change and human health, i.e., CO, PM 2.5 and BC.

Ground-based emissions
The ground-based emissions (GBE) estimations are based on the ground-based dataset (GBD) of Burned Area collected by a fire agency. In the Calabria region, the task of fires census is entrusted to the "Comando Unità Forestali, Ambientali e Agroalimentari dei Carabinieri". In particular, after fire extinguishment, the task of fires collects the fire information (e.g., start and end date, location, Burned Area with vegetation features). Despite some errors due to unavoidable inaccuracies (e.g., errors in reporting the ignition start time and the fire intensity), the GBD is considered a dataset with a reasonable confidence level for fires in FL and GL (Chiriaco et al. 2013;Galizia et al. 2021). The native spatial resolution of GBD is very high due to the sampling of Burned Area perimeter through GPS. In addition to ground-based Burned Area concerning the Calabria region, we analyzed the corresponding groundbased Dry Matter. GBE are provided by the Italian greenhouse gases (GHG) inventory emissions, within the Framework Convention on Climate Change. This inventory includes emissions from fires occurring on FL and GL subcategories, estimated by ISPRA with the For-fires model (ISPRA 2022), a model based on the approach developed by Bovio (2007), amount, originating from all the source activities resulting in pollutant emissions, is also evidenced. ASA* stands for All Sources Activities without FL+OWL+GL* subcategories following the IPCC Guidelines (Aalde et al. 2009). Among GBE, we considered Carbon (C), CO, PM 2.5 , and BC for our case study.
At first, the Mass of Available Fuel for combustion is calculated based on biomass volume from the Italian National Forest Inventory (INFC 2022). However, the Mass of Available Fuel is not frequently updated, although it varies due to tree growth, effects of wildfires, and land use change. For this reason, Scrinzi et al. (2017) revised the 2005 INFC data through a LIDAR campaign. The differences between INFC and the more recent parameters range from approximately -58% to +106% (see Table 4 in the Appendix). The updated parameters were estimated mainly through the observations of the beech forests in the Pollino area, the chestnut coppices in the Coastal Chain, the pine forests in Sila, and the Abetine in Serre (Fig. 1c). To assess the emission uncertainty due to the biomass volume errors, we re-calculated the emissions for the period going from the observation date (2017) onwards (2019) using the updated Mass of Available Fuel for the observed areas.
The For-fires model was applied to estimate FL and GL for 2008-2019 with INFC parameters, and for 2017-2019 with both biomass volume parameters.

Remote sensing-based emissions
Regarding the Remote Sensing Emissions (RSE), we used a widely applied global Remote Sensing Dataset (RSD) for fires by modelling communities, that is the last version of the Global Fire Emissions Database (GFED) (Pan et al. 2020), called GFED4s (Van Der Werf et al. 2017). The remote sensing dataset is based on the Burned Area estimated from active fires detection during the satellite transit (Giglio et al. 2013). Unlike the previous version, the GFED3, characterized by a spatial resolution of 0.5 • , the GFED4s version, having a resolution of 0.25 • , includes the detection of small fires (< 100 ha) that enhance the considered Burned Area, given by the product MCD64A1 provided by MODIS.
The validated GFED4s products are available at the link http://www.globalfiredata.org (last access on 29th of November 2022) for the period included between 1997 and 2016, after which only preliminary emission results are available. Using GFED4s data in hdf5 format, we calculated Burned Area, Dry Matter, C, CO, PM 2.5 , and BC for the Calabria region. While Burned Area is reported as aggregated data, the contribution of Dry Matter and emissions is disaggregated depending on the land cover of the 500 m resolution MODIS Collection-5 product (MCD12Q1). According to the land cover of our case study, we considered the SAVA (savanna, grassland, and shrubland fires) and TEMF (temperate forest fires) classes.

Methodology
We assessed the ability of remote sensing-based data to simulate the ground-based data, representing the most accurate data available. At first, we compared remote sensing Burned Area, Dry Matter, and emissions, and the corresponding ground-based values calculated with INFC volume on the common period available for this case study (i.e., from 2008 to 2016). In particular, we analyzed the temporal evolution through least-squares linear regressions and different metrics, such as Pearson correlation and the mean relative error (MEr): While for Burned Area we considered all types of vegetation burnt, for Dry Matter and emissions we analyzed FL and GL for GBE corresponding to TEMF and SAVA for RSE. We performed the spatial assessment for Burned Area using Qgis Software. To compare the two datasets spatially, showing different spatial resolutions, we reduced the GBD data to the same resolution of RSD (0.25 • ), aggregating the GBD Burned Area into the same grid of the satellite product.

