Fire-fighting fire characteristic data acquisition and classification identification processing system

Abstract

The invention discloses a fire-fighting characteristic data acquisition and classification identification processing system, and the system comprises a data acquisition module which is used for obtaining original spectral reflectivity data and original thermal infrared radiation data; the preprocessing module is used for generating a hyperspectral data cube and temperature distribution data of the target fire scene area based on the data; the data analysis module comprises a spectral analysis sub-module and a grade judgment sub-module, and the spectral analysis sub-module outputs combustible material types and proportion weights thereof based on a hyperspectral data cube; the grade judgment sub-module is used for determining a corresponding fire intensity grade based on the temperature distribution data; the comburent judgment module is used for outputting the dominant comburent type of the pixel region; and the situation generation module is used for generating a comprehensive situation map for representing the comburent type and the fire intensity. When fire identification is carried out, the influence of comburent types can be fully considered, the identification precision is effectively improved, and the fire behavior trend in the future preset time period is accurately predicted.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a system for collecting, classifying, and identifying fire-related characteristic data. Background Technology

[0002] Forest fires, grassland fires, and large open-air storage yard fires pose significant challenges to fire monitoring and fighting efforts due to their rapid spread, diverse and complex types of combustibles, and the coexistence of multiple combustion forms such as surface fires, crown fires, and underground fires.

[0003] Currently, fire monitoring mainly relies on traditional methods such as satellite remote sensing and lookout tower observation. While these technologies can obtain basic information such as the fire area and the approximate location of open flames at a macroscopic level, they have significant technical limitations. Satellite remote sensing is limited by its temporal and spatial resolution, making it difficult to capture detailed changes in the burning materials in localized areas of a fire in real time, and its accuracy is easily affected by clouds and smoke. Lookout tower observation is limited by its field of view and terrain, and cannot cover large-scale fire areas with complex terrain, thus limiting its monitoring accuracy and range.

[0004] Effective firefighting relies heavily on a precise understanding of the types and states of combustibles at the fire site. Different types of combustibles, such as coniferous forests, broad-leaved forests, shrubs, meadows, and peat, exhibit significantly different fire behavior characteristics during combustion, and their corresponding firefighting strategies and suitable resources also vary. For example, high-intensity coniferous forest crown fires require large fixed-wing aircraft to drop fire retardants, while mixed grass and shrub fires are more suitable for helicopter bucket operations, and smoldering peat areas require targeted underground fire suppression measures.

[0005] Existing technologies also include using drones to collect thermal infrared data for fire monitoring. However, thermal infrared data processing is mostly limited to simple temperature threshold judgment, failing to combine three-dimensional temperature field distribution to achieve refined classification of combustion intensity. Furthermore, it lacks the ability to collaboratively process multi-source data and has not established a correlation analysis mechanism between combustible type and combustion intensity, thus making it difficult to support accurate fire prediction and fire-fighting resource scheduling decisions. Summary of the Invention

[0006] This application aims to at least partially solve one of the technical problems in the aforementioned technologies.

[0007] To achieve the above objectives, the first aspect of this application proposes a fire scene feature data acquisition and classification system, comprising: a data acquisition module for acquiring raw spectral reflectance data and raw thermal infrared radiation data of each spatial sampling point within a target fire scene area scanned by an unmanned aerial vehicle (UAV); a preprocessing module for generating a hyperspectral data cube of the target fire scene area based on the raw spectral reflectance data; and generating temperature distribution data of the target fire scene area based on the raw thermal infrared radiation data; and a data analysis module comprising a spectral analysis submodule and a level judgment submodule, wherein the spectral analysis submodule is used to output each [data point] based on the hyperspectral data cube using a first deep learning model. The module defines the combustible material type and its weighting for each pixel; the level determination submodule determines the fire intensity level corresponding to each spatial sampling point based on temperature distribution data and a preset temperature-intensity mapping relationship; the combustible material determination module outputs the dominant combustible material type for the pixel area based on the combustible material type and its weighting, combined with combustible material combustion characteristic parameters; and the situation generation module spatially aligns and merges the dominant combustible material type with the fire intensity level, labeling each geographic unit in the target fire area with corresponding combustible material type and combustion level labels to generate a comprehensive situation map characterizing combustible material type and fire intensity.

