Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
METHOD THAT DETECTS AREAS OF ACTIVE FIRE HOTSPOTS IN REAL-TIME, CALCULATES THE MOST PROBABLE IGNITION POINT AND ASSESSES FIRE PROBABILITY INDICATORS, USING SATELLITE IMAGES AND FUEL DATA.
Document Type and Number:
WIPO Patent Application WO/2016/132161
Kind Code:
A1
Abstract:
In brief, it is a method of monitoring wildfires and detecting ignition points in real- time using satellite images and vegetation fuel data. The method belongs to the fields of remote sensing and digital image analysis. The method addresses the issue of the low spatial resolution of satellite images that are used to monitor fires by combining satellite images with vegetation fuel data, as well as data of topography and meteorology. The satellite image is separated using an automated algorithm implemented for digital processing, into burning and non-burning sub-pixels. Subsequently, based on fuel data, vegetation density and topography datasets that have better spatial resolution than the satellite image, pixels of satellite imagery are divided into sub-pixels of smaller dimension, thus illustrating and representing sub-areas of fire observation with higher spatial resolution for which a fire probability indicator is being calculated. Then, taking into account the flammability information of each sub-pixel as well as meteorological data, the most probable ignition point for each fire event is estimated. After applying a fire dispersion model initialized from the calculated fire ignition point using topography, vegetation fuel and meteorology data, the fire dispersion is computed and based on that, the fire probability indicator is re-estimated as well.

