Data-driven flood grading method and system

Abstract

The invention discloses a data-driven flood grading method and system, and particularly relates to the field of flood grading, and the method comprises the steps: obtaining a first grading reference through a multi-source heterogeneous data interface, and extracting a first reference feature through a convolutional neural network; meanwhile, the load pressure value of each edge computing node is obtained in real time to serve as a second grading reference, and the extraction tasks are dynamically distributed to each edge computing node; and inputting the first reference feature into a flood grade classification model to obtain a third grading reference so as to achieve flood grading. According to the data-driven flood grading method and system, real-time matching of computing resource requirements and node processing capacity is realized through time sequence modeling capacity of load data in combination with a dynamic allocation mechanism, and response delay of an actual measurement task is reduced; through a dual-channel architecture, collaborative optimization of spatial features and time sequence features is realized; by establishing a confidence-driven incremental learning mechanism, the recognition accuracy of the novel disaster mode is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of flood classification, and more specifically, to a data-driven flood classification method and system. Background Art

[0002] With the development of water conservancy projects and intelligent monitoring technologies, it is particularly important to build an accurate and efficient flood warning and classification system. Traditional flood classification adopts a three-layer architecture of sensor network - data center - manual analysis. Hydrological data is collected in real time through terminal devices such as water level gauges and rain gauges, and after being transmitted to the central processing unit through a communication network, professional personnel determine the risk level based on historical experience and statistical models. However, this solution has significant drawbacks, including strong dependence on human experience leading to subjective biases, insufficient density of hardware sensor deployment resulting in monitoring blind spots, and static thresholds being unable to dynamically reflect the comprehensive risks of the basin.

[0003] To overcome the static defects of traditional methods, the current trend is towards the data-driven direction. A typical improved solution uses a machine learning model to replace manual rules. A classification architecture based on the random forest algorithm adds interfaces for meteorological radar and satellite remote sensing data, introduces a neural network model to build the mapping relationship between multi-dimensional features and flood levels, and improves the adaptability of the model by dynamically updating the training set. To address the latency defects of the traditional architecture, distributed computing frameworks and machine learning algorithms are currently introduced for optimization. The improved system adopts an architecture of collaborative processing between edge computing nodes and the cloud. Embedded intelligent terminals are deployed at the front end to preprocess the original data, and convolutional neural networks are used to extract feature parameters such as water level rise and flow velocity mutation in real time, and then the classification model is updated in the cloud through a federated learning mechanism.

[0004] However, in actual use, there are still some drawbacks, such as the lack of a dynamic load balancing mechanism between nodes, resulting in processing delays for edge devices in some high-concurrency areas due to overloaded computing resources; the dependence of model training on the completeness of historical disaster data, and when encountering extreme climate events beyond historical extremes, the insufficient coverage of the feature space leads to the failure of classification. Summary of the Invention

[0005] To overcome the above-mentioned defects of the prior art, the present invention provides a data-driven flood classification method and system, through the following solutions, to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A data-driven flood classification method, comprising:

[0008] S1: Obtain a first classification benchmark through a multi-source heterogeneous data interface, where the first classification benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media public opinion data;

[0009] S2: Perform feature extraction tasks on the first classification benchmark through a convolutional neural network to generate the first benchmark features that are spatio-temporally synchronized.

[0010] S3: Obtain the second classification benchmark of each edge computing node in real time. The second classification benchmark is the load pressure value predicted for each edge computing node based on a deep learning model.

[0011] S4: According to the second classification benchmark, dynamically allocate the extraction tasks of the first benchmark features to each edge computing node through a deep reinforcement learning algorithm.

[0012] S5: Pass the first benchmark features through a flood level classification model to obtain the third classification benchmark.

[0013] S6: Establish a classification warning trigger mechanism based on the third classification benchmark and feedback it to the classification decision terminal.

[0014] Preferably, in S2, the first benchmark features include a spatial feature vector group, a time feature matrix, and multi-source fusion features. The spatial feature vector group consists of spectral texture features, terrain gradient features, and hydrological feature maps.

[0015] Preferably, in S4, the implementation method of the deep reinforcement learning algorithm specifically includes:

[0016] Decompose the second classification benchmark into a multi-dimensional state vector. The multi-dimensional state vector includes node hardware status, network status, and task characteristics. Among them, the node hardware status is a physical index reflecting the load pressure, the network status is an index for evaluating the data transmission efficiency between nodes, and the task characteristics are used to quantify the degree of demand of the task for node resources.