[0008] In addition, the fire situation characteristic data acquisition and classification identification processing system proposed in this application may also have the following additional technical features:

[0009] As a further description of the above technical solution: the spectral analysis submodule includes: a data partitioning unit, used to partition the hyperspectral data cube into multiple local three-dimensional data blocks; a feature extraction unit, wherein the first deep learning model is a three-dimensional convolutional neural network model, the feature extraction unit is used to sequentially input each local three-dimensional data block into the feature extraction layer of the three-dimensional convolutional neural network, and extract the joint features of spectrum and space through convolution and pooling operations; a probability output unit, used to input the extracted joint features of spectrum and space into the fully connected layer and classification layer of the three-dimensional convolutional neural network, and output the probability value of each pixel belonging to various preset combustible endmembers through the classification layer; and a weight generation unit, used to determine at least one type of combustible material based on the preset combustible endmembers whose probability values ​​exceed a preset threshold, and use the normalized probability values ​​as the corresponding proportion weights.

[0010] As a further description of the above technical solution: the level determination submodule includes: a threshold comparison unit, used to compare the temperature value corresponding to each spatial sampling point in the temperature distribution data with a first temperature threshold and a second temperature threshold, wherein the second temperature threshold is greater than the first temperature threshold; and a level determination unit, used to perform a determination operation based on the comparison result: if the temperature value is lower than the first temperature threshold, the fire intensity level is determined to be level one; if the temperature value is between the first temperature threshold and the second temperature threshold, the fire intensity level is determined to be level two; and if the temperature value is higher than the second temperature threshold, the fire intensity level is determined to be level three.

[0011] As a further description of the above technical solution: the combustible material determination module includes: a combustible material combustion calculation unit, used to calculate the combustion index of the combustible material type for each pixel based on the proportion weight of the multiple combustible material types contained therein and the corresponding combustible material combustion characteristic parameters obtained from the database; and a determination unit, used to determine the combustible material type with the largest combustion index as the dominant combustible material type of the pixel area.

[0012] As a further description of the above technical solution: the situation generation module includes: an alignment and fusion unit, used to perform coordinate system and grid resampling on the spatial distribution data corresponding to the combustible material type and the fire intensity level, and for each geographically overlapping pixel pair, combine the corresponding dominant combustible material type and combustion state into a structured semantic tag pair; and an identification mapping unit, used to map the semantic tag pair into a classification identifier displayed on the comprehensive situation map according to a preset semantic fusion rule.

[0013] As a further description of the above technical solution: the preprocessing module includes: a spectral correction unit, used to perform geometric fine correction on the original spectral reflectance data based on the real-time positioning and attitude data of each UAV and the interior and exterior orientation elements of the hyperspectral imager, mapping the corrected data to a unified geographic coordinate system and performing correction to generate a spectral data array; a data cube construction unit, used to integrate and interpolate the spectral data array according to a preset spatial grid and band order to generate the hyperspectral data cube of the target fire area; and a temperature reconstruction unit, used to perform multi-view matching and three-dimensional point cloud reconstruction on the original thermal infrared radiation data based on the real-time positioning and attitude data of each UAV and the installation parameters of the multi-angle thermal infrared sensor array to generate the temperature distribution data of the target fire area.

[0014] As a further description of the above technical solution: the fire situation feature data acquisition and classification identification processing system of this application also includes: a fire prediction module, used to extract spatiotemporal features of the target fire area, including the dominant combustible type, the combustion state and its spatial distribution, based on the comprehensive situation map; and input the spatiotemporal features into a preset second deep learning model; the second deep learning model is configured to output the fire spread direction prediction and fire spread speed prediction of each geographical unit in the target fire area within a preset time period in the future, based on the combustion characteristics and spread patterns corresponding to different combustible types.

[0015] As a further description of the above technical solution: the fire prediction module includes: a graph structure construction unit, used to construct a spatial graph structure with each geographical unit in the comprehensive situation map as a node and the spatial adjacency relationship between geographical units as edges; and to encode the combustible type label and combustion state label of each geographical unit at multiple consecutive times into feature vectors to form the time series features of each node on the spatial graph structure; and a spatiotemporal prediction unit, wherein the second deep learning model is a spatiotemporal graph convolutional network model, the spatiotemporal prediction unit is used to input the spatial graph structure with the time series features into the second deep learning model, capture spatial dependencies through graph convolutional layers, capture temporal dependencies through its temporal convolutional layers, and finally generate the fire spread direction prediction and the fire spread speed prediction through its output layer.

[0016] As a further description of the above technical solution: the fire situation feature data acquisition and classification identification processing system of this application also includes: a dispatch suggestion generation module, used to receive available fire extinguishing resource information from the fire extinguishing resource database, the available fire extinguishing resource information including resource type, quantity and performance parameters; and to match and analyze the comprehensive situation map with the available fire extinguishing resource information, and output fire extinguishing resource dispatch suggestions corresponding to different areas in the comprehensive situation map based on a preset fire situation feature and fire extinguishing resource mapping table.