Inventors:
KONTOES CHARALAMPOS (GR)
CHAIREKAKIS THEMISTOKLIS (GR)
Application Number:
PCT/GR2016/000005
Publication Date:
August 25, 2016
Filing Date:
February 16, 2016
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
KONTOES CHARALAMPOS (GR)
CHAIREKAKIS THEMISTOKLIS (GR)
International Classes:
G06K9/00
Domestic Patent References:
WO2008051207A22008-05-02
Foreign References:
JP2011024471A2011-02-10
US20080027649A12008-01-31
Other References:
STEFAN VOIGT ET AL: "Integrating satellite remote sensing techniques for detection and analysis of uncontrolled coal seam fires in North China", INTERNATIONAL JOURNAL OF COAL GEOLOGY, vol. 59, no. 1-2, 1 July 2004 (2004-07-01), AMSTERDAM, NL, pages 121 - 136, XP055276125, ISSN: 0166-5162, DOI: 10.1016/j.coal.2003.12.013
LANEVE G ET AL: "SIGRI Project: Products Validation Results", IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, IEEE, USA, vol. 7, no. 3, 1 March 2014 (2014-03-01), pages 895 - 905, XP011542939, ISSN: 1939-1404, [retrieved on 20140314], DOI: 10.1109/JSTARS.2013.2271379
CALLE A ET AL: "Latest algorithms and scientific developments for forest fire detection and monitoring using MSG/SEVIRI and MODIS sensors", RECENT ADVANCES IN SPACE TECHNOLOGIES, 2005. RAST 2005. PROCEEDINGS OF 2ND INTERNATIONAL CONFERENCE ON ISTANBUL, TURKEY JUNE 9-11, 2005, PISCATAWAY, NJ, USA,IEEE, 9 June 2005 (2005-06-09), pages 118 - 123, XP010838587, ISBN: 978-0-7803-8977-9
RONALD GRAML ET AL: "Bushfire Hotspot Detection Through Uninhabited Aerial Vehicles and Reconfigurable Computing", 2008 IEEE AEROSPACE CONFERENCE; 1-8 MARCH 2008; BIG SKY; MT, USA, IEEE, PISCATAWAY, NJ, USA, 1 March 2008 (2008-03-01), pages 1 - 13, XP031256345, ISBN: 978-1-4244-1487-1
IPHIGENIA KERAMITSOGLOU ET AL: "Downscaling Geostationary Land Surface Temperature Imagery for Urban Analysis", IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IEEE SERVICE CENTER, NEW YORK, NY, US, vol. 10, no. 5, 1 September 2013 (2013-09-01), pages 1253 - 1257, XP011515211, ISSN: 1545-598X, DOI: 10.1109/LGRS.2013.2257668
AGAM N.; W. P. KUSTAS; M. C. ANDERSON; F. LI; C. M. U. NEALE: "A vegetation index based technique for spatial sharpening of thermal imagery", REMOTE SENSING OF ENVIRONMENT, vol. 107, no. 4, 2007, pages 545 - 558, XP002758314
JEFF DOZIER: "A method for satellite identification of surface temperature fields of subpixel resolution", REMOTE SENSING OF ENVIRONMENT, 1 January 1981 (1981-01-01), pages 221 - 229, XP055277036, Retrieved from the Internet DOI: 10.1016/0034-4257(81)90021-3
AGAM N.; W. P. KUSTAS; M. C. ANDERSON; F. LI; C. M. U. NEALE: "A vegetation index based technique for spatial sharpening of thermal imagery", REMOTE SENSING OF ENVIRONMENT, vol. 107, no. 4, 2007, pages 545 - 558
DENEKE, H.; R. ROEBLING: "Downscaling of METEOSAT SEVIRI 0.6 and 0.8 micron channel radiances utilizing the high-resolution visible channel", ATMOSPHERIC CHEMISTRY AND PHYSICS, vol. 10, 2010, pages 9761 - 9772
"Active fire monitoring with MSG algorithm - theoretical basis document", DARMSTADT
FINNEY, MARK A.: "Fuels Management-How to Measure Success: Conference Proceedings", FORT COLLINS, CO: U.S. DEPARTMENT OF AGRICULTURE, article "An Overview of FlamMap Fire Modeling Capabilities", pages: 213 - 220
HUTENGS C.; M. VOHLAND: "Downscaling land surface temperatures from MODIS data to mesoscale resolution with Random Forest regression", DGPF PROCEEDINGS, vol. 23
KERAMITSOGLOU, I.; C.T. KIRANOUDIS; Q. WENG: "Geoscience and Remote Sensing Letters", vol. 10, 2013, IEEE, article "Downscaling Geostationary Land Surface Temperature Imagery for Urban Analysis", pages: 1253 - 1257
KERAMITSOGLOU, IPHIGENIA: "Global Urban Monitoring and Assessment through Earth Observation", 2014, article "Investigations of the Diurnal Thermal Behavior of Athens, Greece, by Statistical Downscaling of Land Surface Temperature Images and Pattern Analysis", pages: 249 - 265
KERAMITSOGLOU, IPHIGENIA: "Advanced earth observation methodologies for the study of the Thermal Environment of Cities", PROCEEDINGS OF SECOND INTERNATIONAL WORKSHOP ON EARTH OBSERVATION AND REMOTE SENSING APPLICATIONS (EORSA, 2012, pages 11 - 15
KOLIOS S.; G. GEORGOULAS; C. STYLIOS: "Achieving downscaling of Meteosat thermal infrared imagery using artificial neural networks", INTERNATIONAL JOURNAL OF REMOTE SENSING, vol. 34, no. 21, 2013, pages 7706 - 7722
KUSTAS, W. P.; J.M. NORMAN; M.C. ANDERSON; A.N. FRENCH: "Estimating subpixel surface temperatures and energy fluxes from the vegetation index-radiometric temperature relationship", REMOTE SENSING OF ENVIRONMENT, vol. 85, 2003, pages 429 - 440
LIU, D.; R. PI: "Downscaling thermal infrared radiance for subpixel land surface temperature retrieval", SENSORS, vol. 8, no. 4, 2008, pages 2695 - 2706
MERLIN, 0.; D. DUCHEMIN; O. HAGOLLE; F. JACOB; B. COUDERT; G. CHEHBOUNI; G. DEDIEU; J. GARATUZA; Y. KERR: "Disaggregation of MODIS surface temperature over an agricultural area using a time series of Formosat-2 images", REMOTE SENSING OF ENVIRONMENT, vol. 114, 2010, pages 2500 - 2512
PYNE S.J.; P.L. ANDREWS; R.D. LAVEN: "Introduction to Wildland Fire", 1996, JOHN WILEY & SONS, INC
STATHOPOULOU M.; C. CARTALIS: "Downscaling AVHRR land surface temperatures for improved surface urban heat island intensity estimation", REMOTE SENSING OF ENVIRONMENT, vol. 113, no. 12, 2009, pages 2592 - 2605
WENG, Q.; D. LU; J. SCHUBRING: "Estimation of land surface temperature-vegetation abundance relationship for urban heat island studies", REMOTE SENSING OF ENVIRONMENT, vol. 89, 2004, pages 467 - 483
YANG, G.; R. PU; C. ZHAO; W. HUANG; J. WANG: "Estimation of subpixel land surface temperature using an endmember index based technique: A case examination on ASTER and MODIS temperature products over a heterogeneous area", REMOTE SENSING OF ENVIRONMENT,, vol. 115, no. 5, 2011, pages 1202 - 1219
YANG, G.; R. PU; W. HUANG; J. WANG; C. ZHAO: "A Novel Method to Estimate Subpixel Temperature by Fusing Solar-Reflective and Thermal-Infrared Remote-Sensing Data With an Artificial Neural Network", IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, vol. 48, 2010, pages 2170 - 2178
ZAKšEK, K.; K. OSTIR: "Downscaling land surface temperature for urban heat island diurnal cycle analysis", REMOTE SENSING OF ENVIRONMENT, vol. 117, 2012, pages 114 - 124
ZAKšEKA, K.; M. SCHROEDTER-HOMSCHEIDT: "Parameterization of air temperature in high temporal and spatial resolution from a combination of the SEVIRI and MODIS instruments", ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING, vol. 64, 2009, pages 414 - 421
ZHOU, J.; Y. CHEN; J. WANG; W. ZHAN: "Maximum nighttime urban heat island (UHI) intensity simulation by integrating remotely sensed data and meteorological observations", IEEE J. SEL. TOP. APPL. EARTH OBS. REMOTE SENS., vol. 4, 2011, pages 138 - 146
DIMITRAKOPOULOS, A.P.; V. MATEEVA; G. XANTHOPOULOS: "Fuel models for Mediterranean vegetation types in Greece", GEOTECHNICAL SCIENTIFIC ISSUES (GEOTECHNICAL CHAMBER OF GREECE): SERIES VI, vol. 12, no. 3, 2001, pages 192 - 206
CABALLERO, D.; G. XANTHOPOULOS; D. KALLIDROMITOU; G. LYRINTZIS; M. BONAZOUNTAS; P. PAPACHRISTOU; O. PACIOS: "FOMFIS: Forest fire management and fire prevention system", PROCEEDINGS OF THE INTERNATIONAL SYMPOSIUM ON ''FOREST FIRES: NEEDS AND INNOVATIONS, 18 November 1999 (1999-11-18), pages 93 - 98
Download PDF:
Claims:
It is a method implemented on a computer to detect, in real time, areas with active fire hotspots by analyzing satellite images which include:

a. An acquisition step of satellite images which are composed by pixels, b. A step combining the above data with vegetation fuel data, comprising at least the vegetation density data and the vegetation fuel type.

It is a method, according to claim 1, which comprises a step of subdividing each pixel of the satellite images into sub-pixels of smaller dimension due to the smaller size of fuel data aiming to significantly improve the spatial resolution of the satellite observation.

It is a method computing the fire probability indicator comprising the method of claim 2 and further comprising:

a. a step of classifying each pixel of satellite images data either to a burning region pixel or to a non-burning region pixel comprising:

i. a sub-step in which a channel with a wavelength in the spectral range of the infrared, is being isolated in each satellite image,

ii. a sub-step comparing the temperatures of the pixels of the previous channel with a set of dynamically adaptive thresholds resulting to the classification of the pixel as burning or non-burning,

b. a step of computing, in each sub-pixel, the fire probability indicator which is a numerical value that is a function of at least the vegetation fuel parameter of the sub-pixel and of the classification of this sub-pixel, calculated in the previous step.

It is a method for computing the most probable fire ignition point comprising the method of claim 3 and further comprising a spatio-temporal clustering procedure which in each timestamp of the satellite dataset creates groups of pixels of burnt areas forming fire ^snapshots that are interlinked to time creating clusters, each cluster representing a discrete fire.

