Flood prediction method based on multi-sensor data fusion

Abstract

The invention provides a flood prediction method based on multi-sensor data fusion, and the method comprises the steps: obtaining multi-source monitoring data of a monitoring region, the multi-source monitoring data comprises meteorological data, hydrological data and remote sensing data, the meteorological data comprises rainfall information, wind speed information, wind direction information, temperature information and humidity information, and the hydrological data comprises hydrological data and remote sensing data; the hydrological data comprises overground water level information and underground water level information; performing preprocessing and feature extraction on the multi-source monitoring data to form input data corresponding to the monitoring area; inputting the input data into the trained flood prediction model, and determining a prediction result through the flood prediction model. In the model construction process, reliable prediction of flood disasters can be realized by expanding a sample set, performing clustering by using a clustering algorithm, extracting feature vectors as input and adopting LSTM as a model architecture to perform flood prediction.

Description

Technical Field

[0001] The present application relates to the technical field of flood monitoring, and in particular to a flood prediction method based on multi-sensor data fusion. Background Art

[0002] Natural disasters occur frequently around the world, among which floods have a serious impact on human society and economy. Floods are usually accompanied by extreme weather conditions such as heavy rainfall, rainstorms, and flash floods, which cause the water levels of rivers, lakes, reservoirs and other water bodies to rise sharply, and even flood. Due to the suddenness and destructiveness of floods, a rapid and accurate prediction system has become the key to reducing the impact of disasters and improving social response capabilities.

[0003] Traditional flood prediction mainly relies on limited meteorological and hydrological observation stations, which limits the effective monitoring of large areas. Moreover, the data of a single sensor is difficult to fully reflect the complex flood situation. In order to overcome this problem, multi-sensor fusion technology came into being. Multi-sensor fusion can not only improve the spatial and temporal resolution of observation data, but also make up for the deficiencies between individual sensors and improve the reliability and comprehensiveness of information.

[0004] The use of multi-data fusion solutions does have great advantages for flood monitoring. This solution is not only conducive to more accurate prediction of flood events, but can also be combined with artificial intelligence technology to achieve early warning, providing more adequate response time for all sectors of society and minimizing disaster losses. However, when artificial intelligence technology is applied in this regard, due to the large number of data types required, and the relevant data on flood disasters is not too much, there are even fewer data samples that meet the conditions. The reliability of the application of artificial intelligence technology in flood monitoring and early warning depends to a large extent on the number of training samples. There are many samples in normal times, but few flood disaster samples that meet the conditions. The imbalance of the number of samples leads to poor model training results. Therefore, in order to realize the reliable application of artificial intelligence in flood monitoring and early warning based on sensor data, sample imbalance is a problem that must be solved. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a flood prediction method based on multi-sensor data fusion, so as to solve the sample imbalance problem and improve the reliability of flood prediction by expanding flood disaster samples.

[0006] In order to achieve the above purpose, the embodiments of the present application are implemented in the following ways: In a first aspect, an embodiment of the present application provides a flood prediction method based on multi-sensor data fusion, comprising: obtaining multi-source monitoring data of a monitoring area, wherein the multi-source monitoring data includes meteorological data, hydrological data, and remote sensing data, the meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information, and the hydrological data includes ground water level information and groundwater level information; preprocessing and feature extraction of the multi-source monitoring data to form input data corresponding to the monitoring area; inputting the input data into a trained flood prediction model, and determining the prediction result through the flood prediction model.

[0007] In combination with the first aspect, in a first possible implementation method of the first aspect, the construction process of the flood prediction model is: obtaining a historical monitoring data set of the monitoring area, the historical monitoring data set includes a number of historical monitoring data, each historical monitoring data contains meteorological data, hydrological data, and remote sensing data for a period of time, the meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, humidity information, and the hydrological data includes ground water level information and groundwater level information; preprocessing each historical monitoring data to form a sample set, each sample corresponds to a sample label, and the sample label reveals whether a flood disaster occurs; based on the sample label, the sample set is divided into a first-class sample set and a second-class sample set, wherein the first-class sample set The sample set includes samples that have experienced flood disasters, and the second sample set includes samples that have not experienced flood disasters; the first sample set is expanded to obtain an expanded first sample set, and the expanded first sample set and the second sample set are merged to form a training sample set, wherein the training sample set contains N training samples; the training sample set is clustered, and after the clustering is completed, the feature vector corresponding to the training sample is constructed based on the distance between the training sample in the training sample set and each cluster center; a neural network model is constructed, and the training sample set is divided into a training set and a test set, and the neural network model is trained and tested through the feature vectors corresponding to the training samples in the training set and the test set to obtain a trained flood prediction model.