[0017] According to the fire situation characteristic data acquisition and classification identification processing system of this application, when identifying a fire, it can output the dominant combustible type at different locations of the target fire scene. By spatially aligning and fusing the dominant combustible type with the fire intensity level, a more realistic comprehensive situation map can be obtained, which effectively improves the identification accuracy and accurately predicts the fire trend within a preset time period. Compared with traditional experience models, this prediction method fully considers the influence of combustible type. The prediction results can provide forward-looking data support for the formulation of fire fighting tactics, help to deploy fire fighting forces in advance, and curb the spread of fire.

[0018] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0020] Figure 1 This is a system block diagram of a fire situation feature data acquisition, classification, identification, and processing system according to an embodiment of this application;

[0021] Figure 2 This is a system block diagram of a data analysis module according to an embodiment of this application;

[0022] Figure 3 This is a system block diagram of a preprocessing module according to an embodiment of this application;

[0023] Figure 4 This is a flowchart of a fire scene feature data acquisition, classification, and identification processing system according to an embodiment of this application;

[0024] Figure 5 This is a flowchart of obtaining the dominant combustible type according to one embodiment of this application. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0026] The fire-fighting feature data acquisition, classification, and identification processing system of this application embodiment is described below with reference to the accompanying drawings.

[0027] like Figure 1 and Figure 4 As shown in the embodiment of this application, the fire-fighting feature data acquisition and classification identification processing system may include a data acquisition module for acquiring the original spectral reflectance data and original thermal infrared radiation data of each spatial sampling point within the target fire area scanned by the UAV.

[0028] It should be noted that the raw spectral reflectance data was acquired using a hyperspectral imager mounted on the UAV. The hyperspectral imager has wide-band, high-resolution spectral detection capabilities, which can capture the differences in reflectance characteristics of different burning materials (such as coniferous forests, broad-leaved forests, shrubs, meadows, peat, etc.) in different spectral bands in the target fire area. The raw thermal infrared radiation data was acquired using a multi-angle thermal infrared sensor array mounted on the UAV. It can capture the thermal radiation signals of spatial sampling points from multiple perspectives, effectively avoiding the interference of smoke and terrain obstruction on the data from a single perspective, while also capturing the temperature distribution differences in the fire area. The UAV swarm can adopt a grid-based deployment strategy, flexibly adjusting the swarm size, flight altitude, and scanning path according to the scope of the target fire area, the complexity of the terrain, and the urgency of the fire, to ensure no blind spots in the fire area.

[0029] The preprocessing module generates a hyperspectral data cube of the target fire area based on the original spectral reflectance data; and generates temperature distribution data of the target fire area based on the original thermal infrared radiation data.

[0030] For clarity, in the embodiments of this application, such as Figure 3 As shown, the preprocessing module includes a spectral correction unit, which performs geometric fine correction on the original spectral reflectance data based on the real-time positioning and attitude data of each UAV and the interior and exterior orientation elements of the hyperspectral imager. The corrected data is then mapped to a unified geographic coordinate system and corrected to generate a spectral data array.

[0031] The data cube construction unit is used to integrate and interpolate the spectral data array according to the preset spatial grid and band order to generate a hyperspectral data cube of the target fire area, mathematically represented as Cube(x, y, λ), where x and y are spatial coordinates and λ is the spectral band, fully preserving the band spectral characteristics of each spatial location.

[0032] The temperature reconstruction unit is used to generate three-dimensional temperature distribution data T(x, y, z) based on the real-time positioning and attitude data of the UAV and the installation parameters of the multi-angle thermal infrared sensor array, using the RANSAC algorithm to perform point cloud matching and stitching, where x, y, and z are three-dimensional spatial coordinates. This enables a three-dimensional representation of the temperature field of the fire scene and provides accurate data for subsequent fire intensity classification.

[0033] like Figure 2 As shown, the data analysis module includes a spectral analysis submodule and a grade judgment submodule.

[0034] The spectral analysis submodule is used to output the type of combustible material and its proportion weight for each pixel based on the hyperspectral data cube through the first deep learning model.

[0035] Specifically, the spectral analysis submodule includes a data partitioning unit, which is used to divide the hyperspectral data cube Cube(x, y, λ) into multiple local three-dimensional data blocks. The block size is set to M×M×K (M is the side length of the spatial dimension; K is the number of spectral bands) to ensure that each data block contains complete spectral features and retains local spatial correlation information, thereby avoiding overfitting during model training.

[0036] The feature extraction unit, wherein the first deep learning model is a three-dimensional convolutional neural network model, is used to sequentially input each local three-dimensional data block into the feature extraction layer of the three-dimensional convolutional neural network, and extract the joint features of spectral and spatial data through convolution and pooling operations.

