A method for predicting forest fire risk

Abstract

The present invention proposes a method for predicting forest fire risks, which relates to the field of machine learning. The forest fire risk prediction process proposed by the present invention includes constructing a forest fire risk assessment data set, generating data sequence feature vectors, and then sequentially passing through a data feature information interaction module and a spatial neighborhood feature interaction module. Then, the data sequence features are converted into time series features, and a time series feature capture module is used and a fully connected neural network is utilized to obtain predicted values. The data feature information interaction module proposed by the present invention captures the correlation between different types of data. The spatial neighborhood feature interaction module aggregates the feature information of adjacent regions. The time series feature capture module processes a unified time series feature representation, realizes accurate modeling of long time series data. By using the above three modules, the model can consider the information in the three dimensions of data, neighborhood and time, further enhancing the prediction accuracy of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and particularly relates to a method for predicting forest fire risk. Background Art

[0002] Forest fires refer to fire incidents occurring in forests or forest areas, which have a significant negative impact on the ecosystem, including extensive damage to vegetation, destruction of animal habitats, and deterioration of soil quality. This may further lead to soil erosion and pose a threat to the life and property safety of nearby residents. Therefore, predicting forest fire risk is particularly important. Through scientific risk prediction, potential high-risk areas can be effectively identified, and appropriate preventive and response measures can be taken in a timely manner to reduce the probability of fire occurrence and the losses it brings.

[0003] Machine learning has shown significant advantages in forest fire prediction due to its ability to process multiple data types simultaneously, handle new data in real time, and have good scalability. Using it for forest fire prediction can automatically learn and identify potential fire risks from complex data, thus achieving more accurate prediction. However, existing machine learning-based forest fire prediction methods simply splice different types of data and embed them into a feature space, ignoring the multivariate correlation between them, thereby limiting the performance and generalization ability of the model.

[0004] Forest fire risk assessment data usually exhibits highly nonlinear and complex relationships. Embedding each data sequence independently into a feature space helps the machine learning model learn and utilize the potential connections between different data more effectively, focus on spatial neighborhood features, and then convert the data sequence information into time series information. Further using the information in the time dimension enables the model to more comprehensively and accurately warn of forest fire risks, help formulate effective fire prevention measures, and reduce the losses caused by forest fires. Summary of the Invention

[0005] The present invention provides a method for predicting forest fire risk, aiming to learn more representative data feature representations by utilizing the potential connections between different data sources at the same location. Subsequently, a graph convolutional network is used to capture and aggregate the feature information of the spatial neighborhood, enabling the model to identify and learn the correlation and dependence between adjacent regions. Finally, the data sequence features are converted into time series features, combined with the capture and processing of time series information, to dynamically track the changes in fire risk and achieve accurate prediction of forest fire risk.

[0006] The technical method adopted by the present invention to achieve the above object specifically includes the following steps:

[0007] S1. Collect relevant monitoring data for each forest area, including temperature, precipitation, wind speed, wind direction, relative humidity, vegetation type, the accumulation of combustible substances on the ground, terrain orientation, and smoke content. Clean the data to improve its quality and construct a forest fire risk assessment dataset;

[0008] S2. Embed each type of data sequence generated in a forest area within a certain time interval into a feature space and splice them into a feature vector of this type of data;

[0009] S3. Construct a data feature information interaction module. Use the multi-head attention mechanism to enable the interaction of this feature information with the feature information of other different types of data, capture the correlation between different types of data, and thus generate richer and more representative feature embeddings;

[0010] S4. Construct a spatial neighborhood feature interaction module. Through the graph convolutional network (Graph Convolutional Networks, GCN), learn the correlation weights between the current forest area and its adjacent areas, and use these weights to aggregate the feature information of adjacent areas, so that the feature representation of the current forest area contains rich comprehensive spatial neighborhood information;

[0011] S5. Extract and splice the information at each moment in the data sequence information in sequence over time to obtain time series information, where the information at each time step contains all the data information of a specific forest area at the same time point;

[0012] S6. Construct a time series feature capture module. Through the LSTM (Long Short-Term Memory) network, extract and process the dependencies in long time series data to enhance the model's understanding and expression ability of time dynamic features;

[0013] S7. Construct a fully connected neural network to calculate the final forest fire risk prediction value from the finally obtained feature representation.