It is a method, according to claim 4, which comprises in each timestamp:

a. a step of computing the geometric center for each cluster snapshot,

b. a step, in each cluster, of computing a position-weighted geometric center, based on all the geometric centers of the different instances of the same cluster, with weights to decrease as far as the value of the timestamp of each snapshot increases,

c. a step, in each snapshot of each cluster, of finding the maximum distance of the geometrical centers of the sub-pixels that constitute the snapshot from the position computed in the previous step,

d. a step, in each sub-pixel, of normalization of its distance from the position of step b to the value of step c,

e. a step, in each sub-pixel, of detecting the fuel polygon with the maximum area which touches on that sub-pixel,

f. a step, in each snapshot of each cluster, of finding the maximum of the areas of the previous step,

g. a step, in each instance of each cluster, of normalization of the areas calculated in step e with the maximum area of the previous step, h. a step, in each snapshot, of computing the most probable fire ignition point, after calculating the maximum value of a function applied to each sub-pixel that at least accepts as parameters the following:

i. the fire probability indicator calculated in the second step of the third claim, ii. the normalized value of the abovementioned step d,

iii. the normalized value of step g,

iv. and meteorological data concerning this sub-pixel.

It is a method, according to claim 5, wherein in each cluster the most probable fire ignition point is selected only by those sub-pixels that were sub-pixels of the clusters initial snapshot.

It is a method, according to claim 6, where the meteorological data includes at least the wind vector data.

It is a method for computing the fire probability which comprises the method of claim 7 and further comprises in each sub-pixel an estimation step of fire dispersion, and a step of comparison of the results of the previous step with satellite observations.

Description:
Method that detects areas of active fire hotspots in real-time, calculates the most probable ignition point and assesses fire probability indicators, using satellite images and fuel data. Technical Field

The present invention belongs to the fields of remote sensing, digital image analysis and geographic information systems. The invention proposes a method which identifies in realtime areas with active fire hotspots and calculates the most probable fire ignition point, as well as, it assesses fire probability indicators on subdivisions of the region, using satellite images and fuel data. More specifically, it is an innovative method of dynamic detection of forest areas with active fire hotspots; in particular areas which emit significant amount of heat thus can be detected by applying the appropriate pixel based processing on satellite images. This method is implemented for digital processing and besides satellite images, it also makes use of vegetation fuel data, weather forecasting and terrain (in particular topography) data in order to estimate the fire probability indicators, which are expressed as numerical values indicating the probability of a fire hotspot occurrence in areas which are subdivisions of the original pixels of satellite images, that from now on will be referred to as sub-pixels.

Vegetation fuel is considered to be the whole living or dead organic vegetative material that is either on the ground (as ground foliage, needles, twigs, logs, grass, bushes, shrubs and trees) or on trees (such as branches, foliage, standing dead trees) which is prone to spread or start a fire after ignition (Pyne et al.). The vegetation fuel data that the present invention utilizes come along with information which has derived from laboratory measurements of the physical and chemical properties of representative vegetation samples, covering a wide range of forest ecosystems in Greece. Dimitrakopoulos, Mateeva and Xanthopoulos, resulted in twelve models of vegetation fuel in Greece. According to Dimitrakopoulos, Mateeva and Xanthopoulos, the vegetation fuel data model is defined as "a typical cluster of vegetation fuel, whose quantification of physical and chemical parameters is representative of the typical situation of a particular type of vegetation."

From the analysis of the above data, indicators are derived related to the vegetative material and its density, which in turn quantifies the flammability of a geographical area. More specifically, flammability is defined as the quantity which indicates the vulnerability of a geographical area to wild fire hazards and is related to the density and the vegetation fuel type (that in turn is related to vegetation type), since the same amount of different species of trees, characterized by different fuel type, has different level of flammability. On the contrary, areas characterized by the same species of vegetation may have a different flammability levels because of vegetation density, quantity and morphology (e.g. plant height) or moisture content. The type, quantity, chemical composition, size and shape, as well as the spatial consistency of vegetation fuel are important factors that contribute to the ignition and spread of a fire outbreak.

State of the art

Real-time fire detection applications make use of satellite data derived from geostationary satellites (e.g. MSG-SEVIRI). The advantage of geostationary satellites is that they provide images at high temporal frequency. The disadvantage is the large satellite imagery pixel dimension thus the low observation resolution, not allowing the precise spatial detection and monitoring of a forest fire as well as the estimation of the most probable fire ignition point. Recent studies about the improvement of the spatial analysis of satellite imagery used by fire monitoring applications are mentioned below.

Deneke and Roebling make use of the high resolution MSG-SEVIRI spectral channel to improve the lower resolution channels. Kolios et al. use Artificial Neural Networks to improve the spatial resolution of MSG-SEVIRI images.

Several studies have been conducted (e.g. by Kustas et al., Weng et al., Agam et al., Zaksek and Shroedter-Homescheidt, Merlin et al., Stathopoulou and Kartalis, Zhou et al., Hutengs and Vohland, Liu and Pi, Keramitsoglou, Keramitsoglou et al.) to improve the analysis of digital imagery data related to the Earth's surface temperature (LST - Land surface temperature), through satellite processing of images and the use of auxiliary satellite and non- satellite sources of high resolution data.

In some studies (e.g. Yang et al., Zaksek and Ostir), the improvement of digital images spatial resolution which encode the temperature of the earth's surface, use auxiliary data related to vegetation indicators. However, it is noted that those indicators are quantitative expressions resulting from the difference of visible and near infrared reflectance, used for the classification of vegetated and non- vegetated areas, the estimation of the plant health state, as well as the detection of phenological changes, thus they do not relate the vegetation fuel data to the active fire hotspot detection.

Management and Prevention of Forest Fires FOMFIS (Lyrintzis et al.) system, uses also vegetation fuel data (data of categorized fuels), but in contrast to the present invention, FOMFIS does not make use of satellite data to identify areas with active fire hotspots. FOMFIS is a computer-based system for the prediction and extinction of forest fires, that uses fuel data (only vegetation fuel type data but not fuel density data), which are generated by semi-automatic photo interpretation of satellite images.