[0037] It should be noted that a 3D-CNN model is used, and spectral-spatial joint features are extracted through convolution and pooling operations. The core convolution operation formula is as follows:

[0038]

[0039] Where F is the feature value on the feature map after convolution; P×Q×R is the 3D convolution kernel size (spatial dimension P×Q, spectral dimension R); W is the convolution kernel weight; b is the bias term; σ is the activation function (using the ReLU function, σ(x)=max(0,x)); Cube is the input local 3D data block, which is subsequently compressed in terms of feature dimension through max pooling to retain key features.

[0040] The probability output unit is used to input the extracted joint spectral and spatial features into the fully connected layer and classification layer of the 3D convolutional neural network. The classification layer outputs the probability value of each pixel belonging to a preset combustible material endmember. In other words, the extracted joint features are input into the fully connected layer and the Softmax classification layer, and the output is the probability value of each pixel belonging to a preset combustible material endmember (e.g., coniferous forest, broadleaf forest, etc., with a total of n classes). The Softmax formula is:

[0041]

[0042] Where: P(c) is the probability that a pixel belongs to the c-th type of combustible endmember; is the c-th type feature value output by the fully connected layer (i.e., the feature vector component output by the fully connected layer for the c-th type of combustible endmember, generated by the extracted spectral and spatial joint features through linear transformation and activation function processing of the fully connected layer); n is the preset total number of combustible endmembers; the denominator is the exponent sum of all category feature values, ensuring that the sum of all probability values ​​is 1, which is completely consistent with the mathematical definition of the Softmax function.

[0043] The weight generation unit is used to determine at least one type of combustible material based on the preset combustible material end-member whose probability value exceeds a preset threshold, and to use the normalized probability value as the corresponding proportion weight.

[0044] It should be noted that a preset probability threshold α is set (usually 0.1, which can be adjusted according to actual accuracy requirements), and combustible endmembers with P(c) > α are screened out. Then, the probability values ​​after screening are normalized and used as the proportion weights. The normalization formula is:

[0045]

[0046] Where: w is the proportion weight of the c-th type of combustible material; S is the set of combustible endmembers for P(c)>α. Through this step, the proportion of each type of combustible material in each pixel is determined, thus solving the identification problem in mixed combustion scenarios.

[0047] The level determination submodule is used to determine the fire intensity level corresponding to each spatial sampling point based on temperature distribution data and through a preset temperature-intensity mapping relationship.

[0048] Specifically, the grade judgment submodule includes a threshold comparison unit, which compares the temperature value corresponding to each spatial sampling point in the temperature distribution data with a first temperature threshold and a second temperature threshold, wherein the second temperature threshold is greater than the first temperature threshold.

[0049] Understandably, the first temperature threshold 373K, or 100℃, can be taken as the critical temperature for smoldering; the second temperature threshold. We can choose 873K, which is 600℃, corresponding to the critical temperature for vigorous combustion, and this meets the requirements. > .

[0050] The fire intensity level determination unit is used to perform a determination operation based on the comparison results: if the temperature value is lower than the first temperature threshold, the fire intensity level is determined to be Level 1, i.e., smoldering state, corresponding to peat smoldering, embers, and other scenarios; if the temperature value is between the first temperature threshold and the second temperature threshold, the fire intensity level is determined to be Level 2, i.e., open flame state, corresponding to grass and shrub burning, broadleaf forest burning, and other scenarios; if the temperature value is higher than the second temperature threshold, the fire intensity level is determined to be Level 3, i.e., intense burning state, corresponding to coniferous forest crown fire, large-area meadow fire, and other scenarios.

[0051] The combustible material determination module is used to output the dominant combustible material type of a pixel region based on the type of combustible material and its proportion weight, combined with combustible material combustion characteristic parameters.

[0052] Specifically, such as Figure 5As shown, the combustible material determination module includes: a combustible material combustion calculation unit, used to calculate the combustion index of the combustible material type for each pixel based on the proportion weight of the multiple combustible material types it contains and the corresponding combustible material combustion characteristic parameters obtained from the database; and a determination unit, used to determine the combustible material type with the largest combustion index as the dominant combustible material type of the pixel area.

[0053] It should be noted that a database containing various combustible combustion characteristic parameters is pre-installed in the local storage unit. The database is stored according to the type of combustible end-member, corresponding to various types of combustibles such as coniferous forests, broad-leaved forests, shrubs, meadows, and peat. Each type of combustible entry is associated with at least two sets of core combustion characteristic parameters: combustion rate and calorific value. A unique mapping relationship is established between the parameters and the combustible end-member tags, which can be directly matched and called through the combustible type tags output by the spectral analysis submodule, avoiding parameter calling errors.