[0008] In combination with the first possible implementation of the first aspect, in a second possible implementation of the first aspect, a sample set of a class is expanded to obtain an expanded sample set of a class, including: calculating a sample set of a class Middle Samples and a class of samples The distances of other samples in : , in, A sample set Middle Samples and a class of samples Middle The distance of samples, A sample set Middle samples, A sample set Middle samples, is the dimension of the sample; based on the Samples With a sample set The distance of other samples in the , offset synthesis, get the Samples The corresponding synthetic samples.

[0009] In combination with the second possible implementation manner of the first aspect, in a third possible implementation manner of the first aspect, based on Samples With a sample set The distance of other samples in the , offset synthesis, get the The synthetic samples corresponding to the samples include: Samples , the following formula is used for sample synthesis: , in, For sample To the sample The synthetic samples obtained by offsetting, is the offset coefficient, satisfying: , , , A sample set Middle Samples and a class of samples The maximum distance to other samples in The corresponding value range is [0.25, 0.5).

[0010] In combination with the third possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, determining the sample Can it be used as a sample? The process of offsetting objects is as follows: With sample Is the distance less than the sample With a class of sample sets The mean distance of all samples in If the sample With sample The distance is less than the mean distance , determine the sample As a sample The offset object; if the sample With sample The distance is not less than the mean distance , determine the sample Not available as sample The offset object.

[0011] In combination with the first possible implementation manner of the first aspect, in a fifth possible implementation manner of the first aspect, clustering the training sample set includes: calculating the training sample set The similarity between every two training samples in the matrix is ​​obtained ; Construct the attraction matrix and the membership matrix ; For the attraction matrix and the membership matrix Perform iterative updates, and update the cluster center of the current iteration number each time until the set number of iterations is reached, or the The cluster centers obtained in the first iteration are The cluster centers obtained in the iterations are the same, and clustering is completed. Cluster centers.

[0012] In combination with the fifth possible implementation of the first aspect, in a sixth possible implementation of the first aspect, the training sample set is calculated The similarity between every two training samples in the matrix is ​​obtained , including: using the following formula to calculate the training sample set The similarity between every two training samples in :

[0013] , , in, The training sample set Middle training samples With training samples The similarity between Represents training samples With training samples The distance between For the training samples With training sample set The mean distance between all training samples in , For the training samples With training sample set The maximum distance between all training samples in For the training samples With training sample set The minimum distance between all training samples in For training samples As the reference degree of cluster center, The training sample set The total number of training samples in; Based on this, the similarity matrix formed : , in, is the similarity matrix.

[0014] In combination with the sixth possible implementation of the first aspect, in a seventh possible implementation of the first aspect, the attraction matrix and the membership matrix Perform iterative updates, including: Update the attraction matrix using the following formula and the membership matrix : , , in, For the The attraction matrix of the iteration In the sample As a sample The fitness of the cluster center, For the The membership matrix of the iteration In the sample Select Sample The fitness of the cluster center.

[0015] In combination with the seventh possible implementation manner of the first aspect, in an eighth possible implementation manner of the first aspect, constructing a feature vector corresponding to the training sample based on the distance between the training sample in the training sample set and each cluster center includes: for the training sample set Middle training samples :Calculate training samples and The distance between the cluster centers is obtained distance values; based on distance values ​​form training samples The corresponding eigenvector .

[0016] In combination with the seventh possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, the neural network model adopts LSTM.