Other studies refer to the pixel-based classification of satellite images into burning and non-burning classes. An example is the classification algorithm of EUMETSAT (an international organization that manages the family of the METEOSAT geostationary satellites). EUMETSAT algorithm uses two spectral channels of METEOSAT geostationary satellites located in the infrared part of the electromagnetic radiation spectrum and calculates their correlation, in order to track those satellite image pixels that correspond to burning areas. EUMETSAT algorithm, unlike the present invention, does not use fuel, topography and weather forecast data, nor does it estimate the fire probability in subdivisions of the original pixels of satellite images, or the most probable fire ignition point.

Invention summary.

The invention proposes a method implemented for digital processing, to detect in real time, areas with active fire hotspots at a resolution higher than that of the initially available satellite imagery, including:

a. An acquisition step of satellite images which are composed by pixels,

b. A step combining the above data with fuel data, comprising at least the density and the vegetation fuel type (that in turn is related to vegetation type), topography and weather forecasting data. The method includes a step of subdividing each pixel of satellite images into sub-pixels of smaller dimension, due to the smaller dimension of the fuel data, thus intending to significantly improve the spatial resolution of the initial satellite observation.

The invention also includes a method of computing the fire probability indicator, a method of computing the most probable fire ignition point and a method of reassessing the fire probability indication in real-time, by taking into account its most probable fire ignition point.

Brief description of schemes

Figure 1 depicts the implementation of the individual computation steps of the present invention by making use of four different clusters of shapes.

The ellipses symbolize the beginning and the end of the processing chain depicted on the figure.

The square shapes symbolize each major and secondary algorithmic process of the invention. The side rectangular shapes represent the output datasets generated from the execution of the invention's algorithmic processes, as well as the satellite data used as their input.

The shapes that resemble the logic "OR" gate symbolize all the other algorithmic processes input data, i.e. meteorological, topography, and vegetation fuel data. More specifically:

Dl, in Figure 1, corresponds to Meteorological Data.

D2, in Figure 1, corresponds to Topography Data.

D3, in Figure 1, corresponds to Vegetation Fuel Data.

Letter A in Figure 1, corresponds to the "Algorithm for fire detection based on the classification of satellite data imagery pixels. "

Letter B in Figure 1, corresponds to the "Sub-routine for applying the spatio-temporal correction criteria to each pixel classified at step A."

Letter C in Figure 1, corresponds to the "Algorithm for calculating the fire probability indicator for each sub-pixel, resulting in the significant improvement of the initial spatial resolution."

Letter D in Figure 1 , corresponds to the "Algorithm for reassessment of the fire probability for each sub-pixel with evidence derived from the fire dispersion model".

Letter F in Figure 1, corresponds to the "Algorithm of identifying the most probable fire ignition point."

Letter G, in Figure 1, corresponds to step "Computation of the fire dispersion model." Letter S in Figure 1, indicates the 'Start' of the automated implementation of the invention. Letter E in Figure 1 , indicates the "End" of the automated implementation of the invention. Number 1 in Figure 1 symbolizes the "Satellite imagery datasets" that are used at step A for the active fire detection and monitoring.

Number 2 in Figure 1 , symbolizes the "Geospatial data resulting from the classification of satellite data pixels in burning/non-buming" at step B.

Number 3 in Figure 1, symbolizes the "Geospatial data encoding the values of the fire probability indicators at sub-pixel resolution" resulting from step C.

Number 4 in Figure 1, symbolizes the "Geospatial data encoding the improved fire probability indicators accrued after the second stage reassessment of fire probability indicator at sub-pixel resolution" applied at step D. Number 5 in Figure 1 corresponds to the "Geospatial data encoding the results derived after applying the fire dispersion model" at step G.

Disclosure of the invention

As already mentioned under chapter "state of the art", there is no method that at the same time makes use of satellite data images and vegetation fuel data for real-time fire monitoring at sub-pixel resolution. In its optimal implementation, the invention is designed to additionally use weather forecasting data, as well as area specific topographic data, in combination with fire dispersion models, thus targeting to improve the initial spatial resolution of the satellite data. In any case, the purpose of this invention is first to identify in real-time areas containing active fire hotspots and then to monitor the dispersion of those fire events, also in real time. While the invention covers this need, it additionally calculates the fire existence probability in any sub-area of the affected geographical area, i.e. a fire probability indicator as well as the most probable fire ignition point.

In summary, after classifying satellite images in burning and non-burning areas, the invention exploits the high spatial resolution of the vegetation fuel data and based on that, it applies a subdivision of each pixel of satellite data into sub-pixels of smaller dimension, i.e. of higher spatial accuracy. The aforementioned procedure aims to calculate and map the fire probability indicator to a sub-pixel spatial resolution level of each initial pixel comprising the satellite image.

At the same time, besides calculating the fire probability indicator, the invention computes the most probable fire ignition point through a process which comprises a complex spatio-temporal analysis of satellite observations as well as the clustering of burning pixels. This process also performs the calculation of the geometric center of each cluster snapshot, the combination of these clusters with the fire probability indicator of each sub-pixel that was estimated before, the available fuel material ahead of each burning sub-pixel, in wind direction with the simultaneous use of meteorological forecast datasets. It is pointed out that all of the above concur in realtime.

In conclusion, the invention primarily addresses the issue of the initial low spatial resolution of geostationary satellite images which are used for real-time fire monitoring and at the same time achieves a reliable estimation of the most probable fire ignition point, aiming to redefine the fire indicator, based on that estimation.

Subsequently, details of the invention are being further analyzed:

Given that fire incidents are dynamic events which change over time, the real-time satellite monitoring is crucial. This need is met by the use of geostationary satellites in combination with polar orbiting satellites. Geostationary satellites are characterized by synchronous rotation to the earth, with constant speed and altitude, therefore having the advantage of always facing the Earth from the same perspective, providing continuous and real-time data collection for the same spatial coverage. The disadvantage of geostationary satellites is their great distance from Earth, that leads to satellite images with low spatial resolution, which makes the spatially accurate detection of fire outbreaks almost impossible.

