Forest fire risk assessment method based on causal graph network

The causal graph neural network model leverages multi-source data and an encoder-decoder structure to address the limitations of fixed graph topologies in existing forest fire risk assessment, enhancing prediction accuracy and adaptability across varied geographical regions.

CN118536388BActive Publication Date: 2024-12-06合肥恒宝天择智能科技有限公司
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202410548341.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-06
Publication Date
2024-12-06
Estimated Expiration
2044-05-06

AI Technical Summary

Technical Problem

Existing forest fire risk assessment methods based on graph neural networks suffer from limited universality and application scope due to differences in geography, climate, and vegetation across regions, as they rely on fixed graph topologies and are trained on specific regional data.

Method used

A causal graph neural network model using ground and aerial sensors to collect multi-source data, combined with an encoder-decoder structure for causal discovery, allowing for flexible adaptation to different regions by learning and inferring potential causal relationships.

Benefits of technology

Enhances the universality and accuracy of fire risk prediction by adapting to varying regional conditions, improving model flexibility and applicability across diverse terrains and climates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118536388B_ABST
    Figure CN118536388B_ABST
Patent Text Reader

Abstract

The present invention discloses a forest fire risk assessment method based on a causal graph network, the method comprising: fusing ground data with aerial data to form a comprehensive data set for forest fire risk assessment; dividing the relevant area into grids according to the location information of ground sensors; mapping DEM data and normalized difference vegetation index NDVI data to the divided grids, and counting the average value of the data in each grid; building a forest fire risk assessment model based on a causal graph neural network; inputting the time series data of the area to be assessed into the trained causal graph neural network model; obtaining future state change information through model prediction; and assessing the forest fire risk according to the prediction results. The present invention significantly improves the accuracy and flexibility of fire risk assessment by optimizing model design, realizing cross-regional data sharing and model training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of forest fire risk assessment, and in particular to a forest fire risk assessment method based on a causal graph network. Background Art

[0002] Methods for forest fire risk assessment mainly rely on data-driven approaches. These methods are usually based on historical sensor data and fire records in a specific area, and build models to fit the relationship between the probability of fire in the area and sensor data. The core of this method is to use the principles of statistics and machine learning to analyze large amounts of data to reveal potential risk patterns.

[0003] Among the many data-driven methods, methods based on graph neural networks are particularly prominent. Graph neural networks can use graph structures to represent entities and their relationships in data, and are particularly suitable for dealing with node network problems with complex relationships. In forest fire risk assessment, graph neural networks can well capture the mutual influence between regional grids, thereby more accurately predicting fire risks. By building a graph neural network model and training it with historical data, it is possible to predict future fire risks.

[0004] However, although the forest fire risk assessment method based on graph neural network has achieved certain success, it also has obvious limitations. Specifically, the graph network structure of sensor data in different regions may be different. This means that the graph neural network model trained in one region may not be directly applied to other regions. This difference between regions leads to limited universality of the model, which limits the scope of application of the model.

[0005] In addition, the graph neural network models in the prior art are often modeled based on a fixed graph topology. This means that the model can only be trained on historical data of a specific region and is only applicable to future fire risk predictions in that region. Since factors such as geography, climate, and vegetation may vary greatly between regions, this model with a fixed graph topology cannot adapt to the actual conditions in different regions. This further limits the scope of application of the model, making it difficult to promote and reuse the model.

[0006] In summary, although the forest fire risk assessment method based on graph neural network in the existing technology has certain advantages, it still has obvious limitations. The universality and application scope of the model are limited, making it unable to meet the needs of forest fire risk assessment in different regions. Therefore, it is necessary to explore a more flexible and universal forest fire risk assessment method to overcome the shortcomings of the existing technology. Summary of the invention

[0007] In view of this, the present invention proposes a forest fire risk assessment method based on a causal graph network to solve the problem of insufficient universality of forest fire risk assessment methods in the prior art.

