TECHNICAL FIELD

The field of this invention is that of the assessment of wildfire risks for a given geographical area.

BACKGROUND

Despite the significant uptake in wildfire events over the past years, wildfire risk remains largely under-modeled. The lack of appropriate risk modeling tools to assess the vulnerability of sites at risk has serious implications.

In 2020, the total wildfire insured loss amounted to 14 USD bn for the United States.

Services already exist which provide evaluation of the risk of wildfire for a given area. These services either rely on human experts or use computer tools which rely on past data and are costly in term of data processing.

These tools usually use probabilistic calculations only taking into account past events.

They provide a one-shot estimation and do not allow any monitoring.

They are used as "black boxes" and are limited as to the understanding they provide to the end-user.

They are further usually limited in term of scalability (country, state).

SUMMARY OF THE INVENTION

The invention provides a reliable tool which does not present the drawbacks of the tools known in the prior art.

According to one aspect, the invention provides a method for assessment of wildfire risk for a given geographical area, wherein the following steps are performed:

- acquisition from various data sources of various raw measured data relating at least to said geographical area;

- processing said raw data to extract features;

- processing said features to determine hazard attributes values;

- combining the hazard attributes values to determine at least a risk probability score; - outputting a risk probability assessment for said geographical area based on said risk probability score, and wherein these steps are regularly repeated with refreshed raw measured data to allow a monitoring of said geographical area.

In particular, the output can be a mapping of the risk probability score for the geographical area analysed, with a risk probability assessment associated to each of the pixels of the mapping, the steps been performed for each of said pixels.

Also, alerts can be output depending on the score or scores for the geographical area.

In preferred implementations, the hazard attributes values determined can be discretized; the risk probability score can also be discretized.

As an example, it can be computed through a logistic regression, with weights determined taking into account from past fires. Other calculation/classification can also be contemplated in particular using machine learning techniques.

The steps can be performed by a calculation server for all the geographical regions covered by the data sources and wherein said server outputs a risk probability assessment for a selected geographical area in response to a request including information as to the selected geographical area, said request been transmitted to said server by an end user through an end user terminal.

Selected thresholds to discretize the risk probability score can be modifiable by an end user through an end user terminal, to adapt said end-user needs.

At least a rating characterizing the importance of at least a feature within the risk score is calculated.

Data sources can include sources from the following group: Satellite imagery, Digital Elevation Model, Geolocation, Weather & Climate, Public Data.

By way of example, hazard attributes can include measured parameters from the following group: Vegetation Composite, Water Stress Composite, Land Cover classification. Activity-weighted Roads, Urban Areas, Temperature, Lightning, Vapor Pressure Deficit, Past Precipitations, Power Lines, Rail Network, Camp Fire, Fire History, Slope, Solar Irradiation, etc...

This list is no way to be considered as exhaustive and can be adapted depending on the data sources available and/or depending on the end-user need. BRIEF DESCRIPTION OF THE DRAWINGS

The above and other objects, features and advantages of this invention will be apparent in the following detailed description of an illustrative embodiment thereof, which is to be read in connection with the accompanying drawings wherein: figure 1 is a schematic diagram illustrating a system implementing a possible method according to the invention ; figure 2 illustrates main steps of a possible method according to the invention.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The method proposed allows the monitoring of wildfire risk for a given geographical area.

It can be run on a system as illustrated on figure 1 which comprises a server 1 able to perform the calculations as described hereunder. Server 1 is also adapted to exchange with remote end user terminals 2 and with various data sources 3 though a network 4, which can be an internet network or any other network allowing.

An end-user terminal 2 can be any type of computing device that includes a processor, a memory, and network communication capabilities: mobile phones, personal digital assistants, tablets, laptops, desktops, etc. It includes interface means such as a keyboard and a screen. It can run a specific downloaded application allowing to dialog with server 1. The end-user can also be a server. As an alternative, the dialog with server 1 is directly run by said server 1 through an API allowing the connection and exchanges.

More particularly, server 1 implements the following steps illustrated on figure 2:

- acquisition of the raw data from various sources 3 (step 11 - raw data layer- such as a satellite image or the localisation of the power lines in the region analysed - other examples of attributes are provided in the description which follows) ;

- extraction of features using the raw data of the data sources 3 (step 12 - features can be e.g. water stress or distance from power lines for instance - other examples of features are provided in the description which follows);

- determination of hazard attributes values (step 13);

- determination of a risk model (step 14).

In a possible implementation, the output is a mapping of the risk probability, with a risk probability score associated to each of the pixels of the mapping. The probability score can be converted with a colour matching. The coloured mapping thus obtained can be output through a display on an graphical interface of an end user terminal, with the possibility for the end-user to scroll and zoom on said mapping.