[0017] Based on the multi-dimensional state vector, with the goal of maximizing the overall processing efficiency of edge computing nodes and minimizing task processing latency, generate a dynamic allocation strategy through discrete actions and continuous actions. Among them, the discrete action is to output a probability distribution according to the load pressure value and output optional edge nodes. The continuous actions include proportionally splitting the feature extraction tasks to adjacent low-load nodes for high-load nodes and preferentially allocating tasks with high timeliness requirements to nodes with moderate load pressure and stable network.

[0018] Preferably, in S5, obtaining the third classification benchmark through the flood level classification model specifically includes:

[0019] S501: Fuse the first benchmark features through a federated learning framework and construct a three-dimensional flood evolution in combination with a hydrodynamic model to obtain a flood classification feature set. The flood classification feature set is the training data set input into the flood level classification model.

[0020] S502: Learn the flood classification feature set using a dual-channel neural network and adjust the parameters of the flood level classification model through the dynamic weight averaging algorithm;

[0021] S503: Incrementally optimize the classification layer of the flood level classification model;

[0022] S504: Construct the disaster chain topological relationship of the target basin and feedback it to the flood level classification model;

[0023] S505: Construct a multi-level verification system and perform real-time update and correction on the flood level classification model.

[0024] Preferably, in S5, fuse the first benchmark features through the federated learning framework and construct a three-dimensional flood evolution in combination with the hydrodynamic model to obtain the flood classification feature set, specifically including:

[0025] Based on the differential privacy federated learning framework, aggregate the first benchmark features of each edge computing node and the high-precision digital elevation model;

[0026] Couple the hydrodynamic model with multi-spectral remote sensing data for three-dimensional dynamic flood evolution;

[0027] Generate a spatio-temporal labeled training set through historical data backtracking in the system operation database.

[0028] Preferably, in S5, learn the flood classification feature set using a dual-channel neural network and adjust the parameters of the flood level classification model through the dynamic weight averaging algorithm, specifically including:

[0029] Construct multiple dual-channel feature extractors to parallelly process various types of flood feature data;

[0030] Use the dynamic weight averaging algorithm to aggregate multiple dual-channel feature extractors and adjust the aggregation weights according to the KL divergence.

[0031] Preferably, in S5, adjust the aggregation weights according to the KL divergence, specifically expressed as:

[0032]

[0033] where, w i represents the aggregation weight of the i-th dual-channel feature extractor, β represents the preset sensitivity parameter, D KL represents the KL divergence, θ i and θ j respectively represent the model parameters of the i-th dual-channel feature extractor and the model parameters of the j-th dual-channel feature extractor, θ gdenoted as the average parameters of all feature extractors, N denoted as the total number of dual-channel feature extractors, i and j respectively denoted as the indices of the dual-channel feature extractors, D KL (θ i ||θ g ) is used to measure the distribution difference between the model parameters of the i-th dual-channel feature extractor and the average parameters of all feature extractors.

[0034] Preferably, for the S5, the disaster chain topological relationship of the target basin includes disaster nodes and edge weights calculated based on the Pearson correlation coefficient.

[0035] To achieve the above object, the present invention provides the following technical solution: A data-driven flood classification system, including a system operation database, a system central processing module, and a classification decision terminal. Implementing the above data-driven flood classification method further includes:

[0036] Multi-source data acquisition module: used to obtain the first classification benchmark through a multi-source heterogeneous data interface, and the first classification benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media public opinion data;

[0037] Feature extraction module: used to perform feature extraction tasks on the first classification benchmark through a convolutional neural network to generate spatio-temporally synchronized first benchmark features;

[0038] Computing load acquisition module: used to obtain the second classification benchmark of each edge computing node in real time, and the second classification benchmark is the load pressure value predicted for each edge computing node based on a deep learning model;

[0039] Load balancing module: used to dynamically allocate the extraction tasks of the first benchmark features to each edge computing node according to the second classification benchmark through a deep reinforcement learning algorithm;

[0040] Flood level analysis module: used to pass the first benchmark features through a flood level classification model to obtain a third classification benchmark;

[0041] Classification decision module: based on the third classification benchmark, establish a classification warning trigger mechanism and feedback it to the classification decision terminal;

[0042] The system operation database includes all data texts of a data-driven flood classification system and real-time collects the information texts output by each module. The system central processing module is used to control the information text instructions output by each module in the system, and the classification decision terminal is an information output device that receives the data-driven flood classification system.