[0008] The specific technical solutions of the present invention are as follows:

[0009] A forest fire risk assessment method based on causal graph neural network, comprising:

[0010] Step 1, using ground sensors deployed in forest areas to obtain ground data; obtaining aerial data through satellite remote sensing technology; the aerial data includes digital elevation model DEM data and multispectral remote sensing data; calculating the normalized difference vegetation index NDVI based on satellite multispectral remote sensing data to reflect vegetation growth conditions; fusing the ground data with the aerial data to form a comprehensive data set for forest fire risk assessment;

[0011] Step 2: Divide the relevant area into grids according to the location information of the ground sensor; map the DEM data and the normalized difference vegetation index NDVI data to the divided grids, and calculate the average value of the data in each grid; normalize the ground data, the DEM average value, and the NDVI average value data to form a feature vector for model input;

[0012] Step 3: Build a forest fire risk assessment model based on a causal graph neural network, which includes an encoder for processing time series data and a decoder for predicting future states; train the model using a training data set containing data from multiple representative regions, calculate the loss through forward propagation, and update the model parameters using back propagation until the model performance meets the preset requirements;

[0013] Step 4: Input the time series data of the area to be evaluated into the trained causal graph neural network model; obtain future state change information through model prediction; and evaluate the forest fire risk based on the prediction results.

[0014] Specifically, in step 1, the ground data includes three-dimensional multi-dimensional perception data of air temperature and humidity, ambient light intensity, wind speed and direction, rainfall, soil moisture content, surface litter moisture content and phenological fire risk factors.

[0015] Specifically, in step 3, the loss function is used for optimization during the training process of the causal graph neural network model.

[0016] Specifically, the loss function includes the negative log-likehood function of the decoder and the KL divergence of the encoder, which are used to measure the difference between the model prediction results and the actual results.

[0017] Specifically, in step 1, the ground sensor data focuses on two key indicators: surface temperature and humidity and moisture content of ground combustibles. The surface temperature and humidity sensor includes a power module, a transmitter module, a drift and temperature compensation module, and a data processing module. These modules work together to obtain surface temperature and humidity data, and capture weak signal changes through signal sampling and amplification functions. A multispectral camera is used to capture multispectral images of ground vegetation, and the chlorophyll content and moisture content of the vegetation are inverted based on an empirical model to obtain the moisture content index of ground combustibles.

[0018] Specifically, in step 1, the calculation formula of the normalized difference vegetation index NDVI is NDVI=(NIR-RED) / (NIR+RED), wherein NIR is near infrared light reflection and RED is red light reflection.

[0019] Specifically, in step 3, the encoder is designed to adopt a fully connected graph neural network structure fenc, which transforms and processes the feature vector of the time series data, outputs the feature vector of the corresponding edge relationship, and converts it into a probability prediction of the edge relationship through the softmax function.

[0020] Specifically, in step 3, the decoder design includes receiving the feature vector of the causal relationship and time series data predicted by the encoder as input, performing forward propagation through the neural network, and predicting the change value of the current feature vector and the feature vector of the next time stage, so as to obtain the estimated mean of the feature vector of the next time stage.

[0021] Specifically, the decoder further calculates the probability distribution of the feature vector value based on the estimated mean, and represents the prediction result of the future time series data in the form of a normal distribution.

[0022] Specifically, step 4 also includes setting a fixed number of training cycles, and calculating and monitoring the training loss and validation loss of the model at the end of each cycle to evaluate the convergence and performance of the model.

[0023] The beneficial effects of the present invention are:

[0024] (1) Model design and universality improvement:

[0025] The present invention has made a significant breakthrough in model design, successfully retaining the advantages of graph neural networks in describing node relationships, while breaking through the reliance of traditional graph neural networks on fixed graph structure data. This innovation enables the model to flexibly respond to differences in the graph network structure of sensor data in different regions, thereby significantly enhancing the universality and application scope of the model. In addition, by introducing the idea of ​​causal discovery, the present invention successfully transforms the fire risk assessment problem into a more general and flexible problem, further improving the applicability and accuracy of the model.