Temporal and spatial correlation between GBD and RSD burned area
Annual Burned Area estimated by both GBD and RSD shows a similar temporal evolution (R 2 of 0.87) with an almost systematic overestimation of RSD, except for 2009, 2015, and 2016 (Fig. 3). While the year with the maximum Burned Area corresponds to 2012 for both methodologies (22,144 ha and 20,967 ha for RSD and GBD, respectively), the minimum Burned Area is recorded in different years: 2655 ha in 2013 for GBD and 3326 in 2015 for RSD, respectively. The GBD-based Burned Area yearly mean values are lower than RSD (Table 1). The overestimation calculated in our case study (MEr = +18%) is lower than the results of a study performed for the period 2001-2011 on a continental (European) scale (Turco et al. 2019a), which calculated an overestimation of +56%. This result reflects the typical RSD overestimation due to the cropland fires in Europe detected by remote sensing but generally not reported by groundbased agencies (Giglio et al. 2013). Despite the missing cropland fires in GBD-based Burned Area, the GBD method is more reliable compared to the RSD data, affected by detection problems (e.g., fire extension, fire numbers) that cause an underestimation in comparable (non-cropland) land covers (Galizia et al. 2021;Turco et al. 2019a).
The comparison between the GBD and RSD methodologies highlights a different spatial distribution for the Burned Area values. While for GBD, the largest Burned Area is placed in the north of the Calabria region, for RSD, it is shifted eastward (Fig. 4a, b). Additionally, the GBD and RSD Burned Area annual trends show a low spatial  correlation, while the MEr calculated over the whole period is significantly high (Fig. 4c, d). In particular, satellite revealed difficulties of detection along the coasts as shown by recurring annual null Burned Areas in the coastal pixels. This satellite deficit may limit its use in similar areas surrounded by sea. At the GFED4s spatial resolution (equal to 0.25 • ), the spatial intercomparison emphasizes the low RSD ability to simulate the GBD spatial patterns. Generally, reducing the spatial resolution (i.e., averaging values over larger areas), better correlations are obtained (Turco et al. 2019a) as shown by the R 2 in Fig. 3 calculated for the whole region.

Comparison of GBD and RSD dry matter and emissions
After the Burned Area comparison, we analyzed the time evolution of Dry Matter related to TEMF and SAVA for RSD, whereas FL and GL for GBD. The comparison highlights, as opposed to Burned Area, an underestimation by remote sensing systems (Fig. 5). This underestimation is  Table 1 (MEr = -71%). The correlation between Dry Matter values by GBD and RSD is stronger than Burned Area analysis (R 2 = 0.96). The maximum Dry Matter occurred in the same year for both, which was 2012, with 386 kt and 901 kt for RSD and GBD, respectively, while the minimum values were recorded in 2013 with approximately 24 kt and 94 kt for RSD and GBD, respectively (Fig. 5). The temporal evolution of the annual estimate of GBE and RSE shows a similar trend to Dry Matter due to the calculation method. The RSE trend shows an excellent ability to describe the temporal variation, as confirmed by the high correlation for all emissions (Fig. 6). Nevertheless, similarly to the Dry Matter case, the RSE estimations are significantly underestimated compared to GBE values, as showed by negative values of the average yearly MEr (Fig. 7). This underestimation is even more pronounced for CO and BC, which reveal a significant difference in absolute values (i.e., orders of magnitude) ( Table 1).

Uncertainty assessment of ground-based and remote sensing methodologies
Although we considered the ground-based method as a reference, both methodologies are affected by several errors that should be evaluated while using these estimations, especially for modelling purposes. Regarding the GBE, the possible sources of errors could be: i) neglecting tiny fires (< 1 ha), ii) errors of fire agency personnel in recording or reporting the Burned Areas and their vegetation types, iii) the uncertainties of applied parameters, such as the Mass of Available Fuel, and iv) the combustion factors (depending on the scorch height). On the other hand, the RSE uncertainties are linked mainly to spatial resolution, fire detection, and the ability to distinguish the vegetation type.
Both methodologies are affected by uncertainties of Emission Factors (EFs) parameters, which were extracted from the recommendation of EMEP /EEA 2009(San-Miguel-Ayanz et al. 2009), based heavily on Andreae and Merlet (2001) and the updated values revised by Akagi et al. (2011). These EFs could be updated in agreement with a recent work of Andreae (2019) that revised their values according to 370 published studies and underlined their significant-high uncertainty due mainly to the specific burning conditions causing a large variability of emissions.

