A forest fire danger dynamic early warning system and method

Abstract

The application discloses a forest fire danger dynamic early warning system and method, and belongs to the technical field of forest fire early warning. The method continuously collects meteorological parameters, vegetation water content, terrain characteristics and historical lightning point distribution data through multiple source sensing nodes, inputs the multiple source data into a recursive convolution network after aligning the multiple source data according to the same time index, uses a time attention module in the recursive convolution network to dynamically allocate weights to the aligned data, obtains fire danger inducing feature sequences of each sensing node at different time steps, maps the fire danger inducing feature sequences to a pre-constructed fire danger grade atlas, determines an initial fire danger grade at the current moment by comparing the distance between the feature sequence and each grade node center vector, and generates a fire danger dynamic distribution grid covering the whole monitoring area according to the initial fire danger grade and a fire danger propagation speed parameter shared between adjacent sensing nodes, and the fire danger dynamic distribution grid is used as real-time early warning output.

Description

Technical Field

[0001] This invention relates to the field of forest fire early warning technology, specifically to a dynamic forest fire risk early warning system and method. Background Technology

[0002] Forest fire risk early warning relies on the dynamic perception and fusion analysis of multiple environmental factors. In existing technologies, meteorological, vegetation, and topographical information collected by multi-source sensors is typically processed in segments, lacking in-depth alignment and correlation modeling at the same time reference. This results in insufficient utilization of time-series information and an inability to accurately capture the evolution of fire-inducing factors over continuous time. Some solutions use convolutional neural networks or recurrent neural networks to process single-type time-series data, but assign equal importance to data at different times, ignoring the significant differences in the contribution of different time steps to the current fire risk level. This makes the model slow to respond to sudden weather changes or rapid changes in vegetation moisture content, resulting in insufficient real-time performance and accuracy of early warnings.

[0003] Furthermore, existing early warning methods typically only output discrete fire risk levels for monitoring points, without considering the spatial propagation characteristics of fire between adjacent areas. In real-world scenarios, once a forest fire occurs, the fire risk is not an isolated point distribution, but a dynamic process of continuous spread along terrain, vegetation, and wind fields. Discrete level early warnings cannot reflect the continuous changes and spread trends of fire risk between regions, making it impossible to guide prevention and control resources to critical paths where the fire may spread rapidly.

[0004] To address the aforementioned issues, adaptive attention modeling of multi-source time-series data is needed. This would enable the model to automatically focus on key time steps based on the temporal dependencies of the data, resolving the early warning delay problem caused by fixed weights. Simultaneously, the point-like fire risk level needs to be expanded into a spatially continuous dynamic distribution of fire risk, accurately describing the propagation process of fire between adjacent areas and addressing the problem that discrete early warning systems cannot characterize the spread trend. Summary of the Invention

[0005] This invention provides a dynamic early warning system and method for forest fire risk, aiming to achieve continuous spatiotemporal dynamic early warning of forest fire risk and improve the timeliness and spatial coverage of the early warning.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] This invention discloses a method and system for dynamic early warning of forest fire risk. The method includes: continuously collecting meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data through multi-source sensor nodes deployed within a monitoring area; aligning the meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data according to the same time index and inputting them into a recurrent convolutional network; using the time attention module in the recurrent convolutional network to dynamically assign weights to the aligned data to obtain a fire risk induction feature sequence for each sensor node at different time steps; mapping the fire risk induction feature sequence to a pre-constructed fire risk level map; determining the initial fire risk level at the current moment by comparing the distance between the fire risk induction feature sequence and the center vector of each level node in the fire risk level map; generating a dynamic fire risk distribution grid covering the entire monitoring area based on the initial fire risk level and the fire risk propagation speed parameter shared between adjacent sensor nodes; and outputting the dynamic fire risk distribution grid as a real-time early warning. This method integrates multi-source heterogeneous data through recurrent convolutional networks and temporal attention mechanisms, which can accurately extract the temporal features of fire risk induction and realize the spatial dynamic extrapolation of fire risk level by using the propagation speed parameter, significantly improving the spatiotemporal resolution and real-time performance of forest fire risk early warning.

[0008] As a preferred embodiment of the present invention, the multi-source sensing node is equipped with a wind speed sensor, an air temperature sensor, an air humidity sensor, a rainfall sensor, and a spectral reflectance meter, which are used to collect meteorological parameters and vegetation moisture content, respectively; terrain features are extracted from a pre-stored digital elevation model of the monitoring area; historical lightning strike point distribution data are retrieved from a historical fire record database, showing the locations of lightning strike events within a specified radius around each sensing node within the last three years. During data alignment, various types of data are timestamped at a fixed sampling period, data with the same timestamp are combined into one input sample, and arranged into an input sequence according to the order of the timestamps, which are then fed sequentially into the input layer of the recurrent convolutional network.

[0009] As another preferred technical solution of the present invention, the process of obtaining the fire hazard induction feature sequence using the time attention module includes: extracting the hidden state vector corresponding to each time step in the hidden layer of the recurrent convolutional network; inputting the hidden state vector into the scoring function in the time attention module to obtain the original attention score for each time step; converting the original attention scores of all time steps into attention weights after normalization; and multiplying the hidden state vector of each time step with the corresponding attention weight and concatenating them in the order of time steps to form the fire hazard induction feature sequence. Preferably, when allocating dynamic weights, the time attention module assigns higher attention weights to hidden state vectors closer to the current time step and lower attention weights to hidden state vectors farther from the current time step, thereby strengthening the contribution of recent environmental changes to fire hazard.

[0010] As another preferred technical solution of the present invention, the construction method of the pre-constructed fire risk level map includes: collecting historical fire risk cases in the monitoring area over the past several years, each historical fire risk case containing multiple fire risk inducing feature samples of consecutive time steps and the corresponding actual fire risk level; grouping the fire risk inducing feature samples in all historical fire risk cases according to the actual fire risk level; calculating the mean of each dimension for all samples within the same fire risk level group, and combining the mean values ​​of each dimension into the central vector corresponding to that fire risk level; arranging the central vectors corresponding to all fire risk levels from low to high, thus forming the fire risk level map. When determining the initial fire risk level, the feature vector corresponding to the last time step is extracted from the fire risk inducing feature sequence as the current feature vector, and the Euclidean distance between the current feature vector and the central vector of each level node in the fire risk level map is calculated respectively. The fire risk level corresponding to the level node with the smallest Euclidean distance is selected as the initial fire risk level at the current moment.

