Midterm Fire Danger Prediction Using Satellite Imagery and Auxiliary Thematic Layers

: Wildﬁres constitute a signiﬁcant environmental pressure in Europe, particularly in the Mediterranean countries. The prediction of ﬁre danger is essential for sustainable forest ﬁre management since it provides critical information for designing effective prevention measures and for facilitating response planning to potential ﬁre events. This study presents a new midterm ﬁre danger index (MFDI) using satellite and auxiliary geographic data. The proposed methodology is based on estimations of a dry fuel connectivity measure calculated from the Moderate Imaging Spectrometer (MODIS) time-series data, which are combined with biophysical and topological variables to obtain accurate ﬁre ignition danger predictions for the following eight days. The index’s accuracy was assessed using historical ﬁre data from four large wildﬁres in Greece. The results showcase that the index predicted high ﬁre danger ( ≥ 3 on a scale within [ 1, 4 ] ) within the identiﬁed ﬁre ignition areas, proving its strong potential for deriving reliable estimations of ﬁre danger, despite the fact that no meteorological measurements or forecasts are used for its calculation.


Introduction
Forest fires are considered a rather complex natural process and one of the most crucial environmental issues in Europe [1], resulting in severe ecological, economic, and social consequences [2]. Most of the total European fire-affected areas are located in the Mediterranean countries during the summer season [2][3][4]. For example, Portugal, Spain, France, Italy, and Greece accounted for 78% of the total burnt area and 84% of the total number of fires registered in Europe in the 2000-2013 period, according to Burned Areas Perimeters dataset of the European Forest Fire Information System (EFFIS) [4]. For the five larger southern European countries most affected by wildfires (Portugal, Spain, France, Italy, and Greece), the number of fires has increased from 1980 to 2000 (partly due to improvements in monitoring systems) and has decreased afterwards (which is partly attributed to the improvement of the fire suppression mechanism), albeit with considerable variability in both number of fires and total area burned each year [5,6]. Nevertheless, climate change is expected to result in increased temperatures and prolonged droughts that may, in turn, lead to a major increase in the frequency and severity of fires [5,7], not only in the Mediterranean countries but also in northerner European countries as well.
Reliable fire danger forecast constitutes one of the most significant components of integrated forest fire management [8] and it is highly dependent on a variety of factors affecting fire ignition and spread. These factors comprise meteorological variables, vegetation condition, topography, The aforementioned limitations and uncertainties about the results' accuracy can be alleviated with the use of satellite remote sensing. Satellite remote sensing technologies have the advantage of capturing the spatial-spectral variability on a regular basis and over extended and/or remote areas [32][33][34]. In the last decade, satellite remote sensing has been investigated for its applicability in reliable forecasting of fire danger conditions, either by employing the remotely-sensed data alone or combining them with meteorological variables [33,35,36].
With respect to fire danger estimation methodologies that incorporate meteorological variables, Yu et al. [37] employed both remote sensing and meteorological variables for fire risk forecast in Cambodia. They combined MODIS-derived products with Shuttle Radar Topography Mission (SRTM)-derived DEM and Tropical Rainfall Measuring Mission (TRMM)-derived precipitation data reporting good agreement between the predicted risk and MODIS-derived fire spots. Moreover, Landsat satellite imagery was employed by Mitri et al. [38] for the generation of a land use/cover map, which was combined with other biophysical and WorldClim data (e.g., average monthly climate data for minimum, mean, and maximum temperature, and precipitation) [39] in order to reliably predict wildfire risk in the Lebanon region. The methodology was later enhanced by combining the initial fire risk data with drought spatial characteristics extracted from the WorldClim database, topographic, socio-economic, and fire data, leading to the development of a new daily fire danger index for the Lebanon region [26]. In another study [40], weather variables provided by 45 meteorological stations, vegetation data derived from Landsat and IKONOS satellite images, as well as other anthropogenic, topographic, and fire data were fed as inputs into a multi-layer perceptron, in order to accurately map regional forest fire probability in a Mediterranean ecosystem. DaCamara et al. [41] introduced a methodology for operational generation of fire danger maps over the Mediterranean Europe using vegetation cover maps obtained from Global Land Cover 2000 (GLC2000) dataset, meteorological data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) operational model, and fire activity based on the Fire Detection and Monitoring (FD&M) product from the Satellite Application Facility on Land Surface Analysis (LSA SAF) [42]. Chowdhury and Hassan [34] used MODIS-derived environmental variables, such as daily perceptible water (PW), Ts, and Normalized Difference Vegetation Index (NDVI) for the development of a daily fire danger forecasting system.
