The frequency and intensity of disasters have been growing in the last decades due to rising concentration of people and assets in regions exposed to hazards of natural (e.g., earthquakes, tsunamis, floods) or anthropogenic (e.g. fires, pollution) origin. Rapid urbanization, the increasing vulnerability of modern industrial societies and a combination of economic, political, cultural, and social pressures have led to millions of people living in conditions of high risk. The combination of this urban growth and climate change have resulted in a tendency of increasingly damaging hazards of mixed meteorological and man-made origin. Fire, for example, is a natural phenomenon shaping most ecosystems on Earth and plays essential roles regulating atmospheric oxygen, carbon cycle, and climate. However, wildfire is also a hazard, and it becomes a disaster when it propagates from the wildland-urban interface (WUI) to towns or cities and results in loss or human lives and valuable assets.

In Chile, forest fire has become a large and recurrent problem that affects the fringes of populated areas, resulting in disasters like the Valparaíso fires of 2014 and 2015, or the 2017 forest fire that completely burned the town of Santa Olga. In Australia, on the other hand, bushfire incidents in the WUI have a longer history and have been well researched in academia and technical commissions; reports indicate that more than 11000 dwellings have been lost due to this hazard and the number of exposed housed grows quickly as proportionally greater amounts of land are being developed as fragmented, low density peri-urban settlements. To ensure the long term resilience of such settlements, several studies have focused on the analysis of observable urban design and landscape factors and their implications for predicting house loss in unplanned bushfires. In this work, multispectral satellite imagery and geographic data are used to quantify and compare parameters of the built environment of two towns devastated by WUI fires in recent years, Wye River in Australia and Santa Olga in Chile. The severity and distribution of fire damage is also characterized from remote sensing data, and correlated with urban morphology parameters in order to identify common vulnerability indicators of settlements exposed to WUI fires.

To investigate the correlation between settlements' characteristics and vulnerability to fires originated in the WUI, a combination of MSI/Sentinel 2 imagery and local georeferenced databases were used to quantify a set of morphological and landscape parameters. First, pre-fire images are used to characterize the distribution and density of vegetation in each town through the Normalized Difference Vegetation Index (NDVI). Terrain slopes are calculated using topographic data generated by the Shuttle Radar Topography Mission and for analysis of the towns' morphological features, property lot footprints were obtained from public and individual building footprints were manually digitalized based on Google Maps satellite imagery. Finally, fire damage in each town was assessed from comparison of pre- and post-event satellite imagery, which are used to calculate the differential Normalized Burn Ratio (ΔNBR), and from inspection of optical imaging and damage cadasters.

A comparison of the general morphological properties of Wye River and Santa Olga reveals significant differences in the layout and building density of each town. The studied areas are similar in extension, comprising 58 and 51 hectares respectively, but the pre-fire number of houses in Santa Olga was more than double than in Wye River, and correspondingly the average lot size was much smaller (274.6 versus 1250.2 m2). However, the average site occupancy ratio, calculated as the average of each property's built footprint area over the lot footprint area, is higher in Santa Olga, with a value of 0.26 versus 0.19 in Wye River.

The NDVI maps obtained for Santa Olga and Wye River (Fig. 1) were used to study the distribution of vegetation relative to houses. Given that NDVI values range from 0 for no vegetation to +1 for the highest possible density of green leaves, pixels with NDVI>0.5 are considered to represent relatively dense vegetation, and the distance to the closest pixel of dense vegetation is calculated for each property. The average minimum distances obtained are 63 m in Santa Olga and 15 m in Wye River, so although the latter had a lower building density, its constructions were more closely surrounded by vegetation. Regarding slope, analysis of SRTM digital elevation models shows that Wye River was built on steeper terrain, with and average property inclination angle of 15.9º, that more than doubles the 7.0º average slope obtained for Santa Olga.