[0011] As a further preferred technical solution of the present invention, the process of generating a dynamic fire risk distribution grid includes: taking the location of each sensor node as the center point of the grid cell, and assigning the initial fire risk level to the grid cell where the corresponding center point is located; obtaining the fire risk propagation speed parameter shared between two adjacent sensor nodes; calculating the propagation time required for the fire to spread from one sensor node to another based on the fire risk propagation speed parameter and the straight-line distance between two adjacent sensor nodes; performing linear interpolation of the fire risk level on the intermediate grid cell between two adjacent sensor nodes based on the propagation time; and stitching all grid cells together according to their respective geographical coordinates to form a dynamic fire risk distribution grid. The linear interpolation is specifically implemented as follows: determine the first fire hazard level of the first grid cell corresponding to the adjacent first sensing node, and the second fire hazard level of the second grid cell corresponding to the second sensing node; determine the total straight-line path distance between the first grid cell and the second grid cell, and the number of intermediate grid cells passed through the path; for any intermediate grid cell, calculate the ratio of the distance from the intermediate grid cell to the first grid cell to the total straight-line path distance, multiply the first fire hazard level by a factor of 1 and subtract the ratio, add the result of multiplying the second fire hazard level by the ratio, and obtain the fire hazard level of the intermediate grid cell.

[0012] Preferably, the fire spread velocity parameter shared between adjacent sensor nodes is obtained through historical data statistics: multiple pairs of adjacent sensor nodes are selected within the monitoring area; for each pair of adjacent sensor nodes, the time difference between when the pair of adjacent sensor nodes detected the same forest fire and the actual distance between the pair of adjacent sensor nodes are obtained; the actual distance is divided by the time difference to obtain the single fire spread velocity between the pair of adjacent sensor nodes; the single fire spread velocity corresponding to all adjacent sensor nodes is subjected to median filtering, and the median velocity obtained after filtering is used as the fire spread velocity parameter shared between all adjacent sensor nodes. As another optional implementation, the fire spread velocity parameter can also be dynamically calculated based on the average wind speed, vegetation continuity, and terrain slope between two adjacent sensor nodes to reflect the impact of the environment on fire spread in real time.

[0013] As a supplementary technical solution of the present invention, before sending the aligned data into the time attention module for dynamic weight allocation, the data is preprocessed: continuous numerical values ​​are normalized and discrete category labels are one-hot encoded; the normalized and one-hot encoded data are reassembled into a standardized input tensor according to the original time order, and the standardized input tensor is used as the actual processing object of the time attention module, thereby eliminating the difference in dimensions and improving the network processing efficiency.

[0014] This invention also includes a dynamic forest fire risk early warning system. The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the aforementioned dynamic forest fire risk early warning method. This system can automatically collect and fuse multi-source sensor data, dynamically assess fire risk levels using recurrent convolutional networks and temporal attention mechanisms, and generate a high-resolution dynamic fire risk distribution grid by combining propagation speed parameters, providing accurate and timely decision support for forest fire prevention and management.

[0015] The beneficial effects of this invention are:

[0016] On one hand, a recurrent convolutional network combined with a temporal attention module is used to dynamically assign weights to multi-source sensor data. After aligning meteorological parameters, vegetation moisture content, topographic features, and historical lightning strike distribution data according to the same time index, the recurrent convolutional network extracts the latent state vector for each time step. The temporal attention module generates an initial attention score for each time step using a scoring function and normalizes it to obtain attention weights. The latent state vectors of each time step are then weighted and fused according to the attention weights to form a fire hazard induction feature sequence. This approach enables the model to automatically learn the different contributions of different time steps to the fire hazard state, assigning greater weight to latent states that are closer to the current time step and have a higher correlation with fire hazard triggering, while suppressing time steps that are far away or have weak correlation. This allows for rapid response when meteorological elements change abruptly or vegetation moisture content decreases rapidly, shortening the delay from environmental changes to early warning output, and significantly improving the timeliness of early warnings and sensitivity to dynamic changes.

[0017] On the other hand, after obtaining the initial fire risk level of the sensor nodes, the fire propagation velocity parameter shared between adjacent sensor nodes is introduced to generate a dynamic fire risk distribution grid covering the entire monitoring area. The initial fire risk level of each sensor node is assigned to the corresponding grid cell. The propagation time is calculated using the shared propagation velocity parameter between adjacent sensor nodes and their straight-line distance. Based on this, the fire risk level of the intermediate grid cells between adjacent nodes is linearly interpolated. All grid cells are then stitched together according to geographical coordinates to form a dynamic fire risk distribution grid. This process transforms the discrete point-like fire risk level into a spatially continuous risk map, accurately reflecting the transmission and spread characteristics of fire between adjacent areas. This allows the early warning results to show the direction, rate, and intensity changes of fire spread along terrain and wind direction, providing continuous decision-making basis for accurately delineating evacuation areas and deploying containment forces. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of the dynamic early warning method for forest fire risk;

[0020] Figure 2 This is a flowchart of the method for determining the initial fire hazard level based on the fire hazard rating map;

[0021] Figure 3 This is a flowchart of the method for generating dynamic fire hazard distribution grids. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1 This invention provides a dynamic early warning method for forest fire risk. The method includes: continuously collecting meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data from multi-source sensor nodes deployed within a monitoring area; aligning these data according to the same time index and inputting them into a recurrent convolutional network; dynamically assigning weights to the aligned data using a time attention module within the recurrent convolutional network to obtain a fire risk induction feature sequence for each sensor node at different time steps; mapping the fire risk induction feature sequence to a pre-constructed fire risk level map; determining the initial fire risk level at the current moment by comparing the distance between the fire risk induction feature sequence and the center vectors of each level node in the fire risk level map; generating a dynamic fire risk distribution grid covering the entire monitoring area based on the initial fire risk level and the fire propagation speed parameters shared between adjacent sensor nodes; and using the dynamic fire risk distribution grid as a real-time early warning output.

[0024] By aligning multi-source data over time and feeding it into a recurrent convolutional network, the temporal attention module dynamically highlights the contribution of different time steps to fire risk prediction. Then, the high-dimensional feature sequence is matched with a fire risk level map to a specific initial fire risk level. Finally, the initial fire risk level of discrete nodes is expanded into a dynamic fire risk distribution grid in continuous space by combining the fire risk propagation speed parameter, ensuring that the early warning results can reflect the spatial diffusion trend of fire risk in the monitoring area.