[0017] Beneficial effects:

[0018] 1. By constructing a flood prediction model, we obtain historical monitoring data sets (multi-sensor data, including several historical monitoring data, each of which contains meteorological data, hydrological data, and remote sensing data for a period of time. Meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. Hydrological data includes ground water level information and groundwater level information). After preprocessing, we form a sample set, divide the sample set (by label division, so as to expand the flood disaster samples with a small number), and further expand the sample set to form a training sample set. Then, we use the clustering algorithm to cluster, extract the feature vector of the training sample as the input of the neural network model, use LSTM as the model architecture, and use the training set and test set for training and testing to obtain a trained flood prediction model. In the actual application process, we can input the multi-source monitoring data into the flood prediction model after preprocessing and feature extraction, and determine the prediction results through the flood prediction model to achieve reliable prediction of flood disasters. This method can effectively improve the prediction effect of the flood prediction model, provide more sufficient response time for all sectors of society, and minimize disaster losses.

[0019] 2. In the process of sample set expansion, the offset expansion strategy is adopted to introduce samples that are different from the basic samples (such as samples ) and offset objects (such as samples ), and at the same time, the distance between the basic sample and other samples can be taken into consideration to design an offset strategy and realize sample synthesis, so that the synthetic samples have better feasibility and are closer to reality, thereby ensuring the training effect. The larger the relative distance, the smaller the offset coefficient, and the smaller the relative distance, the larger the offset coefficient, which can effectively balance the errors of the synthetic samples and ensure the availability of the synthetic samples. In addition, when determining whether a sample can be used as an offset object, the distance between the two is used as a screening condition to control the quality of the synthetic samples. Of course, in order to reduce the distribution density of synthetic samples, some other offset object selection strategies can also be designed, such as dividing them into levels according to different distances, and assigning different offset object selection ratios according to different levels (for example, less than the distance mean). , select three quarters of them as the offset objects, not less than the mean distance , select one quarter of them as offset objects), thereby reducing the distribution density of synthetic samples to further improve the training effect.

[0020] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 This is a flow chart of constructing a flood prediction model in this embodiment.

[0023] Figure 2 Flow chart of the flood prediction method based on multi-sensor data fusion in this embodiment. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0025] In order to facilitate the understanding of this solution, the construction process of the flood prediction model is first introduced here. Figure 1 , Figure 1 This is a flow chart of constructing a flood prediction model in this embodiment. The process of constructing a flood prediction model may include step S11, step S12, step S13, step S14, step S15, and step S16.

[0026] First, step S11 may be executed.

[0027] Step S11: Obtain a historical monitoring data set of the monitoring area. The historical monitoring data set includes several historical monitoring data. Each historical monitoring data contains meteorological data, hydrological data, and remote sensing data for a period of time. The meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. The hydrological data includes ground water level information and groundwater level information.

[0028] In this embodiment, historical monitoring data (several pieces) of the monitoring area (for example, XX province, or XX region) can be collected to form a historical monitoring data set. Each piece of historical monitoring data contains meteorological data (including meteorological information at multiple time points in the time period), hydrological data (including water level information at multiple time points in the time period), and remote sensing data (usually one time point in the time period, such as surface topography and elevation information) for a period of time (for example, 24 hours, or 12 hours as a period). Meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. Hydrological data includes ground water level information and groundwater level information. Due to the region as the boundary, the collection of this information is usually distributed to multiple monitoring points, and how to arrange the monitoring points in practice needs to be determined according to the actual situation of the monitoring area, which will not be elaborated here.

[0029] After the historical monitoring data set is obtained, step S12 may be executed.

[0030] Step S12: Preprocess each piece of historical monitoring data to form a sample set, each sample corresponds to a sample label, and the sample label reveals whether a flood disaster occurs.

[0031] In this embodiment, for each piece of historical monitoring data, preprocessing can be performed, such as data alignment (which can be performed based on time, space or other key variables. In this embodiment, time is used as the alignment reference to divide the data into data for each period, and the data for each period is used as a piece of historical monitoring data). After alignment, data cleaning, outlier processing, missing value filling, etc. can be performed to ensure the quality of the data, and the sample corresponding to each piece of historical monitoring data is obtained, thereby forming a sample set. Of course, each sample also corresponds to a sample label, which reveals whether a flood disaster has occurred in this monitoring area (for example, whether the hydrological information of the historical data exceeds the warning value can be defined as whether a flood disaster has occurred).

