Method for intelligent forecasting of the wildfire probability

Abstract

The invention relates to means of intelligent forecasting of wildfire probability. The method comprises obtaining geospatial data set comprising weather measurements for 7 days up to the time of wildfire probability forecasting, data about the vegetation cover of the Earth (type of vegetation cover, vegetation indices of vegetation cover, and data on the fraction of absorbed photosynthetically active radiation), and geospatial data (elevation, slope direction, gradient, population density, and distance from roads). The obtained data set is processed with a convolutional neural network (CNN) trained on archival data sets containing all the above groups of attributes, as well as a reference markup with coordinates of actually recorded wildfire outbreaks. Based on such processing with the CNN a regular spatial grid is built for a given area of forecasting the probability of new wildfire outbreaks. Then the forecasting of wildfire probability is carried out for each of the following five days based on such regular grid, which results in 5 maps. Said method provides increasing efficiency of forecasting the probability of occurrence of natural fires for five forecast days.

Description

[0001] METHOD FOR INTELLIGENT FORECASTING OF THE WILDFIRE PROBABILITY

[0002] The patent application has been prepared as part of Activity 3.4.1 of the Research Center for Artificial Intelligence in the area of Optimizing Management Decisions to Reduce Carbon Footprint with the support of ANO Analytical Center under the Government of the Russian Federation (Subsidy Agreement ID 000000D730321P5Q0002, Agreement No. 70-2021-00145 of November 02, 2021).

[0003] FIELD OF THE INVENTION

[0004] In general, the claimed technical solution relates to the field of computer science, and in particular, to a method and system for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN) for a five-day forecast.

[0005] The presented solution can be used at least by the national ministries of civil defense, emergencies and disaster management (e.g. EMERCOM of Russia, etc.), regional monitoring services, forestry enterprises, etc. to make prompt decisions on measures to suppress fires, build fire breaks, evacuate the population at risk, etc.

[0006] PRIOR ART

[0007] Traditional technologies for preventing emergencies and minimizing their implications are often ineffective: they are not always accurate, cannot quickly process large arrays of data, etc. In this regard, wildfire risk management is an urgent task not only for specialized organizations and companies, but also for IT developers who create alternative solutions, which are often based on artificial intelligence, to help study large volumes of geospatial, meteorological and temporal data in detail to forecast the fire outbreaks.

[0008] In 2019, Jae Seung Lee of Hongik University in South Korea, together with his students, created an algorithm that can predict the fire outbreaks and fire behavior with an accuracy of 90%. The technology is based on high-resolution satellite imagery and Microsoft's Azure Machine Learning algorithms. The creators of the system claim that it can help firefighters allocate resources more efficiently by regrouping crews and sending them where the resources are lacking. This allows emergency services to respond faster to calls and minimize potential damage from fires. Jae Seung Lee hopes that this tool can also be used for other purposes, such as fighting crime.

[0009] In 2019, CrowdAI introduced FireNet, a software program that uses artificial intelligence to predict forest fires. The algorithm analyzes images from digital drones that circle over forest ranges and take photos at 20 frames per second. The accuracy of fire detection is 92%. The data is processed in real time. The sites with a high probability of fire outbreak are identified by using the drone's GPS coordinates. According to Vice, a publication, the creation of FireNet was inspired by software for analyzing medical imagery, which helps doctors differentiate healthy tissues from unhealthy ones. The system was trained on video footage of forest fires obtained from unmanned aerial vehicles.

[0010] In Russia, an intelligent video surveillance system, that monitors the fire hazards and was implemented by MTS, operates in the Sebezhsky National Park in the Pskov region. The system comprises fourteen cameras with special lenses and sensors. Online cameras monitor the fire safety of forested areas by detecting hazardous temperature rises, which enables the park rangers to prevent fires early, even before there are any open flames. The thermal imaging cameras also report to the monitoring center about any instances of campfires, bonfires or smoking in the protected areas.

[0011] The prior art discloses a solution [1] that describes a model for predicting the level of wildfire hazard. The approach is based on simulating the level of wildfire hazard using Poisson regression.

[0012] However, unlike the claimed solution, the solution known from the prior art [1] uses data from individual weather stations, which is further interpolated over the entire surveyed area. Also, the solution known from the prior art [1] does not consider the impact of factors other than weather measurements.