[0043] Preferably, for the classification decision module, when the third classification benchmark reaches a preset warning trigger threshold, trigger the corresponding warning level;

[0044] The preset warning trigger threshold constructs a threshold matrix using the quantile regression algorithm based on historical flood events in the system operation database, and calculates the eigenvalue distribution interval corresponding to each warning level as the warning trigger threshold.

[0045] Technical effects and advantages of the present invention:

[0046] 1. Through the ability of time series modeling of load data and combined with the dynamic allocation mechanism, the present invention realizes the real-time matching of computing resource requirements and node processing capabilities, effectively eliminates the processing delay problem caused by resource overload of edge devices in high-concurrency areas, and reduces the measured task response delay;

[0047] 2. Through the dual-channel architecture and combined with the dynamic weight averaging algorithm based on KL divergence, the present invention realizes the collaborative optimization of spatial features and time series features, and improves the feature space coverage rate in flood events exceeding historical extremes;

[0048] 3. By establishing a confidence-driven incremental learning mechanism, the present invention improves the recognition accuracy of new disaster patterns. Brief Description of the Drawings

[0049] Figure 1 It is a flowchart of the steps of a data-driven flood classification method provided according to an embodiment of the present application.

[0050] Figure 2 It is a structural block diagram of a data-driven flood classification system provided according to an embodiment of the present application.

[0051] Figure 3 It is a flowchart of obtaining the third classification benchmark through a flood level classification model in a data-driven flood classification method provided according to an embodiment of the present application. Detailed Embodiments

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0053] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular forms "a", "an", "the", "above-mentioned", "said", and "this" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.

[0054] Hereinafter, the terms "first", "second", and "third" are only used for descriptive purposes and should not be construed as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first", "second", and "third" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0055] As shown in the Figure 1 accompanying drawings, a data-driven flood grading method includes obtaining a first grading benchmark through a multi-source heterogeneous data interface and extracting first benchmark features using a convolutional neural network; simultaneously obtaining the load pressure values of each edge computing node in real time as a second grading benchmark and dynamically allocating extraction tasks to each edge computing node; inputting the first benchmark features into a flood level classification model to obtain a third grading benchmark so as to achieve flood grading. The specific implementation steps are as follows:

[0056] S1: Obtain a first grading benchmark through a multi-source heterogeneous data interface, and the first grading benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media public opinion data;

[0057] S2: Perform a feature extraction task on the first grading benchmark through a convolutional neural network to generate spatio-temporally synchronized first benchmark features;

[0058] S3: Obtain the second grading benchmark of each edge computing node in real time, and the second grading benchmark is the load pressure value of each edge computing node predicted based on a deep learning model;

[0059] S4: Dynamically allocate the extraction task of the first benchmark features to each edge computing node according to the second grading benchmark through a deep reinforcement learning algorithm;

[0060] S5: Pass the first benchmark features through a flood level classification model to obtain a third grading benchmark;

[0061] S6: Establish a grading warning trigger mechanism based on the third grading benchmark and feedback it to the grading decision terminal.

[0062] Specifically, in S1, satellite remote sensing data, ground sensor monitoring data, and social media public opinion data transmitted by multi-source heterogeneous data interfaces are collected, and a verification processing operation is performed. The verification processing operation is used to preprocess the data transmitted by the multi-source heterogeneous data interfaces and transmit the verified data.

[0063] Furthermore, the multi-source heterogeneous data interfaces include, but are not limited to, satellite remote sensing data interfaces, message queue interfaces, and file transfer interfaces; the satellite remote sensing data obtained through the multi-source heterogeneous data interfaces includes, but is not limited to, optical remote sensing images, radar remote sensing images, etc., the ground sensor monitoring data includes, but is not limited to, water level information, flow range, meteorological data, etc., and the social media public opinion data includes flood-related text, pictures, and video information with geographical tags in multiple social platforms; the verification processing operation includes, but is not limited to, topological correction, denoising, detecting abnormal or missing values, etc.

[0064] In this embodiment, the WMS / WCS protocol based on the OGC standard is adopted to obtain microwave image data and optical image data with a spatial resolution of up to 10 meters at a fixed period of 15 minutes; at the same time, the Internet of Things terminals deployed in the pre-set flood control key areas are connected through the Message Queuing Telemetry Transport protocol to receive encrypted data packets of key environmental parameters in real time, including but not limited to accurate water level information with a measurement accuracy of ±0.5 cm and flow velocity information with a measurement range covering 0.1 to 5 m / s. The sampling frequency is dynamically adjusted in the range of 1 to 5 minutes according to the flood season; the social media public opinion data uses the Scrapy framework to build a distributed crawler cluster, and real-time collects flood-related text and picture data with geographical tags in multiple social platforms through a semantic filtering algorithm.