[0054] When the combustible combustion calculation unit is working, it can directly call the corresponding combustible combustion characteristic parameters and calculate the combustion index of each type of combustible using the weighted product method. The specific formula is as follows:

[0055]

[0056] in, The flammability index of Class C combustibles, The weighting of Class c combustible materials is determined by the normalized probability value output from the spectral analysis submodule. The combustion rate parameters for Class C combustibles were obtained from the database; The calorific value parameters of Class C combustibles obtained from the database. and These represent the weighting coefficient and the combustion index, respectively. The impact of combustibles on fire behavior was comprehensively quantified. Higher values ​​indicate a stronger dominant role of the combustible in the fire situation, resulting in assessments that better reflect actual fire conditions. The assessment unit selected the combustion index. The largest type of combustible material, serving as the dominant combustible material type in this pixel region.

[0057] By calculating the combustion index, the problem of some combustibles with high combustion rate and high calorific value but low weight being overlooked in the determination of the dominant type due to their small physical quantity is effectively avoided. This ensures that the system can accurately identify and warn of potential high-risk sources in the fire scene, thereby improving the accuracy of risk assessment of the comprehensive situation map and the safety of fire extinguishing decisions.

[0058] The situation generation module is used to spatially align and merge the dominant combustible types with the fire intensity level, and to label each geographic unit in the target fire area with the corresponding combustible type label and fire intensity label, so as to generate a comprehensive situation map that characterizes combustible types and fire intensity.

[0059] Specifically, the situation generation module includes: an alignment and fusion unit, which performs coordinate system and grid resampling on the spatial distribution data corresponding to combustible material types and fire intensity levels, and combines the dominant combustible material type and combustion state into a structured semantic label pair for each geographically overlapping pixel pair; and an identification and mapping unit, which maps the semantic label pairs into classification labels displayed on the comprehensive situation map according to preset semantic fusion rules.

[0060] It should be noted that the alignment and fusion unit first unifies the coordinate system of the dominant combustible type data and the fire intensity level data, and then uses grid resampling (for example, using the nearest neighbor interpolation method to ensure consistent data spatial resolution) to make the pixels of the two types of data geographically overlap. Subsequently, a structured semantic label pair including the dominant combustible type and the fire intensity level is generated for each overlapping pixel, such as (coniferous forest, level 3) and (peat, level 1).

[0061] The label mapping unit, based on preset semantic fusion rules, maps label pairs to visual classification labels (such as color + symbol combinations: a solid red circle represents "Level 3 - Coniferous Forest", a hollow yellow circle represents "Level 2 - Grass and Shrub", and a gray triangle represents "Level 1 - Peat"), ultimately generating a comprehensive situation map. This map fully presents the spatial relationship between "combustion type - fire intensity," which, compared to traditional monitoring that only provides information on the location of fire points, is closer to the actual situation at the fire scene, significantly improving identification accuracy and providing precise spatial basis for firefighting tactics.

[0062] In another embodiment of this application, the fire prediction module of the fire situation feature data acquisition and classification identification processing system further includes a fire prediction module, which is used to extract spatiotemporal features of the target fire area, including the dominant combustible type, combustion state and its spatial distribution, based on the comprehensive situation map; and input the spatiotemporal features into a preset second deep learning model; the second deep learning model is configured to output the fire spread direction prediction and fire spread speed prediction of each geographical unit in the target fire area within a preset time period in the future, based on the combustion characteristics and spread patterns corresponding to different combustible types.

[0063] The fire prediction module includes a graph structure construction unit, which is used to construct a spatial graph structure with each geographic unit in the comprehensive situation map as a node and the spatial adjacency relationship between geographic units as an edge; and to encode the combustible type label and combustion status label of each geographic unit at multiple consecutive times as a feature vector to form the time series features of each node on the spatial graph structure.

[0064] The spatiotemporal prediction unit, in which the second deep learning model is a spatiotemporal graph convolutional network model, is used to input the spatial graph structure with time series features into the second deep learning model, capture spatial dependencies through graph convolutional layers, capture temporal dependencies through its temporal convolutional layers, and finally generate fire spread direction prediction and fire spread speed prediction through its output layer.

[0065] Specifically, each geographic unit in the comprehensive situation map is used as a node in the spatial map structure. The resolution of the geographic unit remains the same as before (3m×3m). Each node corresponds to a unique geographic coordinate (x, y) and is associated with the core attributes of the geographic unit (dominant combustible type label, combustion status label). If the target fire area is divided into N geographic units after grid division, the spatial map structure contains N nodes. The node number is mapped one-to-one with the geographic unit coordinates, which facilitates subsequent result backtracking.