[0032] Step S13: Divide the sample set into a first-class sample set and a second-class sample set based on the sample labels, wherein the first-class sample set includes samples that have experienced flood disasters, and the second-class sample set includes samples that have not experienced flood disasters.

[0033] In this embodiment, the sample set can be divided into a first-class sample set and a second-class sample set based on the sample labels. Specifically, samples that have experienced flood disasters (of course, they can include various stages of floods, such as the initial stage, flood stage, and end stage) are divided into the first-class sample set, and samples that have not experienced flood disasters are divided into the second-class sample set. Due to the small number of flood disaster events, the number of samples in the first-class sample set is very small compared to the number of samples in the second-class sample set. If no processing is performed, there will be a problem of sample imbalance.

[0034] Therefore, it is necessary to execute step S14.

[0035] Step S14: expanding the sample set of the first category to obtain an expanded sample set of the first category, and merging the expanded sample set of the first category with the sample set of the second category to form a training sample set, wherein the training sample set includes N training samples.

[0036] In this embodiment, in order to expand the samples of a type of sample set, the following method may be used: First, calculate a class of sample sets Middle Samples and a class of samples The distances of other samples in : , (1) in, A sample set Middle Samples and a class of samples Middle The distance of samples, A sample set Middle samples, A sample set Middle samples, is the dimension of the sample. From this, we can calculate a class of sample sets The distance between each sample and other samples in .

[0037] Furthermore, based on Samples With a class of sample sets The distance of other samples in the , offset synthesis, get the Samples The corresponding synthetic samples.

[0038] For example, for Samples , the following formula can be used for sample synthesis: , (2) in, For sample To the sample The synthetic samples obtained by offsetting, is the offset coefficient, satisfying: , (3) , (4) , A sample set Middle Samples and a class of samples The maximum distance to other samples in The corresponding value range is [0.25, 0.5).

[0039] The above is a sample synthesis method designed with an offset strategy. In order to further improve the quality of the synthesized samples, we can select the offset object (such as sample ) to get started.

[0040] For example, determine the sample Can it be used as a sample? The process of offsetting an object is as follows: First, judge the sample With sample Is the distance less than the sample With a sample set The mean distance of all samples in :If the sample With sample The distance is less than the mean distance , determine the sample As a sample The offset object; if the sample With sample The distance is not less than the mean distance , determine the sample Not available as sample The offset object.

[0041] Of course, this strategy has better quality control for synthetic samples. If you need to balance the density of synthetic samples, you can also design the following offset object selection strategy: By sample For example, you can follow other samples and samples The distance between them is divided into distance levels (for example, the average distance is used as the boundary to divide into adjacent sample groups and distant sample groups), and different offset object selection ratios are assigned according to different levels (for example, less than the distance mean A group of neighboring samples, randomly select three quarters of them as offset objects, not less than the mean distance of far-away sample combinations, and randomly select one quarter of them as offset objects), thereby reducing the distribution density of synthetic samples to further improve the training effect.

[0042] After completing the sample expansion of the first-class sample set in this way, the expanded first-class sample set can be obtained. Then it can be compared with the number of the second-class sample set. If the number of samples in the expanded first-class sample set is at the same order of magnitude as that in the second-class sample set, no further processing is required. If the number of samples in the second-class sample set is one order of magnitude higher than that in the first-class sample set, the second-class sample set is sampled to reduce the number of samples in the second-class sample set to balance the samples.

[0043] Then, the expanded first-category sample set and the second-category sample set (determining whether to sample according to actual needs) can be merged to form a training sample set, where the training sample set contains N training samples.

[0044] After obtaining the training sample set, step S15 may be executed.

[0045] Step S15: clustering the training sample set, and after clustering is completed, constructing a feature vector corresponding to the training sample based on the distance between the training sample in the training sample set and each cluster center.

[0046] In this embodiment, the training sample set can be clustered. The clustering process is as follows: First, we can calculate the training sample set The similarity between every two training samples in the matrix is ​​obtained Specifically, the following formula can be used to calculate the training sample set: The similarity between every two training samples in : , (5) , (6) , (7) in, The training sample set Middle training samples With training samples The similarity between Represents training samples With training samples The distance between For the training samples With training sample set The mean distance between all training samples in , For the training samples With training sample set The maximum distance between all training samples in For the training samples With training sample set The minimum distance between all training samples in For training samples As the reference degree of cluster center, The training sample set The total amount of training samples in .