[0025] In practice, multi-source sensor nodes deployed within the monitoring area continuously collect meteorological parameters, vegetation moisture content, topographic features, and historical lightning strike distribution data. Each sensor node is equipped with a wind speed sensor, air temperature sensor, air humidity sensor, and rainfall sensor to collect meteorological parameters. A spectral reflectance meter is also installed at each sensor node to collect vegetation moisture content. Topographic features are extracted from a pre-stored digital elevation model of the monitoring area. The locations of lightning strikes within a specified radius around each sensor node within the past three years are retrieved from a historical fire record database and used as historical lightning strike distribution data. The specified radius is determined based on an assessment of the historical lightning impact range of the monitoring area and can be selected as 500 meters.

[0026] In practice, meteorological parameters, vegetation moisture content, topographic features, and historical lightning strike distribution data are timestamped using a fixed sampling period. The fixed sampling period is set to 10 minutes to ensure that the temporal granularity can capture short-term changes in meteorological and vegetation conditions. Meteorological parameters, vegetation moisture content, topographic features, and historical lightning strike distribution data with the same timestamp are combined into a single input sample. The input samples are arranged into an input sequence according to the order of their timestamps, and each input sample in the input sequence is sequentially fed into the input layer of the recurrent convolutional network.

[0027] The recurrent convolutional network consists of a data preprocessing module, a one-dimensional convolutional layer, a long short-term memory recurrent layer, and a temporal attention module, connected sequentially. The data preprocessing module is connected between the input layer and the one-dimensional convolutional layer. Before the aligned data is fed into the temporal attention module for dynamic weight allocation, the data preprocessing module normalizes the continuous values ​​in the aligned data and performs one-hot encoding on the discrete class labels in the aligned data.

[0028] Continuous numerical data include meteorological parameters such as wind speed, air temperature, air humidity, and rainfall; vegetation moisture content; topographic features such as slope and aspect; and the planar distance from lightning strike points to sensor nodes from historical lightning strike point distribution data. Normalization is performed using the following formula:

[0029]

[0030] in, This represents any of the above continuous numerical values ​​to be normalized. This represents the minimum value among all continuous values ​​of the same type collected by all sensor nodes within the monitoring area at the current time step. This represents the maximum value of the same type of continuous data collected by all sensor nodes within the monitoring area at the current time step. express The normalized values. The minimum and maximum values ​​mentioned above are updated in real time after aggregating and statistically analyzing the data collected by all sensor nodes at the same time step at each time step.

[0031] Discrete category labels include landform category labels in terrain features and binary labels in historical lightning strike distribution data indicating the presence of lightning strikes within a specified radius. When performing one-hot encoding on discrete category labels, each possible landform category is mapped to a one-hot vector dimension, and the lightning strike presence label is mapped to another one-hot vector dimension; if a category appears, the dimension corresponding to that category is set to 1, and the other dimensions are set to 0. All data after normalization and one-hot encoding are reassembled into a standardized input tensor according to the original timestamp order. The shape of the standardized input tensor is the number of time steps multiplied by the total feature dimension, where the total feature dimension is the sum of the continuous numerical dimension and the one-hot encoded dimension. The standardized input tensor is output from the data preprocessing module and fed into a one-dimensional convolutional layer, then sequentially through a long short-term memory recursive layer, and finally, the time attention module uses the information generated from the hidden state of the standardized input tensor as the actual processing object.

[0032] In some embodiments, the one-dimensional convolutional layer contains 32 convolutional kernels, each with a size of 3 along the time dimension and a stride of 1, using a linear rectified function as the activation function. The one-dimensional convolutional layer performs a time-dimensional convolution operation on the normalized input tensor, outputting a convolutional feature sequence. The long short-term memory recursive layer contains 64 hidden units, returning a sequence of hidden state vectors for all time steps. The long short-term memory recursive layer receives the convolutional feature sequence as input, models the temporal dependencies between multi-source data, and outputs a sequence of hidden state vectors. The hidden state vector sequence is passed to a temporal attention module, which dynamically assigns weights to the hidden state vector sequence to obtain a fire hazard induction feature sequence.

[0033] The recurrent convolutional network (RCN) needs to be trained before being used for real-time early warning. The training process uses a dataset constructed from historical fire risk cases in the monitored area. Each training sample consists of a continuous sequence of multi-source data aligned across multiple time steps and the corresponding actual fire risk level label. The multi-source data aligned sequence in the training samples is timestamped and preprocessed to obtain a standardized input tensor, which is then fed into the RCN. The time attention module outputs a fire risk induction feature sequence and extracts the feature vector of the last time step. The Euclidean distance between the feature vector of the last time step and the center vector of the corresponding actual fire risk level in the pre-constructed fire risk level map is calculated as the loss function value. An adaptive moment estimation (IME) optimization algorithm is used to update the trainable parameters in the one-dimensional convolutional layer, the long short-term memory (LSTM) recurrent layer, and the time attention module. The initial learning rate of the IEM optimization algorithm is set to 0.001, the first-order moment decay rate is set to 0.9, the second-order moment decay rate is set to 0.999, and the batch size is set to 32. Iterative training is performed for multiple epochs. Training stops when the loss function value on the validation set no longer decreases after five consecutive training epochs, thus obtaining a recurrent convolutional network for real-time inference. During real-time inference, the trained recurrent convolutional network is input into the standardized input tensor, which is aligned and preprocessed from the current time step and all previous time steps. The input tensor is then processed sequentially by a data preprocessing module, a one-dimensional convolutional layer, a long short-term memory recurrent layer, and a time attention module to obtain the fire hazard induction feature sequence of the current sensor node at different time steps.

[0034] In the implementation, the temporal attention module in the recurrent convolutional network (RCN) dynamically assigns weights to the aligned data, obtaining the fire hazard induction feature sequence for each sensor node at different time steps. The RCN consists of a data preprocessing module, a one-dimensional convolutional layer, a long short-term memory (LSM) recurrent layer, and the temporal attention module, connected in sequence. The one-dimensional convolutional layer contains 32 convolutional kernels, each with a size of 3 along the time dimension and a stride of 1, using a linear rectified function as the activation function. The LSM recurrent layer contains 64 hidden units, returning the hidden state vector sequence for all time steps. The temporal attention module consists of a global average pooling layer, a first fully connected layer, a second fully connected layer, and a normalized exponential function layer, connected in sequence. The first fully connected layer contains 32 neurons and uses a linear rectified function as the activation function. The second fully connected layer contains one neuron and does not use an activation function.