[0013] The prior art discloses a system [2] for assessing the risks of wildfires. However, this system does not use convolutional neural networks to capture the spatial characteristics of land cover.

[0014] In addition, the prior art discloses ISDM-Rosleskhoz, an information system for remote monitoring of forest fires of the Russian Federal Forestry Agency, which allows estimating the wildfire probability based on weather data and a deterministic -probabilistic approach [3]. The paper assessed the possibility of determining the threshold of the fire hazard index (based on the methodology proposed by V.G. Nesterov), at which there was a fire outbreak in the past according to the statistics.

[0015] The use of data from remotely sensed Earth observation to predict the probability of fire outbreaks was proposed in paper [4]. For more reliable results, the paper identified the areas that differ in terms of fire outbreaks. However, the proposed approach provides for the usage of NOAA data and weather data from the Russian Federal Service for Hydrometeorology and Environmental Monitoring (Rosgidromet) without considering other characteristics.

[0016] The prior art discloses a patent application filed in the USA (US20220398840A1 Smoke and Fire Recognition, Fire Forecasting, and Monitoring, patent holder: KNOETIK SOLUTIONS INC, published on December 15, 2022), which describes a method and system for acquiring images and information related to environmental weather and Earth-related information to automatically recognize signals related to smoke or fire; and predicting, by computation, the existence of a fire, or the growth and spread of a fire.

[0017] Unlike the above method which detects the fire outbreak and fire spread, the claimed solution is implemented to predict the fire outbreak probability up to several days in advance.

[0018] The prior art discloses a patent RU2486594C2 “METHOD OF FOREST FIRE MONITORING AND INTEGRATED SYSTEM FOR EARLY DETECTION OF FOREST FIRES BASED ON THE PRINCIPLE OF MULTI-SENSOR PANORAMIC TERRAIN VIEW WITH THE FUNCTION OF HIGH-PRECISION DETECTION OF THE FIRE OUTBREAK AREA” (patent holder: Videophone MV Closed Joint-Stock Company, published on June 27, 2013). This patent for invention describes a system for enhancing the reliability and accuracy of forest fire detection. A feature of the system is that the terrain is monitored from at least two points located on the cellular communication masts by means of a thermal imaging camera and a video camera mounted so that their axes are parallel and fixed on a scanning platform placed on each cellular communication mast, wherein the images received in thermal and video channels, along with the data of angular and azimuthal direction of camera axes, are transmitted to a central server, wherein the images received from thermal and video cameras and the data from angular and azimuthal meters located on cellular communication masts are converted into a system of geographic coordinates, and wherein the fire outbreak areas are georeferenced to geographic coordinates and displayed on an electronic map of the terrain, and the video image is superimposed on the image from the thermal imaging camera and the resulting images are displayed as three separate images on the operator's monitor and / or sent to a storage device.

[0019] However, the known solution allows detecting the fire outbreaks immediately at the time of the event, but does not allow to predict the fire outbreak probability up to several days in advance. Moreover, the known solution does not use convolutional neural networks to capture the spatial characteristics of land cover.

[0020] The prior art discloses a patent application US20210110136A1 Fire Detection via Remote Sensing and Mobile Sensors (patent holder: IBM, published on April 15, 2021). This application for invention describes a method and system for detecting events. A satellite image is obtained and processed using a first convolutional neural network (CNN) to generate a satellite vector that identifies at least one fire. The mobile sensor is automatically directed to the mobile location based on the satellite vector.

[0021] Unlike the known system, the claimed solution allows predicting the fire outbreak probability up to several days in advance. The prior art discloses an international patent application WO2018116966A1 FIRE MONITORING SYSTEM (patent holder: HOCHIKI CO. et al., published on June 28, 2018), which describes a multilayer neural network that is efficiently trained and improves the accuracy of fire detection, even if some characteristics of the fire on the overall image are limited to a narrow range. The monitor image of the monitored area that was captured by the monitor camera is segmented into a multitude of block images. The images for units are fed into a fire detector, which is configured for fire detection using a multilayer neural network. The apparatus for generating the training image segments the training image, where a fire has been detected in the monitored area, into a multitude of unit images and, depending on the presence of flame and / or smoke in each of the unit images and the portions of each of the unit images occupied by flame and / or smoke, classifies and causes the unit images to be stored as either fire images, training unit images, training unit images in a normal situation, or non-training unit images. A training control unit (26) feeds to the fire detector the images for fire training unit and images for normal situation training unit, which were stored in the apparatus for generating the training image, and causes a multilayer neural network (30) to learn through deep learning.