[0047] Based on this, the similarity matrix formed : , (8) in, is the similarity matrix.

[0048] Get the similarity matrix After that, we can further construct the attraction matrix and the membership matrix , in this embodiment, the attraction matrix and the membership matrix All are initialized to 0. Construct the attraction matrix and the membership matrix After that, we can calculate the attraction matrix and the membership matrix Perform iterative updates.

[0049] Specifically, the attraction matrix is ​​updated using the following formula: and the membership matrix : , (9) , (10) in, For the The attraction matrix of the iteration In the sample As a sample The fitness of the cluster center, For the The membership matrix of the iteration In the sample Select Sample The fitness of the cluster center.

[0050] In this way, iterative updates are performed until the set number of iterations (for example ), or The cluster centers obtained in the first iteration are If the cluster centers obtained in the iterations are the same, clustering can be terminated, and a total of Cluster centers.

[0051] After clustering is completed, the feature vector corresponding to the training sample can be constructed based on the distance between the training sample in the training sample set and each cluster center. Middle training samples :You can calculate the training samples and The distance between the cluster centers is obtained distance values, and then based on distance values ​​form training samples The corresponding eigenvector .For example Each cluster center has a corresponding unique number, and the training samples are calculated in the order of the numbers. and The distance between cluster centers and The distance value is based on The cluster centers have corresponding unique numbers and are arranged in sequence to form a A one-dimensional vector of element values, as training samples The corresponding eigenvector .

[0052] Then, a neural network model is constructed (LSTM is selected in this embodiment, and other embodiments may also select other models, such as a variant model of LSTM, or RNN, attention mechanism, etc.), and the training sample set is divided into a training set and a test set (7:3). The neural network model is trained and tested through the feature vectors corresponding to the training samples in the training set and the test set, and finally a trained flood prediction model is obtained.

[0053] After completing the construction of the flood prediction model, the flood prediction model can be installed on the server, and then the flood prediction method based on multi-sensor data fusion can be run through the server's built-in program to achieve flood prediction.

[0054] See also Figure 2 , Figure 2 is a flow chart of the flood prediction method based on multi-sensor data fusion in this embodiment. In this embodiment, the flood prediction method based on multi-sensor data fusion may include step S21, step S22 and step S23.

[0055] First, the server may execute step S21.

[0056] Step S21: Acquire multi-source monitoring data of the monitoring area, wherein the multi-source monitoring data includes meteorological data, hydrological data, and remote sensing data. The meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. The hydrological data includes ground water level information and groundwater level information.

[0057] In this embodiment, the server can obtain multi-source monitoring data (collected by various sensors arranged in the monitoring area). The multi-source monitoring data includes meteorological data, hydrological data, and remote sensing data. Meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. Hydrological data includes ground water level information and groundwater level information. It should be noted that due to the complexity of the actual situation, the monitoring area may be different. The several types of data listed in the multi-source data are only examples, and the data merging (and integrated calculation) of multiple monitoring points is not described in detail (the monitoring areas are diverse in actual situations). Factors such as quantity and location are uncertain. This embodiment only uses the abstracted integrated data (for example, groundwater data, an abstracted parameter may be calculated by integrating multiple monitoring points) as an example for explanation, and should not be regarded as a limitation of this application. Of course, this situation needs to be consistent with the processing of historical monitoring data in the construction of the flood prediction model.

[0058] After obtaining the multi-source monitoring data, the server may execute step S22.

[0059] Step S22: pre-processing and feature extraction of multi-source monitoring data to form input data corresponding to the monitoring area.

[0060] In this embodiment, the server may perform preprocessing on the multi-source monitoring data, such as data cleaning, outlier processing, missing value filling, etc. After the preprocessing is completed, feature extraction may be further performed.

[0061] Exemplarily, the feature extraction process mainly calculates the distance between the multi-source monitoring data and M cluster centers (the clustering process is described in detail in the construction process of the flood prediction model in the previous article, and will not be repeated here), and then arranges them in sequence according to the unique numbers of the M cluster centers to obtain a one-dimensional vector containing M element values, realizes feature extraction, and obtains the input data corresponding to the monitoring area.