[0065] In this embodiment, for satellite remote sensing data, accurate projection conversion is performed using WGS84, and the Douglas-Peucker algorithm is used to perform topological structure correction on the vector map boundary to eliminate geometric errors; for ground sensor monitoring data, the Kalman filtering algorithm is applied to eliminate the interference noise introduced by environmental factors; for social media public opinion data, the BERT-NER technology is used to extract entity information related to the affected locations from the text, and then the geographical location described in the text is converted into standard longitude and latitude coordinates by combining with the geocoding application programming interface service.

[0066] Specifically, in S2, the convolutional neural network adopts a depthwise separable convolution structure. When performing feature extraction tasks, it captures multi-scale features in the first hierarchical benchmark by setting convolutional kernels of different scales, and adds a batch normalization layer and an activation function layer after the convolutional layer to improve the stability of feature extraction and the non-linear expression ability, generating spatio-temporally synchronized first benchmark features. The first benchmark features include a spatial feature vector group, a temporal feature matrix, and multi-source fusion features. The spatial feature vector group consists of spectral texture features, terrain gradient features, and hydrological feature maps.

[0067] In this embodiment, the spatial feature vector group is composed of 128-dimensional spectral texture features extracted by Sentinel-2, 64-dimensional terrain gradient features generated by processing with the Sobel operator, and 256-dimensional hydrological feature maps formed by ground sensors through a spatial interpolation field; the temporal feature matrix constructs a time series tensor based on a 72-hour sliding window, with a time resolution of 10 minutes and including dynamic indicators of the change rate of hydrological parameters and the cumulative amount of meteorological elements, and generates 32-dimensional temporal feature vectors through encoding by an LSTM time series encoder; the multi-source fusion features integrate a social media public opinion heat map based on the spatial distribution of geographical tags TF-IDF, a cross-modal attention weight matrix representing the correlation between satellite, sensor, and public opinion features, and a spatial alignment residual map composed of differential features after multi-source data registration.

[0068] The spatio-temporally synchronized first benchmark features generated in this embodiment include: performing multi-scale feature extraction using a depthwise separable convolution architecture, configuring three groups of parallel extraction channels composed of large-scale channels, medium-scale channels, and micro-scale channels. The large-scale channels use a 9×9 convolutional kernel with a dilation rate of 2 to capture basin-level hydrological features, the medium-scale channels use a 5×5 convolutional kernel with a dilation rate of 1 to extract river channel morphology features, and the micro-scale channels use a 3×3 convolutional kernel and a dilation rate of 1 to focus on local water level mutation features; the outputs of each channel are processed by a batch normalization layer, and the activation function selects an improved ELU.

[0069] Specifically, in S3, when obtaining the second hierarchical benchmark of each edge computing node in real time, the deep learning model is a long short-term memory network. By learning and predicting data such as the historical load data, the current task queue length, and the network bandwidth of the edge computing node, the load pressure value of each edge computing node is obtained.

[0070] In this embodiment, a four-layer bidirectional LSTM network topology is designed. The input gate uses the Sigmoid function to control the memory strength of time series features, the forget gate introduces an adaptive decay factor to optimize long-term dependencies, and the output gate combines Peephole connections to enhance state visibility; the output layer is mapped to the load pressure value interval [0, 1] through a fully connected layer and uses the Tanh activation function to achieve non-linear normalization.

[0071] Specifically, in S4, the deep reinforcement learning algorithm adopts the Proximal Policy Optimization (PPO) algorithm. With the goal of maximizing the overall processing efficiency of edge computing nodes and minimizing task processing latency, it dynamically allocates the extraction tasks of the first benchmark features to each edge computing node according to the second classification benchmark.

[0072] Furthermore, the implementation method of the deep reinforcement learning algorithm includes:

[0073] Decompose the second classification benchmark into a multi-dimensional state vector, where the multi-dimensional state vector includes node hardware state, network state, and task characteristics. Among them, the node hardware state is a physical indicator reflecting the load pressure, the network state is an indicator for evaluating the data transmission efficiency between nodes, and the task characteristics are used to quantify the degree of demand for node resources by the task; in this embodiment, the node hardware state includes CPU utilization rate, memory occupancy rate, and GPU video memory usage, which are collected at intervals of 50 ms; the network state includes bandwidth occupancy rate and round-trip delay; the task characteristics include the amount of data to be processed, the complexity of feature extraction, and the node load pressure value, which quantify the degree of demand for node resources by the task;

[0074] Based on the multi-dimensional state vector, with the goal of maximizing the overall processing efficiency of edge computing nodes and minimizing task processing latency, generate a dynamic allocation policy through discrete actions and continuous actions. Among them, the discrete action is to output a probability distribution according to the load pressure value and output optional edge nodes, and the continuous actions include proportionally splitting the feature extraction task to adjacent low-load nodes for high-load nodes and preferentially allocating tasks with high timeliness requirements to nodes with moderate load pressure and stable network; in this embodiment, set the high-load node as the edge computing node with a load pressure value greater than 0.7, and set the low-load node as the edge computing node with a load pressure value less than 0.3.

