A data-driven flood staging method and system

Abstract

The application discloses a data-driven flood grading method and system, and particularly relates to the field of flood grading, which comprises the following steps: acquiring a first grading benchmark through a multi-source heterogeneous data interface, and extracting first benchmark features by using a convolutional neural network; simultaneously, acquiring load pressure values of each edge computing node in real time as a second grading benchmark, and dynamically distributing extraction tasks to each edge computing node; inputting the first benchmark features into a flood grade classification model to acquire a third grading benchmark, so as to achieve flood grading. The data-driven flood grading method and system realize real-time matching of the computing resource demand and the node processing capability by the time series modeling capability of the load data in combination with a dynamic distribution mechanism, reduce the measured task response delay, realize the collaborative optimization of the spatial features and the time series features by a double-channel architecture, and improve the new disaster mode recognition accuracy by establishing a confidence-driven incremental learning mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flood grading, more particularly, the present application relates to a data-driven flood grading method and system. BACKGROUND

[0002] With the development of water conservancy projects and intelligent monitoring technology, it is particularly important to build a precise and efficient flood warning grading system. The traditional flood grading adopts a three-layer architecture of sensor network-data center-manual analysis. The water level meter, rain gauge and other terminal devices are used to collect hydrological data in real time. The data is transmitted to the central processing unit through the communication network, and then the professional personnel determines the risk level according to the historical experience and statistical model. However, this scheme has significant defects, including strong dependence on manual experience leading to subjective bias, insufficient hardware sensor layout density causing monitoring blind area, and static threshold unable to dynamically reflect the comprehensive risk of the river basin.

[0003] To overcome the static defects of the traditional method, the current direction is data-driven. A typical improved scheme uses a machine learning model to replace manual rules. A grading architecture based on a random forest algorithm adds a weather radar and satellite remote sensing data interface. A neural network model is introduced to build a mapping relationship between multi-dimensional features and flood levels. The adaptability of the model is improved by dynamically updating the training set. To overcome the time delay defect of the traditional architecture, a distributed computing framework and machine learning algorithm are introduced for optimization. The improved system uses an edge computing node and a cloud-based collaborative processing architecture. An embedded intelligent terminal is deployed at the front end to preprocess the original data. Convolutional neural networks are used to extract real-time feature parameters such as water level rise and flow rate sudden change. The federated learning mechanism is used to update the grading model in the cloud.

[0004] However, in actual use, there are still some shortcomings, such as the lack of dynamic load balancing mechanism between nodes, which causes the edge devices in some high-concurrency areas to have processing delays due to overloaded computing resources. The model training relies on the completeness of historical disaster data. When extreme weather events beyond historical extremes occur, the feature space coverage is insufficient, leading to grading failure. SUMMARY

[0005] To overcome the above-mentioned defects of the prior art, the present application provides a data-driven flood grading method and system, which solves the problems raised in the background art by the following scheme.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

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

[0008] S1: Obtain a first grading reference through a multi-source heterogeneous data interface. The first grading reference includes satellite remote sensing data, ground sensor monitoring data and social media public opinion data.

[0009] S2: performing a feature extraction task on the first hierarchical reference through a convolutional neural network to generate first reference features in space-time synchronization;

[0010] S3: obtaining a second hierarchical reference of each edge computing node in real time, the second hierarchical reference being a load pressure value of each edge computing node predicted based on a deep learning model;

[0011] S4: dynamically allocating the extraction task of the first reference features to each edge computing node through a deep reinforcement learning algorithm according to the second hierarchical reference;

[0012] S5: classifying the first reference features through a flood level classification model to obtain a third hierarchical reference;

[0013] S6: establishing a hierarchical early warning triggering mechanism based on the third hierarchical reference and feeding back to a hierarchical decision terminal.

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

[0015] Preferably, in the S4, the deep reinforcement learning algorithm includes the following embodiments:

[0016] The second hierarchical reference is decomposed into a multi-dimensional state vector, the multi-dimensional state vector including a node hardware state, a network state, and a task feature, wherein the node hardware state is a physical index reflecting the load pressure, the network state is an index evaluating the data transmission efficiency between nodes, and the task feature is a quantitative demand degree of the task for node resources;

[0017] Based on the multi-dimensional state vector, a dynamic allocation strategy is generated through discrete actions and continuous actions, with the goal of maximizing the overall processing efficiency of the edge computing nodes and minimizing the task processing delay, wherein the discrete action is a probability distribution output according to the load pressure value, outputting the selected edge node, and the continuous action includes proportionally dividing the feature extraction task to adjacent low-load nodes for the high-load node and preferentially allocating the task with high time efficiency requirement to the node with moderate load pressure and stable network.