To address this contradiction, this invention refers to a method implemented for digital processing, in order to detect, in real-time, areas with active fire hotspots through analyzing satellite images. This method contains a step of acquisition of satellite imagery consequently pixel consisted data and a step of combining the above data with fuel vegetation datasets, which include at least vegetation density data, as well as classified as fuel vegetation data, based on the type of vegetation from which they derive and also topographical and meteorological data. Considering the finer resolution of the fuel data and aiming to significantly improve the initial spatial resolution of satellite observation data, this method includes a step which subdivides each pixel of satellite images into sub-pixels of smaller size. In more technical terms and in order to implement the above steps, the invention refers to a method for calculating the fire probability indicator which comprises a method for identifying active fire hotspots, as follows:

a. a classification step of each pixel of satellite images data (Figure 1.1) either to a burning region pixel or to a non-burning region pixel which includes:

i. a sub-step in which a channel with a wavelength in the spectral range of the infrared, is being isolated in each satellite image,

ii. a sub-step of comparing the temperatures of the pixels of the previous channel with a set of dynamically adaptive thresholds resulting to the classification of the pixel as burning or non-burning,

b. a step of calculating, for every sub-pixel, the fire probability indicator, which is a numerical value depending at least from the vegetation fuel index characterizing the sub-pixel and from the classification calculated at the previous step.

Following, step a is analyzed in detail (Figure 1A). The infrared part of the electromagnetic radiation is being used for the detection of fire events. The initial separation into burning and non-burning pixels can be implemented with only a single channel which is in the mid-infrared region of the electromagnetic spectrum and enables the ability to detect thermal irregularities within the pixel of the satellite image. However, the instructions of the EUMETSAT organization could be followed in the present invention, by using two channels from the infrared region of the electromagnetic radiation. The abovementioned channel which is located in the mid-infrared region part of the electromagnetic spectrum, and one more, located in the thermal-spectrum of the electromagnetic spectrum.

The pixels of satellite images acquired, are classified into burning and non-burning by using a variation of the proposed by the EUMETSAT classification algorithm. The classification is based on using four thresholds which concern:

the temperature of the pixel as it is estimated by the channel in the mid- infrared region of the electromagnetic spectrum,

the difference between pixel temperatures resulting from the channels in the mid-infrared region and in the thermal-spectrum electromagnetic radiation region,

■ and the standard deviation of the temperature measured by the channel in mid-infrared region, for a 3x3 pixels area window around the examined pixel, and ■ the standard deviation of the temperature measured by the channel in the thermal infrared, for a 3x3 pixels area window around the examined pixel as well.

A variation of the EUMETSAT classification algorithm that could be utilized, concerns its adaptation to the geographical conditions of Greece. More specifically,

New thresholds of fire hotspots detection were estimated, that are adjusted to the surface coverage, to the land use and to the climate conditions that indicate the vulnerability of a Greek area to fire outbreaks, as well as regions of the Mediterranean with similar climatic conditions, using at the same time the analysis of time-series of fire events and the empirical method of trial and error in order to minimize false alarm event and miss-detected real fire events.

Moreover, dynamic adjustment processes for the above defined thresholds were adopted, in order to adapt them for each new satellite image and to each pixel of that image. The dynamic threshold redefinition is a function of the geographical position of the area (φ, λ), of the zenith angle of the sun and at the same time of the height and angle of the sun at the satellite acquisition time. In conclusion, thresholds change/adjust depending on the geographical coordinates of the pixel and the time of the satellite image acquisition.

It follows a description of the optimization process applied to the thematic and geospatial information derived after the primary classification of the pixels (Figure 1.B), in order to complete step a and obtain the final improved product of the classification of the satellite image pixels to burning and non-burning (Figure 1.2).

The segmentation between the burning and the non-burning pixels can result in three classes, the burning with 100% confidence level, the non-burning also with a confidence level of 100%) and the doubtful burning pixels, with a confidence level of 50%.

Some pixels are displayed as burning in one timestamp, as non-burning in the next timestamp, again as burning in the next and so on so forth. To avoid this phenomenon, for each timestamp, the time sequence of occurrence of each pixel is being examined as burnable or non-burnable in the past, in order to decide whether this will be eventually classified as burning or non-burning for the current timestamp. For this purpose, a moving average filter is applied to the burnt pixels with a fixed time step of 5 minutes and with a 30 minute time window width. More specifically, the filter is applied to the pixels classified as burning, at least once in the last 30 minutes. The final assessment for each pixel as burning, results from the average of observations within the 30 minutes time window, in which it was characterized as burning by the classification algorithm. After this, an additional time filter validating the observations as true is applied, only for those which the entire pixel appears as burning, in at least two of the timestamps within a time frame of 15 minutes.

The large size of the pixels is an obstacle to thematic and spatial accuracy of the classified pixels and introduces errors in the initial classification of the pixel. It is possible that part of a burning pixel could geographically coincide with the sea area. To avoid such errors, there is applied an additional step in order to improve the spatio- temporal classification of the satellite images. To this end, only the part of the pixels which geographically coincide with land is saved, by applying a Sea/Land "mask" to the original satellite images. Similarly, the pixels on land which happen to be characterized by land use with zero or small fire proneness are being removed based on land use maps.

Next Step b is analyzed in detail, which is implemented firstly to improve the low spatial resolution of geostationary satellite images and to achieve the increase of the spatial resolution of observations to sub-pixel level (Figure l .C), and secondly, to estimate the locations of that most likely fire hotspots occur (Figure l .F). To achieve this, the present invention combines the acquired satellite images with vegetation fuel data (Figure 1.D3) as well as with meteorological (Figure 1.D1) and topographical data (Figure 1.D2), that are described in detail below.