[0022] Unlike the said source, in the claimed solution, a convolutional neural network predicts the wildfire probability up to several days in advance.

[0023] The prior art also discloses a patent for invention KR101869442B1 Fire Detecting Apparatus and the Method Thereof (patent holder: NAT UNIV KONGJU IND UNIV COOP FOUND, published on June 20, 2018). This solution describes a fire detecting apparatus, that uses an algorithm for automatic smoke detection with video surveillance and image processing technology to detect smoke and track a moving object, and then applies a cascade classification model using a convolutional neural network (CNN) that differentiates between smoke and nonsmoke objects to detect the fire outbreak. According to this invention, a fire detecting apparatus comprises a unit for detecting the smoke area, which detects a change in a background pixel relative to the image of target fire detection area fed from a fire detection camera and groups the related pixels to classify a candidate smoke area; and a unit for classifying the smoke area, which receives and classifies the candidate smoke area as a smoke area or a non-smoke area, so as to identify and output the information on the smoke area when there is such a smoke area.

[0024] The above solution detects the fire outbreak, i.e. the fire, while the claimed solution predicts the wildfire outbreak probability up to several days in advance.

[0025] The prior art discloses a solution RU2645179 C2 PROBABILISTIC SATELLITE SYSTEM FOR MONITORING FOREST FIRES (patent holder: Federal State Budgetary Military Educational Institution of Higher Education A.F. Mozhaisky Military Space Academy of the Ministry of Defense of the Russian Federation, Russian Federation, represented by the Ministry of Defense of the Russian Federation, published on February 16, 2018), which describes satellitebased Earth observation systems covering vast regions of the Earth. The system satellites, placed in circular orbits, are equipped with a scanning wide-angle infrared optoelectronic system with a linear photodetector to detect forest fire hotspots. The satellites also carry an optronic IR tracking system that is redirected by targeting guidance received from a scanning system. This tracking system is made wide-angle (with a fisheye IR lens) and has multiple matrix photodetectors to detect and identify the parameters of forest fire hotspot, as well as to generate an alert signal about it. The technical result of the invention is aimed at expanding the functional capabilities of the system, reducing the weight and dimensional characteristics of the system satellites and reducing the costs of the system’s creation and operation.

[0026] While the said known solution just detects a forest fire hotspot, the claimed solution predicts the wildfire outbreak probability up to several days in advance.

[0027] The proposed technical solution is aimed at eliminating the gaps in the state of the art and differs from previously known solutions in that the proposed solution utilizes geospatial data and convolutional neural network (CNN) to intelligently predict the wildfire probability up to five days in advance. This results in preventing the wildfire outbreaks and the spread of wildfires.

[0028] SUMMARY OF THE INVENTION

[0029] The claimed technical solution proposes a new approach to forecasting the wildfire probability, which allows preventing the wildfire outbreaks and the spread of wildfires.

[0030] This technical solution addresses the technical problem of creating an automated method and system for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN) for a five-day forecast.

[0031] The technical result achieved when addressing the above technical problem is the enhanced effectiveness of forecasting the wildfire probability for a five-day forecast.

[0032] This solution will allow to adopt optimal decisions for timely preparation and positioning of manpower and resources in areas with high fire hazard probability.

[0033] The forestry and agricultural businesses can use these forecasts to plan their activities, for example, to carry out prescribed burns or to determine the most appropriate time for agricultural works.

[0034] The insurance companies can use the forecasts to assess risks and manage fire -related insurance policies.