[0018] Preferably, in the S5, the third hierarchical reference is obtained through the flood level classification model, and specifically includes the following steps:

[0019] S501: fusing the first reference features through a federated learning framework, and combining a hydrodynamic model to construct a three-dimensional flood evolution to obtain a flood classification feature set, the flood classification feature set being a training data set input to the flood level classification model;

[0020] S502: learning the flood classification feature set by using a double-channel neural network, and adjusting the flood grade classification model parameters by a dynamic weight average algorithm;

[0021] S503: incrementally optimizing the classification layer of the flood grade classification model;

[0022] S504: constructing a disaster chain topological relationship of the target basin and feeding back to the flood grade classification model;

[0023] S505: constructing a multi-level verification system and updating and correcting the flood grade classification model in real time.

[0024] Preferably, in the S5, the first benchmark features are fused by using a federal learning framework, and a three-dimensional flood evolution is constructed by combining a water power model to obtain a flood classification feature set, specifically including:

[0025] Based on a differential privacy federal learning framework, the first benchmark features of each edge computing node are aggregated with a high-precision digital elevation model;

[0026] The water power model is coupled with multispectral remote sensing data to perform three-dimensional dynamic flood evolution;

[0027] The spatiotemporal label training set is generated by backtracking the historical data in the system operation database.

[0028] Preferably, in the S5, the double-channel neural network is used to learn the flood classification feature set, and the dynamic weight average algorithm is used to adjust the flood grade classification model parameters, specifically including:

[0029] A plurality of double-channel feature extractors are constructed to process multiple types of flood feature data in parallel;

[0030] The dynamic weight average algorithm is used to aggregate the plurality of double-channel feature extractors, and the aggregation weight is adjusted according to the KL divergence.

[0031] Preferably, in the S5, the aggregation weight is adjusted according to the KL divergence, and is specifically represented as:

[0032]

[0033] wherein w i represents the aggregation weight of the i-th double-channel feature extractor, β represents a preset sensitivity parameter, D KL represents the KL divergence, θ i and θ j respectively represent the model parameters of the i-th double-channel feature extractor and the model parameters of the j-th double-channel feature extractor, and θ gN represents the total number of two-channel feature extractors, i and j represent the indexes of two-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 two-channel feature extractor and the average parameters of all feature extractors.

[0034] Preferably, 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 purpose, the present application provides the following technical scheme: a data-driven flood grading system, including a system operation database, a system central processing module and a grading decision terminal, implementing the above-mentioned data-driven flood grading method, further comprising:

[0036] A multi-source data acquisition module is used to obtain a first grading reference through a multi-source heterogeneous data interface, and the first grading reference includes satellite remote sensing data, ground sensor monitoring data and social media public opinion data.

[0037] A feature extraction module is used to perform a feature extraction task on the first grading reference through a convolutional neural network to generate a first reference feature with time and space synchronization.

[0038] A computing load acquisition module is used to acquire a second grading reference of each edge computing node in real time, and the second grading reference is a load pressure value of each edge computing node predicted based on a deep learning model.

[0039] A load balancing module is used to dynamically allocate the extraction task of the first reference feature to each edge computing node through a deep reinforcement learning algorithm according to the second grading reference.

[0040] A flood level analysis module is used to pass the first reference feature through a flood level classification model to obtain a third grading reference.

[0041] A grading decision module establishes a grading early warning triggering mechanism based on the third grading reference and feeds back to the grading decision terminal.

[0042] The system operation database includes all data texts of the data-driven flood grading system, and real-time collects 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 grading decision terminal is an information output device receiving the data-driven flood grading system.

[0043] Preferably, the grading decision module triggers the corresponding early warning level when the third grading reference reaches a preset early warning triggering threshold.

[0044] The preset early warning trigger threshold is constructed by using a quantile regression algorithm based on historical flood events in a system operation database to construct a threshold matrix, and the characteristic value distribution interval corresponding to each early warning level is calculated as the early warning trigger threshold.