Focusing on methodologies not relying on meteorological variables directly, Abdollahi et al. [43] employed MODIS-derived products and historical fire ignition data for the performance enhancement of an existing fire danger forecasting system (FFDFS) that is solely based on remote sensing data. MODIS vegetation and canopy water content-related indices were analyzed by Arganaraz et al. [44] for live fuel moisture estimation and the subsequent prediction of fire danger conditions in Argentina. Matin et al. [45] integrated several factors related to land cover, surface temperature, topography, and spatial proximity to human activities using MODIS satellite imagery and Geographic Information System (GIS) to generate a fire risk map covering the area of Nepal. MODIS data were also employed for land cover mapping of the Uttarakhand state of India [46]. The produced vegetation map was used to produce a fuel type danger index, which was subsequently combined with various topographic variables (i.e., elevation, slope, aspect, and terrain ruggedness) for the generation of the final so-called Static Fire Danger Index, the accuracy of which approached 95%. Babu et al. [47] developed a fire danger index mainly based on MODIS-derived parameters (e.g., Land Surface Temperature, Normalized Multiband Drought Index, Modified Normalized Difference Fire Index), which is currently operationally employed in Kazakhstan. Liu et al. [48] used MODIS surface temperature (Ts) and reflectance data in combination with Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER)-derived DEM and historical lightning-caused fire data for fire risk forecast over a mountainous area in China. MODIS 16-day enhanced vegetation index (EVI) and daily Ts were integrated by Bisquert et al. [49] to predict fire danger conditions over the Galicia region in Spain for the period 2001-2006. Akther and Hassan [50] evaluated different satellite-based variables/indices to predict fire danger over boreal forests of Alberta for the period 2006-2008, proving that the combined use of Ts, normalized multiband drought index (NMDI) and temperature vegetation wetness index (TVWI) results in accurate (91.63%) fire danger predictions. Ahmed et al. [51] introduced a new model for forest fire danger prediction in northern Alberta on a four-day basis using three different Terra MODIS products (i.e., surface reflectance, land surface temperature, land cover map). The results showcased high agreement with the historical fire data revealing its' strong potential for operational use.
Most of the aforementioned fire danger forecasting systems provide daily fire danger estimations, i.e., they forecast the next day's fire danger conditions. Fire danger assessment on a midterm basis is also of high importance and could be synergistically employed with the daily estimations, aiding stakeholders to apply the appropriate prevention management practices. Here, we consider as midterm the fire danger forecasts that cover a period of at least one week ahead from the day they are issued. The number of such midterm indices that do not rely on meteorological measurements or predictions is still limited in the literature [34].
This study presents a new midterm fire danger index (MFDI) for a Euro-Mediterranean region (Greece), which is based on Earth Observation and ancillary geographic data and provides reliable estimations of fire ignition danger for a period of eight days ahead without using any meteorological data. Our methodology builds upon a previously proposed automated method for monitoring the spatial dry fuel connectivity [52] and combines it with additional biophysical and topographic variables for fire danger forecast on a midterm basis. Considering the spatial heterogeneity of vegetation patterns, the diversity of tree species, and the topographical variability characterizing the Mediterranean ecosystems [53], the spatial resolution of the fire danger maps is crucial for the accurate representation of fire danger conditions. The dry fuel connectivity estimates are considered as the main component of MFDI and provide information of 500 m spatial resolution, whereas all the other integrated variables have a spatial resolution of 30 m. Figure 1 provides an overview of the proposed methodology. Time-series of MODIS satellite data-spanning the 10 previous years-are used to identify extrema in vegetation moisture content. This is achieved by first calculating a moisture-sensitive spectral index from the original MODIS bands and then deriving a so-called fuel connectivity measure, taking into consideration the index's deviation relative to the 10 years history as well as the respective values from its spatial neighbors. The fuel connectivity measure is combined with other static variables (fuel type, topographic features, and distance from specific land cover/use classes) by means of the analytic hierarchy process [54] multi-criteria analysis, thus deriving the proposed fire danger index. The accuracy of the latter is finally assessed considering a number of test cases in Greece and various sources of information. The rest of this section describes in detail each individual step.