[0035] The environmental authorities can use the data to monitor environmental conditions and take measures to preserve natural resources and biodiversity. The said technical result is achieved by using a method for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN) for a five- day forecast, wherein such method is implemented with at least one processor and comprises the steps, where:

[0036] - a geospatial data set is obtained comprising:

[0037] • Weather measurements for 7 days up to the time of wildfire probability forecasting;

[0038] • Data about the vegetation cover of the Earth comprising: information on the type of vegetation cover, vegetation indices of vegetation cover and data on the fraction of absorbed photosynthetically active radiation (PAR);

[0039] • Geospatial data comprising: information on elevation, slope direction, gradient, population density, and distance from roads;

[0040] - the obtained data set is processed with a CNN trained on archival data sets containing all the above groups of attributes, as well as a reference markup with coordinates of actually recorded wildfire outbreaks, and

[0041] - based on such processing with a CNN:

[0042] • a regular spatial grid is built for a given area of forecasting the probability of new wildfire outbreaks;

[0043] • and the forecasting of wildfire probability is carried out for each of the following five days based on such regular grid, which results in 5 maps.

[0044] In a particular embodiment of the described solution, the vegetation indices of vegetation cover are obtained from a MODIS satellite.

[0045] In a particular embodiment of the proposed solution, the vegetation indices of vegetation cover include the Leaf Area Index (LAI), Normalized Difference Vegetation Index (ND VI), and Enhanced Vegetation Index (EVI).

[0046] In another particular embodiment of the proposed solution, before feeding a data set to the input of the CNN, the data is brought to the same spatial resolution and the data is normalized.

[0047] In another particular embodiment of the proposed solution, the data on the vegetation cover of the Earth additionally contains at least the boundaries of water bodies, types of vegetation cover, terrain data and distance from population centers.

[0048] In another particular embodiment of the proposed solution, the weather data contains at least the data on temperature, humidity, and precipitation.

[0049] In another particular embodiment of the proposed solution, the archival data sets are first submitted to a verification and filtering step. In another particular embodiment of the proposed solution, the fire hazard index is additionally calculated, wherein such index evaluates the dynamics of changes in the attributes of the environment over a certain period of time.

[0050] In addition, the claimed technical result is achieved with a system for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN) for a five-day forecast, wherein such system comprises:

[0051] - at least one processor;

[0052] - at least one memory connected to a processor and containing machine-readable instructions which, when executed by at least one processor, ensure the execution of a method for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN).

[0053] BRIEF DESCRIPTION OF THE DRAWINGS

[0054] An embodiment of the invention will be described below in accordance with the accompanying figures, which are presented to explain the essence of the invention and in no way limit the scope of the invention. The following figures are attached to the application:

[0055] Fig. 1 shows an embodiment of intelligent forecasting of the wildfire probability.

[0056] Fig. 2 shows an example of the approach to forecasting the wildfire probability using convolutional neural networks and geospatial data.

[0057] Fig. 3 presents a general view of a system for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN) for a five-day forecast.

[0058] DETAILED DESCRIPTION OF THE INVENTION

[0059] The following detailed description of the embodiment of the invention provides numerous details of the embodiment to ensure a clear understanding of the present invention. However, to those skilled in the art, it will be obvious how the present invention could be used, whether with or without these details of its embodiment. In other cases, the well-known methods, procedures, and components have not been described in detail to avoid overcomplicating the understanding of this invention’s features.

[0060] In addition, the provided presentation will make it clear that the invention is not limited to the presented embodiment. Numerous potential modifications, changes, variations, and substitutions, that retain the essence and form of this invention, will be obvious to those skilled in the art. This technical solution may be implemented on a computer, in the form of an automated information system (AIS) or a machine-readable medium containing the instructions for executing the above method.

[0061] The technical solution can be implemented as a distributed computer system.

[0062] In this solution, a system means a computer system, electronic computer (computer), computer numerical control (CNC), programmable logic controller (PLC), computerized control systems and any other devices capable of performing a specified, well-defined sequence of computational operations (actions, instructions).

[0063] A command processing device means an electronic unit or integrated circuit (microprocessor) that executes machine instructions (programs).

[0064] The command processing device reads and executes machine instructions (programs) from one or more data storage devices, such as random access memory (RAM) and / or read-only memory (ROM). RAM devices may include, but are not limited to, hard disk drives (HDD), flash memory, solid state drives (SSD), optical media (CD, DVD, BD, MD, etc.) and others.

[0065] A program is a sequence of instructions designed to be executed by a control device of a computing machine or command processing device.

[0066] As shown in Fig. 1, the claimed method (100) for intelligent forecasting of the wildfire probability using geospatial data n convolutional neural network (CNN) for a five-day forecast comprises several steps.