[0045] The technical effects and advantages of the present application are as follows:

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

[0047] 2. The present application realizes the cooperative optimization of spatial features and time sequence features by a double-channel architecture combined with a dynamic weight average algorithm based on KL divergence, and improves the feature space coverage in flood events exceeding historical extreme values.

[0048] 3. The present application improves the new disaster mode recognition accuracy by establishing a confidence degree driven incremental learning mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A step flow chart of a data-driven flood grading method according to an embodiment of the present application.

[0050] Figure 2 A structural block diagram of a data-driven flood grading system according to an embodiment of the present application.

[0051] Figure 3 A flow chart of obtaining a third grading reference by a flood grade classification model in a data-driven flood grading method according to an embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this 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 used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0055] As attached Figure 1 The data-driven flood classification method shown includes obtaining a first classification benchmark through a multi-source heterogeneous data interface and extracting the first benchmark features using a convolutional neural network; simultaneously, acquiring the load pressure values ​​of each edge computing node in real time as a second classification 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 classification benchmark, thereby achieving flood classification. The specific implementation steps are as follows:

[0056] S1: Obtain the first-level benchmark through a multi-source heterogeneous data interface. The first-level benchmark includes satellite remote sensing data, ground sensor monitoring data, and social media sentiment data.

[0057] S2: Perform feature extraction on the first hierarchical benchmark through a convolutional neural network to generate spatiotemporally synchronized first benchmark features;

[0058] S3: Real-time acquisition of the second-level benchmark for each edge computing node, which is the load pressure value of each edge computing node predicted based on a deep learning model;

[0059] S4: Based on the second hierarchical benchmark, the extraction task of the first benchmark feature is dynamically assigned to each edge computing node through a deep reinforcement learning algorithm;

[0060] S5: Use the first benchmark feature to obtain the third classification benchmark through the flood level classification model;

[0061] S6: Establish a graded early warning triggering mechanism based on the third graded benchmark and feed it back to the graded decision-making 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 to preprocess the data transmitted by the multi-source heterogeneous data interfaces and transmit the verified data.

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

[0064] In this embodiment, WMS / WCS protocols based on OGC standards are 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, through a message queue telemetry transmission protocol, Internet of Things terminals deployed in the pre-set key flood control areas are connected to receive encrypted data packets of key environmental parameters including, but not limited to, accurate water level information with a measurement accuracy of ±0.5 cm and range flow information with a measurement range covering 0.1 to 5 meters per second, etc., and the sampling frequency is dynamically adjusted to the interval of 1-5 minutes according to the flood season; social media public opinion data adopts a Scrapy framework to build a distributed crawler cluster, and through a semantic filtering algorithm, flood text and picture data containing geographical tags in multiple social platforms are collected in real time.

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

[0066] Specifically, in S2, the convolutional neural network adopts a deep separable convolution structure, and when performing a feature extraction task, different scale convolution kernels are set to capture multi-scale features in the first hierarchical reference, and a batch normalization layer and an activation function layer are added after the convolution layer to improve the stability and non-linear expression ability of feature extraction, generate a first reference feature with space-time synchronization, and the first reference feature includes a spatial feature vector group, a time feature matrix, and multi-source fusion features, wherein the spatial feature vector group is composed 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 from Sentinel-2, 64-dimensional terrain gradient features generated by processing with a Sobel operator, and 256-dimensional hydrological feature maps formed by spatial interpolation fields of ground sensors; the time feature matrix is constructed based on a 72-hour sliding window to form a time series tensor, the time resolution is 10 minutes, and contains dynamic indicators of hydrological parameter change rate and meteorological element cumulative amount, and a 32-dimensional time feature vector is generated by encoding with an LSTM time series encoder; the multi-source fusion features include a social media public opinion heat map based on the spatial distribution of geographic labels 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 difference features after multi-source data registration.

[0068] The first reference feature with space-time synchronization generated in this embodiment includes: multi-scale feature extraction using a deep separable convolution architecture, configuring three groups of parallel extraction channels composed of large-scale channels, medium-scale channels and micro-scale channels, wherein the large-scale channel uses a 9x9 convolution kernel with a hole rate of 2 to capture watershed-level hydrological features, the medium-scale channel uses a 5x5 convolution kernel with a hole rate of 1 to extract river channel shape features, and the micro-scale channel uses a 3x3 convolution kernel with a hole rate of 1 to focus on local water level mutation features; each channel output is processed by a batch normalization layer, and the activation function uses an improved ELU.