In the past, before the implementation of this proposed methodology, i.e. the combined use of the satellite images data and fuel data with high spatial resolution (ideally also including data of meteorology and terrain morphology), the sub-division of the pixel data of satellite images into sub-pixels had no value. Now, due to the higher spatial resolution of fuel data and the above mentioned data which import separate indications concerning fire occurence, the process of subdivision into sub-pixels becomes meaningful and has special value since it provides us with indications of higher accuracy for the ability or disability of a sub-area to be burned, therefore increasing the spatial resolution of the observation. However, as the fuel data might be at very high spatial resolution, an intermediate level of subdivision between the pixel dimension and the fuel data dimension should be chosen, in order to avoid excessive analysis to an ineffective extent. In case of real-time forest fire hotspots detection, the optimal resolution that has been used in this invention is 500 meters.

More specifically, in order to achieve optimized spatial resolution than that of the observation of the initially received satellite images, each burning or possibly burning pixel is divided into sub-pixels, for each of which an indicator of fire probability is calculated as a first step. That is a new numerical value derived from the correlation of a) the level of confidence of the initial classification for the entire burning pixel, with b) additional quantified indicators in sub-pixel level, which are related to vegetation and terrain morphology related parameters. It is pointed out that fuel vegetation data used to improve the spatial resolution refer to the category of vegetation fuel data as well as on its density.

The geospatial distribution of vegetation fuel data models as already mentioned is used as input data to the invention in the form of an up-to-dated digital map, created for this purpose, and which illustrates in detail the vegetation fuel data in high spatial resolution. The vegetation fuel data map comprises of seven segments, in each of which a weight factor is attributed that indicates the degree to which each corresponding class is prone to fire and affects the fire transmission at a particular sub-pixel as well as its neighborhood. Because of the fact that each sub-pixel may contain more than one category of vegetation, the weight factor that is eventually attributed to the sub-pixel is calculated as the weighted average of the weight factors corresponding to each individual category of fuel data contained within. It is calculated as the fraction that has the sum of the individual products of the surface of each of the categories contained as a numerator, for the factor of the vegetation of the category and as a denominator the total surface of the sub-pixel.

Similarly to the case of fuel vegetation category, for each vegetation density category is assigned another weight factor. The density value assigned to each sub-pixel results as before, by the weighted average of the weight coefficients of each fuel density class contained within.

As the fire expands mainly on the ground, alterations in topography can cause drastic changes to the behavior of a fire. Consequently, the topographic data are used as additional evidence for the assessment of the fire probability indicator in each of the sub-pixels of the satellite imagery. More specifically, the topographic parameters taken into consideration are the following:

Slope: An example of the influence of terrain slope in the evolution of the fire is the case of fire moving on topographic acclivity. High temperature developed in the lower topographical points causes the drying of vegetation located in the upper parts, thus facilitating the fire dispersion.

■ Altitude: Higher the altitude, lower the temperature and higher the relative humidity, which reduces the probability of incidence and progression of fire at higher altitudes.

■ Orientation of slope: The slope orientation is divided to five sub-categories: i) North, ii) South, iii) West, iv) East and v) Flat region. Different concentration and quality of vegetation fuel is noticed depending on the exposure of the slope regarding the horizon. More specifically, on the northern slopes, which receive less solar radiation, the fuel is cooler and moister than on the other slopes. Fuel dries faster in the southern and southwestern slopes, as hotter and dryer climatic conditions prevail. Eastern slopes are heated significantly during forenoon, while western slopes receive solar radiation until sunset.

The topographic data ensue from the appropriate processing of the digital terrain model, which comes from the ASTER satellite, and has a 30m pixel dimension (i.e. spatial resolution).

For each of the above topography data, the initial dimension of the 30m pixel is readjusted to the dimension of the examined sub-pixel by the process of image resampling, in which:

i. The slope of the reconstructed sub-pixel is calculated as the average of inclination angles of the 30m pixels which constitute it.

ii. The orientation of the reconstructed sub-pixel corresponds to the orientation of the majority of the 30m pixels which constitute it.

iii. The altitude of the reconstructed sub-pixel results from the average of the altitudes of the 30m pixels which constitute it.

As in the case of vegetation fuel data, for each of the aforementioned categories of topographic data, is assigned a weight factor representing the facilitation of occurrence, preservation, and expansion of the fire to the nearby areas.

Up to this point, the calculation of the fire probability indicator in each sub-pixel uses a series of weight factors, arising from the supplementary fuel data, as well as from topographic data. Specifically, in each sub-pixel, the fire probability indicator is being calculated in the first stage (Figure 1.3). Moreover, it results in conjunction to the level of confidence computed at the initial classification of the satellite image with evidence derived from the parameters of slope, altitude, slope orientation, fuel category and fuel density.

Subsequently, the invention computes the most probable fire ignition point (Figure 1.Z). This geographical location is necessary for the calculation and capture of the fire- event's spatio-temporal evolution, by applying a dispersion model providing information which is then correlated with the observations of the satellites. This process is an independent method of the present invention and aims to the computation of the most probable fire ignition point. It comprises the computation method of the aforementioned fire probability indicator, and also implements a spatiotemporal clustering process that creates groups of pixels of burning areas for each timestamp of satellite data, which in turn comprise fire snapshots and are interlinked in time creating clusters, while each cluster represents a discrete fire. This method additionally comprises a step for the selection of the most probable fire ignition point in each cluster, and only of those sub-pixels that were also sub-pixels of the clusters initial snapshot. Furthermore, the method includes at least the wind vector as meteorological data.

The first step of the method is the identification of the geographical distribution of burnt pixels. Each set of adjacent burning pixels, which appear for the first time and are less than a threshold distance, are grouped in a cluster with specific geographical area and spreading and are considered to constitute an independent fire event under a common timestamp of beginning and ending. There are several clustering methods but in the present invention direct spatio-temporal clustering (online spatio-temporal clustering) was used.

In case of a group of pixels classified as burning during the observation of the current timestamp is evaluated, with appropriate spatio-temporal criteria, to be intersecting with a fire event corresponding to an observation of a previous timestamp within a predetermined time offset, then the method requires merging of their geometric imprints and updating of the timestamp of the fire event to the time of the last observation. In case of a burning pixel group that is not spatially and temporally associated with another fire event, then this group defines and constitutes a new fire event.