[0035] The temporal attention module extracts the hidden state vector corresponding to each time step in the hidden layer. The hidden layer is the output layer of the Long Short-Term Memory (LSTM) recursive layer, and the hidden state vector is a 64-dimensional vector output by the LSM recursive layer at each time step. Each dimension of the 64-dimensional vector corresponds to the state value of a hidden unit in the LSM recursive layer at the current time step. The hidden state vector is then input into the scoring function in the temporal attention module. The scoring function is implemented by performing forward calculations on the hidden state vector sequentially through a global average pooling layer, a first fully connected layer, and a second fully connected layer.

[0036] For the At each time step, the hidden state vector is denoted as... , It is a 64-dimensional real-valued vector. Global average pooling layer pairs The average value is calculated across 64 dimensions to obtain the pooled scalar. The first fully connected layer will pool the scalar... After multiplying with the weight matrix of the first fully connected layer and adding the bias vector of the first fully connected layer, the result is activated by a linear rectified function to obtain the output vector of the first fully connected layer. , It is a 32-dimensional real vector. The second fully connected layer takes the output vector of the first fully connected layer. The scalar value of the second fully connected layer's output is obtained by performing a dot product operation with the weight vector of the second fully connected layer and then adding the single bias value of the second fully connected layer. Scalar value For the first The original attention scores for each time step are obtained. The above forward computation operation is performed on the hidden state vectors of all time steps to obtain a sequence of original attention scores equal to the number of time steps.

[0037] In some embodiments, after the scoring function outputs the raw attention scores, the normalized exponential function layer receives the sequence of raw attention scores. ,in, This represents the total number of time steps. This represents the raw attention score at the first time step. This represents the raw attention score at the second time step. Indicates the first The raw attention scores for each time step are obtained. The normalized exponential function layer normalizes the raw attention scores for all time steps using the following formula to obtain the attention weights for each time step:

[0038]

[0039] in, Indicates the first The original attention scores at each time step. The value range is 1 to , Indicates the first The original attention scores at each time step. The summation index ranges from 1 to 1. , Represented by natural constant as the base, with The value of the exponential function of the exponent. Represented by natural constant as the base, with The value of the exponential function of the exponent. Indicates from arrive For all The value obtained by summing, Indicates the first Attention weights at each time step. It is a real number between 0 and 1, and the sum of the attention weights at all time steps is 1.

[0040] When allocating dynamic weights, the temporal attention module assigns higher attention weights to hidden state vectors closer to the current time step and lower attention weights to hidden state vectors farther away from the current time step. This characteristic is automatically learned by the temporal attention module during training. During training, the training dataset for the recurrent convolutional network is constructed from historical fire risk cases in the monitored area. Each training sample consists of a continuous sequence of multi-source data aligned across multiple time steps and the corresponding actual fire risk level label. The hidden state vector sequence obtained after data preprocessing, one-dimensional convolutional layers, and long short-term memory recurrent layers is input into the temporal attention module. The temporal attention module outputs a fire risk induction feature sequence and extracts the feature vector of the last time step in the fire risk induction feature sequence. The Euclidean distance between the feature vector of the last time step and the center vector of the corresponding actual fire risk level in the pre-constructed fire risk level map is calculated as the loss function value. An adaptive moment estimation optimization algorithm is used to update the trainable parameters of the two fully connected layers after the global average pooling layer in the temporal attention module, as well as the trainable parameters of other modules in the recurrent convolutional network. The initial learning rate of the adaptive moment estimation optimization algorithm was set to 0.001, the first-order moment decay rate to 0.9, the second-order moment decay rate to 0.999, and the batch size to 32. Iterative training was performed for multiple epochs, stopping when the loss function value on the validation set no longer decreased after five consecutive training epochs. During the training process where the loss function value continuously decreased, the temporal attention module automatically adjusted the weights and biases in the global average pooling layer, the first fully connected layer, and the second fully connected layer through backpropagation. This ensured that the original attention score corresponding to the hidden state vector closer to the current time step gradually increased, thus increasing the corresponding attention weight; conversely, the original attention score corresponding to the hidden state vector farther from the current time step gradually decreased, thus decreasing the corresponding attention weight.

[0041] The hidden state vector at each time step is multiplied by the corresponding attention weight and then concatenated in time step order to form a fire hazard induction feature sequence. For the th At each time step, calculate the weighted hidden state vector. ,in, For the first A 64-dimensional hidden state vector at each time step For the first Attention weights at each time step This represents the multiplication operation between a scalar and a vector, i.e., attention weights. With the hidden state vector Multiplying each dimension together yields a weighted latent state vector. , It is also a 64-dimensional real vector. The above weighting operation is performed at all time steps to obtain a weighted sequence of hidden state vectors. The weighted hidden state vectors are arranged in time step order. By joining the two ends together, a length of [length missing] is formed. The one-dimensional vector is the fire hazard induction feature sequence of the corresponding sensor node.

[0042] In practice, the fire hazard induction characteristic sequence is mapped to a pre-constructed fire hazard level map, and the initial fire hazard level at the current moment is determined. (See also...) Figure 2 The fire hazard induction feature sequence is output by the temporal attention module. This sequence contains a weighted latent state vector equal to the number of time steps, each a 64-dimensional real-valued vector. The weighted latent state vector corresponding to the last time step is extracted from the fire hazard induction feature sequence as the current feature vector. The current feature vector is a 64-dimensional real-valued vector. The first dimension of the current feature vector corresponds to the state value of the first hidden unit in the Long Short-Term Memory recursive layer after attention weighting at the last time step; the second dimension corresponds to the state value of the second hidden unit in the Long Short-Term Memory recursive layer after attention weighting at the last time step; and so on up to the 64th dimension.

[0043] The pre-constructed fire hazard rating map contains multiple rating nodes, each corresponding to a preset fire hazard level, and each rating node has a central vector. There are five fire hazard levels, from lowest to highest: low fire hazard, medium fire hazard, high fire hazard, very high fire hazard, and extremely high fire hazard. The fire hazard rating map arranges the rating nodes from top to bottom in the order of low fire hazard, medium fire hazard, high fire hazard, very high fire hazard, and extremely high fire hazard. The central vector of each rating node is a 64-dimensional real vector, and the dimension of the central vector is consistent with the dimension of the current feature vector.