[0026] (2) Application of Encoder-Decoder structure:

[0027] In order to achieve more accurate fire risk prediction, the present invention innovatively uses the encoder-decoder structure to model the causal discovery problem. This structure can not only better capture the mutual influence between regional grids, but also learn and infer the potential causal relationship. Compared with traditional methods, this structure shows higher flexibility and accuracy in dealing with complex relationships, providing new ideas and methods for forest fire risk assessment.

[0028] (3) Cross-regional data utilization and model training:

[0029] The present invention makes full use of the historical fire data of various regions to train the same model, realizing data sharing and reuse. This cross-regional data utilization method not only improves the training efficiency and stability of the model, but also enables the model to learn the fire risk characteristics of more regions. By learning and integrating data from different regions, the model can better adapt to the actual conditions of different regions and further improve the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 It is a flow chart of the forest fire risk assessment method based on the cause-effect graph network of the present invention. DETAILED DESCRIPTION

[0032] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] The present invention proposes a forest fire risk assessment method based on a causal graph network, comprising the following steps:

[0034] Step 1: Data acquisition:

[0035] In the data acquisition stage, the ground and air sensor technologies are used in combination to obtain multi-source data and improve the prediction accuracy by fusing these data.

[0036] Step 1.1: Acquisition of ground data:

[0037] Ground data is obtained through carefully deployed fire risk factor integrated detection stations. These detection stations are composed of intelligent sensors, transmission controllers and installation accessories, which work together to sense and transmit various fire risk factor data in real time.

[0038] Smart sensors are the core equipment for acquiring ground data. They have highly sensitive and accurate perception capabilities. These sensors can sense key fire risk factors such as air temperature and humidity, ambient light intensity, wind speed and direction, rainfall, soil moisture content, surface litter moisture content, and phenology in a three-dimensional and multi-dimensional manner. Through the built-in signal sampling and amplification functions, the sensor can capture weak signal changes and amplify them to a measurable range. At the same time, the drift and temperature compensation module can automatically correct the impact of environmental factors on the measurement data to ensure the accuracy and stability of the data.

[0039] These real-time sensed data are transmitted through the transmission controller to ensure that the data can be accurately and timely transmitted to the data center for further processing and analysis.

[0040] Details of ground sensor data acquisition: In terms of ground sensor data acquisition, the main focus is on two key indicators: surface temperature and humidity and moisture content of ground combustibles.

[0041] The surface temperature and humidity sensor consists of a power module, a transmitter module, a drift and temperature compensation module, and a data processing module. These modules work together to ensure that the sensor can accurately and stably measure the surface temperature and humidity. The built-in signal sampling and amplification functions of the sensor can capture weak signal changes and convert them into measurable electrical signals. At the same time, the drift and temperature compensation module can automatically correct the impact of environmental factors on the measured data, improving the accuracy and reliability of the data.

[0042] To obtain the moisture content of ground combustibles, a multispectral camera is used to capture multispectral images of ground vegetation. These images contain the reflectance information of vegetation in different bands. By inverting the chlorophyll content and moisture content of vegetation based on empirical models, the moisture content index of ground combustibles can be obtained. This index is of great significance for assessing fire risks, predicting fire risk trends, and formulating fire prevention measures.

[0043] Step 1.2: Aerial data acquisition:

[0044] Aerial data mainly comes from satellite remote sensing technology, which provides rich multi-source data.

[0045] First, the satellite DEM data is used to obtain terrain data. DEM data is the abbreviation of Digital Elevation Model. It contains the elevation information of the surface and is an important basic data for studying terrain, watersheds and identifying land objects. By parsing and processing DEM data, a large amount of surface morphological information can be extracted, such as contour lines, elevation maps, slope maps, aspect maps, etc. This information is of great value for terrain analysis, landform identification and other related tasks in geoscience applications.