Raw data layer

Data sources 3 are data bases with which the server can exchange.

They can be of various types and cover various regions, including the geographical area of interest for the client. The coverage of said sources is typically worldwide while the calculation performed by the server is also worldwide.

Once the results of this calculation are available at server 1, a request including information as to the geographical area of interest for the end user is transmitted to said server by an end user through an end user terminal 2.

Examples of typical sources are as follows: Satellite imagery, Digital Elevation Model, Geolocation, Weather & Climate, Public Data, etc.

Satellite imagery can be of low or medium resolution and/or of high resolution. Various Kayrros processing can be implemented to enhance the quality of the images such as the one described in the article of Grompone von Gioi, R., Hessel, C., Dagobert, T., Morel, J. M., and de Franchis, C. entitled "CLOUD DETECTION BY INTER-BAND PARALLAX AND A-CONTRARIO VALIDATION" (XXIV ISPRS Congress (2022 edition), 6-11 June 2022, Nice, France) which relates to cloud detection and is hereby incorporated by reference.

Features extracted and Hazard attributes

Features extracted are values of physical parameters which can be extracted for a given resolution from the raw data.

These physical parameters characterize hazard attributes linked to the risk of fire.

In a typical example, around 20 different hazard attributes or more can be processed.

An example of list of hazard attributes is provided hereunder in Table 1. Table 1

Detailed examples of attributes and features are provided in Annex 1 below. The resolution of the mapping output for the Risk model can be chosen to correspond to the smallest resolution for the various attributes or can be higher.

Hazard Attribute Discretization

The features thus extracted from the raw data sources are further processed by server 1 to be discretized on a common scale of values shared for all the hazard attributes.

Through this pre-processing, the hazard attribute values provided to the risk model processing are standardized and comparable.

Decision trees are used to split hazard attribute distributions, using feature values observed for the past based on the location of past fires for the region. By way of example, each feature can be split into five different categories that indicate the level of risk: 1 is very low and 5 is very high.

It is worth highlighting that these levels are relative to the feature itself and not a global measure of the risk.

Machine learning technologies can be used to determine this categorization for each feature.

Risk probability score

After generating the discrete hazard attribute values, server 1 combines them to estimate local probabilities. This step brings together all the discrete hazard attributes previously computed for a given area of the region in a single probability score that directly indicates the level of risk for said area.

Past fires are taken into account to estimate the right parameters for this model. In particular, a certain weight is allocated to each hazard attribute based on their respective impact on fire behaviour in the past.

The local probability score can be computed through a logistic regression which estimates said probability based on the following formula:

With p(x) being the probability per pixel, Hi a discrete hazard attribute value and Wj the weight given to the hazard attribute. 3 is a constant for a given calculation and a given mapping.

It will be further noted that the hazard attribute weight can be adapted by any end-user (via a connection to server 1 through a terminal 2) depending on its own perception of the relevancy of the hazard attribute in the past local fire situations. The end-user has therefore full understanding and control on the computation of a probability for each pixel in the observed area.

Other calculation/classifications than the one based on logistic regression here above described can also be contemplated in particular using machine learning techniques such as Decision Tree/XGBoost.

Interpretation of the model can be further enhanced through the use of Shapley values to highlight the importance of the various features calculated in the score (calculation of an importance rating for each feature).

An example of Shapley values calculation is e.g. described in the article of Benedek

Rozemberczki, Lauren Watson, Peter Bayer, Hao-Tsung Yang, Oliver Kiss, Sebastian Nilsson, Rik Sarkar entitled "The Shapley Value in Machine Learning" (available at: which is hereby incorporated by reference.

Risk Model Discretization

The Risk profiling model provides a continuous score between 0 % and 100 %. To enable comparison and benchmark against the Wildfire Hazard Potential (WHP) distributed by the US Department of Agriculture (USDA), the model is discretised. Further, these provides an easy-to-read output for a human end-user.

The selected thresholds to discretize the score are indicative and can be modified according to the end-user needs. For instance, the definition of the threshold can be adapted to reflect a certain level of risk for a specific business application. An example of thresholds used are detailed below:

0. 00 < r < 0. 05 : Added to not burnable mask 0. 05 < r < 0. 15 : Category 1 0. 15 < r < 0. 30 : Category 2 0. 30 < r < 0. 45 : Category 3 0. 45 < r < 0. 60 : Category 4 0. 60 < r : Category 5

Where r is the computed risk value

This scoring breakdown allows an easy interpretation of the mapping by the end used.

Other thresholds and categories can be of course be contemplated.