[0044] In practice, the Euclidean distance between the current feature vector and the center vector of each level node in the fire hazard rating map is calculated. The Euclidean distance is calculated using the following formula:

[0045]

[0046] in, This represents the th feature vector of the current feature vector. The values ​​of each dimension The value range is from 1 to 64; The fire hazard rating map indicates the number of fire hazards. The center vector of the nth level node The values ​​of each dimension The values ​​are 1, 2, 3, 4, and 5, which correspond to the nodes of low fire risk, medium fire risk, high fire risk, very high fire risk, and extremely high fire risk levels, respectively. Indicates the current feature vector of the th The value of the first dimension and the first The center vector of the node of the level is... The square of the difference between the values ​​of each dimension; This indicates that the square of the above difference is from arrive The value obtained by summing; This indicates taking the arithmetic square root of the summation result; Indicates the relationship between the current feature vector and the first feature vector. The Euclidean distance between the center vectors of the nodes at each level. After calculation, five Euclidean distance values ​​are obtained. Compare the five Euclidean distance values ​​and select the fire hazard level corresponding to the level node with the smallest Euclidean distance as the initial fire hazard level for the current moment. When there are two or more identical minimum Euclidean distance values, select the level node with the highest fire hazard level from the level nodes corresponding to the identical minimum Euclidean distance values ​​as the initial fire hazard level.

[0047] The pre-constructed fire risk level map is built using the following steps: Historical fire risk cases from the past several years are collected for the monitoring area, with each period defined as five years. Each historical fire risk case contains multiple fire risk inducing feature samples from consecutive time steps and the corresponding actual fire risk level. The fire risk inducing feature samples are 64-dimensional weighted latent state vectors for the corresponding time step, weighted by a time attention module. The actual fire risk level is the fire risk level verified manually or confirmed by historical records, and is one of five levels: low fire risk, medium fire risk, high fire risk, very high fire risk, and extremely high fire risk.

[0048] All historical fire hazard cases were grouped according to the actual fire hazard level into five groups: low fire hazard, medium fire hazard, high fire hazard, very high fire hazard, and extremely high fire hazard. For all fire hazard characteristics within the same fire hazard group, the mean of each dimension was calculated. Taking the low fire hazard group as an example, the low fire hazard group contains... Each fire hazard inducing feature sample is a 64-dimensional real vector, representing the fire hazard inducing feature sample in the low fire hazard group. Mean of each dimension By The first fire-induced characteristic sample Sum of the values ​​in each dimension and divide by The same operation is performed on other fire risk level groupings to obtain the mean of each fire risk level group in each dimension. The means of all 64 dimensions calculated for the same fire risk level group are combined into a 64-dimensional real vector, which is the center vector corresponding to that fire risk level. This process is repeated to obtain the center vectors for low fire risk level nodes, medium fire risk level nodes, high fire risk level nodes, very high fire risk level nodes, and extremely high fire risk level nodes. The five center vectors are arranged in ascending order of fire risk level, with the center vector of the low fire risk level node first and the center vector of the extremely high fire risk level node last, forming a fire risk level map. The fire risk level map is constructed and stored in the system memory before each fire risk warning.

[0049] In specific implementation, please refer to Figure 3 After obtaining the initial fire risk level, a dynamic fire risk distribution grid covering the entire monitoring area is generated based on the initial fire risk level and the fire propagation velocity parameters shared between adjacent sensor nodes. This dynamic fire risk distribution grid is then used as the real-time early warning output. Multiple sensor nodes are deployed within the monitoring area, and their geographical coordinates are known. The geographical coordinates of each sensor node are used as the center point of the grid cell, and the initial fire risk level corresponding to the sensor node is directly assigned to the grid cell centered on the sensor node's location. The grid cell is a square grid, and the side length of the square grid is determined based on the total area of ​​the monitoring area and the sensor node deployment density; the side length of the square grid is set to 50 meters.

[0050] The fire propagation velocity parameter shared between two adjacent sensor nodes is obtained. Two adjacent sensor nodes are defined as a pair of sensor nodes that are geographically directly adjacent, meaning the line connecting the two nodes does not pass through any other sensor nodes. The determination of adjacent sensor nodes is achieved by performing Delaunay triangulation on the geographical coordinates of all sensor nodes. Sensor nodes that form the same side of the triangle after Delaunay triangulation are considered adjacent sensor nodes.

[0051] In some embodiments, the fire propagation speed parameter is dynamically calculated based on the average wind speed, vegetation continuity, and terrain slope between two adjacent sensing nodes. The average wind speed is obtained by taking the arithmetic mean of the wind speed values ​​collected by two adjacent sensing nodes at the current time. The wind speed value collected by the first of the two adjacent sensing nodes is measured by the wind speed sensor deployed at the first sensing node within the current sampling period, and the wind speed value collected by the second of the two adjacent sensing nodes is measured by the wind speed sensor deployed at the second sensing node within the current sampling period. Vegetation continuity is calculated by determining the proportion of the length of a segment with a normalized vegetation index (NVI) continuously greater than 0.4 along the straight path between two adjacent sensor nodes. The NVI is extracted from multispectral remote sensing imagery of the monitoring area with a resolution of 10 meters. A higher proportion of this NVI-greater value results in a vegetation continuity value closer to 1, while a lower proportion results in a vegetation continuity value closer to 0. Topographic slope is extracted from a pre-stored digital elevation model (DEM) of the monitoring area. Topographic slope is the arctangent angle of the ratio of the elevation difference to the straight-line distance along the straight path between two adjacent sensor nodes. The fire propagation speed parameter is calculated by multiplying the average wind speed, vegetation continuity, and a factor related to topographic slope. The factor related to terrain slope is determined based on the terrain slope value. When the terrain slope value is between 0 and 15 degrees, the factor related to terrain slope is 1.0. When the terrain slope value is between 15 and 30 degrees, the factor related to terrain slope is 1.5. When the terrain slope value is greater than 30 degrees, the factor related to terrain slope is 2.0.