Datasets
The EO datasets required for MFDI are eight-day composite surface reflectance products from the MODIS sensor (MOD09A1). MOD09A1 provides an estimate of surface spectral reflectance values of Terra MODIS bands 1-7 (at a spatial resolution of 500 m), corrected for atmospheric conditions, along with a quality flag layer and four observation bands [35]. The products were downloaded from the MODIS Global Subsets Tool of the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) (https://modis.ornl.gov/cgi-bin/MODIS/global/subset.pl) for a ten-year period for each case study. Official perimeters of the burned areas within the study area are also necessary for the application of the proposed method. These data were provided by the Greek Ministry of Environment and Energy and the Laboratory of Forest Management and Remote Sensing (FMRS) of the Aristotle University of Thessaloniki in Greece.
The midterm prediction of fire danger also necessitates spatial information regarding the type and distribution of fuels, which constitutes one of the most important parameters that directly affect fire ignition and propagation danger. A national fuel type map (with 30 m spatial resolution) was developed in the framework of the NOFFi project (http://epadap.web.auth.gr/?lang=en), using the Landsat 8 OLI satellite imagery and ancillary spatial layers, such as the official national vegetation layer provided by the Hellenic Ministry of Environment and Energy [36]. The mapping process was performed through the development of an object-based Image Analysis (OBIA ) model [55] based on fuzzy logic rules, which was applied in all regional units to derive a country-wide map. The validity of the map was assessed using control points from the land use and land cover area frame survey (LUCAS) provided by Eurostat and the overall accuracy reached 92.59% [36].
The boundaries of urban and agricultural areas extracted from the Land Parcel Identification System (LPIS) geodatabase are also required and have been provided by the Hellenic Agricultural Payments Organization (OPEKEPE) in accordance to the European Union's (EU) Common Agricultural Policy (CAP). In addition, the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM) (https://asterweb.jpl.nasa.gov/gdem.asp) (30 m spatial resolution) and the OpenStreetMap (OSM) road network are also considered as fire-influencing parameters for the MFDI computation.
Finally, the validation procedure required the use of additional satellite data, namely the MODIS thermal anomalies and fire products and Sentinel-2 satellite images. The MODIS active fire product (MOD14A1) is a composite of fire pixels at 1 km spatial resolution detected in each cell of a predefined grid with a daily period [56]. MOD14A1 is produced every eight days but includes daily information, i.e., complete per-pixel information for each of the corresponding eight days [56]. The pixels are classified in ten categories according to Table 1 [56]. Sentinel-2 is a European satellite mission within the Copernicus EO program, providing high-resolution multi-spectral imagery from two identical sensors with a revisit frequency of five days. The Sentinel-2 product consists of 13 spectral bands, four of which are at 10 m spatial resolution, six at 20 m and three at 60 m [57]. Sentinel-2 images were used in two test cases, at which the imager captured the fire ignition areas. Tables 2 and 3 provide a brief description of all datasets used in the present study. Table 1. MOD14 fire mask pixel classes [56].

0
Not processed (missing input data) 1 Not processed (obsolete; not used since Collection 1) 2 Not processed (other reason) 3 Non-fire water pixel 4 Cloud (land or water) 5 Non-fire land pixel 6 Unknown (land or water) 7 Fire (low confidence, land or water) 8 Fire (nominal confidence, land or water) 9 Fire (high confidence, land or water)  Table 3. Vector data used in this study.