Regarding fire impact and house damage, analysis of ΔNBR maps (Fig. 2) indicate that only 10% and 8% of the Santa Olga and Wye River study areas, respectively remained unburnt. Generally, the more severely burnt areas correlated with denser vegetation, and location of houses registered as “not damaged” coincide with pixels classified as “unburnt” according to ΔNBR. In terms of lost houses, the situation was significantly more disastrous in Santa Olga, where 82.6% of existing buildings were completely destroyed, compared to 22.4% in Wye River. The extent of this damage is largely explained by the predominance of wood in small Chilean towns like Santa Olga. In fact, the few buildings that at least partly survived the 2017 fire were those made of concrete, like the local police headquarters.

The parameters described above provide insight into the probable sequence and mechanisms of fire propagation in these towns. Although in both cases the fire started in the bushland, the factor that determined its spread across the populated areas varies. In Wye River, for example, the primary fire progression mechanisms involved spotting from the fire front outside the perimeters of the townships, and surface spread through fine surface fuels and near surface vegetation onto built elements. Large distances between buildings implied less house-to-house spread, although it was still a factor due to low wind speeds during and after the fire, which reduced cooling of radiant. The type of structural design also proved determinant in the degree of damage, and for example houses with a steel support structure and non-combustible subfloor, cladding, window frames and doors were effective in resisting ignition. In the case of Santa Olga, however, the rapid advance of fire across town was mostly house-to-house, due to the high building density and the material vulnerability, which was mostly timber. In both cases, fire propagation was also favored by a deficiency in suppression actions, which only consisted of water bombing in the case of Wye River, and were practically inexistent in Santa Olga.

In conclusion, the use of remote sensing in combination with property and geographic data and fire damage cadasters provide insight into the main morphological and landscape parameters that determine a settlement's vulnerability to fires originated in the WUI. A comparative analysis of the fires occurred in 2015 in Wye River and in 2017 in Santa Olga shows that fire damage is related to proximity to the density of vegetated areas, to the concentration and relative proximity of built elements, and to their materiality and structural design. Hence, urban planning and construction regulations may play a relevant role in the prevention and mitigation of bushfire effects, by regulating the use of different terrain classes, limiting the number and density of buildings in a given area, promoting the use of fire-resistant materials and structures, and planning suppression and defense activities ahead of time.

Forest fires have an important ecological role, but they can also threaten public safety, destroy timber, and reduce sustainable harvest levels in managed forests. In recent years, wildfires of large size and high intensity have occurred across many parts in the world. Record heat and dry weather across many regions in 2018 led to an unprecedented fire season with more than 500,000 hectares and 1.3 million hectares burned respectively in California, United States of America, and in British Columbia, Canada, and exceptionally high wildfire activity in Sweden with, at its peak, 45 active fires that burned approximately 26,000 hectares.

With forest fire frequency and intensity likely to increase due to climate change, forest fire management agencies would benefit from improved techniques to map fire progression and obtain accurate and timely information on burned areas. Data collected by Earth observation techniques provide valuable information on forest fires. Sensors such as Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging Radiometer Suite (VIIRS), and Sea and Land Surface Temperature Radiometer (SLSTR) onboard Sentinel-3 can provide daily detection of fire hotspots. Medium resolution optical sensors such as the MSI on Sentinel-2 provide more detailed but less frequent data, and can be affected by clouds, haze, and smoke.

Exploitation of Synthetic Aperture Radar (SAR) data using machine learning methods is a promising technique to address some of the limitations of optical and thermal approaches, with its capability to operate in presence of clouds, haze and smoke, and sensitivity to forest structure. In this study, we evaluated the ability of machine learning methods with radar metrics derived from dual polarization and fully polarimetric data to complement optical/thermal data to identify burned areas over several locations in Western Canada affected by large fires from 2016 to 2018: Fort McMurray (Alberta) and Williams Lake and Fraser Lake (British Columbia).