[0052] In some embodiments, another method for obtaining the fire propagation velocity parameter shared between two adjacent sensor nodes includes: selecting multiple pairs of adjacent sensor nodes within the monitoring area, and for each pair of adjacent sensor nodes, obtaining the time difference between their historical detections of the same forest fire and the actual distance between them. The historical detection time difference is obtained by retrieving the first alarm time of the corresponding pair of adjacent sensor nodes for the same forest fire from the historical fire record database, subtracting the two first alarm times, and taking the absolute value as the time difference. The actual distance is calculated using Euclidean distance based on the geographical coordinates of the pair of adjacent sensor nodes. Dividing the actual distance by the time difference yields the single fire propagation velocity between the pair of adjacent sensor nodes. This calculation is performed on all adjacent sensor node pairs within the monitoring area, resulting in multiple single fire propagation velocity values. All single fire propagation velocity values ​​corresponding to adjacent sensor nodes are arranged in ascending order, and the value in the middle of the arrangement is selected as the median velocity. When the median velocity has an odd number of values, the median velocity is selected as the (N+1) / 2th value after the permutation, where N is the total number of single fire propagation velocity values. When the median velocity has an even number of values, the median velocity is selected as the arithmetic mean of the N / 2th and N / 2+1th values ​​after the permutation. The median velocity is used as a shared fire propagation velocity parameter among all adjacent sensor nodes.

[0053] In practice, after obtaining the fire propagation speed parameters, the propagation time required for the fire to spread from one sensor node to another is calculated based on these parameters and the straight-line distance between two adjacent sensor nodes. The straight-line distance is calculated using Euclidean distance based on the geographical coordinates of the two adjacent sensor nodes, and the propagation time is the quotient obtained by dividing the straight-line distance by the fire propagation speed parameters.

[0054] Linear interpolation of fire hazard levels is performed on intermediate grid cells between two adjacent sensor nodes based on propagation time. The two adjacent sensor nodes are identified as the first and second sensor nodes, respectively, with the center point grid cell corresponding to the locations of the first and second sensor nodes. The first fire hazard level of the first grid cell corresponding to the first sensor node and the second fire hazard level of the second grid cell corresponding to the second sensor node are obtained. The straight-line path and its total distance between the first and second grid cells are determined, with the total distance equal to the straight-line distance between the two adjacent sensor nodes. The number of intermediate grid cells traversed on the straight-line path is determined; intermediate grid cells are all grid cells on the straight-line path except for the first and second grid cells. Intermediate grid cells are determined by whether the intersection of the straight-line path segment and each square grid cell is non-empty. If the straight-line path segment passes through the internal region of a square grid cell, then that square grid cell is identified as an intermediate grid cell. For any intermediate grid cell, the ratio of the distance from that intermediate grid cell to the first grid cell to the total distance of the straight-line path is calculated.

[0055] In some embodiments, the fire hazard level of the intermediate grid cell is determined using the following formula:

[0056]

[0057] in, Indicates the first line path. The fire hazard level of each intermediate grid cell. The value ranges from 1 to positive integers, This indicates the total number of intermediate grid cells along a straight path; This represents the numerical value corresponding to the first fire risk level of the first grid cell. Low fire risk is assigned a value of 1, medium fire risk is assigned a value of 2, high fire risk is assigned a value of 3, very high fire risk is assigned a value of 4, and extremely high fire risk is assigned a value of 5. This represents the numerical value corresponding to the second fire hazard level of the second grid cell, and the correspondence between the assigned values ​​and... The assignment method is the same; Indicates the first The straight-line distance between each intermediate grid cell and the center point of the first grid cell Total distance from the straight path The ratio, i.e. , The unit is meters. The unit is meters. The value of is between 0 and 1, excluding the open interval of 0 and 1; This represents the first fire hazard level value of the first grid cell. with weight The product; This indicates the second fire hazard level value of the second grid cell. with weight The product; symbol Represents arithmetic addition; symbol This represents arithmetic multiplication. When the calculated result... When the value is not an integer, for Rounding is performed to the nearest integer, with the rounded integer value corresponding to the fire hazard level: 1 for low fire hazard, 2 for medium fire hazard, 3 for high fire hazard, 4 for very high fire hazard, and 5 for extremely high fire hazard. This linear interpolation calculation is then performed sequentially on all intermediate grid cells to obtain the fire hazard level for each intermediate grid cell.

[0058] In implementation, all grid cells are stitched together according to their respective geographical coordinates to form a dynamic fire risk distribution grid. The grid cells are stitched together in a square grid format, with adjacent cells in each row arranged sequentially along the longitude and latitude directions of their geographical coordinates, forming a dynamic fire risk distribution grid covering the entire monitoring area. Each grid cell in the dynamic fire risk distribution grid carries a fire risk level value. The dynamic fire risk distribution grid serves as a real-time early warning output, presented as a two-dimensional fire risk level distribution map, with different fire risk levels identified by different colors.

[0059] In practical implementation, a dynamic forest fire risk early warning system includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory is a combination of non-volatile storage medium and random access memory. The non-volatile storage medium stores the computer program, a digital elevation model of the monitoring area, a historical fire record database, and a pre-constructed fire risk level map. The processor executes the computer program to implement all the steps of the dynamic forest fire risk early warning method.

[0060] The system also includes a cluster of multi-source sensor nodes deployed within the monitoring area. Each multi-source sensor node integrates a wind speed sensor, an air temperature sensor, an air humidity sensor, a rainfall sensor, a spectral reflectance meter, a global positioning module, and a wireless communication module. The wind speed sensor collects wind speed data, the air temperature sensor collects air temperature data, the air humidity sensor collects air humidity data, and the rainfall sensor collects rainfall data; these four types of data together constitute meteorological parameters. The spectral reflectance meter collects vegetation moisture content data. The global positioning module obtains the geographic coordinates of the multi-source sensor nodes, providing spatial location information for each node. The wireless communication module transmits the collected meteorological parameters, vegetation moisture content, along with geographic coordinates and timestamps, to the processor via a wireless network.

[0061] The computer program is organized in memory in the form of functional modules, including a data receiving and time alignment module, a data preprocessing module, a recursive convolutional network operation module, an initial fire hazard level determination module, a fire hazard propagation speed calculation module, a fire hazard dynamic distribution grid generation module, and an early warning output module.

[0062] In practical implementation, the processor wirelessly receives meteorological parameters and vegetation moisture content uploaded by all multi-source sensor nodes via a data receiving and time alignment module. Each data point carries a timestamp added by the multi-source sensor node. Simultaneously, the data receiving and time alignment module extracts terrain features from the digital elevation model of the monitoring area stored in memory. These features include aspect, slope, and landform category labels. It also retrieves the locations of lightning strikes within a specified radius around each multi-source sensor node within the past three years from a historical fire record database, using these locations as historical lightning strike distribution data. The specified radius is set to 500 meters. The data receiving and time alignment module combines meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike distribution data into an input sample using the same timestamp, and arranges them into an input sequence according to the order of the timestamps.