Name Extracted Information
National fuel type map Spatial distribution of forest fuels Burned areas Historical fire perimeters Land Parcel Identification System (LPIS) Agricultural and urban areas OpenStreetMap (OSM) Road network

Methods
The proposed method includes two main parts (Figure 1), namely the estimation of the dry fuel spatial connectivity and the combination of these estimates with ancillary data for the calculation of the MFDI fire danger rating. Both parts of the method have been implemented in the free and open-source software R, compiled into an automated workflow.

Dry Fuel Spatial Connectivity Estimation
The dry fuel connectivity constitutes a precipitation-driven variable and its estimation follows the automated procedure introduced in [52]. The measure is based on the Normalized Difference Infrared Index-band 6 (NDIIb6), which is calculated using the first short-wave infrared (SWIR) band (Band 6) of the MOD09A1 product: where Band2 and Band6 are the MODIS near-infrared (841-876 nm) and SWIR (1628-1652 nm) reflectance values, respectively. NDIIb6 is highly correlated with canopy water content, which is one the most important fire-affecting vegetation characteristics, since moisture constrains flammability and fire spread [52,[58][59][60]. It is a normalized index and the values theoretically range between −1 and 1, with the lower values (especially below zero) indicating canopy water stress [60]. The NDIIb6 is calculated for the whole MODIS time-series (ten-year period). Data anomalies (e.g., clouds) are masked out using the quality assurance (QA) metadata (state QA flag) [35]. Any residual noise and atmospheric disturbances affecting the time-series is eliminated by means of the adaptive Savitzky-Golay smoothing filter [61].
The smoothed time-series is subsequently used for calculating the z-score of the NDIIb6 values, expressed as the deviation from the mean value over the examined ten-year period (NDIIb6 anomaly data). If y is the current year, m kx is the average NDIIb6 value over the previous ten years for image pixel k and for the specific eight-day composite x (x ∈ [1,46], covering a whole year), i.e., m kx = 0.1 · ∑ y−1 i=y−10 NDIIb6 kxi , and σ kx is the corresponding standard deviation, the z-score is defined as: The rationale for using the z-scores instead of the absolute NDIIb6 values is that fire danger conditions are linked to these NDIIb6 anomaly data (i.e., relative deviations), rather than the absolute index values [28,52,62,63], since the former represent the drought-induced anomalies, at least as quantified by the pixel's spectral response [28]. In other words, the higher the drought anomalies (low NDIIb6 z-score values), the higher the fuel flammability potential is and, consequently, the greater the danger for a fire to occur is. The normalization of these deviations (division with the standard deviation) allows the subsequent categorization of the z-score values into six dryness classes, with an integer fuel weight being assigned to each one of them (high fuel weights representing drier fuels), as reported in Table 4. Table 4. Dryness classes and fuel weights based on NDIIb6 z-score ranges [52].

Dryness Class
NDIIb6 z-Score Fire danger is not only affected by the drought status of a small area such as that of a pixel, but also from the state of neighboring pixels. Imagine, for example, that a pixel has a very low z-score value (i.e., abnormally dry), but all neighboring pixels have medium or high z-score values. In this case, the probability of a fire to start and spread from the central pixel is lower than that implied by its very low z-score value. Following this reasoning, the aforementioned fuel weight image is processed in order to estimate the dry fuel connectivity map, considering a 3 × 3 pixel neighborhood around each pixel. The weight of each connection between the central pixel and one of its neighbors is computed by summing their fuel weight values. For the four diagonal connections, this sum is further divided by √ 2, accounting for the longer distance [52,64]. Finally, the fuel strength (connectivity measure) of each pixel is calculated by summing the weights of all connections with its eight neighboring pixels.
In other words, areas with high fuel strength represent more strongly connected dry fuels, which leads to higher fire danger.
Mathematically, the calculation of the dry fuel connectivity is performed as follows: where W i,j is the dry fuel connectivity for the pixel located at row i and column j of the fuel weight image, 1 is a column vector of ones, 1 T is its transpose, • denotes the Hadamard product (also known as Schur or entry-wise product), P i,j is the matrix of pixel values in the neighborhood of the pixel located at row i and column j of the fuel weight image: and F i,j is the following weighting matrix: The dry fuel connectivity estimates (W i,j ) are used as the component related to vegetation dryness in the proposed MFDI. This is subsequently combined with the other input parameters through multi-criteria analysis detailed in the following section.