[0067] In the step (101), a geospatial data set is obtained comprising three groups of attributes:

[0068] • Weather measurements for 7 days up to the time of wildfire probability forecasting;

[0069] • Data on the vegetation cover of the Earth comprising: information on the type of vegetation cover, vegetation indices of vegetation cover and data on the fraction of absorbed photosynthetically active radiation (PAR);

[0070] • Geospatial data comprising: information on elevation, slope direction, gradient, population density, and distance from roads;

[0071] The weather measurements are selected from such information resource as the FTP server of the National Oceanic and Atmospheric Administration (NOAA) of the United States. This is a public resource, where the weather data is based on the information received from remotely sensed Earth observation with reference to coordinate grid (Glenn et al., 2006), and it also applies its model to generate forecasts for upcoming periods. The periods are separated from each other using forecast model verification, with the implication that the shorter the forecast time, the more accurate the model. In addition, the weather data from Rosgidromet is used.

[0072] This weather data contains at least the data on temperature, humidity, and precipitation. The statistics on fires are selected from the database of detected heat points (https: / / fires.ru / ). The monitoring area includes the entire Russia and neighboring territories, where the satellite data required for fire detection can be promptly received. The global coverage is based on data from NASA FIRMS (Fire Information for Resource Management System of the National Aeronautics and Space Administration) (Davies et al., 2012).

[0073] The fire information is published on an interactive map, along with the necessary viewing and navigation tools. The sample includes only confirmed heat points of the "wildfire" category.

[0074] The weather data is a critical attribute in predicting the probability of a new wildfire, as it determines the fire susceptibility of the Earth's vegetation cover at any given time.

[0075] The claimed solution does not use weather forecast data, which allows improving its reliability and reduce the error associated with the building of predictions for weather characteristics.

[0076] The claimed solution uses only current weather measurements without forecast values. This includes the data for several previous days on total and potential evaporation, the easterly and northerly wind components (westerly and southerly wind components are unambiguously determined from these two), air temperature at 2m height, dew point temperature, and precipitation amounts. The relationship between total evaporation and bioproductivity of ecosystems, as well as knowledge of the temperature regime at a particular observation point, allows assessing whether the thresholds for ignition of different vegetation types can be reached. In turn, the wind speed allows estimating the susceptibility of the environment to fire propagation, i.e. the probability that an ignited spark will not be extinguished but will grow into a fire.

[0077] The data on the vegetation cover of the Earth is obtained from satellites using the methods of remotely sensed Earth observation. Such attributes include the type of vegetation cover, vegetation indices of vegetation cover and data on the fraction of absorbed photosynthetically active radiation (PAR). Moreover, the data on the vegetation cover of the Earth can additionally contain at least the boundaries of water bodies, types of vegetation cover, terrain data and distance from population centers.

[0078] The maps containing these characteristics are a product of a MODIS satellite. In addition to the Earth's surface classes, the vegetation indices of vegetation cover are also calculated, where such indices contain additional information about the condition of the vegetation cover and have a significant impact on the risk of wildfire outbreaks. The vegetation indices represent a mathematical model that allows obtaining a quantitative assessment of some vegetation cover characteristics by translating spectral values into a target value. Such indices include the Leaf Area Index (LAI), Normalized Difference Vegetation Index (NDVI), and Enhanced Vegetation Index (EVI), and the measurements of the fraction of absorbed photosynthetically active radiation (PAR) are used to describe the condition of the vegetation cover. These characteristics are critical for determining the wildfire probability because, at different growth stages and in different general conditions, the plants have different burning properties and ignition thresholds. The geospatial data also include information on elevation, slope direction, gradient, population density, and distance from roads.

[0079] The characteristics of terrain have an impact on the risk of wildfire outbreak. For example, steep mountain slopes are less prone to fire outbreaks. Additional information on the distance from infrastructure facilities (roads and population density) characterizes the chances that the human may provoke wildfires as a result of careless handling of fire during the fire hazard season. All measurements are stored in a shared database, from where the data is collected and further processed.

[0080] In addition, the claimed solution provides an option to additionally compute the fire hazard indices, which are calculated only on the basis of weather data for the previous days and allow to consider the cumulative effect (e.g. following a prolonged drought or rains).