The 2016 fire of Fort McMurray burned up to 590,000 hectares of coniferous and broadleaf forests interspersed with treed wetlands. Most burned areas exhibited moderate to high level of severity. The Williams Lake and Fraser Lake study sites were largely dominated by coniferous forest with small patches of regenerating forest due to forestry activities. The 2017 Williams Lake fire burned approximately 13,000 hectares compared to 92,000 hectares for the Fraser Lake fire. The forests at both sites were impacted with a relatively equal proportion of low, medium, and high level of severity. All three test sites are characterized by relatively flat terrain with gentle slopes.

RADARSAT-2 and Sentinel-1 C-band SAR data were acquired pre-fire and during the fires and analyzed to characterize burned areas over the three study sites. Priority was given to pre-fire SAR datasets acquired in the same season and under dry weather conditions, given that canopy moisture may affect the sensitivity of SAR to detect burn scars. Optical-derived burn severity maps from Landsat 8 and Sentinel-2 satellite data time series were used together with MODIS and VIIRS thermal data, land cover, weather data, and field survey plot data, to train, calibrate, and assess the accuracy of a machine learning model using SAR data for burned area mapping. Polarimetric decomposition was applied to RADARSAT-2 data to partition the radar response into contributions from different scattering mechanisms. In addition, several amplitude metrics that could be sensitive to changes in forest structure due to fire were also derived from RADARSAT-2 and Sentinel-1 data, including the Radar Vegetation Index (RVI), and Radar Forest Degradation Index (RFDI). The Radar Burn Ratio (RBR) approach was used to examine the degree of change of the different metrics with respect to the pre-fire status by using post- to pre-fire ratios. The RBR calculated for all metrics were used as inputs to a Random Forest (RF) machine learning algorithm to evaluate their contribution to identify burned areas.

The RF models using polarimetric and amplitude variables provided burned area information broadly consistent with information from satellite optical data. When including all RADARSAT-2 variables (polarimetric and amplitude) in the RF model over the Fort McMurray site, we found that the most important variables for burn area classifications were: HV/HH, RVI, RFDI, HH, HV/VV and volume scattering, suggesting that amplitude variables are most important for burn area classification.

To assess the accuracy of the RF models parametrized with Sentinel-1 or RADARSAT-2 data, we compared the results to burn severity maps derived from Landsat 8 and Sentinel-2 imagery over the Williams Lake fire. Results showed that RF models were effective in classifying burned areas with moderate to high burn severity at accuracy ranging from 67.5% to 94.4% for Sentinel-1 and from 57.3% to 86.9% for RADARSAT-2, but areas of low burn severity were not classified accurately, with 16.9% and 16.1% accuracy for Sentinel-1 and RADARSAT-2, respectively (Figure 1).

The ability of a SAR RF model to accurately detect forest severely and moderately affected by fire, but not forest affected by low burn severity, is due to the effect the varying level of burn severity has on the forest canopy and the C-band wavelength. C-band SAR scattering from forests is generally dominated by volume scattering from the top of the canopy (branches/leaves/needles) due to its relatively short wavelength. As the fire severity increases in a forest stand, a greater percentage of the branches and leaves at the top of the canopy are destroyed, which decreases volume scattering. Low severity fires are expected to cause limited changes to the upper canopy layer, which does not significantly alter volume scattering.

While we identified the level of burn severity as impacting the performance of burned areas classification for Sentinel-1 and RADARSAT-2, further work is needed to understand the parameters causing the variability in the SAR signal that is not related to the fire effect. Parameters such as the variation in canopy and soil moisture, incident angle, and polarization should be investigated further. The RF model method should also be tested in different forest ecosystems, e.g., Sweden, Russia, and over sites with different forest stand characteristics (tree species, tree height, stand density, etc.). Several upcoming SAR missions, including the C-band RADARSAT Constellation Mission (RCM), L-band SAOCOM, X-band ICEYE should provide further opportunities to access SAR data to pursue research to complement optical/thermal data for burned area mapping, improving current forest fire response capabilities.

Figure 1: Burned area and burned severity classifications over the Wildwood fire discovered on 07 July 2017 in proximity to Williams Lake, BC, Canada: (left panel) RADARSAT-2 burned area classification from 29 July 2017; (centre panel) Sentinel-1 burned area classification from 31 July 2017; and (right panel) Landsat 8 burn severity classification from 22 August 2017.