In more technical terms, and in order to identify the most probable ignition point of the fire, the results of direct spatio-temporal clustering are being utilized, and the above method further includes, in each timestamp, the following steps:

a. a step of computing the geometric center for each cluster snapshot,

b. a step, in each cluster, of computing a position-weighted geometric center, based on all the geometric centers of the different instances of the same cluster, with weights to decrease as far as the value of the timestamp of each snapshot increases,

c. a step, in each snapshot of each cluster, of finding the maximum distance of the geometrical centers of the sub-pixels that constitute the snapshot from the position computed in the previous step, d. a step, in each sub-pixel, of normalization of its distance from the position of step b to the value of step c,

e. a step, in each sub-pixel, of detecting the fuel-on polygon with the maximum area which touches on that sub-pixel,

f. a step, in each snapshot of each cluster, of finding the maximum of the areas of the previous step,

g. a step, in each instance of each cluster, of normalization of the areas calculated in step e with the maximum area of the previous step,

h. a step, in each snapshot, of computing the most probable fire ignition point, after calculating the maximum value of a function applied to each sub-pixel that at least accepts as parameters the following:

i. the fire probability indicator calculated above,

ii. the normalized value of the abovementioned step d,

iii. the normalized value of step g,

iv. and meteorological data concerning this sub-pixel.

Next, the aforementioned steps of the method are being analyzed in detail. For all the above defined, distinct from each other, fire events their geometric center is recorder for the current timestamp. A new weighted position with geometric center p is calculated for the current timestamp considering all the recorded geometric centers of each fire event (cluster). The normalized distance A of each sub-pixel that composes the cluster to the position p is also computed for the current timestamp.

For each sub-pixel, polygons with fuel, that geometrically touch it, are specified. The one with the largest area e is isolated. The maximum of e is computed, denoted as e^, and then for each sub-pixel of the current timestamp, a normalized area E is computed by the fraction e/ e max .

The calculation of the distance A and the area E is necessary in order to determine the most probable ignition point.

Beyond fuel and topography, prevailing weather conditions at a burning area, influence significantly the fire spreading. The wind is the dominant factor determining the rate of dispersion and the fire direction. Moreover, when computing the most probable ignition point, the prevailing wind conditions in the burning area must be taken into account. Therefore, the methodology takes into account the available fuel vegetation ahead each burning sub-pixel according to the wind direction. For this reason, in each sub-pixel, and for each timestamp, a matrix of spatial weights having specific spatial form (specific spatial arrangement of its elements), in a way that it is oriented according the wind direction, is applied. The central element of this matrix coincides spatially with the geographic center of each sub-pixel to which it applies. When applying the matrix, and based on its elements that constitute the weights in the calculation, a new weighted average, which represents a new weighted fire probability indicator for the sub-pixel in the first stage , is calculated. The values of the elements of the matrix must follow certain rules in order to properly function as weights and to support the observations that are located towards the front of the fire. Specifically, they decrease as these elements are by distance eliminated from the application point of the wind vector which coincides with the geometric center of the central element of the matrix. The values of these elements are also reduced as these elements are removed at an angle from the direction of the wind vector and are set to zero when the last angle exceeds ninety degrees.

Based on the above, the computing of the most probable fire ignition point is done for each snapshot, based on the maximum value of a function, which is applied to each sub-pixel and considers as parameters the following:

i. The first stage fire probability indicator of the sub-pixel as calculated above.

ii. The normalized distance d.

iii. The normalized area E.

iv. The first stage new weighted fire probability indicator for the sub-pixel

In each independent fire event, the most probable ignition point IP belongs to one of those sub-pixels which spatially belong to the first snapshot of the fire event, have high rates of the first stage fire probability indicator, have high availability of vegetation fuel max E, according the wind direction max f, and are their distance from the weighted geometric center p, is minimal min A.

Next, the invention implements a method (Figure 1.D) of redefinition of the fire indicator (Figure 1.4) based on the most probable location of the fire ignition point and on the estimation of its spatial dispersion (Figure 1.5) and the fire estimated arrival time, using, among other data, the location of this point.

To implement this process, the invention comprises a computing method of fire probability, which includes the above method of computing the most probable fire ignition point, then a step of estimating the fire dispersion considering that point (the most probable fire ignition point) and finally, for each sub-pixel a step of comparing the results of the previous step with satellite observations.

Therefore, it becomes clear that a key step of the present invention also relates to the estimation of the fire dispersion. To assess the fire dispersion FlamMap software (Finney) was chosen. The FlamMap (Figure l .H) computes the fire dispersion at each timestamp and for each independent fire event using the following data:

i. Speed and direction of wind.

ii. The most likely ignition point IP.

iii. The aforementioned survey data.

iv. Vegetation fuel data.

The result of FlamMap is a vector dataset which encodes the arrival times of the fire fronts at defined areas, as predicted by the software FlamMap or any other software that has the ability to export relevant information. These data are used to redefine the fire probability indicators of each sub-pixel at a second step. That is for each fire instance detected by the satellite, the values of fire probability indicators computed at first stage are initially spatially and temporally compared with the results of FlamMap.

The initial value of the fire probability indicator at first stage is degraded or enhanced based on the model by using two factors. The first factor enhances the fire possibility in the sub-pixels which spatially and temporally coincide with the areas estimated by the fire dispersion model. The second factor determines the level of enhancement or degradation of the fire probability in each sub-pixel of a snapshot, by estimating the temporal distance between the timestamp of the snapshot from the initial timestamp of used by the dispersion model. This means that the reliability of the model decreases exponentially with time. For example, the prediction of the model is considered reliable 15 minutes after the start of the calculations, and up to 1 - 2 hours ahead, but a forecast is not considered reliable when referring to 300 minutes (-4-5 hours) ahead. Based on the above and through advanced image analysis and exploitation of evidence derived from external sources, the invention achieves to dynamically disseminate to the crisis management authorities, information on wildfire evolution at sub regions of 500m x 500m, every 5 minutes, increasing by approximately 50 times the spatial resolution provided, taking into account the spatial resolution of the initial satellite observation by geostationary satellites.