[0081] It should be noted that the computed fire hazard indices include the Nesterov index. This index considers the cumulative effect and allows evaluating the dynamics of changes in the attributes of the environment over a certain period of time. Therefore, the algorithm can consider not only individual weather indicators for each day prior to the start of forecasting, but also the effect that certain weather characteristics over an interval of two weeks may have on the general level of fire hazard. It also should be noted that the fire hazard indices have dimensionless values and, when the value of the index increases, the fire hazard level increases, too; but, for example, when precipitation exceeds a given threshold, the indices are reset to zero and further accumulation starts from the zero point. In the proposed algorithm, the indices are used along with other groups of attributes, because the indices alone do not consider other significant factors affecting the wildfire probability.

[0082] Next, on the step (102), the obtained data set is processed with a CNN trained on archival data sets containing all the above groups of attributes, as well as a reference markup with coordinates of actually recorded wildfire outbreaks.

[0083] Before feeding a data set to the input of the CNN, the data is brought to the same spatial resolution and the data is normalized.

[0084] CNN is trained on archival data that is specifically collected for the territory of the Russian Federation. CNN is trained using the method of error back propagation, where a tensor comprising geospatial attributes is fed to the input of the model; and a prediction map is obtained at the output of the model, and such map is used, along with the true value map, to calculate the prediction error; next, the weights of the CNN are updated based on the obtained values using the method of error back propagation. The archival data, which is used as the true values on the fire outbreak, is submitted first to a verification and filtering step. This step consists in selecting the fires, that have actually started, and excluding thermal anomalies that were detected by sensors but did not result in outbreak and spread of fires. The resulting processed data set represents filtered and verified coordinates and start dates of fires brought to a common format. This solution was developed specifically for the territories of the Russian Federation and provides a significant advantage of the approach compared to the approaches developed for the territories and natural and climate conditions of other countries. The proposed solution can also be adapted to other territories for better quality of forecasting. The general diagram of the approach is presented in Fig. 2.

[0085] To select the optimal hyperparameters of the neural network algorithm, the training is performed on the prepared archival data sets containing all of the above groups of attributes, as well as the reference markup, i.e. the coordinates of actually detected wildfire outbreaks. The tuning of the algorithm involves the selection of the target loss function. Since the task is to determine the wildfire probability on each of the five subsequent days, it is proposed to use the binary cross entropy loss function, which is calculated between a vector with a length of 5 values (5 prediction days), using the reference values, and a vector of the same size containing the wildfire probabilities on each of the 5 subsequent days, with the predictions of the neural network model. Therefore, the model optimization is based on the errors obtained for each individual example of fire outbreak, as well as for the examples where there were no fire. The algorithm is tuned, and the number of training epochs, batch size and learning rate are selected.

[0086] All three groups of attributes are brought to one common raster format with a geospatial resolution of less than 1km per pixel and form a single attribute space for training the neural network. This generates a tensor with the size of N*M*K, where N and M represent the height and width of the image corresponding to an individual prediction cell (with a pixel’s geospatial resolution of up to 1km); and K is the total number of attributes for the above three groups of input data. The archival data on fire outbreaks is used as reference data to train the model. This data is based on verified heat points (thermal anomalies in a given area) and corresponds to the initial moment when a new fire was detected. In addition to the examples with fire outbreaks, the data set also contains examples corresponding to geographic coordinates and specific dates of observation where no new fire outbreak or spread of existing fires was detected. All collected and processed data is geo-referenced.

[0087] In the step (103), based on the processing of the obtained data set with a CNN, a regular geospatial grid with a resolution of 0.1 to 0.2 degrees is built for a given area of forecasting the new wildfire probability, which allows pinpointing with a sufficiently high precision the sites with high risk of fire outbreak. The approaches known from the prior art are used to prepare a fire hazard forecast for the entire region as a whole, which substantially complicates the task of decision-making on preparing the manpower and resources before the fire outbreak.

[0088] In the step (104), the forecasting of wildfire probability is carried out for each of the following five days based on such regular grid, which results in 5 maps.