Managing error source of the ground-based methodology: mass of available fuel
We analyzed the effect of the Mass of Available Fuel (MB) parameters on the uncertainty associated with the emissions related to all the investigated pollutants. The MB parameters are subject to evolution in time due to trees' growth, wildfires, and changes in land use. As described in the paragraph about the methods ("Ground-based emissions"), the suggested parameters, reported by INFC and involved in international reports performed by ISPRA, were compiled in 2005. Following studies showed updated parameters (Scrinzi et al. 2017) that should be used. We analyzed the effects of these parameters on emissions related to the revised areas (highlighted in Fig. 1c) for the period after the study by Scrinzi et al. (2017), i.e., 2017-2019. In Table 2 we show the results for C, CO, PM 2.5 , and BC, with INFC volume (V) and updated volume (V*), and the resulting MEr (which is equal for all considered emissions due to their proportionality to C). Despite the significant uncertainty in the Mass of Available Fuel deriving from different estimated volumes, the emission errors are relatively low (Table 2). In particular, the maximum bias was achieved in 2017, a record year for fires in this region (GBD Burned Area equal to 30,359 ha). Therefore, it can be inferred that the application  Table 2 Early emissions of C (kt), CO (kt), PM 2.5 (kt), and BC (kt) related to the revised areas shown in Fig. 1   of not updated parameters is not the major drawback to the final estimations.

Sources of errors for remote sensing methodology and data validation
The primary source of uncertainty in remote sensing emissions methodology is given by the spatial and temporal satellite resolution. The fire detection was limited to the satellite overfly, the scan angle, and atmospheric condition (Giglio et al. 2013;Saide et al. 2015;Wang et al. 2018). Additionally, the satellite's inability to define the burning stage of fire increases the uncertainty of emission estimates, which are heavily dependent on the flaming or smouldering stage (Pan et al. 2020). Finally, the lower spatial resolution than GBE and the active biome types enhance the uncertainty estimates. The GFED4s RSE are available until 2016. After that, preliminary emissions are available for 2017 onwards because the complex fire count and Burned Area validation system require a long process. Table 3 shows the estimations of C, CO, PM 2.5 , and BC emissions for the common period between available GBE and preliminary RSE, i.e., 2017-2019. In the same Table, the robustness of the regression equations achieved comparing GBE andRSE in 2008-2016 (Fig. 6) was assessed by calculating the expected emissions values for 2017-2019, named RSE*. For 2017, there was a considerable underestimation of preliminary RSE for all emissions even compared to the expected RSE (Table 3). The RSE probably could not detect a large number of small fires, causing a significant reduction in emission estimations. On the other hand, preliminary results for both remaining years (2018, 2019) highlighted an overestimation of preliminary RSE in respect to the expected RSE (Table 3).

Conclusions
The ever-increasing climate change conditions ongoing in the Mediterranean region are exacerbating the wildfires phenomenon due to greater drought and a higher frequency of heatwaves (Turco et al. 2018). Reliable estimates of fire emission inventories are fundamental for implementing effective adaptation strategies to these climate-dependent phenomena. Nevertheless, the fundamental methodologies are still affected by significant uncertainties. Indeed, while global inventories, based on remote sensing wildfire data, are affected by the satellite source errors, the ground-based inventories are characterized by more accurate wildfire data, even if they are available for small areas (Galizia et al. 2021).
This work assesses the uncertainties of a broadly used remote sensing (i.e., GFED) compared to a ground-based emission inventory in a case study in southern Italy (Calabria region), whose available ancillary data allows the assessment of several parameters involved in the estimation. Regards Burned Area, the comparison between RSD and GBD estimated for 2008-2016 highlighted a similar temporal trend but, on the other hand, a scarce spatial agreement. The amount of Burned Area is overestimated by satellite, probably due to cropland fires detected only remotely and not reported by ground-based agencies (Turco et al. 2019a). Nevertheless, the numerous RSD wildfires detection problems (e.g., spatial resolution, fire extension, fire numbers, interferences) render the satellite Burned Area less reliable than GBD. Moreover, although the temporal trends are strongly correlated, the Dry Matter and the emissions for C, CO, PM 2.5 , and BC showed a heavy underestimation by remote sensing (from -66% to -97%). This underestimation suggests a strong need to improve the satellite-based method and, meanwhile, careful use in their application to atmospheric models. Concerning the GBE inventory, our analysis of the Mass of Available Fuel indicates that compared to other uncertainties, the accuracy of its value does not influence the GBE estimation particularly. Finally, this study highlighted that the preliminary RSE does not follow the same discrepancies of the analyzed period with validated data. Therefore, we discourage their preliminary application without any corrections. Future improvements will concern the application of revised EFs (Andreae 2019) and specific studies in the field focusing on the Calabrian and Italian fire characteristics.   Funding Open access funding provided by University of Calabria within the CRUI-CARE Agreement. This work was supported by PAC CALABRIA 2014-2020-Asse Prioritario 12, Azione B) 10.5.12.

Data availability
The raw datasets involved for the current study are available at the link http://globalfiredata.org/pages/data/ for remote sensing-based and at http://emissioni.sina.isprambiente.it/ serie-storiche-emissioni/ for ground-based data (last access on 29th of November 2022).

Competing interests
The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons. org/licenses/by/4.0/.