[0046] In addition, multispectral remote sensing data from satellites is also used to obtain vegetation data. Multispectral data contains reflection information of different bands. By processing and analyzing this information, the Normalized Difference Vegetation Index (NDVI) can be obtained. NDVI is a numerical indicator that reflects the growth status of vegetation. It uses the reflection information of red light and near-infrared light to evaluate the growth status of green vegetation in the target area. By calculating the NDVI index, vegetation can be preliminarily classified to identify different types of vegetation such as shrubs, trees, wetlands, and grasslands. The calculation method of NDVI is to use the reflection of red light and near-infrared light, which can show information such as plant growth, vitality and productivity of the ecosystem. The larger the value, the more plant growth. The formula is as follows:

[0047]

[0048] Among them, NIR is near infrared light reflection; RED is red light reflection. The value of NDVI is between -1 and 1. When RED = 0, the maximum value is 1; conversely, when NIR = 0, the minimum value is -1.

[0049] In the process of acquiring aerial data, remote sensing data also needs to be preprocessed and calibrated to eliminate the impact of atmospheric interference, geometric distortion and other factors on data quality. At the same time, professional image processing technology is also needed to enhance and extract features from remote sensing data in order to better utilize these data for analysis and prediction.

[0050] In summary, by combining ground and air sensor technologies, comprehensive and accurate multi-source data can be obtained, which will help us better understand and analyze the surface environment.

[0051] Step 2: Model building:

[0052] The probabilistic approach is used to model the causal discovery process. The model design aims to deeply explore the causal relationships and their dynamic evolution in time series data. The core of the model consists of two main parts: the encoder and the decoder. The encoder is mainly responsible for extracting and predicting the edge relationships in the causal graph from the time series data, while the decoder models the dynamics of the time series data based on these predicted edge relationships.

[0053] Step 2.1, data preprocessing:

[0054] Data preprocessing is a key step in model building, which directly affects the accuracy of subsequent feature extraction and model training. According to the precise location of the ground sensor, the relevant geographical area is carefully divided into grids. The center point of each grid corresponds to the actual location of a sensor, ensuring that the spatial correspondence of the data is accurate.

[0055] Subsequently, the DEM data and NVDI data acquired by remote sensing are accurately mapped to these grids. The average values ​​of DEM and NVDI in each grid are calculated through advanced interpolation algorithms and statistical analysis methods. These values ​​reflect the terrain and vegetation characteristics of the area.

[0056] Next, the DEM average, NVDI average, ground temperature and humidity sensor value, and ground combustible moisture content value of each grid area are normalized. The purpose of normalization is to eliminate the dimensional differences between different features so that they can be compared and calculated on the same scale. After normalization, these feature values ​​are spliced ​​into a feature vector containing multi-dimensional feature information, providing rich input data for subsequent model training.

[0057] Step 2.2, encoder design:

[0058] Encoder q Φ (z|x) uses a fully connected graph neural network structure, fenc,Φ, which can effectively process time series data and extract potential relationships among them. In the encoder, each time series data is regarded as a node (Vertex) of the graph, and the relationship between time series data is represented by edges (Edge).

[0059] The core of the encoder is to transform and process the feature vectors of time series data. After passing through a series of network layers (such as fenc and Φ), the feature vector will output a set of feature vectors corresponding to the edge. These feature vectors are further processed by the softmax function and converted into probabilistic predictions of edge relationships. In this way, the encoder can mine potential causal relationships from time series data and provide important input information for the subsequent decoder. The specific expression is as follows:

[0060] ψ ij =fenc,Φ(x) ij

[0061] q Φ (z ij |x)=Softmax(ψ ij / τ)

[0062] Among them, i and j represent the index of the node in the graph structure; x represents multi-source data; z represents the edge in the graph structure; τ is an artificially set hyperparameter used to adjust the probability distribution curve.

[0063] Step 2.3, decoder design:

[0064] Decoder p θ (x|z) is the key part to model the dynamics of temporal data based on the causal graph predicted by the encoder. The decoder receives two main inputs: one is the causal relationship z predicted by the encoder ij , which reveals the interaction between different time series data; the second is the feature vector x of the time series data up to the current time stage t t ={x t1 ,...x t N}, which reflects the current state information. N represents the total number of nodes in the graph structure.