Multi-Criteria Analysis
The second part of the methodology comprises the application of the Analytical Hierarchy Process (AHP) for estimating the final fire danger rating. AHP is a multi-criteria decision-making method that employs pairwise criteria comparisons in order to determine a scale of preferences among different sets of alternatives [54]. Numerical values representing the relative importance of one factor against all others are assigned to each factor, according to a judgment scale (for example, linear), so that they can be compared to each other in a pairwise comparison matrix [54,65].
Saaty's pairwise comparison scale (Table 5) [54,66] was used for the development of the pairwise comparison matrix. It is a nine-valued scale that quantifies the relative importance between any two criteria in a multi-criteria decision system, based on the linguistic description of an expert. If the importance of a criterion over another has been graded with some value by the expert, the reciprocal value is used by definition when comparing the criteria in the reverse order. The criteria compared to each other include the fuel connectivity estimates, fuel types, distance from agricultural areas, urban areas and road network, elevation, slope, and aspect. The relative importance values have been determined through expert knowledge and the results are reported in Table 6. The validity of the latter has been confirmed by calculating the so-called consistency ratio (CR), the most well-known measure employed to quantify the consistency of the comparison matrix in AHP [66,67]. In our case, the CR was calculated as 0.17, which is considered acceptable [68][69][70][71]. Therefore, no further reevaluation of the expert's judgments was performed, and the comparison matrix was considered as acceptable.
The relative factor priorities (weights) are derived by first normalizing the comparison matrix's columns (by dividing each column with its sum) and then calculating averages for each row [66]. The resulting weights in our case are reported in Table 7. Obviously, the fuel connectivity factor received the highest relative weight, since it was rated as the most important among all factors by the expect (values greater than one in the first row of Table 6). Fuel type received the second-highest weight, closely followed by proximity to agricultural areas, since a substantial number of wildfires in the Mediterranean area start from human activities in agricultural areas [72,73].
In order to derive the final MFDI estimations, all spatial datasets used to calculate the factors of Table 7 (see Section 2.1) must be co-registered into a common spatial grid and converted into a common metric scale. For the latter, we chose to employ a four-valued ordinal scale from 1 to 4 (i.e., 1 = low, 2 = moderate, 3 = high, 4 = very high), frequently used in fire danger forecasting systems [46,50,74].
The slope and aspect rasters are generated by the DEM raster (which provides the elevation information). The linear vector features of the road network and the polygons of the agricultural and urban areas are used for the derivation of the raster datasets (with a spatial resolution of 30 m, co-registered with the DEM raster) representing the Euclidean distance to the closest source, i.e., proximity to roads, agricultural, and urban areas. Finally, the dry fuel connectivity raster (500 m spatial resolution) is resampled to 30 m (following the nearest neighbor resampling approach) and co-registered with the DEM raster as well.in order to match the spatial resolution of all other raster datasets. All input variables are subsequently classified to the aforementioned four-valued ordinal scale, following the reclassification procedure reported in Table 8. Following the example of other relevant studies (e.g., [38,43,[75][76][77]), the thresholds used to categorize continuous variables into fire ignition danger classes or the reclassification of nominal variables are based on expert fire knowledge and experience taking into account the index objective-which is to classify fire ignition danger-as well as the ecosystem type for which the index is developed (i.e., Mediterranean). More specifically, the reasoning followed for each parameter is summarized in the following points:

Reciprocals of above zero
If activity i has one of the above nonzero numbers assigned to it when compared with activity j, then j has the reciprocal value when compared with i Table 6. Pairwise comparison matrix developed using the Saaty's pairwise comparison scale (C1 = fuel connectivity, C2 = fuel type, C3 = elevation, C4 = slope, C5 = aspect, C6 = proximity to roads, C7 = proximity to urban areas, C8 = proximity to agricultural areas). • The fuel connectivity measure is confined in the range [0, 68.28], which can be obtain if Equations (3)-(5) are employed with all elements of matrix F i,j having the minimum (0) or maximum (5) z-score fuel weight value (see Table 4). This range is simply linearly mapped to the (continuous) range [1,4].