[0066] Using the spatial adjacency relationship between geographic units as edges, the "four-neighbor adjacency rule" is adopted for determination—that is, a node (geographic unit) is only connected to nodes directly adjacent in the four directions of up, down, left, and right, without considering indirect adjacency in the diagonal direction, to avoid overgeneralization of spatial dependencies. The weight of the edge is set to a fixed value of 1.0. If there are differences in combustion state between adjacent geographic units (such as open flame area and smoldering area being adjacent), the weight can be dynamically adjusted to 1.2 to strengthen the spatial association of areas with abrupt changes in combustion state. The construction of edges is automatically generated by calculating the distance of geographic coordinates, and the adjacency determination threshold is 3m (consistent with the edge length of the geographic unit) to ensure accurate adjacency relationships.

[0067] The final constructed spatial graph structure is represented as G=(V, E), where V is the set of nodes (V={v1,v2,...,v...). n E is the set of edges (E is an N×N adjacency matrix, where the adjacency matrix element aᵢⱼ=1 indicates that nodes vᵢ and vⱼ are adjacent, aᵢⱼ=0 indicates that they are not adjacent, and aᵢⱼ=1.2 indicates adjacent nodes with a sudden change in burning state). The adjacency matrix is ​​stored in a sparse matrix format to reduce memory usage.

[0068] A comprehensive situational map of multiple consecutive moments was selected as the time-series data source. The time interval Δt was set to 5 minutes (determined by considering the drone scanning cycle and the fire spread rate; too short a Δt would increase computational load, while too long a Δt would result in the loss of time-series details). The number of consecutive moments T was set to 12 (i.e., covering 60 minutes of historical fire data, balancing the completeness of time-series information with model training efficiency). If there were fewer than 12 historical moments, forward padding was used to supplement the data (filling in sequentially based on the earliest moment data, labeled as "interpolated supplementary data" to avoid time-series breaks).

[0069] The combustible material type label and combustible state label of each geographic unit at T consecutive time points are encoded into a fixed-dimensional feature vector, which serves as the time series feature of the corresponding node. The encoding rules are disclosed as follows:

[0070] Combustion type label coding: The unique thermal coding method is adopted. For example, there are 5 types of combustibles + 1 type of non-combustibles, a total of 6 types of labels, which are coded as a 6-dimensional vector (e.g., coniferous forests correspond to [1,0,0,0,0,0], broad-leaved forests correspond to [0,1,0,0,0,0], and non-combustibles correspond to [0,0,0,0,0,1]).

[0071] The numerical coding method is adopted, with the first level (smoldering) coded as 1, the second level (open flame) coded as 2, the third level (vigorous burning) coded as 3, and no combustible material coded as 0;

[0072] The encoding results at each time step are concatenated into a 7-dimensional vector (6-dimensional combustible type vector + 1-dimensional combustion state value). The vectors at T time steps are arranged in chronological order to form a T×7 time series feature matrix. This matrix is ​​then flattened into a 7T-dimensional feature vector (T=12 in this embodiment, and the feature vector dimension is 84-dimensional) as the final temporal feature of the node.

[0073] The encoded feature vectors are normalized (using Min-Max normalization to compress the values ​​to the [0,1] range) to avoid interference from differences in feature values ​​of different dimensions on model training (in the system, the unique heat vector of the combustion type does not need to be normalized).

[0074] The model input consists of a node feature matrix and an adjacency matrix, and the output consists of the predicted spread direction and speed of each geographic unit. The structure and core parameters are as follows:

[0075] Input layer: dimension N×84, synchronously inputting an N×N adjacency matrix to represent spatial association.

[0076] Spatiotemporal feature extraction module: 2 stacked ST-GCN units, each unit contains graph convolutional layers (ChebNet operator, 1×1 convolutional kernel, 64 / 128 output channels), temporal convolutional layers (1D convolution, 3×1 kernel, stride 1), batch normalization layers (momentum 0.9) and ReLU activation function.

[0077] Dropout layer: Located between two ST-GCN units, with a dropout rate of 0.5, to prevent overfitting.

[0078] Output prediction module: 1 fully connected layer with 256 neurons (ReLU activation) + output layer, output dimension 2×6×N, linear activation function directly outputs the predicted value.

[0079] In another embodiment of this application, the fire prediction module of the fire situation feature data acquisition and classification identification processing system further includes: a dispatch suggestion generation module, used to receive available fire extinguishing resource information from an embedded online synchronously updated fire extinguishing resource database. The available fire extinguishing resource information includes resource type (such as fixed-wing aircraft, helicopters, water tanker fire trucks, dry powder fire trucks, etc.), quantity (such as available units / persons for a single type of resource), and performance parameters (such as aircraft load capacity, fire truck water supply capacity, flame retardant coverage area, etc.), wherein the parameter format is uniformly preset to facilitate matching and analysis; the comprehensive situation map is matched and analyzed with the available fire extinguishing resource information, and based on the preset fire situation feature and fire extinguishing resource mapping table, fire extinguishing resource dispatch suggestions corresponding to different areas in the comprehensive situation map are output.