[0069] Specifically, in S3, when the second hierarchical reference of each edge computing node is acquired in real time, the deep learning model is a long short-term memory network, which learns and predicts the historical load data, current task queue length and network bandwidth of the edge computing node to obtain the load pressure value of each edge computing node.

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

[0071] Specifically, in S4, the deep reinforcement learning algorithm adopts a proximal policy optimization algorithm to dynamically allocate the extraction tasks of the first benchmark features to each edge computing node according to the second hierarchical benchmark, aiming to maximize the overall processing efficiency of the edge computing nodes and minimize the task processing delay.

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

[0073] The second hierarchical benchmark is decomposed into a multi-dimensional state vector, which includes node hardware state, network state and task feature, wherein the node hardware state is a physical indicator reflecting the load pressure, the network state is an indicator evaluating the data transmission efficiency between nodes, and the task feature is a quantitative indicator of the demand degree of the task for node resources; in this embodiment, the node hardware state includes CPU utilization, memory occupancy and GPU memory usage, which are collected at an interval of 50 ms; the network state includes bandwidth occupancy and round-trip time; the task feature includes the amount of data to be processed, the feature extraction complexity and the node load pressure value, which quantifies the demand degree of the task for node resources.

[0074] Based on the multi-dimensional state vector, a dynamic allocation strategy is generated through discrete actions and continuous actions, aiming to maximize the overall processing efficiency of the edge computing nodes and minimize the task processing delay, wherein the discrete action is a probability distribution output according to the load pressure value, outputting the selected edge node, and the continuous action includes proportionally dividing the feature extraction tasks from the high-load node to the adjacent low-load node and preferentially allocating the high-time-efficiency requirement task to the node with moderate load pressure and stable network; in this embodiment, the high-load node is set as the edge computing node with a load pressure value greater than 0.7, and the low-load node is set 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 feature into the flood level classification model, the flood level classification model obtains the third hierarchical benchmark according to the first benchmark feature.

[0076] Further, the step of obtaining the third hierarchical benchmark through the flood level classification model is as follows:

[0077] S501: Fusion of the first benchmark feature through a federated learning framework, and construction of a three-dimensional flood evolution combined with a hydrodynamic model to obtain a flood classification feature set, which is a training data set input into the flood level classification model;

[0078] Further, the implementation steps of S501 include:

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

[0080] In this embodiment, the ε-δ differential privacy federated learning framework is used to aggregate the first benchmark features and high-precision digital elevation models distributed on each edge computing node, and the privacy parameters ε=0.5 and δ=10 are set. -5 A local differential privacy processor is deployed on each edge computing node to apply differential privacy technology to the obtained first benchmark features, collect and fuse encrypted gradient parameters from all edge computing nodes to form a feature data base by injecting Gaussian noise with a standard deviation σ=0.3 and implementing gradient clipping, and the feature data base is used as the input of the water power model;

[0081] S501-2: Coupling water power model and 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 water power simulation environment, HEC-RASv6.3 is selected as the core water power model, and multi-spectral remote sensing data is coupled to build a three-dimensional dynamic flood evolution, which specifically includes:

[0083] In the terrain data layer of the three-dimensional dynamic flood evolution, the basic 0.5-meter resolution DEM data and 10-meter resolution Sentinel-2 multi-spectral image are fused, and an enhanced super-resolution generative adversarial network algorithm is applied to improve the effective accuracy of the terrain data to 0.3 meters to support more detailed flood simulation;

[0084] In the hydraulic parameter layer of the three-dimensional dynamic flood evolution, a dynamic adjustment mechanism of Manning roughness coefficient is implemented, and adaptive adjustment is performed according to the vegetation coverage retrieved from the remote sensing image to more accurately simulate the water flow resistance;

[0085] In the boundary condition layer of the three-dimensional dynamic flood evolution, the long short-term memory network model is used to predict the spatial distribution of rainfall in the next 6 hours, and the prediction results are used as the key input conditions to drive the water power model to run;