Criteria
• Regarding fuel types, agricultural areas were assigned the highest risk value, since a substantial number of wildfires in the Mediterranean area start from human activities in agricultural areas [72,73]. Low vegetation, shrubs, and sparse evergreen broadleaved vegetation were assigned a high-risk value, because they dry easier than other vegetation types during the summer months, increasing their flammability. Conversely, dense or very dense deciduous broadleaved forests received the minimum risk value, since they are characterized by increased moisture content throughout the whole fire season. All other fuel type categories received a moderate risk value. This classification of fuel types with respect to fire ignition risk is consistent with the results of various studies in Mediterranean ecosystems (e.g., [73,[78][79][80]). We should note that water and non-vegetated areas are not assigned any risk level, so they proposed MFDI is not calculated for those areas (since they are not burnable). • Regarding topography, lower altitudes are significantly more prone to fires, with the risk becoming negligible above 1000 m [81]. Although the fire spread risk increases at steeper slopes, we chose to assign decreasing wildfire ignition risk values as the terrain's slope increases, because human activities are concentrated mainly in flat terrains (urban interfaces, road network, agricultural activities, etc.) [82,83]. As for terrain aspect, south-facing slopes exhibit the highest ignition risk-since they receive higher direct solar energy during the day [84]-and conversely, northern aspects the lowest risk, whereas eastern and western aspect have an intermediate (to low) risk [79,85,86]. • A buffer of 500 m around the road network is frequently considered as a high-risk zone for wildfires ignitions [43,73,81,87] and the same threshold was also applied for proximity to urban areas. With respect to agricultural land, the higher the proximity to these areas the greater the danger of fire ignition [88]. Following this general rule and considering the fire ignition drivers in Greece, empirical thresholds based on expert knowledge and experience were considered for defining the three distance zones, see Table 8. • In assigning the discrete values, each parameter is considered independently from the others and the minimum (1) and maximum (4) values are always assigned to the extreme categories. When a parameter has been split into three categories (e.g., proximity to agricultural areas), the intermediate category receives the value closer to the more appropriate extreme, as decided by the expert.
Having transformed all parameters into a common scale, the final MDFI value is derived via a weighted sum of all variables, using the weights obtained by the AHP (Table 7).

Accuracy Assessment
The validation of MFDI was based on historical fire events of four large wildfires in Greece, namely, in the Attica region (fire of June 2007), in two areas at Euboea Island (Karistos and Farakla for the fires of July 2016), and one at Kythira Island (fire of August 2017). The ten-year analysis period varies in each case study according to the year of fire occurrence, ranging from the MODIS eight-day composite ten years prior to the fire event until the one before the eight days period for which fire danger is being estimated. Thus, MODIS time-series data for the period 2007-2016 were employed for Karistos and Farakla areas at the Euboea island and 2008-2017 for the Kythira island. As for the Attica region, the lack of availability of a full MODIS time-series before 2001 led to the analysis of the period 2001-2010. The decision of using the specific time range was based on the assumption that the trend of drought under current climate conditions remains unchanged between the two time periods (1998-2007 and 2001-2010).
Ideally, MFDI's validation would have to be performed considering a large number of fire ignition points from historical wildfires. However, these data could not be provided to us by the Greek Fire Service that maintains this database, due to legal restrictions. As such, we used the aforementioned four historical fire events instead, for which we were able to identify the ignition points by other means. In the case of Karistos (Euboea region), we employed the MODIS active fire product (MOD14A1), comparing the average MFDI value within all MOD14A1 pixels flagged as fire (high confidence) with the average MFDI value over the rest of the burned area. Although the use of the MOD14A1 may seem as a systematic approach in validating MFDI, in practice it exhibits limitations, due to the Terra satellite overpass time being long after the fire ignition in most cases. Actually, the case of Karistos was the only case we could find that the Terra overpass time was close to the fire ignition time, with the latter being confirmed from press reports. In the cases of Farakla and Kythira Island, the fire ignition area was detected using Sentinel-2 satellite images, which were acquired around the time of the fire eruption. Finally, press reports facilitated the identification of the fire ignition area for the fire event of 2007 in Attica with good accuracy, since the fire started near one of the few residential areas nearby. In all MFDI cases, the official fire perimeters were used for the empirical/indirect MFDI accuracy assessment, as well as for the calculation of statistics within the burned areas. Figure 2 presents the MFDI maps of the four aforementioned test cases, calculated for the eight-day period that each fire erupted.