[0075] Specifically, in S5, the flood level classification model is a pre-constructed learning model. By inputting the first benchmark features into the flood level classification model, the flood level classification model obtains the third classification benchmark according to the first benchmark features.

[0076] Furthermore, the steps to obtain the third classification benchmark through the flood level classification model are as follows:

[0077] S501: Integrate the first benchmark features through the federated learning framework and combine with the hydrodynamic model to construct a three-dimensional flood evolution to obtain a flood classification feature set, where the flood classification feature set is the training data set input into the flood level classification model;

[0078] Even further, the implementation steps of S501 include:

[0079] S501-1: Aggregate the first benchmark features of each edge computing node and the high-precision digital elevation model based on the differential privacy federated learning framework;

[0080] In this embodiment, the ε-δ differential privacy federated learning framework is adopted to aggregate the first benchmark features and the high-precision digital elevation model distributed in each edge computing node, and the privacy parameters are set as ε = 0.5 and δ = 10 -5 ; Deploy a local differential privacy processor on each edge computing node, apply differential privacy technology to the obtained first benchmark features, inject Gaussian noise with a standard deviation of σ = 0.3 and perform gradient clipping, and collect and fuse the encrypted gradient parameters from all edge computing nodes to form a feature data basis, which serves as the input of the hydrodynamic model;

[0081] S501-2: Couple the hydrodynamic model with multi-spectral remote sensing data for three-dimensional dynamic flood evolution;

[0082] In this embodiment, the aggregated feature data in S501-1 is imported into the hydrodynamic simulation environment, HEC-RASv6.3 is selected as the core hydrodynamic model, and multi-spectral remote sensing data is coupled to construct three-dimensional dynamic flood evolution, specifically including:

[0083] In the terrain data layer of the three-dimensional dynamic flood evolution, fuse the basic DEM data with a resolution of 0.5 meters and the Sentinel-2 multi-spectral image with a resolution of 10 meters, and apply the enhanced super-resolution generative adversarial network algorithm to improve the effective accuracy of the terrain data to 0.3 meters to support more refined flood simulation;

[0084] In the hydraulic parameter layer of the three-dimensional dynamic flood evolution, implement a dynamic adjustment mechanism for the Manning roughness coefficient, which is adaptively adjusted according to the vegetation coverage retrieved from the remote sensing image to more accurately simulate the flow resistance;

[0085] In the boundary condition layer of the three-dimensional dynamic flood evolution, use the long short-term memory network model to predict the spatial distribution of rainfall in the next 6 hours, and use the prediction result as the key input condition to drive the operation of the hydrodynamic model;

[0086] S501-3: Generate a spatio-temporal marker training set by backtracking the historical data in the system operation database;

[0087] In this embodiment, the three-dimensional dynamic flood evolution in S501-2 is utilized to perform the retrospective simulation of historical flood events, and extreme flood events are selected for refined simulation to generate a spatio-temporal labeled dataset. The spatio-temporal labeled dataset includes, but is not limited to, spatial features, temporal features, and composite features. Among them, the spatial features include the inundation depth with an inversion verification error less than 5 cm through lidar point cloud data, and the flow velocity field distribution verified by particle image velocimetry technology; the temporal features include the flood peak propagation rate calculated based on timestamp difference, and the recession curve describing the water accumulation process; the composite features include the dike stress index obtained through finite element analysis, and the population exposure evaluated by integrating mobile phone signaling data; the obtained spatio-temporal labeled dataset constitutes the flood classification feature set required for this embodiment and serves as the training data for the flood level classification model.