[0014] Preferably, in step S1, the monitoring data for each forest area includes temperature, precipitation, wind speed, wind direction, relative humidity, vegetation type, the accumulation of combustible substances on the ground, terrain orientation, and smoke content. Preprocess the collected data, including removing error values and outliers to ensure the accuracy of the data, filling in missing values to ensure the integrity of the data, and avoiding analysis biases caused by incomplete data.

[0015] Preferably, in step S2, the input monitoring data for the forest area is , where is the number of data types, is the time series length of the monitoring data related to the forest area, is the type of data in the monitoring value sequence of this data type obtained within the length of time, where is the monitoring value obtained at the moment. The corresponding feature vector is obtained through the look-up table function:

[0016] ;

[0017] In the formula, is the look-up table function, is the new feature vector obtained through the look-up table. After all the data in are mapped through the look-up table and concatenated in chronological order, the specific process is as follows:

[0018] ;

[0019] In the formula, is the unified feature representation of the monitoring value sequence of the type of data obtained within the length of time , is the data dimension after passing through the look-up table, and the symbol is the concatenation operation.

[0020] Preferably, since different types of data have completely different physical meanings, mapping each type of data independently into a feature vector and concatenating them according to the time series can more effectively capture and learn the correlation between different data, avoiding the problems of information loss and weakening of correlation that may be caused by mixing multiple data in one feature vector, thus realizing the flexible extraction of data features and the retention of time-dependent relationships, providing a solid foundation for the subsequent model to learn and model the multivariate correlation between data.

[0021] Preferably, in step S3, the multi-head attention mechanism with the number of attention heads being is used to capture the correlation between different types of data, thereby generating richer and more representative feature embeddings. First, the input feature vector matrix is respectively passed through three linear transformations to generate a query matrix , a key matrix and a value matrix , and the specific calculation process is as follows:

[0022] ;

[0023] ;

[0024] ;

[0025] Wherein 、 and are the query matrix, key matrix, and value matrix respectively, 、 and are trainable weight matrices for generating queries, keys, and values respectively, is the dimension of the vector. Subsequently, the dot product between the query and the key is calculated to obtain the attention score matrix. The specific calculation process is as follows:

[0026] ;

[0027] Wherein is the attention score matrix of the th attention head, is the function, 、 and are the query, key, and value of the th attention head, is the dimension of the key vector, is the scaling factor used to prevent the problem of gradient disappearance caused by too large dot product values, is the transpose operation. After calculating the attention for each head respectively, the outputs of all heads are concatenated to obtain the output of the multi-head attention mechanism. The specific calculation process is as follows:

[0028] ;

[0029] Wherein is the final feature output, where is the feature representation of the rd class of data obtained through the multi-head attention mechanism, is the weight matrix for the output transformation after concatenation, is the concatenation operation.

[0030] Preferably, the multi-head attention mechanism projects the current data feature representation into multiple subspaces, calculates the attention weights between it and other input data in each subspace respectively, so as to dynamically select more relevant parts of other input data for aggregation. Finally, the outputs of each subspace are concatenated together to capture the diverse feature representations in the data.

[0031] Preferably, in step S4, the graph convolutional network is used to learn the correlation weights between the current forest area and its adjacent areas, and these weights are used to aggregate the feature information of the adjacent areas. The specific calculation process is as follows:

[0032] ;

[0033] ;

[0034] wherein is the feature matrix obtained after passing through the graph convolutional network, is the adjacency matrix between forest regions, where is the number of divided forest regions, is the identity matrix, is the adjacency matrix with self-loops added, is 's degree matrix, and its diagonal elements are the sum of each row element of, is the trainable weight matrix.

[0035] Preferably, through the graph convolutional network, the representation of each forest region not only retains its own features but also integrates the features of its neighboring regions, thereby being able to more effectively capture the information of spatial neighborhood features. This feature integration method can improve the model's understanding and expression ability of the relationships between regions, and further enhance the ability to recognize the overall spatial pattern, enhancing the robustness and accuracy of the model.

[0036] Preferably, in step S5, first extract the feature representations of various types of data in at the same moment and splice them to obtain the time series feature representation , where is the feature representation at the th moment.