Results
In Attica (Figure 2a), the mean MFDI value within the burned area is approximately 2.7. The index presents moderate to very high values (that is, greater than 2) in most of the area's extent, with the very high values located to the northwest side of the fire-affected area. According to press reports, the specific fire ignited at Dervenoxoria Municipality, near the village Stefani, which is located on the northwest side of Mount Parnitha, where the MFDI values were very high (greater than 3). Figure 3 depicts the individual parameters (i.e., input variables) used for the index calculation. We can observe that the fuel connectivity (Figure 3a) exhibits close to maximum values in approximately half the burned area. Although this parameter is assigned the largest (by far) weight by the AHP (see Table 7), most other parameters denote a medium to small danger within most of the burned area. As such, the final result (Figure 3i) reports medium to low risk in most of the affected regions and pinpoints the dangerous areas on the northwestern part of the fire perimeter, where the fire actually started from. This part of the area exhibited high fuel connectivity values, but moreover was surrounded by highly flammable fuels and was close to agricultural and/or urban areas.
The MFDI calculated for Karistos (Figure 2b) exhibits very high danger values (greater than 3.5) mainly in the northern part of the burned area, whereas the mean index value within the whole fire perimeter is approximately 2.9. The index validation in this specific case was performed using the MOD14A1 product ( Figure 4). More specifically, the region belonging to MOD14A1 pixels of class 9 (fire, high confidence, land or water) within the fire perimeter as compared to the rest of the burned area. The mean MFDI value within the area identified as fire by MOD14A1 (class 9) was 3.4, i.e., high to very high fire danger. On the other hand, the mean MFDI value of all other pixels within the burned area was 2.8, i.e., moderate to high fire danger.
For the Farakla (Euboea region) case (Figure 2c), the index's accuracy was empirically assessed with the use of a Sentinel-2 satellite image, which was acquired around the time that the fire erupted. The ignition point was located in the northern part of the fire-affected region, which is adjacent to agricultural crops ( Figure 5) and characterized by a very high fire danger according to the MFDI map (greater than 3). Indeed, press reports connected the origin of this fire to agricultural burnings. Since the MFDI quantifies the danger for fire ignition and not for spread, high MFDI values are observed only in few areas, one of which is the fire ignition point, contrary to the rest of the burned area. In this case, the strong north wind (clearly identifiable in the Sentinel-2 image from the smoke's direction) was the reason for the spread of the fire to the south, even though the predicted fire danger was much lower for those areas. Figure 5c depicts the fuel connectivity parameter, which highlights the opposite situation of that observed in Parnitha (Figure 3). In the latter, the high fuel connectivity values within most of the burned area were confined by the other parameters in the final MFDI. Here, the fuel connectivity is medium to low in almost all burned areas, but specific points are highlighted as highly dangerous in MFDI due to their proximity to agricultural areas.  A Sentinel-2 image acquired around the fire eruption time was also used for the validation of the MFDI for the Kythira Island wildfire (Figure 2d). A closer examination of the index values within the burned area ( Figure 6) reveals that the fire started in an area where the MFDI exhibited relatively higher values (approximately 3) due to the adjacent of agricultural fields. According to unconfirmed reports, the fire might have started from anthropogenic causes rather than natural ones. This is the reason why most of the fire-affected region comprises low to moderate MFDI values. The presence of intense north winds caused the fire to spread fast and destroy a large forested area.