[0089] The result of the claimed solution is the creation of 5 maps describing the wildfire probability in a given area for 5 days in advance. For each individual day, the map is provided as an image where each pixel corresponds to one geographic grid cell with a size of less than 0.2 degrees. In each cell, the preparation of the forecast does not depend on neighboring cells but is determined by natural and climate indicators exclusively within that cell. A sigmoid function is used as the activation function on the last layer of the neural network. The sigmoid activation function can take real values as input, which it translates to a range of values from 0 to 1 in the output. The resulting values can be interpreted as the algorithm's confidence that an object belongs to class zero or class one. Therefore, the values in the output of the algorithm for each prediction cell vary from 0 to 1, which corresponds to the fire outbreak probability in that cell on each of the following days (the closer the value is to 1, the higher is the fire outbreak probability in that cell on a given day). Based on the obtained maps, it is possible to identify the territories most exposed to fire hazard under given geo-climatic conditions.

[0090] It is important to note that the model predicts the fire probability for 5 days in the future based only on the data for past days and does not use weather forecast data for subsequent days. With this approach, the model learns to identify key factors and patterns that have led to fire outbreaks in the past. The non-use of weather forecasts or other predictive data removes the uncertainty associated with the potential inaccuracy or errors in these forecasts. This means that the model relies exclusively on actual historical data.

[0091] The claimed solution allows identifying areas with the highest risk of new fire outbreaks and to take timely measures for prompt and effective response in case of fire outbreak (preparation of manpower and resources), as well as to mitigate such risk through proactive measures (e.g. artificial moistening of territories).

[0092] In addition, the claimed solution allows obtaining, on a grid down to 0.1 degrees, a forecast of the wildfire probability separately for each of the next five days. As it does not depend on shortterm weather forecasts, the proposed approach is resilient to their potential unreliability and variability by focusing on more stable historical data.

[0093] In addition, the claimed solution implements automatic uploading and processing of all necessary weather and spatial data to create a prediction. The approach relies on the use of a neural network algorithm that considers geospatial characteristics when generating a forecast. The algorithm can be further adapted to new territories to obtain more accurate forecasts based on the climate and geographical characteristics of the area of interest. With the accumulation of data and feedback from previous forecasts, the model can adjust and refine its algorithms, which enhances its accuracy and reliability.

[0094] Fig. 3 presents an example of a general view of a computing system (300), which ensures an embodiment of the claimed method or is a part of a computer system, such as, a server, a personal computer, a part of a computing cluster, which processes the data required to implement the claimed technical solution.

[0095] In a general embodiment, the system (300) includes one or more processors (301), memory facilities such as RAM (302) and ROM (303), input / output interfaces (304), input / output devices (305), and a networking device (306) interconnected by a common information sharing bus.

[0096] The processor (301) (or multiple processors, multi-core processor, etc.) may be selected from a range of commonly used devices, e.g., those from manufacturers such as: Intel™, AMD™, Apple™, Samsung Exynos™, MediaTEK™, Qualcomm Snapdragon™, etc. For the processor or one of the processors used in the system (300), it is also necessary to consider a graphics processor, such as an NVIDIA or Graphcore GPU, which also can be used for implementing the method in whole or in part, as well as for training and applying machine learning models to various information systems.

[0097] RAM (302) represents random access memory and is designed to store machine-readable instructions executable by the processor (301) to perform the necessary operations for logical data processing. RAM (302) typically contains executable instructions of the operating system and related software components (applications, program modules, etc.) In this case, RAM (302) may be the available memory capacity of the graphics card or the graphics processor.

[0098] ROM (303) represents one or more read-only memory devices, such as a hard disk drive (HDD), a solid state drive (SSD), flash memory (EEPROM, NAND, etc.), optical storage media (CD-R / RW, DVD-R / RW, BlueRay Disc, MD), etc.)

[0099] Various types of RO interfaces (304) are used to organize the operation of system components (300) and to organize the operation of external connected devices. The selection of appropriate interfaces depends on the particular embodiment of the computing device, which may comprise, but are not limited to: PCI, AGP, PS / 2, IrDa, FireWire, LPT, COM, SATA, IDE, Lightning, USB (2.0, 3.0, 3.1, micro, mini, type C), TRS / Audio jack (2.5, 3.5, 6.35), HDMI, DVI, VGA, Display Port, RJ45, RS232, etc.

[0100] Various devices (305) for information RO are used to enable user interaction with the computing system (300), e.g., keyboard, display (monitor), touch screen display, touch pad, joystick, mouse manipulator, light pen, stylus, touchpad, trackball, speakers, microphone, augmented reality devices, optical sensors, tablet, light indicators, projector, camera, biometric identification means (retina scanner, fingerprint scanner, voice recognition module), etc.