Wildfire behaviour prediction requires a deep comprehension of propagation-driving parameters and of their spatial and temporal variability. Despite the wide variety of developed wildland fire simulators, a major source of uncertainties still stems from input data inaccuracy as well as from the incompleteness in the scientific understanding of fire behaviour response to treatments. The research aims at designing and implementing a wildfire behaviour modelling methodology effective in wildfire hazard characterization on a local scale, calibrating it to past events recorded in wildland fire databases of Umbria and Sardinia Regions. In line with the study's objective, the study intends to further investigate satellite sensor technologies to acquire updated and reliable input data with a suitable spatial and temporal resolution.
Geospatial data on recent historical wildland fire events and their respective control parameters has been collected from Umbria and Sardinia regional databases in order to build a complete wildfire database to validate and calibrate the wildfire propagation modelling methodology. The wildfire behaviour modelling methodology has been designed employing a simulation toolset implemented in GRASS GIS, available as free and open source software for geospatial data management and analysis, image processing and spatial modelling. Founded on a quasi-empirical mathematical model, the toolset requires investigating the local topographical and meteorological parameters as well as the pyrological and phenological characteristics of the local plant communities to estimate direction and rate of spread and to simulate the flaming front of the elliptical anisotropic propagation from a defined ignition point applying a least-cost path research algorithm. Within the context of this research, a data pre-processing methodology integrating information from land cover maps and weather station with satellite multispectral imagery from Sentinel 2 and Landsat 7 has been defined in order to obtain the spatial variability of propagation-driving parameters. The accuracy of the resulting models has been evaluated in terms of morphological matching between obtained simulations and respective documented historical events boundaries, minimizing both false negative and false positive modelled burned areas. Afterwards, a genetic algorithm, an heuristic local research method, has been implemented to optimize the accuracy of the models and to estimate their suitability to describe the spatiotemporal evolution of the flaming front. The designed algorithm explores over a large set of feasible solutions and, through an iterative process, converges towards a local optimum of the search space, where some of the simulations whose modeling parameters combinations maximize the accuracy are concentrated.
Obtained results indicate preliminary models do not exceed an accuracy of 60% in describing the sample of events selected form the wildfire database. The implementation of the genetic algorithm has ensured an outstanding improvement in model accuracy, reaching a maximum of 75%. Outcomes suggest genetic algorithm has allowed inferring a restricted domain of existence for propagation-driving parameters, potentially reducing uncertainties arising from data inaccuracy. However, the descriptive capacity of the genetic algorithm-based models highlight there still exist opportunities for further enhancements. Latest and forthcoming satellite missions (ADM-Aeolus and GEDI) are going to make available crucial data in addressing spatiotemporal variations of vegetation and weather conditions. Future activities will be focused on the development of remote-sensing-driven methodology to distinguish patterns among different fuel models and to monitor spatiotemporal variations in fuel moisture content as well as in wind speed and directions.