Discussion
The MODIS-derived dry fuel connectivity parameter within the proposed MFDI is used as an indicator of fuel dryness (more precisely, its relative deviation from its historical assumed norm). Since fuel dryness significantly increases fire ignition risk, this parameter was assigned the highest importance among all factors (see Table 7). However, fuel dryness does not only affect fire ignition danger but also the fire spread risk. Therefore, MFDI introduces a number of other parameters, in order to tilt the scale towards the fire ignition danger component rather than the spread risk one. This is evident from the presentation of all individual parameters in the case of the Parnitha wildfire ( Figure 3). There, the dry fuel connectivity parameter exhibited high values in almost all the areas burned, but the MFDI values are decreased in most regions by the other factors, with high values being maintained only in some small regions, in one of which the fire actually started from. Conversely, in the Farakla wildfire ( Figure 5), the fuel connectivity was medium to low in almost all burned areas, but the ignition point was ultimately highlighted with high MFDI values due to its proximity to agricultural areas. In either case, it is evident that the fuel connectivity measure is not sufficient to quantify the fire ignition danger by itself, which concurs with the empirical observation that a high dryness deviation is an important but not decisive factor for a wildfire to erupt.
It is undeniable that a larger database of fire ignition points is essential for performing a rigorous quantitative assessment of the proposed MFDI, which we are currently pursuing. Lacking such data presently, we reported here all cases for which we could find trustworthy information on the fire ignition locations, either through MODIS data, Sentinel-2 imagery, or cross-checked media information. Even though there are only four test cases, we argue that the results showcase the strong potential of the proposed method in deriving reliable estimations of fire danger, at the very least for the Mediterranean ecosystem of Greece.
The proposed MFDI relies on MODIS-derived NDIIb6 anomaly data (the dry fuel connectivity measure) for identifying abnormal vegetation dryness in the previous eight days the index provides estimation for. The latter is expected to increase fuel flammability potential and, therefore, fire ignition danger in the next eight-days period (for which the index is calculated for), provided that the weather conditions do not change abruptly (i.e., heavy rainfall or otherwise increased humidity). Relying on such a satellite-derived proxy of vegetation dryness allows us to avoid using meteorological data (either measurements or predictions), which exhibit the limitations described in the introduction (e.g., ineffective spatial interpolation if no dense network of meteorological stations exists) and may involve substantial costs for obtaining them. However, weather-based fire danger indices do exhibit advantages compared to our approach, such as the estimation of surface fuel humidity or the prediction of future weather conditions (e.g., precipitation probability, wind intensity, or increased lightning activity that is a frequent source of ignitions, especially in remote forested areas). Therefore, MFDI should be considered as a complementary rather than competitive (to weather-based indices) tool in prefire management. Combining MFDI with weather-based indices or otherwise incorporating meteorological information into the index would be of high interest, but this is a future work outside the scope of the current study.
Accordingly, the other factors-save the fuel dryness-incorporated into MFDI try to capture the most frequent causes of wildfire ignition, but not all. Specifically, fuel type and topographic features cover the most important natural causes of fire ignition. The proximity to agricultural areas, roads, and urban areas was introduced to cover the most frequent anthropogenic sources of fire occurrence, which are mostly due to negligence. Arsons are thus not covered, which is a parameter difficult to model and an open research topic. Incorporating recent research outcomes in this direction in the future would also be very interesting.

Conclusions
This paper presented a new method for the reliable estimation of fire ignition danger on a midterm basis. The developed MFDI provides 500m spatial information related to the danger of fire ignition for the next eight days. Its development is primarily based on the availability of satellite data, namely the MODIS eight-day composite products, contrary to other operationally used fire danger indices (for example, the daily index provided by the European Forest Fire Information System-EFFIS), which rely mostly on meteorological predictions. In particular, the proposed method employs MODIS time-series for the spatial estimation of the dry fuel connectivity, which is then combined with ancillary data through a multi-criteria analysis for the derivation of MFDI fire ignition danger estimate.
The index was validated considering the cases of four large wildfires in Greece, namely, the Attica region, the Farakla and Karistos areas at Euboea and the Kythira Island. The validation of the resulting fire danger maps was based on historical fire events and-although not fully quantitative-it showcased that the MFDI may constitute a reliable index for assessing the spatial distribution of wildfire danger. Future investigation of the potential of MFDI in combination with other fire danger indices would be of high interest, in order to effectively support midterm fire prevention planning.