[0101] The networking device (306) ensures the transmission of data via an internal or external computing network, such as Intranet, Internet, LAN, etc. One or more devices (306) may include, but are not limited to, an Ethernet card, a GSM modem, a GPRS modem, an LTE modem, a 5G modem, a satellite communication module, an NFC module, a Bluetooth and / or BLE module, a Wi-Fi module, etc.

[0102] These application materials disclosed the preferred embodiment of the claimed technical solution, which should not be used to limit its other, particular embodiments that are within the scope of the claimed legal protection and are obvious to those skilled in the relevant art.

[0103] References:

[0104] 1. A. Kh. Topkaryan et al. A MODEL FOR PREDICTING THE LEVEL OF WILDFIRE HAZARD USING THE POISSON DISTRIBUTION LAW / / Scientific and Educational Tasks Of Civil Defense, 2022, No. 4 (55), p. 38-48 (in Russian)

[0105] 2. I.E. Zhigalov, M.I. Ozerova. DESIGNING AN AUTOMATED SYSTEM FOR ASSESSING THE RISKS OF WILDFIRE OUTBREAKS / / Scientific and Technical Volga Region Bulletin, 2012, No. 6, p. 234-237 (in Russian)

[0106] 3. A.S. Podolskaya, D.V. Ershov, P.P. Shulyak. Applying the Method for Estimating the Probability of Forest Fires in ISDM-Rosleskhoz / / Current Problems in Remote Sensing of the Earth from Space, 2011, vol. 8, No. 1, p. 118-126.

[0107] 4. E.V. Ivanov et al. On the Model for Predicting the Wildfire Probability Based on Remotely Sensed Earth Observation Data / / Current Problems in Remote Sensing of the Earth from Space, 2022, vol. 19, No. 3, p. 77 (in Russian)

Claims

CLAIMS1. A computer- implementable method for intelligent forecasting of wildfire probability using geospatial data and a convolutional neural network (CNN) for a five-day forecast, wherein such method is executed with at least one processor and comprises the steps, where:- a geospatial data set is obtained comprising:• Weather measurements for 7 days up to the time of wildfire probability forecasting;• Data about the vegetation cover of the Earth comprising: information on the type of vegetation cover, vegetation indices of vegetation cover and data on the fraction of absorbed photosynthetically active radiation (PAR);• Geospatial data comprising: information on elevation, slope direction, gradient, population density, and distance from roads;- the obtained data set is processed with a CNN trained on archival data sets containing all the above groups of attributes, as well as a reference markup with coordinates of actually recorded wildfire outbreaks, and- based on such processing with a CNN:• a regular spatial grid is built for a given area of forecasting the probability of new wildfire outbreaks;V and the forecasting of wildfire probability is carried out for each of the following five days based on such regular grid, which results in 5 maps.

2. A method according to claim 1 characterized in that the vegetation indices of vegetation cover are obtained from a MODIS satellite.

3. A method according to claim 1 characterized in that the vegetation indices of vegetation cover include the Leaf Area Index (LAI), Normalized Difference Vegetation Index (ND VI), and Enhanced Vegetation Index (EVI).

4. A method according to claim 1, characterized in that, before feeding a data set to the input of a CNN, the data is brought to a single spatial resolution and the data is normalized.

5. A method according to claim 1 characterized in that the data on the vegetation cover of the Earth can additionally contain at least the boundaries of water bodies, types of vegetation cover, terrain data and distance from population centers.

6. A method according to claim 1 characterized in that the weather data contains at least the data on temperature, humidity, and precipitation.

7. A method according to claim 1 characterized in that the archival data sets are first submitted to a verification and filtering step;8. A method according to claim 1 characterized in that it additionally calculates the fire hazard index, which evaluates the dynamics of changes in the attributes of the environment over a certain period of time.

9. A system for intelligent forecasting of the wildfire probability using geospatial data and a convolutional neural network (CNN) for a five-day forecast, comprising:- at least one processor;- at least one memory connected to a processor and containing machine -readable instructions that, when executed by at least one processor, ensure the execution of the method described in claims 1-8.

Images