Wildland urban interface (WUI)

Wildland urban interface is defined as the area between unoccupied vegetated land and human developed land. Having an updated mapping of the WUI presents a major upside in the context of wildfires. The WUI is usually considered as a very risky area and furthermore it is an area with potential assets of interest. The presence of vegetation means that there is a lot of fuel to burn and the presence of developed land means that the ignition factors are numerous.

Server 1 uses satellite imagery data, as well as geolocation and geospatial data to define these areas detect areas where vegetation and human activity intersect.

Two types of Will are thus defined on the mapping and are specifically marked on the mapping output display:

• Interface Area: Area that demarcates the urban area from the vegetated one. The interface area refers to vegetation that is directly exposed to human activity. This area can grow with the construction of new urban areas over the years, or with vegetation expansion around already existing urban areas.

• Intermix Area: Area where vegetation and urban area overlap. These are usually small vegetated areas that are inside areas with high human activity.

Assessment and Monitoring

The method proposed allows regular refreshment of the mapping, typically, a monthly refresh.

The monitoring allows to generate alerts depending on the scores for the geographical area analysed by the end use and allows to take actions on the zones within this area which are identified at risk.

It can be used by insurance companies (to define their pricing and underwriting policies), by local authorities to organize preventive measures where possible, by firemen stations to help the men deployment and management on the field, etc... .

On case of risky period, daily updates can be performed.

The modification detected on the mapping can be used for alerting at an early stage and for providing regular situation updates. Annex 1 - Examples of hazard attributes

The data sources, hazard attributes and features here described are only provided to illustrate the method proposed. Other data sources, hazard attributes and features can also be contemplated depending in particular on the data sources available.

1. Data Source: Satellite

1.1. Vegetation Composite

Aggregation of Sentinel 2 Indexes that have a direct relationship with vegetation quality and type.

• Data Sources: Sentinel 2 constellation.

• Resolution: 10 meters.

• Feature: Linear Combination of NDVI (Normalized difference vegetation index), SAVI (Soil adjusted vegetation index) and RVI (Relative vigor index). The weights are determined to preserve the maximum amount of variance from the three input indexes into a single output index.

• Correlation with historical fires: High.

This feature contains very precise (up to 10 meters resolution) and up to date information about the vegetation type, density and quality. It is therefore a direct indicator of the state of the fuel in the observed surface.

1.2. Water Stress Composite

• Aggregation of Sentinel 2 Indexes that have a direct relationship with the moisture content of the vegetation.

• Data Sources: Sentinel 2 constellation.

• Resolution: 10 meters.

• Feature: Linear Combination of NDWI (Normalized difference water index), NDMI (Normalized difference moisture index) and NMDI (Normalized multi drought index). The weights are determined to preserve the maximum amount of variance from the three input indexes into a single output index.

• Correlation with historical fires: High.

This feature contains very precise (10 meters resolution) and up to date information about the dryness level of the vegetation. It is therefore a direct indicator of the state of the fuel in the observed surface. 1.3. Land Cover classification

This feature represents the coverage type of the observed surface as Water, Forest or Urban area.

• Data Sources: Copernicus (PROBA V) between 2015 and 2019.

• Resolution: Up to 10 meters.

• Feature: The classes are filtered and grouped to keep relevant information to fire spread. This Attribute is meant to give an indication of the surface type. Information about the vegetation itself (Density, quality etc.) is captured by satellite data.

1: Unburnable area

2: Herbaceous Vegetation

3: Shrubs

4: Open Forests

5: Closed Forests

• Correlation with historical fires: High.

This feature is an indicator of the degree of fuel present on the terrain and is therefore highly correlated with past fires. Some classes, like Dense Forests, also mean that access to the area is restricted which means that ground containment efforts are limited.

2. Data Source: Geolocation

2.1. Activity-weighted Roads

A proximity to OpenStreetMap (OSM) road network layer, weighted by geolocation activity was created to capture fire risks associated with roads.

• Data Sources: Geolocation (point data, dynamic). OpenStreetMap (vector, static).

• Resolution: NA

• Feature: Activity-weighted distance to the nearest road.

• Correlation with historical fires: High.

2. 2. Urban Areas

A proximity to built-up areas layer was created to capture anthropogenic fire risk factors. • Data Sources: Geolocation (point data, dynamic).

• Resolution: 200 meters.

• Feature: Distance in meters to the nearest built-up area. Geolocation data was processed to determine activity hotspots corresponding to cities and towns.

• Correlation with historical fires: High

(vs. smoking, arson, playing with fire and structure fire records).

3. Data Source: Climate & Weather

3.1. Temperature

Temperature is an important factor during the fire season.