Accurate and rapid mapping of fire damaged areas plays a key role for supporting fire management, accounting for environmental loss, defining planning strategies for environmental monitoring and risk disaster management. Under climate change conditions, drought severity may strengthen wildfire disturbance, increasing number and dimension of fires. Satellite data play a major role in supporting knowledge about fire severity by delivering rapid information to map areas damaged by fire in a accurately and prompt way.
Operational services for burned area mapping have been developed in the past years to provide near-real time information on wildland fires from satellite data, like the fire grading maps generated under the Copernicus Emergency Management Service (EMS) and the European Forest Fire Information System (EFFIS). Despite the large availability of freely available satellite data at high spatial resolution, operational services for high spatial resolution wildfire mapping based on new satellite sensors and novel methodologies still have to be developed.
Objectives of the research study is to map burned areas at national level using multitemporal series of Sentinel-2 MSI data. Wildfires occurred across Italy during year 2017 were used as case study. A total of 5795 Sentinel-2 images, atmospherically and terrain corrected using MACCS-ATCOR Joint Algorithm (MAJA), were processed at 20 m spatial resolution. After applying a proper pixel masking procedure, spectral indices designed to detect burned areas were computed from masked spectral Bottom-of-Atmosphere reflectances, specifically the Normalized Burnt Ratio (NBR), Burned Area Index for Sentinel-2 (BAIS2) and Albedo modified Burned Area Index for Sentinel-2 (BAIS2α). Additionally, Leaf Area Index (LAI) was computed from each Sentinel-2 image using the SNAP biophysical processor, in order to integrate the information related to fire severity with the status of vegetation prior the wildfire disturbance occurred. Finally, a specific algorithm has been developed for the identification of burned areas from multitemporal Sentinel-2 observations. The algorithm makes use of pixel-based time series of fire related spectral indices to detect abrupt changes provoked by wildfire events, later object-based analysis is used to reduce the generation of false positives.
Additionally, the present research study used the SWIR wavelength from Sentinel-2 MSI band centered at 2.185 µm to identify active fires, thanks to the strong radiometric signal resulting from the combination of surface reflectance and high fire emissivity, even visible at wavelengths shorter than 3 µm and still detectable under the fire plume.
Results show that the adopted spectral indices provide different burned areas maps, and report a high number of false positives in arable lands that still require significant improvements of the detection algorithm. Results demonstrate the suitability of the adopted methodology, stimulating the development of new detailed operational downstreaming services at national level based on Copernicus data for post-fire high resolution mapping. Discussion points out the needs to be addressed in future research studies to develop a systematic Sentinel-2 MSI burned area mapping capability.

In 2015 Indonesia was impacted by massive wildfires, mostly of anthropogenic origin and possibly fuelled by the abnormally dry conditions characteristic of El Niño years. These fires destroyed millions of hectares of forest both on mineral soil and peatland. They occurred in a region where cloud cover is frequent and where haze from the fires can significantly impact the observation of the surface. This has enormous implications for current approaches estimating greenhouse gases (GHG) emissions from fires or to map fire scars. These methods employ data acquired from optical sensors, either data collected from visible to middle infrared to map fire scars or thermal data to estimate fire radiative power and consequently derive emissions.
The analysis presented here explores microwave data made available by the operational Sentinel-1 mission, namely that acquired by Sentinel-1A (S-1A) in dual-polarization (VV and VH) interferometric wide scan mode (IW). S-1A IW dual-polarization data was acquired over Central Kalimantan during the 2015 fire season. This dataset was used to generate the first time-resolved map of burnt areas during the 2015 fire season in Indonesia. Burnt areas were mapped in four consecutive periods (20 July-13 August, 13 August-6 September, 6-30 September and 30 September-24 October) using a machine learning algorithm (random forests, RF). The approach that was followed relied on modelling the presence and absence of burnt area against the backscatter intensity of pre-fire and post-fire conditions. The reference dataset was composed of 2,940 observations in unburnt locations and 640 observations over burnt areas, with 70% assigned to training and the remainder for testing.
We set the commission error of the burnt class to 10% as the criterion for defining the threshold to classify an observation as unburnt or burnt. The fitted RF model had an overall accuracy of 96%, omission and commission errors of the burnt class lower than 3% and an omission error of the burnt class of 12%. A total burnt area of approximately 1.8 million ha was mapped in the period 20 July - 24 October, representing 30% of the study area, with approximately 65% mapped in six weeks (6 September - 24 October).
Comparison with burnt area estimates from the Global Fire Emissions Database (GFED4) at 0.5˚ scale showed poor agreement, with consistently much lower estimates produced by GFED4. Additionally, periods with much lower agreement could be tracked to those affected by higher cloud cover, thus suggesting the ability of C-band SAR data to improve the estimates where and when cloud cover is substantial. The method presented in this study offers an alternative to reduce the uncertainty of the estimates of GHG emissions from fires in tropical areas affected by substantial cloud cover.