[0088] S502: Use a dual-channel neural network to learn the flood classification feature set and adjust the parameters of the flood level classification model through the dynamic weight averaging algorithm;

[0089] Furthermore, the implementation steps of S502 include:

[0090] S502-1: Construct multiple dual-channel feature extractors to process various types of flood feature data in parallel;

[0091] The dual-channel feature extractor in this embodiment includes a main channel and an auxiliary channel. The main channel focuses on processing the spatial feature distribution, adopts a pre-trained ResNet-101 network structure, receives raster data of the flood inundation range with 256×256 pixels as input, and effectively extracts spatial pattern information through deep residual blocks, outputting a 1024-dimensional feature vector; the auxiliary channel is responsible for extracting the temporal evolution law, adopts a bidirectional long short-term memory network, with the hidden layer set to 128 units, inputs time series data, and outputs a 64-dimensional temporal evolution feature vector;

[0092] S502-2: Aggregate multiple dual-channel feature extractors using the dynamic weight averaging algorithm and dynamically adjust the aggregation weights according to the KL divergence;

[0093] In this embodiment, dynamically adjusting the aggregation weights according to the consistency between the dual-channel feature extractors includes: calculating the KL divergence between the model parameters of each dual-channel feature extractor. If the parameter divergence is greater than the preset divergence threshold, trigger the weight calibration mechanism;

[0094] Further, adjusting the aggregation weights according to the KL divergence is specifically expressed as:

[0095]

[0096] where, w iis denoted as the aggregation weight of the i-th dual-channel feature extractor, β is denoted as a preset sensitivity parameter, D KL is denoted as the KL divergence, θ i and θ j are respectively denoted as the model parameters of the i-th dual-channel feature extractor and the model parameters of the j-th dual-channel feature extractor, θ g is denoted as the average parameter of all feature extractors, N is denoted as the total number of dual-channel feature extractors, i and j respectively denote the indices of the dual-channel feature extractors; in this embodiment, β is set to 0.8, and when the divergence threshold is greater than 0.15, the weight calibration mechanism is triggered, where D KL (θ i ||θ g ) is used to measure the distribution difference between the model parameters of the i-th dual-channel feature extractor and the average parameters of all feature extractors, and is specifically expressed as:

[0097]

[0098] where x is denoted as the model parameter point of the dual-channel feature extractor, θ i (x) is denoted as the probability density of the i-th dual-channel feature extractor at the parameter point x, θ g (x) is denoted as the probability density of the average parameters of all feature extractors at the parameter point x;

[0099] It should be noted that the larger the sensitivity parameter β, the more significant the impact of the KL divergence difference on the weight. The weights of the dual-channel feature extractors with small model parameter differences are higher, and the weights of the dual-channel feature extractors with large differences are suppressed. The smaller the sensitivity parameter β, the more uniform the weight distribution, and the weaker the impact of the difference on the result;

[0100] S503: Incrementally optimize the classification layer of the flood level classification model;

[0101] Furthermore, the implementation steps of S503 include: embedding a Monte Carlo Dropout layer in the classification layer of the flood level classification model for uncertainty quantification, and setting trigger conditions based on confidence and pattern matching to adjust the flood level classification model and data augmentation;

[0102] In this embodiment, the dropout probability is set to 0.5 and 50 random forward propagations are performed. The confidence index is defined as 1, and when the confidence is lower than 0.75 for 3 consecutive times and the cosine similarity is less than 0.6, the model is updated;

[0103] S504: Construct the disaster chain topological relationship of the target basin and feedback it to the flood level classification model;

[0104] Furthermore, the topological relationship of the disaster chain in the target basin includes disaster nodes and edge weights calculated based on the Pearson correlation coefficient;

[0105] In this embodiment, a topological graph of 12 types of disaster nodes is constructed. The 12 types of disaster nodes include heavy rain, flood, landslide, debris flow, levee breach, urban waterlogging, farmland waterlogging, mountain flood, storm surge, reservoir dam break, river channel siltation, and lightning strike. Strong associations are established for edge weights based on a Pearson correlation coefficient greater than 0.6;

[0106] S505: Construct a multi-level verification system and perform real-time update and correction on the flood level classification model;

[0107] In this embodiment, the multi-level verification system includes, but is not limited to, performing concurrent simulations at 10 6 times per second on the NVIDIA DGX A100 cluster with a latency of less than 50 ms, retrospectively verifying major flood events in the past 20 years for historical disaster cases with an F1 score requirement of greater than 0.87, simulating 72-hour extreme rainfall events, etc.

[0108] Specifically, in S6, the hierarchical early warning trigger mechanism constructs a four-color early warning system based on the third classification benchmark. When the third classification benchmark reaches the preset early warning trigger threshold, the corresponding early warning level is triggered. The early warning levels include blue early warning, yellow early warning, orange early warning, and red early warning. And the early warning information is fed back to the hierarchical decision-making terminal through various methods such as text messages, APP push, and radio.