[0062] After obtaining the input data, the server may execute step S23.

[0063] Step S23: input the input data into the trained flood prediction model, and determine the prediction result through the flood prediction model.

[0064] In this embodiment, the server can input the input data into a trained flood prediction model, perform feature processing through the flood prediction model, and finally determine the prediction result to achieve flood prediction in the monitoring area.

[0065] It should be noted that for flood prediction, the monitoring area can be expandable. For example, historical data may come from multiple regions, and monitoring is not limited to these regions. Of course, in order to achieve high-precision flood prediction, a regional restriction solution can be adopted to conduct special training for areas prone to flood disasters (using the historical data set of this area to form a training sample set to train the model) in order to achieve higher precision and accuracy of flood prediction, which is not limited here.

[0066] In summary, the embodiment of the present application provides a flood prediction method based on multi-sensor data fusion, by constructing a flood prediction model, obtaining a historical monitoring data set (multi-sensor data, including several historical monitoring data, each historical monitoring data contains meteorological data, hydrological data, and remote sensing data of a period of time, meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, humidity information, hydrological data includes ground water level information and groundwater level information), and forming a sample set after preprocessing, and dividing the sample set (dividing according to labels, so as to expand the flood disaster samples with a small number), and further expanding the sample set to form a training sample set. Then, clustering is performed using a clustering algorithm, and the feature vectors of the training samples are extracted as the input of the neural network model. LSTM is used as the model architecture, and training and testing are performed using the training set and the test set to obtain a trained flood prediction model, so that in the actual application process, the multi-source monitoring data can be input into the flood prediction model after preprocessing and feature extraction, and the prediction results are determined by the flood prediction model to achieve reliable prediction of flood disasters. This approach can effectively improve the prediction effect of flood prediction models, provide more sufficient response time for all sectors of society, and minimize disaster losses.

[0067] In the process of sample set expansion, the offset expansion strategy is adopted to introduce samples that are different from the basic samples (such as samples ) and offset objects (such as samples ), and at the same time, the distance between the basic sample and other samples can be taken into consideration to design an offset strategy and realize sample synthesis, so that the synthetic samples have better feasibility and are closer to reality, thereby ensuring the training effect. The larger the relative distance, the smaller the offset coefficient, and the smaller the relative distance, the larger the offset coefficient, which can effectively balance the errors of the synthetic samples and ensure the availability of the synthetic samples. In addition, when determining whether a sample can be used as an offset object, the distance between the two is used as a screening condition to control the quality of the synthetic samples. Of course, in order to reduce the distribution density of synthetic samples, some other offset object selection strategies can also be designed, such as dividing them into levels according to different distances, and assigning different offset object selection ratios according to different levels (for example, less than the distance mean). , select three quarters of them as the offset objects, not less than the mean distance , select one quarter of them as offset objects), thereby reducing the distribution density of synthetic samples to further improve the training effect.

[0068] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0069] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A flood prediction method based on multi-sensor data fusion, characterized in that: include: Acquire multi-source monitoring data of the monitoring area, wherein the multi-source monitoring data includes meteorological data, hydrological data, and remote sensing data. The meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. The hydrological data includes ground water level information and ground water level information. Preprocess and extract features of multi-source monitoring data to form input data corresponding to the monitoring area; The input data is input into the trained flood prediction model, and the prediction results are determined by the flood prediction model.