[0080] It is understandable that the aforementioned fire-fighting resource database is a relational or time-series database, the core of which consists of one or more interrelated data tables. Each record represents an independently dispatchable fire-fighting resource unit (such as an aircraft, a fire truck, or a fire brigade).

[0081] Each record contains at least the following structured fields: Resource Unique Identifier (ID): Used to uniquely identify and track the resource unit.

[0082] Resource types: coded and labeled according to a preset classification system, such as: fixed-wing firefighting aircraft, helicopters (buckets), heavy-duty water tanker fire trucks, dry powder / foam combined fire trucks, long-range water supply systems, special fire extinguishing agent delivery units, front-line fire brigades, etc.

[0083] Resource status: Indicates the current combat readiness status of the resource, such as: standby, in deployment, in operation, returning to base / returning to base, under maintenance, unavailable.

[0084] Real-time location information: latitude and longitude coordinates obtained through GPS / BeiDou module, and the most recent update timestamp.

[0085] Quantity / capacity parameters: These refer to the inherent, quantifiable capacity indicators of the resource unit, such as: aircraft / helicopters: maximum liquid / flame retardant capacity, single-operation coverage area (hectares); fire trucks: water tank capacity, water pump flow rate, dry powder storage capacity; fire brigades: number of personnel that can be dispatched, types of equipment that can be carried.

[0086] Performance parameters: Parameters that describe the characteristics of resource operation, such as: cruising / driving speed, maximum operating range / radius, suitable fire types, and adaptability to terrain / weather conditions.

[0087] It should be noted that the fire characteristics (dominant combustible material, intensity level) of each area in the comprehensive situation map are first integrated with the fire prediction results (spread direction, speed) to form a comprehensive label for the regional fire situation; then the mapping table is called, and combined with the constraints of the number of available resources, suitable resources are selected and priority is assigned (such as priority scheduling for areas with fast spread speed and high intensity).

[0088] Based on a pre-set "fire characteristics - fire extinguishing resources" mapping table, the mapping table divides fire scenarios into three dimensions: "burning material type + fire intensity + spread rate", and matches the optimal resource type and quantity accordingly. For example, a level 3 fire in coniferous forest (fierce burning) is matched with a large fixed-wing aircraft + fire retardant, and a level 2 fire in grass and shrubland is matched with helicopter bucket operations.

[0089] It outputs structured scheduling suggestions, including regional coordinates, suitable resource types and quantities, deployment order, and operation scope, which are overlaid with the comprehensive situation map and fire prediction map, and can be directly synchronized to the command terminal.

[0090] In summary, the fire situation feature data acquisition and classification identification processing system according to the embodiments of this application can output the dominant combustible type at different locations of the target fire scene when identifying a fire. By spatially aligning and fusing the dominant combustible type with the fire intensity level, a more realistic comprehensive situation map can be obtained, which effectively improves the identification accuracy and accurately predicts the fire trend within a preset time period. Compared with traditional experience models, this prediction method fully considers the influence of combustible type. The prediction results can provide forward-looking data support for the formulation of fire fighting tactics, help to deploy fire fighting forces in advance, and curb the spread of fire.

[0091] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0093] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A system for collecting, classifying, and processing fire-related characteristic data, characterized in that, include: The data acquisition module is used to acquire the raw spectral reflectance data and raw thermal infrared radiation data of each spatial sampling point within the target fire area scanned by the UAV; The preprocessing module generates a hyperspectral data cube of the target fire area based on the original spectral reflectance data; and generates temperature distribution data of the target fire area based on the original thermal infrared radiation data. The data analysis module includes a spectral analysis submodule and a grade determination submodule. The spectral analysis submodule is used to output the type of combustible material and its proportion weight corresponding to each pixel based on the hyperspectral data cube through a first deep learning model. The level determination submodule is used to determine the fire intensity level corresponding to each spatial sampling point based on temperature distribution data and through a preset temperature-intensity mapping relationship; The combustible material determination module is used to output the dominant combustible material type of the pixel area based on the combustible material type and its proportion weight, and in combination with combustible material combustion characteristic parameters. The situation generation module is used to spatially align and fuse the dominant combustible type with the fire intensity level, and to label each geographic unit in the target fire area with a corresponding combustible type label and fire intensity label, so as to generate a comprehensive situation map that characterizes combustible type and fire intensity.

2. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, The spectral analysis submodule includes: A data partitioning unit is used to divide the hyperspectral data cube into multiple local three-dimensional data blocks; The feature extraction unit, wherein the first deep learning model is a three-dimensional convolutional neural network model, is used to sequentially input each of the local three-dimensional data blocks into the feature extraction layer of the three-dimensional convolutional neural network, and extract joint spectral and spatial features through convolution and pooling operations; The probability output unit is used to input the extracted spectral and spatial joint features into the fully connected layer and classification layer of the three-dimensional convolutional neural network, and output the probability value of each pixel belonging to various preset combustible endmembers through the classification layer; The weight generation unit is used to determine at least one type of combustible material based on the preset combustible material end-member whose probability value exceeds a preset threshold, and to use the normalized probability value as the corresponding proportion weight.

3. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, The level determination submodule includes: A threshold comparison unit is used to compare the temperature value corresponding to each spatial sampling point in the temperature distribution data with a first temperature threshold and a second temperature threshold, wherein the second temperature threshold is greater than the first temperature threshold. The level determination unit is used to perform a determination operation based on the comparison result: if the temperature value is lower than the first temperature threshold, the fire intensity level is determined to be the first level; If the temperature value is between the first temperature threshold and the second temperature threshold, then the fire intensity level is determined to be the second level. If the temperature value is higher than the second temperature threshold, the fire intensity level is determined to be level three.

4. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, The combustible material determination module includes: The combustible combustion calculation unit is used to calculate the combustion index of the combustible type for each pixel based on the proportion weight of the multiple combustible types it contains and the corresponding combustible combustion characteristic parameters obtained from the database. The determination unit is used to determine the combustible type with the highest flammability index as the dominant combustible type of the pixel area.

5. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, The situation generation module includes: The alignment and fusion unit is used to perform coordinate system one and grid resampling on the spatial distribution data corresponding to the combustible material type and the fire intensity level, and for each geographically overlapping pixel pair, combine the corresponding dominant combustible material type and combustion state into a structured semantic tag pair. The identifier mapping unit is used to map the semantic tag pairs into classification identifiers displayed on the comprehensive situation map according to preset semantic fusion rules.

6. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, The preprocessing module includes: The spectral correction unit is used to perform geometric fine correction on the original spectral reflectance data based on the real-time positioning and attitude data of each UAV and the interior and exterior orientation elements of the hyperspectral imager. The corrected data is mapped to a unified geographic coordinate system and then corrected to generate a spectral data array. The data cube construction unit is used to integrate and interpolate the spectral data array according to a preset spatial grid and band order to generate the hyperspectral data cube of the target fire area. The temperature reconstruction unit is used to perform multi-view matching and three-dimensional point cloud reconstruction on the original thermal infrared radiation data based on the real-time positioning and attitude data of each UAV and the installation parameters of the multi-angle thermal infrared sensor array, so as to generate the temperature distribution data of the target fire area.

7. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, Also includes: The fire prediction module is used to extract spatiotemporal features of the target fire area, including the dominant combustible type, the combustion state and its spatial distribution, based on the comprehensive situation map. The spatiotemporal features are then input into a preset second deep learning model. The second deep learning model is configured to output, based on the combustion characteristics and spread patterns of different types of combustibles, the predicted fire spread direction and fire spread speed of each geographical unit in the target fire area within a preset time period in the future.

8. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 7, characterized in that, The fire prediction module includes: The graph structure construction unit is used to construct a spatial graph structure with each geographic unit in the comprehensive situation map as a node and the spatial adjacency relationship between geographic units as an edge; and to encode the combustible type label and combustion status label of each geographic unit at multiple consecutive times as a feature vector to form the time series features of each node on the spatial graph structure. The spatiotemporal prediction unit, wherein the second deep learning model is a spatiotemporal graph convolutional network model, is used to input the spatial graph structure with the time series features into the second deep learning model, capture spatial dependencies through graph convolutional layers, capture temporal dependencies through its temporal convolutional layers, and finally generate the fire spread direction prediction and the fire spread speed prediction through its output layer.

9. The fire situation characteristic data acquisition, classification, and identification processing system according to claim 1, characterized in that, Also includes: The scheduling suggestion generation module is used to receive available fire-fighting resource information from the fire-fighting resource database, wherein the available fire-fighting resource information includes resource type, quantity and performance parameters; The integrated situation map is then matched and analyzed with the available fire-fighting resource information. Based on a preset fire characteristics and fire-fighting resource mapping table, fire-fighting resource scheduling suggestions corresponding to different areas in the integrated situation map are output.

Images