[0109] Furthermore, the preset early warning trigger threshold constructs a threshold matrix using the quantile regression algorithm based on historical flood events in the system operation database. In this embodiment, by analyzing the spatio-temporal feature matrix of 426 flood events in the past 30 years, the eigenvalue distribution intervals corresponding to each early warning level are calculated as the early warning trigger threshold. Among them, the blue early warning threshold is the P25 quantile value, the yellow early warning threshold is P75, the orange early warning threshold is P95, and the red early warning threshold is P99.9. The threshold matrix is compared and updated in real time using a sliding window mechanism. Among them, the blue and yellow early warnings use a 15-minute rolling window for feature trend analysis, and the orange and red early warnings enable a 5-minute adaptive window to achieve rapid response; it should be noted that the window sliding step size is dynamically adjusted according to the real-time rainfall intensity, and the step size is compressed to 1 minute during a red rainstorm warning.

[0110] As Figure 2 shown, a data-driven flood level classification system includes a system operation database, a system central processing module, and a hierarchical decision-making terminal, and further includes:

[0111] Multi-source data acquisition module: used to obtain the first classification benchmark through a multi-source heterogeneous data interface, and the first classification benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media public opinion data;

[0112] Feature extraction module: used to perform feature extraction tasks on the first classification benchmark through a convolutional neural network to generate spatiotemporally synchronized first benchmark features;

[0113] Computing load acquisition module: used to obtain the second classification benchmark of each edge computing node in real time, and the second classification benchmark is the load pressure value predicted for each edge computing node based on a deep learning model;

[0114] Load balancing module: used to dynamically allocate the extraction tasks of the first benchmark features to each edge computing node according to the second classification benchmark through a deep reinforcement learning algorithm;

[0115] Flood level analysis module: used to pass the first benchmark features through a flood level classification model to obtain the third classification benchmark;

[0116] Classification decision module: based on the third classification benchmark, establish a classification warning trigger mechanism and feedback it to the classification decision terminal;

[0117] The system operation database includes all data texts of a data-driven flood classification system, and real-time collects the information texts output by each module. The system central processing module is used to control the information text instructions output by each module in the system, and the classification decision terminal is an information output device that receives the data-driven flood classification system.

[0118] In this embodiment, when the third classification benchmark reaches the preset warning trigger threshold, the corresponding warning level is triggered; the preset warning trigger threshold constructs a threshold matrix using the quantile regression algorithm based on historical flood events in the system operation database, and calculates the eigenvalue distribution interval corresponding to each warning level as the warning trigger threshold.

[0119] Secondly: In the accompanying drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved. For other structures, reference can be made to the usual designs. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other;

[0120] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A data-driven flood classification method, characterized in that Including: S1: Obtain the first classification benchmark through a multi-source heterogeneous data interface. The first classification benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media public opinion data; S2: Perform feature extraction tasks on the first classification benchmark through a convolutional neural network to generate the first benchmark features that are spatio-temporally synchronized; S3: Real-time obtain the second classification benchmark of each edge computing node. The second classification benchmark is the load pressure value predicted for each edge computing node based on a deep learning model; S4: According to the second classification benchmark, dynamically allocate the extraction tasks of the first benchmark features to each edge computing node through a deep reinforcement learning algorithm; S5: Pass the first benchmark features through a flood level classification model to obtain the third classification benchmark; S6: Establish a classification early warning trigger mechanism based on the third classification benchmark and feedback it to the classification decision terminal.

2. The data-driven flood classification method according to claim 1, wherein: In S2, the first benchmark features include a spatial feature vector group, a time feature matrix, and multi-source fusion features. Among them, the spatial feature vector group consists of spectral texture features, terrain gradient features, and hydrological feature maps.

3. A data-driven flood classification method according to claim 1, wherein: In S4, the implementation method of the deep reinforcement learning algorithm specifically includes: Decompose the second classification benchmark into a multi-dimensional state vector. The multi-dimensional state vector includes node hardware status, network status, and task characteristics. Among them, the node hardware status is a physical indicator reflecting the load pressure, the network status is an indicator for evaluating the data transmission efficiency between nodes, and the task characteristics are indicators for quantifying the demand of tasks for node resources; Based on the multi-dimensional state vector, with the goal of maximizing the overall processing efficiency of edge computing nodes and minimizing task processing latency, generate a dynamic allocation strategy through discrete actions and continuous actions. Among them, the discrete action is to output a probability distribution according to the load pressure value and output optional edge nodes, and the continuous actions include proportionally splitting the feature extraction tasks to adjacent low-load nodes for high-load nodes and preferentially allocating tasks with high timeliness requirements to nodes with moderate load pressure and stable network.