2. The flood prediction method based on multi-sensor data fusion according to claim 1 is characterized in that: The construction process of the flood prediction model is as follows: Obtain a historical monitoring data set of the monitoring area. The historical monitoring data set includes several historical monitoring data. Each historical monitoring data includes meteorological data, hydrological data, and remote sensing data for a period of time. The meteorological data includes rainfall information, wind speed information, wind direction information, temperature information, and humidity information. The hydrological data includes ground water level information and groundwater level information. Each piece of historical monitoring data is preprocessed to form a sample set. Each sample corresponds to a sample label, which reveals whether a flood disaster has occurred. Based on the sample labels, the sample set is divided into a first-class sample set and a second-class sample set. The first-class sample set is the samples that have experienced flood disasters, and the second-class sample set is the samples that have not experienced flood disasters. Perform sample expansion on the first-class sample set to obtain an expanded first-class sample set, and merge the expanded first-class sample set and the second-class sample set to form a training sample set, wherein the training sample set includes N training samples; The training sample set is clustered. After clustering is completed, the feature vector corresponding to the training sample is constructed based on the distance between the training sample in the training sample set and each cluster center; A neural network model is constructed, and the training sample set is divided into a training set and a test set. The neural network model is trained and tested through the feature vectors corresponding to the training samples in the training set and the test set to obtain a trained flood prediction model.

3. The flood prediction method based on multi-sensor data fusion according to claim 2 is characterized in that: Perform sample expansion on a class of sample sets to obtain an expanded class of sample sets, including: Calculate a class of sample sets Middle Samples and a class of samples The distances of other samples in : ， in, A sample set Middle Samples and a class of samples Middle The distance of samples, A sample set Middle samples, A sample set Middle samples, is the dimension of the sample; Based on Samples With a class of sample sets The distance of other samples in the , offset synthesis, get the Samples The corresponding synthetic samples.

4. The flood prediction method based on multi-sensor data fusion according to claim 3 is characterized in that: Based on Samples With a class of sample sets The distance of other samples in the , offset synthesis, get the The synthetic samples corresponding to the samples include: For Samples , the following formula is used for sample synthesis: ， in, For sample To the sample The synthetic samples obtained by offsetting, is the offset coefficient, satisfying: ， ， , A sample set Middle Samples and a class of samples The maximum distance to other samples in The corresponding value range is [0.25, 0.5).

5. The flood prediction method based on multi-sensor data fusion according to claim 4 is characterized in that: Determine the sample Can it be used as a sample? The process of offsetting an object is as follows: Judgment Sample With sample Is the distance less than the sample With a class of sample sets The mean distance of all samples in ; If the sample With sample The distance is less than the mean distance , determine the sample As a sample The offset object; If the sample With sample The distance is not less than the mean distance , determine the sample Not available as sample The offset object.

6. The flood prediction method based on multi-sensor data fusion according to claim 2 is characterized in that: Clustering of training sample sets, including: Calculate the training sample set The similarity between every two training samples in the matrix is ​​obtained ; Constructing the attraction matrix and the membership matrix ; Attraction Matrix and the membership matrix Perform iterative updates, and update the cluster center of the current iteration number each time until the set number of iterations is reached, or the The cluster centers obtained in the first iteration are The cluster centers obtained in the iterations are the same, and clustering is completed. Cluster centers.

7. The flood prediction method based on multi-sensor data fusion according to claim 6 is characterized in that: Calculate the training sample set The similarity between every two training samples in the matrix is ​​obtained ,include: The training sample set is calculated using the following formula: The similarity between every two training samples in : ， ， in, The training sample set Middle training samples With training samples The similarity between Represents training samples With training samples The distance between For the training samples With training sample set The mean distance between all training samples in For the training samples With training sample set The maximum distance between all training samples in For the training samples With training sample set The minimum distance between all training samples in For training samples As the reference degree of cluster center, The training sample set The total amount of training samples in ; Based on this, the similarity matrix formed : ， in, is the similarity matrix.

8. The flood prediction method based on multi-sensor data fusion according to claim 7 is characterized in that: Attraction Matrix and the membership matrix Iterative updates, including: Update the attraction matrix using the following formula and the membership matrix : ， ， in, For the The attraction matrix of the iteration In the sample As a sample The fitness of the cluster center, For the The membership matrix of the iteration In the sample Select Sample The fitness of the cluster center.

9. The flood prediction method based on multi-sensor data fusion according to claim 8 is characterized in that: The feature vector corresponding to the training sample is constructed based on the distance between the training sample in the training sample set and each cluster center, including: For the training sample set Middle training samples : Calculate training samples and The distance between the cluster centers is obtained distance values; based on distance values ​​form training samples The corresponding eigenvector .

10. The flood prediction method based on multi-sensor data fusion according to claim 8, characterized in that: The neural network model uses LSTM.

Images