[0063] The data preprocessing module receives the input sequence and normalizes the continuous values ​​in the input sequence. Continuous values ​​include all sensor values ​​from meteorological parameters, vegetation moisture content, slope and aspect values ​​from topographic features, and the planar distances from lightning strike points to multi-source sensor nodes from historical lightning strike point distribution data. Discrete category labels are one-hot encoded. Discrete category labels include geomorphic category labels from topographic features and binary labels indicating the presence of lightning strike points within a specified radius. The data preprocessing module then reassembles the normalized continuous values ​​and the one-hot encoded discrete category labels in their original chronological order into a standardized input tensor.

[0064] The recurrent convolutional network (RCN) computation module comprises a series of interconnected one-dimensional convolutional layers, long short-term memory (LSM) recurrent layers, and a temporal attention module. The one-dimensional convolutional layer contains 32 convolutional kernels, each with a temporal dimension of 3 and a stride of 1, using a linear rectified function as the activation function. The LSM recurrent layer contains 64 hidden units, returning a sequence of hidden state vectors for all time steps. The temporal attention module comprises a series of interconnected global average pooling layers, a first fully connected layer with 32 neurons using a linear rectified function, a second fully connected layer with 1 neuron and no activation function, and a normalized exponential function layer. The RCN computation module receives a normalized input tensor, extracts local temporal features through the one-dimensional convolutional layer, models long-range dependencies through the LSM recurrent layer to obtain the hidden state vector sequence, and then the temporal attention module dynamically assigns weights to the hidden state vector sequence, concatenates them according to time steps, and outputs the fire hazard induction feature sequence corresponding to each multi-source sensor node at different time steps.

[0065] The initial fire risk level determination module extracts the feature vector corresponding to the last time step from the fire risk induction feature sequence as the current feature vector and loads a pre-constructed fire risk level map from memory. The fire risk level map contains five level nodes, corresponding to low, medium, high, very high, and extremely high fire risk, respectively. Each level node has a 64-dimensional center vector. The initial fire risk level determination module calculates the Euclidean distance between the current feature vector and each of the five center vectors, and selects the fire risk level corresponding to the level node with the smallest Euclidean distance as the initial fire risk level for the current time. When multiple equal minimum Euclidean distances exist, the fire risk level corresponding to the level node with the highest fire risk level is selected as the initial fire risk level.

[0066] The fire propagation velocity calculation module obtains the fire propagation velocity parameters shared between adjacent multi-source sensor nodes. Adjacent multi-source sensor nodes are determined by performing Delaunay triangulation on the geographic coordinates of all multi-source sensor nodes, and two multi-source sensor nodes that form the same side of a triangle are identified as adjacent multi-source sensor nodes. In some embodiments, the fire propagation speed parameter is dynamically calculated based on the average wind speed, vegetation continuity, and terrain slope between two adjacent multi-source sensor nodes. The average wind speed is obtained by taking the arithmetic mean of the wind speed values ​​collected by the two adjacent multi-source sensor nodes at the current moment. The vegetation continuity is obtained by calculating the proportion of the length of the section with a normalized vegetation index continuously greater than 0.4 on the straight path between the two adjacent multi-source sensor nodes to the total length of the straight path. The terrain slope is extracted from the digital elevation model of the monitoring area and calculated as the arctangent angle. The factor related to the terrain slope takes a value of 1.0 when the terrain slope value is between 0 and 15 degrees, a value of 1.5 when it is between 15 and 30 degrees, and a value of 2.0 when it is greater than 30 degrees. The fire propagation speed parameter is the product of the average wind speed, vegetation continuity, and the factor related to the terrain slope. In other embodiments, the fire propagation speed parameter is obtained through historical propagation speed statistics. Specifically, for each pair of adjacent multi-source sensor nodes, the time difference and actual distance between the two nodes detecting the same forest fire in the past are obtained. The actual distance is divided by the time difference to obtain the single fire propagation speed. The single fire propagation speeds corresponding to all adjacent multi-source sensor nodes are subjected to median filtering. The median filtering method is to arrange all single fire propagation speed values ​​from smallest to largest, and take the value in the middle position after arrangement as the median speed. When the median speed is an even number of values, the arithmetic mean of the two middle values ​​is taken. The median speed is used as the fire propagation speed parameter shared among all adjacent multi-source sensor nodes.

[0067] The fire hazard dynamic distribution grid generation module establishes square grid cells centered on the geographic coordinates of each multi-source sensor node. The side length of each square grid cell is set to 50 meters, and the initial fire hazard level is directly assigned to the corresponding grid cell. The module obtains the straight-line distance between two adjacent multi-source sensor nodes and calculates the propagation time required for fire to spread from one multi-source sensor node to another using the fire propagation speed parameter. The propagation time is obtained by dividing the straight-line distance by the fire propagation speed parameter. Subsequently, the module determines the straight-line path and total distance between the corresponding grid cells of two adjacent multi-source sensor nodes, determines the number of intermediate grid cells traversed along the straight-line path, and calculates a linear interpolation of the fire hazard level for each intermediate grid cell. The interpolation method uses the fire hazard level values ​​of the first and second grid cells as endpoints, and performs a linear weighting based on the ratio of the distance from the intermediate grid cell to the first grid cell to the total straight-line path distance. The weighted result is rounded to the nearest integer to obtain the fire hazard level. All grid cells are then stitched together according to their geographic coordinates to generate a dynamic fire hazard distribution grid covering the entire monitoring area.

[0068] The early warning output module maps the dynamic distribution grid of fire risk into a two-dimensional visual image. Different fire risk levels are marked with different colors: low fire risk is marked with blue, medium fire risk with green, high fire risk with yellow, very high fire risk with orange, and extremely high fire risk with red. The visual image is then sent to a display device connected to the system. The display device is an LCD screen, which presents the dynamic early warning results of forest fire risk in real time through a graphical interface.