4. A data-driven flood classification method according to claim 1, characterized in that: In S5, obtaining the third classification benchmark through the flood level classification model specifically includes: S501: Fuse the first benchmark features through a federated learning framework and combine with a hydrodynamic model to construct a three-dimensional flood evolution to obtain a flood classification feature set. The flood classification feature set is the training data set input into the flood level classification model; S502: Use a two-channel neural network to learn the flood classification feature set and adjust the parameters of the flood level classification model through a dynamic weight averaging algorithm; S503: Incrementally optimize the classification layer of the flood level classification model; S504: Construct the disaster chain topological relationship of the target basin and feedback it to the flood level classification model; S505: Construct a multi-level verification system and perform real-time update and correction on the flood level classification model.

5. A data-driven flood classification method according to claim 2, characterized in that: In S5, fusing the first benchmark features through a federated learning framework and combining with a hydrodynamic model to construct a three-dimensional flood evolution to obtain a flood classification feature set specifically includes: Based on the differential privacy federated learning framework, aggregate the first benchmark features of each edge computing node and a high-precision digital elevation model; Couple the hydrodynamic model with multi-spectral remote sensing data to perform three-dimensional dynamic flood evolution; Generate a spatio-temporal marker training set by backtracking historical data in the system operation database.

6. A data-driven flood classification method according to claim 2, characterized in that: In step S5, a dual-channel neural network is used to learn the flood classification feature set, and the parameters of the flood level classification model are adjusted by the dynamic weight averaging algorithm. Specifically, it includes: Construct multiple dual-channel feature extractors to process various types of flood feature data in parallel; Use the dynamic weight averaging algorithm to aggregate multiple dual-channel feature extractors and adjust the aggregation weights according to the KL divergence.

7. A data-driven flood classification method according to claim 6, characterized in that: In step S5, adjusting the aggregation weights according to the KL divergence is specifically expressed as: Among them, w i represents the aggregation weight of the i-th dual-channel feature extractor, β represents a preset sensitivity parameter, D KL represents the KL divergence, θ i and θ j represent the model parameters of the i-th dual-channel feature extractor and the model parameters of the j-th dual-channel feature extractor respectively, θ g represents the average parameter of all feature extractors, N represents the total number of dual-channel feature extractors, i and j represent the indices of the dual-channel feature extractors respectively, D KL (θ i ||θ g ) is used to measure the distribution difference between the model parameters of the i-th dual-channel feature extractor and the average parameters of all feature extractors.

8. A data-driven flood classification method according to claim 2, characterized in that: In step S5, the disaster chain topological relationship of the target basin includes disaster nodes and edge weights calculated based on the Pearson correlation coefficient.

9. A data-driven flood classification system, comprising a system operation database, a system central processing module and a classification decision terminal, applicable to the data-driven flood classification method according to any one of the above claims 1-8, characterized in that, It also includes: Multi-source data acquisition module: used to obtain the first classification benchmark through a multi-source heterogeneous data interface. The first classification benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media public opinion data; Feature extraction module: used to perform feature extraction tasks on the first classification benchmark through a convolutional neural network to generate spatio-temporally synchronized first benchmark features; Computing load acquisition module: used to obtain the second classification benchmark of each edge computing node in real time. The second classification benchmark is the load pressure value predicted for each edge computing node based on a deep learning model; Load balancing module: used to dynamically allocate the extraction tasks of the first benchmark features to each edge computing node according to the second classification benchmark through a deep reinforcement learning algorithm; Flood level analysis module: used to pass the first benchmark features through a flood level classification model to obtain a third classification benchmark; Classification decision-making module: establish a classification warning trigger mechanism based on the third classification benchmark and feedback it to the classification decision-making terminal; The system operation database includes all data texts of a data-driven flood classification system and collects the information texts output by each module in real time. The system central processing module is used to control the information text instructions output by each module in the system. The classification decision-making terminal is an information output device that receives the outputs of a data-driven flood classification system.

10. A data-driven flood classification system according to claim 9, characterized in that: In the classification decision-making module, when the third classification benchmark reaches the preset warning trigger threshold, the corresponding warning level is triggered; The preset warning trigger threshold constructs a threshold matrix using the quantile regression algorithm based on historical flood events in the system operation database, and calculates the eigenvalue distribution interval corresponding to each warning level as the warning trigger threshold.

Images