• Data Sources: PRISM Group for historical dataset

• Resolution: 0.008 degrees (around 4 km) for past data, 0.125 degrees (20 kilometers (TBC)) for forecasted data

• Feature: Two features can be derived from this source

Forecasted temperature change: Mean change between the forecasted temperature and the average of the three previous years.

Mean Values of the previous year.

• Correlation with historical fires: Medium.

The risk is mainly correlated with the forecasted change. However, the correlation level is highly dependent on the quality of the forecasts which tend to be less accurate over long periods of time (i.e. a year in this case). This feature could be highly correlated in case of short term predictions. The mean past temperature value shows no special correlation to the future fires in California.

3.2. Lightning

A historical lightning strike frequency layer was created to capture ignition risk from lightning.

• Data Sources: National Oceanic and Atmospheric Administration (NOAA) Severe Weather Database Inventory (SWDI) lightning tile summaries (dynamic).

• Resolution: 0.1 degrees (around 10 kilometers).

• Feature: Static historical lightning strike frequencies. A ten-year aggregation of lightning strike frequencies is used. • Correlation with historical fires: Medium.

Since this is a pure ignition feature, other layers are important in determining whether a fire takes hold and propagates or not; therefore high lightning strike rates do not always correspond to high lightning fire rates.

3.3. Vapor Pressure Deficit

VPD (Vapor pressure deficit) is the difference between the amount of moisture in the air and how much moisture the air can hold when it is saturated. It measures the "drying power" of air and therefore how much vegetation is going to dry out.

• Data Sources: PRISM Group

• Resolution: 0.008 degrees (around 4 km)

• Feature: Mean of the VPD during winter compared to the average of the 3 previous winters.

• Correlation with historical fires: Medium.

VPD during winter can be an indicator on how this feature will evolve during summer. For long term risk forecasting and without robust weather previsions, the relationship between this feature and the fire risk is not necessarily as significant.

3.4. Past Precipitations (3 years)

Precipitations over the past 3 years.

• Data Sources: PRISM Group

• Resolutions.008 degrees (around 4 km)

• Feature: Sum of the precipitations during the last 3 years, including the last winter. The number of years taken into account for this feature is optimized to maximize the correlation with the fire events.

• Correlation with historical fires: High.

Precipitation has a direct impact on how the vegetation will evolve in the future and therefore has a direct impact on the fuel development in the area.

4. Data Source: Public Data

4.1. Power Lines

A proximity to the powerline layer was created to capture risk of power-grid fires. • Data Sources: OSM (vector, static), Homeland Infrastructure Foundation Level Database (HIFLD) (vector, static). HIFLD data added as supplement due to OSM network missing some powerline fires.

• Resolution: 200 meters.

• Feature: Distance in meters to the nearest powerline network geometry.

• Correlation with historical fires: High (vs. powerline fire records).

4.2. Rail Network

A proximity to rail network layer was created to capture risk of fires related to rail operations.

• Data Sources: OSM (vector, static).

• Resolution: 200 meters.

• Feature: Distance in meters to the nearest rail network geometry.

• Correlation with historical fires: High (vs. rail fire records).

4.3. Camp Fire

A proximity to US Dep. Agriculture Forestry Service (USFS) sites layer was created to capture risk of fires related to outdoor leisure activities such as camp fires.

• Data Sources: USFS recreational facility database. Includes sites like campgrounds, picnic areas, trail heads and also USFS facilities.

• Resolution: 200 meters.

• Feature: Distance in meters to the nearest USFS location.

• Correlation with historical fires: High (vs .campfire records).

4.4. Fire History

Past fire frequency and lag (time since last fire) were investigated to capture historical patterns and stress zones.

• Data Sources: Kayrros analysis

• Resolution: 500 meters.

• Feature: Combined historical fire frequency and lag (time since fire).

• Correlation with historical fires: High (tested on unseen test set).

5. Data Source: Topography 5.1. Slope

Slope level of the terrain.

• Data Sources: SRTM (Shuttle Radar Topography Mission) Mission (2000)

• Resolution: 30 meters.

• Feature: Slope level of the terrain in degrees. Other metrics were considered such as the aspect or the value of the elevation, no significant correlation between fires and these features was found.

• Correlation with historical fires: High.

Slope is an important factor in fire propagation as there is more wind along the slope. The heat can also propagate faster down a slope as it will more easily affect the trees above.

5.2. Solar Irradiation / Aspect

Level of sun exposure during the fire season.

• Data Sources: Aspect (SRTM) and sun position during the year

• Resolution: 30 meters.

• Feature: Average of the normal radiation received during the fire season from the sun.

• Correlation with historical fires: Low.

Solar Irradiation is usually a factor in wildfire risk (although not in some area e.g. California).