[0069] Before system deployment and operation, the trainable parameters in the one-dimensional convolutional layer, long short-term memory recursive layer, and temporal attention module of the recursive convolutional network operation module need to be trained offline. During training, historical fire cases from the past five years in the monitored area are extracted from the historical fire risk record database to construct a training dataset. Each training sample contains a multi-source data alignment sequence of a continuous time step and the corresponding actual fire risk level label. An adaptive moment estimation optimization algorithm is used to update the parameters. The initial learning rate is set to 0.001, the first moment decay rate is set to 0.9, the second moment decay rate is set to 0.999, and the batch size is set to 32. The Euclidean distance between the feature vector of the last time step of the fire risk induction feature sequence and the center vector of the corresponding actual fire risk level is used as the loss function value. Training stops when the validation set loss function value no longer decreases after 5 consecutive training cycles. The obtained parameter values ​​are then stored in memory for loading and use by the recursive convolutional network operation module during real-time inference. The construction of the fire risk level map was also completed offline. By grouping fire risk inducing characteristic samples from historical fire cases according to the actual fire risk level and calculating the mean vector of each dimension, five central vectors were formed and stored in order of fire risk level from low to high. During system operation, all calculations are automatically completed by the processor according to the program instructions and parameters in the memory. Multi-source sensor nodes continuously collect and send data at a fixed sampling period of 10 minutes. In each data processing cycle, the processor generates the latest dynamic fire risk distribution grid and refreshes the early warning display, thereby realizing dynamic early warning of forest fire risk.

[0070] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for dynamic early warning of forest fire risk, characterized in that, include: By continuously collecting meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data through multi-source sensor nodes deployed in the monitoring area, the meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data are aligned according to the same time index and then input into a recursive convolutional network. The time attention module in the recursive convolutional network is used to dynamically assign weights to the aligned data to obtain the fire hazard induction feature sequence of each sensing node at different time steps. The fire hazard induction feature sequence is mapped to a pre-constructed fire hazard level map. The initial fire hazard level at the current moment is determined by comparing the distance between the fire hazard induction feature sequence and the center vector of each level node in the fire hazard level map. Based on the initial fire risk level and the fire propagation speed parameters shared between adjacent sensor nodes, a dynamic fire risk distribution grid covering the entire monitoring area is generated, and the dynamic fire risk distribution grid is used as a real-time early warning output.

2. The method for dynamic early warning of forest fire risk according to claim 1, characterized in that, The process involves continuously collecting meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data through multi-source sensor nodes deployed within the monitoring area. Specifically, this includes: A wind speed sensor, an air temperature sensor, an air humidity sensor, and a rainfall sensor are installed at each sensing node to collect the meteorological parameters. A spectral reflectance meter is installed at each sensing node to collect the vegetation moisture content. The terrain features are extracted from a pre-stored digital elevation model of the monitoring area. The locations of lightning strikes within a specified radius around each sensing node in the past three years are retrieved from the historical fire record database as the historical lightning strike point distribution data.

3. The method for dynamic early warning of forest fire risk according to claim 1, characterized in that, The meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike point distribution data are aligned according to the same time index and then input into a recurrent convolutional network, specifically including: The meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike distribution data are timestamped at a fixed sampling period. Meteorological parameters, vegetation moisture content, terrain features, and historical lightning strike distribution data with the same timestamp are combined into an input sample. The input samples are arranged into an input sequence according to the order of their timestamps. Each input sample in the input sequence is fed sequentially into the input layer of the recurrent convolutional network.

4. The method for dynamic early warning of forest fire risk according to claim 3, characterized in that, The time attention module in the recurrent convolutional network is used to dynamically assign weights to the aligned data, resulting in a fire hazard induction feature sequence for each sensor node at different time steps, specifically including: The hidden state vector corresponding to each time step is extracted from the hidden layer of the recurrent convolutional network. The hidden state vector is input into the scoring function in the temporal attention module. The scoring function outputs an original attention score for each time step. The original attention scores of all time steps are normalized to obtain the attention weight of each time step. The hidden state vector of each time step is multiplied by the corresponding attention weight and then concatenated in the order of time steps to form the fire hazard induction feature sequence.

5. The method for dynamic early warning of forest fire risk according to claim 4, characterized in that, The fire hazard induction feature sequence is mapped onto a pre-constructed fire hazard level map. The initial fire hazard level at the current moment is determined by comparing the distance between the fire hazard induction feature sequence and the center vector of each level node in the fire hazard level map. Specifically, this includes: Extract the feature vector corresponding to the last time step from the fire hazard induction feature sequence as the current feature vector, calculate the Euclidean distance between the current feature vector and the center vector of each level node in the fire hazard level map, and select the fire hazard level corresponding to the level node with the smallest Euclidean distance as the initial fire hazard level at the current moment.

6. The method for dynamic early warning of forest fire risk according to claim 5, characterized in that, Based on the initial fire risk level and the fire propagation velocity parameters shared between adjacent sensor nodes, a dynamic fire risk distribution grid covering the entire monitoring area is generated, and this dynamic fire risk distribution grid is used as a real-time early warning output, specifically including: Using the location of each sensor node as the center point of the grid cell, the initial fire risk level is assigned to the grid cell where the corresponding center point is located. The fire risk propagation speed parameter shared between two adjacent sensor nodes is obtained. Based on the fire risk propagation speed parameter and the straight-line distance between two adjacent sensor nodes, the propagation time required for the fire to spread from one sensor node to another is calculated. Based on the propagation time, the fire risk level of the intermediate grid cell between two adjacent sensor nodes is linearly interpolated. All grid cells are stitched together according to their respective geographical coordinates to form the dynamic fire risk distribution grid.

7. The method for dynamic early warning of forest fire risk according to claim 6, characterized in that, The fire spread speed parameter is dynamically calculated based on the average wind speed, vegetation continuity, and terrain slope between two adjacent sensing nodes.

8. The method for dynamic early warning of forest fire risk according to claim 1, characterized in that, The pre-constructed fire hazard rating map is constructed in the following ways: Historical fire risk cases from the past few years are collected in the monitoring area. Each historical fire risk case contains multiple fire risk inducing feature samples from consecutive time steps and the corresponding actual fire risk level. The fire risk inducing feature samples from all historical fire risk cases are grouped according to the actual fire risk level. The mean of each dimension of all fire risk inducing feature samples within the same fire risk level group is calculated. The mean of each dimension is combined to form the central vector corresponding to the fire risk level. The central vectors corresponding to all fire risk levels are arranged from low to high to form the fire risk level map.

9. A method for dynamic early warning of forest fire risk according to claim 1, characterized in that, Before the temporal attention module in the recurrent convolutional network performs dynamic weight allocation on the aligned data, it also includes: The continuous values ​​in the aligned data are normalized, the discrete category labels in the aligned data are one-hot encoded, and the normalized and one-hot encoded data are reassembled into a standardized input tensor according to the original time order. The standardized input tensor is used as the actual processing object of the time attention module.

10. A dynamic early warning system for forest fire risk, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the dynamic early warning method for forest fire risk as described in any one of claims 1 to 9.

Images