[0037] Preferably, in step S6, use LSTM to capture long-term dependencies and obtain the feature representation containing historical information. The specific calculation process is as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] wherein are learnable weight parameters, is the bias parameter, is function are the input gate, forget gate, and output gate of the LSTM at the time stamp respectively, is the time stamp at the candidate memory cell is the time stamp at the memory cell is the time stamp at the hidden state is element-wise multiplication

[0045] Preferably, by concatenating the feature representations of different types of data within the same time step, a unified time series feature representation is generated, and combined with the long short-term memory mechanism of the LSTM, the long-term dependencies and multivariate correlations in the data are captured. This not only effectively integrates the information of multi-source data and improves the expression ability of the model, but also enhances the prediction accuracy of complex time series data through dynamic memory management and learnable weight parameters, providing a solid foundation for more accurate forest fire risk prediction.

[0046] Preferably, in step S7, the output at the last moment of the LSTM in step S6 is taken as the final feature vector representation, and this feature vector is passed through a fully connected neural network to calculate the final forest fire risk prediction value:

[0047] ;

[0048] wherein is the fully connected neural network, is function, is the output at the last moment of the LSTM, is the final forest fire risk prediction value

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention first obtains the feature representation of data in units of data categories, uses the multi-head attention mechanism to capture the correlations between different types of data, dynamically selects relevant data for aggregation, and generates richer and more representative feature embeddings. Then, through the graph convolutional network, the correlation weights between the forest area and its adjacent areas are learned to aggregate the feature information of adjacent areas, thereby introducing the information of the spatial neighborhood into the feature representation. Subsequently, the data sequence features are converted into time series features, and the LSTM is used to process the unified time series feature representation to achieve accurate modeling of long time series data and further enhance the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a step diagram of a forest fire risk prediction method

[0051] Figure 2 It is a diagram of the data feature information interaction module.

[0052] Figure 3 It is a diagram of the partial spatial neighborhood feature interaction module.

[0053] Figure 4 It is a diagram for converting data sequence information into time series information.

[0054] Figure 5 It is an effect diagram for predicting the possibility of forest fires. Specific implementation mode

[0055] The present invention proposes a method for predicting forest fire risks. This method constructs multi-source data sequence features at the same location, makes full use of the potential correlation between different data sources. Then, it uses a graph convolutional network to capture and aggregate the feature information of the spatial neighborhood, enabling the model to identify and learn the correlation and dependence between adjacent regions. Finally, it converts the data sequence features into time series features, combined with the capture and processing of time series information. This method comprehensively considers the three dimensions of data, neighborhood, and time, thereby obtaining a richer and more comprehensive feature representation and achieving accurate prediction of forest fire risks. The technical solutions in the embodiments of the present invention will be described in detail and completely below, specifically including the following steps, as Figure 1 shown.

[0056] S1. Collect the relevant monitoring data of each forest area for 3 years, including temperature, precipitation, wind speed, wind direction, relative humidity, vegetation type, the accumulation of combustible substances on the ground, terrain orientation, and smoke content. Clean the data, including filling missing values, removing irrelevant data, and scaling some data to improve the quality of the data. Construct a forest fire risk assessment data set, and finally divide the data set into a training set and a validation set according to a ratio of 7:3. The first 14 days of data per day are used as the historical data for method learning.

[0057] Furthermore, the relevant monitoring data of the forest area is , where is the number of data types, is the time series length. For the th type of data, the monitoring value sequence obtained within days is , where is the monitoring value obtained at moment. is obtained through a lookup table function to obtain the corresponding feature vector:

[0058] ;

[0059] In the formula is a lookup table function, is the new feature vector obtained through the lookup table, and after all the data in

[0060] ;

[0061] In the formula is the unified feature representation of the monitoring value sequence obtained by the th type of data within days, and the symbol is the concatenation operation.

[0062] S3. Construct a data feature information interaction module, and through the multi-head attention mechanism, make the feature vector interact with the feature vectors of other types of data, so as to capture the correlation between different types of data and generate a richer and more representative feature embedding.