The invention may comprise by the method that identifies areas with active fire hotspots, additional satellite data. In particular, there is provision that the system is supplied with external additional and reliable information for the fire location, taking advantage of higher spatial resolution satellite observations derived from polar orbiting satellite missions. From these satellite observations, footprints of areas with active fire hotspots will be exported, that will feed the processing chain, further improving the fire probability estimation indicators.

References

Agam N., W. P. Kustas, M. C. Anderson, F. Li, C. M. U. Neale (2007). A vegetation index based technique for spatial sharpening of thermal imagery. Remote Sensing of Environment, vol. 107, no. 4, pp. 545-558.

Deneke, H., R. Roebling, 2010. Downscaling of METEOSAT SEVIRI 0.6 and 0.8 micron channel radiances utilizing the high-resolution visible channel. Atmospheric Chemistry and Physics, vol. 10, pp. 9761 -9772.

EUMETSAT. Active fire monitoring with MSG algorithm - theoretical basis document. Darmstadt, 07. 20

Finney, Mark A. (2006). An Overview of FlamMap Fire Modeling Capabilities. Included in: Andrews, Patricia L.; Butler,Bret W., comps. 2006. Fuels Management- How to Measure Success: Conference Proceedings. 28-30 March 2006; Portland, OR. Proceedings RMRS-P-41. Fort Collins, CO: U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, p. 213-220

Hutengs C, M. Vohland (2014). Downscaling land surface temperatures from MODIS data to mesoscale resolution with Random Forest regression. Gemeinsame Tagung 2014 der DGfK, der DGPF, der GfGI und des GiN (DGPF Proceedings 23 / 2014).

Keramitsoglou, I., C.T. Kiranoudis , Q. Weng (2013). Downscaling Geostationary Land Surface Temperature Imagery for Urban Analysis. Geoscience and Remote Sensing Letters, IEEE, vol. 10 , no. 5, pp. 1253 - 1257.

Keramitsoglou, Iphigenia (2014). Investigations of the Diurnal Thermal Behavior of Athens, Greece, by Statistical Downscaling of Land Surface Temperature Images and Pattern Analysis. In: Global Urban Monitoring and Assessment through Earth Observation, pp. 249-265. Editor(s): Qihao Weng

Keramitsoglou, Iphigenia (2012). Advanced earth observation methodologies for the study of the Thermal Environment of Cities, In proceedings of Second International Workshop on Earth Observation and Remote Sensing Applications (EORSA), pp. 11- 15, 8-1 1 June 2012, Shanghai, China.

Kolios S., G. Georgoulas, C. Stylios (2013). Achieving downscaling of Meteosat thermal infrared imagery using artificial neural networks. International Journal of Remote Sensing, vol. 34, no. 21, pp. 7706-7722.

Kustas, W. P., J.M. Norman, M.C. Anderson, A.N. French (2003). Estimating subpixel surface temperatures and energy fluxes from the vegetation index-radiometric temperature relationship. Remote Sensing of Environment, vol. 85, pp. 429-440.

Liu, D., R. Pi (2008). Downscaling thermal infrared radiance for subpixel land surface temperature retrieval, Sensors, vol. 8, no. 4, pp. 2695 -2706

Merlin, O., D. Duchemin, O. Hagolle, F. Jacob, B. Coudert, G. Chehbouni, G. Dedieu, J. Garatuza, Y. Kerr (2010). Disaggregation of MODIS surface temperature over an agricultural area using a time series of Formosat-2 images. Remote Sensing of Environment, vol. 114, pp. 2500-2512.

Pyne S.J., P.L. Andrews, R.D. Laven (1996). Introduction to Wildland Fire, second edition. John Wiley & Sons, Inc. New York.

Stathopoulou M., C. Cartalis (2009). Downscaling AVHRR land surface temperatures for improved surface urban heat island intensity estimation Remote Sensing of Environment, vol. 1 13, no. 12, pp. 2592 -2605.

Weng, Q., D. Lu, J. Schubring (2004). Estimation of land surface temperature- vegetation abundance relationship for urban heat island studies. Remote Sensing of Environment, vol. 89, pp. 467-483.

Yang, G., R. Pu, C. Zhao, W. Huang, J. Wang (201 1). Estimation of subpixel land surface temperature using an endmember index based technique: A case examination on ASTER and MODIS temperature products over a heterogeneous area. Remote Sensing of Environment, vol. 1 15, no. 5, pp. 1202-1219.

Yang, G., R. Pu, W. Huang, J. Wang, C. Zhao,. (2010). A Novel Method to Estimate Subpixel Temperature by Fusing Solar-Reflective and Thermal-Infrared Remote- Sensing Data With an Artificial Neural Network. IEEE Transactions on Geoscience and Remote Sensing, vol. 48, pp. 2170-2178.

Zaksek, ., . Ostir (2012). Downscaling land surface temperature for urban heat island diurnal cycle analysis. Remote Sensing of Environment, vol. 117, pp. 114-124.

Zakseka, ., M. Schroedter-Homscheidt (2009). Parameterization of air temperature in high temporal and spatial resolution from a combination of the SEVIRI and MODIS instruments. ISPRS Journal of Photogrammetry and Remote Sensing, vol. 64, pp. 414— 421.

Zhou, J. ( Y. Chen, J. Wang, W. Zhan (201 1). Maximum nighttime urban heat island (UHI) intensity simulation by integrating remotely sensed data and meteorological observations. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 4, pp. 138-146.

Dimitrakopoulos, A.P., V. Mateeva, and G. Xanthopoulos. 2001. Fuel models for Mediterranean vegetation types in Greece. Geotechnical Scientific Issues (Geotechnical Chamber of Greece): Series VI, vol. 12(3): 192-206 (in Greek).

Caballero, D., G. Xanthopoulos, D. allidromitou, G. Lyrintzis, M. Bonazountas, P. Papachristou, and O. Pacios. 1999. FOMFIS: Forest fire management and fire prevention system, pp. 93-98. In proceedings of the International Symposium on "Forest Fires: Needs and Innovations". November 18-19, 1999, Athens, Greece. Published by CINAR S.A., Athens, Greece, under the auspices of the European Commission DG XII. 41 p.