[0065] The decoder uses a neural network fe to forward propagate the time series information along the edges predicted by the encoder. At each graph node, the decoder accumulates the input information and predicts the change value of the current feature vector and the feature vector of the next time stage through another neural network fv. This change value reflects the evolution trend of the time series data over time.

[0066] The decoder then adds the predicted change value to the feature vector x at the current time step t t,j Thus, we can get the estimated mean μ of the feature vector at the next time stage t+1 t+1,j This estimated mean not only takes into account the current state information, but also incorporates potential causal relationships, making the prediction results more accurate and reliable.

[0067] Based on the estimated mean, the decoder further calculates a probability distribution of the feature vector value. This probability distribution takes the estimated mean as the mean of the normal distribution and the parameter σ2 as the variance of the normal distribution. In this way, the decoder can provide probabilistic prediction results for future time series data.

[0068] Step 2.4, loss function design:

[0069] In order to evaluate the accuracy of model prediction and guide the optimization direction of the model, a suitable loss function is designed. The loss function of the decoder adopts the negative log-likelihood function, which is:

[0070] Eq Φ (z|x)[logp θ (x|z)]

[0071] It measures the difference between the probability distribution predicted by the model and the real data. By minimizing this loss function, the probability distribution predicted by the model can be made closer to the distribution of the real data.

[0072] On the other hand, the loss function of the encoder is expressed using KL divergence. Specifically:

[0073] -KL[q Φ (z|x)||p(z)]

[0074] The KL divergence measures the similarity between two probability distributions. By minimizing the KL divergence, the edge relationship probability distribution predicted by the encoder can be made closer to the true edge relationship probability distribution.

[0075] The overall loss function is:

[0076] L=Eq Φ (z|x)[logp θ (x|z)]-KL[q Φ (z|x)||p(z)]

[0077] By comprehensively considering the loss functions of the decoder and encoder, the model can be optimized as a whole, enabling it to more accurately mine the causal relationships and dynamic behaviors in time series data.

[0078] Step 3: Training:

[0079] During the model training phase, a rigorous process is followed to ensure that the model can fully learn and utilize the temporal dynamics and causal relationships in multi-source data. The following are the detailed steps of the training process:

[0080] Step 3.1: Data collection and preprocessing:

[0081] First, data is collected from multiple representative regions, often with different climates, terrains, and vegetation types, ensuring that the dataset covers a wide range of possible scenarios and conditions.

[0082] Next, a fixed number of nodes N of the graph is set. The choice of this N value is based on task requirements and understanding of data characteristics. Then, according to this number, a portion of the area is randomly intercepted from the collected regional data. These intercepted areas are spatially connected, and the number of available ground sensors in each area is exactly N, thus ensuring that each area corresponds to a graph with N nodes.

[0083] In this way, a training dataset of M graphs with N nodes is obtained. These graphs not only contain spatial information, but also imply temporal dynamics, providing strong support for subsequent model training.

[0084] After collecting enough training data, preprocess the data according to the method described in step 2.1. This includes steps such as data cleaning, formatting, and normalization to ensure that the data input to the model has a consistent format and quality. After preprocessing, a series of time series feature vectors are obtained, which will be used as the input of the encoder-decoder model.

[0085] Step 3.2, model training and optimization:

[0086] The preprocessed time series feature vector is input into the encoder-decoder model to start model training. Through forward propagation, the model calculates the prediction result based on the current parameter settings and compares it with the true label to obtain a loss value.

[0087] The loss function plays a key role here. The negative log-likehood function is used as the loss function of the decoder to measure the accuracy of the model's prediction of time series data. At the same time, for the encoder part, KL divergence is used as the loss function to evaluate the accuracy of the model's prediction of edge relationships. These two loss functions together constitute the objective function of model training.