[0063] Furthermore, as Figure 2 shown, by using a multi-head attention mechanism containing attention heads, capture the correlation between different types of data, and respectively pass the input feature vector matrix through three linear transformations to generate a query matrix , a key matrix and a value matrix , and the specific calculation process is as follows:

[0064] ;

[0065] ;

[0066] ;

[0067] In the formula , and are the query matrix, the key matrix and the value matrix respectively, , and are the trainable weight matrices used to generate the query, the key and the value respectively. Subsequently, calculate the dot product between the query and the key to obtain the attention score matrix, and the specific calculation process is as follows:

[0068] ;

[0069] In the formula is the attention score matrix of the th attention head, is function, , and are the query, key, and value of the -th attention head, is the dimension of the key vector, where is the data dimension obtained through the lookup table function, is the number of attention heads in the multi-head attention mechanism, is the scaling factor used to prevent the problem of gradient disappearance caused by too large dot product values, is the transpose operation. After calculating the attention for each head respectively, the outputs of all heads are concatenated to obtain the output of the multi-head attention mechanism. The specific calculation process is as follows:

[0070] ;

[0071] In the formula is the final feature output, where is the -th class of data's feature representation obtained through the multi-head attention mechanism, is the weight matrix of the output transformation, is the concatenation operation.

[0072] S4. Construct a spatial neighborhood feature interaction module. Through the graph convolutional network, learn the correlation weights between the current forest area and its adjacent areas, and use the correlation weights to aggregate the feature information of the adjacent areas into the feature information of the current forest area, so that the feature representation of the current forest area contains rich comprehensive spatial neighborhood information.

[0073] Furthermore, as Figure 3 shows, taking the forest areas as nodes and whether the forest areas are adjacent as edges, construct the adjacency matrix between the forest areas, where the adjacent forest areas are marked as 1 and the non-adjacent areas are marked as 0. Through the graph convolutional network, learn the correlation weights between the current forest area and its adjacent areas, and use these weights to aggregate the feature information of the adjacent areas. The specific calculation process is as follows:

[0074] ;

[0075] ;

[0076] In the formula is the feature matrix obtained after passing through the graph convolutional network, is the adjacency matrix between the forest areas, where is the number of divided forest areas, is the identity matrix, is the adjacency matrix with self-loops added, is The degree matrix, whose diagonal elements are the sum of the elements in each row of is the trainable weight matrix.

[0077] S5. Sequentially extract the relevant information at each moment in the data sequence according to the time order, and splice this information to construct a complete time series. In this process, the data information at each time step includes all the observation data of a specific forest area at the same time point, so as to ensure that the time series information can reflect the dynamic changes of this area.

[0078] Furthermore, as Figure 4 shown, first extract the feature representations of various types of data in at the same moment and splice them to obtain the time series feature representation where is the feature representation at the

[0079] S6. Construct a time series feature capture module, input the time series data into the LSTM network, and extract the dependencies in the time dimension to enhance the model's understanding and expression ability of time dynamic features.

[0080] Furthermore, use LSTM to capture long-term dependencies and obtain the feature representation containing historical information. The specific calculation process is as follows:

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] In the formula is the learnable weight parameter, is the bias parameter, is function, are the input gate, forget gate and output gate of the LSTM at the time stamps respectively, is the candidate memory cell at the time stamp and is the memory cell at the time stamp and is the time stamp The hidden state at is element-wise multiplication.

[0088] S7. Construct a fully-connected neural network to calculate the final forest fire risk prediction value from the finally obtained feature representation.

[0089] Furthermore, in step S7, the output at the last moment of the LSTM in step S6 is taken as the final feature vector representation, and this feature vector is passed through a fully-connected neural network to calculate the final forest fire risk prediction value:

[0090] ;

[0091] In the formula is the fully-connected neural network, is function, is the output at the last moment of the LSTM, is the final forest fire risk prediction value.

[0092] Furthermore, this method uses the Python 3.9 language, the PyTorch framework in the CUDA 12 environment, and is trained on an NVIDIA RTX 3090 GPU. The batch size during training is 64, and the learning rate is .

[0093] Furthermore, the prediction effect of this method is as Figure 5 shown. The ordinate is the probability of forest fire (% ), the abscissa is time (days), the asterisks indicate that a forest fire actually occurred, and the black line indicates the predicted probability of forest fire. It can be seen from the figure that when a forest fire actually occurred, the predicted value obtained by this method was about 80%, which can fit well with the real data, proving the effectiveness of this method.