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

[0087] The embodiment utilizes the three-dimensional dynamic flood evolution in S501-2 to perform backtracking simulation of historical flood events, selects extreme flood events for fine simulation, and generates a spatiotemporal labeled data set including but not limited to spatial features, temporal features, and composite features, wherein the spatial features include submerged depths with an error of less than 5 centimeters verified by laser radar point cloud data inversion, and flow velocity field distribution verified by particle image velocimetry; the temporal features include flood peak propagation rate calculated based on timestamp difference, and recession curve describing the process of water accumulation; the composite features include dike stress index derived by finite element analysis, and population exposure degree evaluated by fusion of mobile phone signaling data; the obtained spatiotemporal labeled data set constitutes the flood classification feature set required by the embodiment, and serves as training data of the flood grade classification model.

[0088] S502: learning the flood classification feature set by using a dual-channel neural network, and adjusting parameters of the flood grade classification model by using a dynamic weight average algorithm;

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

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

[0091] The dual-channel feature extractor of the embodiment includes a main channel and an auxiliary channel, wherein the main channel focuses on processing spatial feature distribution, adopts a pre-trained ResNet-101 network structure, receives 256x256 pixel flood submerged range raster data as input, effectively extracts spatial pattern information through a deep residual block, and outputs a 1024-dimensional feature vector; the auxiliary channel is responsible for extracting time evolution law, adopts a bidirectional long short-term memory network, sets the hidden layer to 128 units, inputs time series data, and outputs a 64-dimensional time evolution feature vector;

[0092] S502-2: aggregating multiple dual-channel feature extractors by using a dynamic weight average algorithm, and dynamically adjusting aggregation weights according to KL divergence;

[0093] The embodiment dynamically adjusts the aggregation weights according to the consistency between the dual-channel feature extractors, including: calculating the KL divergence between the model parameters of the dual-channel feature extractors, and triggering a weight calibration mechanism if the parameter divergence is greater than a preset divergence threshold;

[0094] Further, the adjustment of the aggregation weights according to the KL divergence is specifically represented as:

[0095]

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

[0097]

[0098] wherein x denotes a model parameter point of the dual-channel feature extractor, and i (x) denotes the probability density of the i-th dual-channel feature extractor at the parameter point x, and g (x) denotes the probability density of the average parameter of all feature extractors at the parameter point x.

[0099] It should be noted that the greater the sensitivity parameter β, the more significant the influence of the KL divergence difference on the weight, the higher the weight of the dual-channel feature extractor with small model parameter difference, and the weight of the dual-channel feature extractor with large difference is suppressed. The smaller the sensitivity parameter β, the more uniform the weight distribution, and the influence of the difference on the result is weakened.

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

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

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

[0103] S504: Construct the disaster chain topological relationship of the target basin and feed back to the flood grade classification model;

[0104] Furthermore, the disaster chain topology of the target watershed includes disaster nodes and edge weights calculated based on the Pearson correlation coefficient;

[0105] In this embodiment, a topology graph of 12 types of disaster nodes is constructed. The 12 types of disaster nodes include rainstorm, flood, landslide, debris flow, levee breach, urban waterlogging, farmland waterlogging, flash flood, storm surge, reservoir dam failure, river siltation and lightning strike. The edge weights are based on a strong correlation established with a Pearson correlation coefficient greater than 0.6.

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

[0107] In this embodiment, the multi-level verification system includes, but is not limited to, executing 10 on the NVIDIA DGX A100 cluster. 6 The simulations can be performed concurrently for 1000 times per second with a latency of less than 50ms, or backtesting of major flood events in the past 20 years with an F1 score greater than 0.87, or simulating 72-hour extreme rainfall events.