[0088] Once the loss value is obtained, the gradient is calculated through the backpropagation algorithm and the parameters of the encoder and decoder are updated accordingly. This process is achieved through the gradient descent optimization method, aiming to minimize the loss function.

[0089] During the training process, regularization techniques and batch normalization methods are also introduced to prevent model overfitting and improve model stability. Regularization limits the complexity of model parameters by adding additional penalty terms to the loss function, while batch normalization accelerates the training process and improves model performance by normalizing each batch of input data.

[0090] Repeat the process of forward propagation, loss calculation, back propagation, and parameter update until the model achieves satisfactory performance on the training dataset. Usually, a fixed number of training cycles (epochs) is set, such as 100 epochs. At the end of each epoch, the training loss and validation loss of the model are calculated and monitored.

[0091] Step 3.3, training convergence and model evaluation:

[0092] During the training process, we will pay close attention to the changing trends of training loss and validation loss. If both loss values ​​show a trend of gradually decreasing and stabilizing with the increase of epochs, then it can be considered that the model has achieved good convergence.

[0093] After the model training is completed, an independent validation dataset is used to evaluate the model’s performance. By comparing the model’s prediction results with the actual observed data, the model’s accuracy, recall and other indicators can be calculated to comprehensively evaluate the model’s performance in tasks such as fire risk prediction and vegetation monitoring.

[0094] If the model shows good performance on the validation data set, it can be applied to actual geoscience scenarios. By acquiring multi-source data in real time and inputting it into the model, the corresponding prediction results can be obtained, providing technical support and decision-making basis for relevant decisions.

[0095] The beneficial effects of the present invention are:

[0096] The purpose of this invention is to propose an innovative forest fire risk assessment method, which not only retains the significant advantages of graph neural networks in describing node relationships, but also successfully breaks through the limitation of traditional graph neural networks' dependence on fixed graph structure data. Through this improvement, the model can flexibly respond to the differences in the graph network structure of sensor data in different regions, thereby significantly enhancing the universality and application scope of the model.

[0097] Specifically, the present invention transforms the traditional fire risk assessment problem into a more general and flexible problem by introducing the idea of ​​causal discovery. In traditional data-driven methods, fire risk assessment is usually based on a fixed undirected graph structure, whose topological relationship is pre-set. However, in the present invention, the topological structure of the graph is no longer preset, but the problem is transformed into a causal discovery process, which enables the model to adaptively learn and infer the potential causal relationship between nodes.

[0098] In order to achieve this transformation, the present invention innovatively uses an encoder-decoder structure to model the causal discovery problem. This structure can not only effectively capture the mutual influence between regional grids, but also learn and infer potential causal relationships, thereby achieving accurate prediction of future fire risks. Compared with traditional methods, this structure has higher flexibility and accuracy in dealing with complex relationships.

[0099] In addition, the present invention can also make full use of the fire history data of various regions to train the same model. Through data sharing and reuse, not only the training efficiency and stability of the model are improved, but also the model can learn the fire risk characteristics of more regions. This cross-regional data utilization method further improves the prediction ability of the model, enabling it to better adapt to the actual conditions of different regions.

[0100] The trained model can be used to predict future fire risks in various regions, providing strong decision-making support for forest fire prevention. Through the prediction results of the model, relevant departments can formulate effective fire prevention measures in advance, thereby reducing the possibility of fire and reducing its losses to the ecological environment and human society.

[0101] In summary, the present invention successfully solves the problems existing in the existing technology in forest fire risk assessment by introducing causal graph neural network and making full use of historical data from various regions, so that a single model can be applied to all relevant regions. This method not only improves the accuracy and flexibility of the assessment, but also provides more efficient and scientific tools and methods for forest fire prevention.