[0094] The above is only the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the creative concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A forest fire risk prediction method, characterized in that: The following steps are involved: S1. Collect 9 types of monitoring data in each forest area, including temperature, precipitation, wind speed, wind direction, relative humidity, vegetation type, accumulation of combustible materials on the ground, terrain orientation and smoke content, and clean the data, including filling missing values, removing irrelevant data, and scaling some data to improve data quality and build a forest fire risk assessment dataset; S2, embed each type of data sequence generated by the forest area within a certain time interval into a feature space, and concatenate them into a feature vector of this type of data; S3. Build a data feature information interaction module, model the complex correlation between different data types through a multi-head attention mechanism, so that the feature information of the current data interacts and merges with the features of other data, thereby generating richer and more representative feature information embedding; S4. Construct a spatial neighborhood feature interaction module to learn the correlation weights between the current forest area and other adjacent areas through a graph convolutional network. These weights are used to aggregate the feature information of adjacent areas, thereby enhancing the spatial neighborhood information represented by the current forest area feature, so that it contains richer spatial comprehensive features. S5, extracting data information at each moment in the data sequence in chronological order, and splicing the information to form time series data, in which each time step contains all relevant data information of a specific forest area at the same time point, thereby retaining the data characteristics in the time dimension; S6. Build a time series feature capture module to extract and process complex dependencies in long time series data through LSTM to enhance the model's ability to understand and express time dynamic features; S7. Build a fully connected neural network and further calculate and process the final feature representation to generate the final forest fire risk prediction value.

2. A forest fire risk prediction method according to claim 1, characterized in that: In step S2, the forest area monitoring data input is ,in is the number of data types, is the time series length of the forest area related monitoring data, For the The data in The monitoring value sequence of this data type obtained within a certain length of time, where for The monitoring value obtained at any time will be After all the data in the table are obtained through the lookup table function to obtain the corresponding feature vectors, they are spliced ​​in chronological order. The specific process is as follows: ; ; In the formula is a lookup table function, for The new feature vector obtained by looking up the table, For the Type of data in length time The unified feature representation of the monitoring value sequence obtained in is the data dimension after the lookup table, symbol For splicing operation.

3. A forest fire risk prediction method according to claim 2, characterized in that: In step S3, first, the input feature vector matrix Generate the query matrix through three linear transformations respectively , key matrix Sum Matrix , using the number of attention heads as The multi-head attention mechanism captures the correlation between different types of data. The specific calculation process is as follows: ; ; ; In the formula , and are trainable weight matrices used to generate queries, keys, and values, respectively. , and are query matrix, key matrix and value matrix respectively, is the dimension of the vector. Then, the dot product between the query and the key is calculated to obtain the attention score matrix. The specific calculation process is as follows: ; In the formula For the The attention score matrix of the attention heads, for function, , and For the The query, key, and value of each attention head, is the dimension of the key vector, is the scaling factor, For the transposition operation, the outputs of all heads after calculating the attention are concatenated to obtain the output of the multi-head attention mechanism. The specific calculation process is as follows: ; In the formula is the final feature output, where For the The feature representation of class data is obtained through the multi-head attention mechanism. is the weight matrix of the output transformation after splicing, For splicing operation.

4. A forest fire risk prediction method according to claim 3, characterized in that: In step S4, the graph convolution network is used to learn the correlation weights between the current forest area and its adjacent areas, and these weights are used to aggregate the feature information of the adjacent areas. The specific calculation process is as follows: ; ; In the formula is the feature matrix obtained after the graph convolutional network, is the adjacency matrix between forest regions, where is the number of forest areas divided, is the adjacency matrix with self-loops added, for The degree matrix of The sum of each row of elements, is the identity matrix, is the trainable weight matrix.

5. A forest fire risk prediction method according to claim 4, characterized in that: In step S5, Extract the feature representation of various types of data at the same time, and concatenate these features to generate a time series feature representation ,in For the Feature representation of the moment.

6. A forest fire risk prediction method according to claim 5, characterized in that: In step S6, LSTM is used to capture long-term dependencies and obtain feature representations containing historical information. The specific calculation process is as follows: ; ; ; ; ; ; In the formula is the learnable weight parameter, is the bias parameter, for function, Timestamp The input gate, forget gate and output gate at and Timestamp The candidate memory cells and memory cells at For timestamp The hidden state of is element-wise multiplication.

7. A forest fire risk prediction method according to claim 6, characterized in that: Step S7, taking the output of the LSTM at the last moment in step S6 As the final feature vector representation, the feature vector is passed through a fully connected neural network to calculate the final forest fire risk prediction value: ; In the formula is a fully connected neural network, for function, is the output of LSTM at the last moment, is the final forest fire risk prediction value.

Images