[0102] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A forest fire risk assessment method based on a causal graph network, characterized in that: include: Step 1, using ground sensors deployed in the forest area to obtain ground data; Acquiring aerial data through satellite remote sensing technology; The aerial data includes digital elevation model DEM data and multispectral remote sensing data; The Normalized Difference Vegetation Index (NDVI) is calculated based on satellite multispectral remote sensing data to reflect vegetation growth conditions; ground data and aerial data are fused to form a comprehensive data set for forest fire risk assessment; Step 2, dividing the relevant area into grids according to the location information of the ground sensor; Map the DEM data and the Normalized Difference Vegetation Index (NDVI) data onto the divided grids, and calculate the average data value in each grid; normalize the ground data, the DEM average value, and the NDVI average value data to form a feature vector for model input; Step 3: Build a forest fire risk assessment model based on a causal graph neural network, which includes an encoder for processing time series data and a decoder for predicting future states; The model is trained using a training dataset containing data from multiple representative regions, the loss is calculated through forward propagation, and the model parameters are updated using back propagation until the model performance meets the preset requirements; Step 4: Input the time series data of the area to be evaluated into the trained causal graph neural network model; obtain the future state change information through model prediction; and evaluate the forest fire risk based on the prediction results; In step 3, the encoder is designed to use a fully connected graph neural network structure fenc, which transforms and processes the feature vector of the time series data, outputs the feature vector of the corresponding edge relationship, and converts it into a probability prediction of the edge relationship through a softmax function; The decoder design includes receiving the feature vector of the causal relationship and time series data predicted by the encoder as input, performing forward propagation through the neural network, and predicting the change value of the current feature vector and the feature vector of the next time stage, so as to obtain the estimated mean of the feature vector of the next time stage.

2. The forest fire risk assessment method based on the causal graph network according to claim 1, characterized in that: In step 1, the ground data includes three-dimensional multi-dimensional perception data of air temperature and humidity, ambient light intensity, wind speed and direction, rainfall, soil moisture content, surface litter moisture content and phenological fire risk factors.

3. The forest fire risk assessment method based on causal graph network according to claim 1, characterized in that: In step 3, a loss function is used for optimization during the training process of the causal graph neural network model.

4. The forest fire risk assessment method based on causal graph network as claimed in claim 3 is characterized in that: The loss function includes the negative log-likehood function of the decoder and the KL divergence of the encoder, which is used to measure the difference between the model prediction result and the actual result.

5. The forest fire risk assessment method based on causal graph network according to claim 1, characterized in that: In step 1, the ground sensor data focuses on two key indicators: surface temperature and humidity and moisture content of ground combustibles; the surface temperature and humidity sensor includes a power module, a transmitter module, a drift and temperature compensation module, and a data processing module, which work together to obtain surface temperature and humidity data, and capture weak signal changes through signal sampling and amplification functions; a multispectral camera is used to capture multispectral images of ground vegetation, and the chlorophyll content and moisture content of the vegetation are inverted based on an empirical model to obtain a ground combustible moisture index.

6. The forest fire risk assessment method based on causal graph network according to claim 1, characterized in that: In the step 1, the calculation formula of the normalized difference vegetation index NDVI is NDVI=(NIR-RED) / (NIR+RED), wherein NIR is near infrared light reflection and RED is red light reflection.

7. The forest fire risk assessment method based on causal graph network according to claim 1, characterized in that: The decoder further calculates the probability distribution of the feature vector value based on the estimated mean and represents the prediction results of future time series data in the form of normal distribution.

8. The forest fire risk assessment method based on causal graph network according to claim 1, characterized in that: The step 4 also includes setting a fixed number of training cycles, and calculating and monitoring the training loss and validation loss of the model at the end of each cycle to evaluate the convergence and performance of the model.

Citation Information

Patent Citations

  • Forest fire prediction method based on landform and remote sensing data

    CN114898138A

  • Learner emotional evolution analysis method and system based on causal graph neural network

    CN115374790A

  • Visual dialogue generation method and device based on anti-fact common sense causal reasoning

    CN116739056A

  • Forest fire risk assessment method and model construction method and device thereof

    CN117057497A

  • Forest fire risk prediction method based on graph convolutional neural network

    CN117575320A