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

[0109] Furthermore, the preset warning trigger threshold is constructed using a quantile regression algorithm based on historical flood events in the system's operating database. In this embodiment, by analyzing the spatiotemporal feature matrix of 426 flood events over the past 30 years, the distribution range of feature values ​​corresponding to each warning level is calculated as the warning trigger threshold. The blue warning threshold is the P25 quantile, the yellow warning threshold is P75, the orange warning threshold is P95, and the red warning threshold is P99.9. The threshold matrix is ​​compared and updated in real time using a sliding window mechanism. Blue and yellow warnings use a 15-minute rolling window for feature trend analysis, while orange and red warnings use a 5-minute adaptive window for 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 attached Figure 2 The data-driven flood classification system shown includes a system operation database, a system central processing module, and a classification decision terminal, and also includes:

[0111] Multi-source data acquisition module: used for acquiring first hierarchical reference through multi-source heterogeneous data interface, the first hierarchical reference including satellite remote sensing data, ground sensor monitoring data and social media public opinion data;

[0112] Feature extraction module: used for performing feature extraction task on the first hierarchical reference through convolutional neural network to generate first reference feature in time and space synchronization;

[0113] Computing load acquisition module: used for acquiring second hierarchical reference of each edge computing node in real time, the second hierarchical reference being load pressure value of each edge computing node predicted based on deep learning model;

[0114] Load balancing module: used for dynamically distributing extraction task of the first reference feature to each edge computing node through deep reinforcement learning algorithm according to the second hierarchical reference;

[0115] Flood level analysis module: used for acquiring third hierarchical reference by classifying the first reference feature through flood level classification model;

[0116] Hierarchical decision module: establishing hierarchical early warning triggering mechanism based on the third hierarchical reference and feeding back to hierarchical decision terminal;

[0117] The system operation database is all data texts of the data-driven flood hierarchical system, and information texts output by each module are collected in real time, the system central processing module is used for controlling information text instructions output by each module in the system, and the hierarchical decision terminal is an information output device receiving the data-driven flood hierarchical system.

[0118] In the embodiment, when the third hierarchical reference reaches a preset early warning triggering threshold, the corresponding early warning level is triggered; the preset early warning triggering threshold is a threshold matrix constructed based on historical flood events in the system operation database by using quantile regression algorithm, and a feature value distribution interval corresponding to each early warning level is calculated as the early warning triggering threshold.

[0119] Secondly, only structures related to the disclosed embodiment are involved in the drawings of the disclosed embodiment, other structures can be referred to the general design, and the same embodiment and different embodiments of the present application can be combined with each other under the condition of no conflict;

[0120] Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A data-driven flood staging method, characterized in that, The method comprises the following steps: S1: obtaining a first hierarchical reference through a multi-source heterogeneous data interface, the first hierarchical reference comprising satellite remote sensing data, ground sensor monitoring data and social media public opinion data; S2: performing a feature extraction task on the first hierarchical reference through a convolutional neural network to generate a first reference feature in time and space synchronization; the first reference feature comprises a spatial feature vector group, a time feature matrix and a multi-source fusion feature, wherein the spatial feature vector group is composed of spectral texture features, terrain gradient features and hydrological feature maps; S3: obtaining a second hierarchical reference of each edge computing node in real time, the second hierarchical reference being a load pressure value of each edge computing node predicted based on a deep learning model; S4: dynamically distributing the extraction task of the first reference feature to each edge computing node according to the second hierarchical reference through a deep reinforcement learning algorithm; the implementation of the deep reinforcement learning algorithm comprises: decomposing the second hierarchical reference into a multi-dimensional state vector, the multi-dimensional state vector comprising a node hardware state, a network state and a task feature, wherein the node hardware state is a physical index reflecting the load pressure, the network state is an index evaluating the data transmission efficiency between nodes, and the task feature is a quantitative index reflecting the demand degree of the task on the node resources; based on the multi-dimensional state vector, generating a dynamic distribution strategy through discrete actions and continuous actions, wherein the discrete action is a probability distribution output according to the load pressure value, and the continuous action comprises proportionally dividing the feature extraction task from a high-load node to adjacent low-load nodes and preferentially distributing a high-time-efficiency requirement task to a node with moderate load pressure and stable network; S5: obtaining a third hierarchical reference by classifying the first reference feature through a flood level classification model, comprising: S501: fusing the first reference feature through a federated learning framework, and combining a water power model to construct a three-dimensional flood evolution to obtain a flood classification feature set, which is a training data set input into the flood level classification model; S502: learning the flood classification feature set through a double-channel neural network, and adjusting the parameters of the flood level classification model through a dynamic weight average algorithm; S503: incrementally optimizing the classification layer of the flood level classification model; S504: constructing a disaster chain topological relationship of a target basin and feeding back to the flood level classification model; S505: constructing a multi-level verification system and updating and correcting the flood level classification model in real time; S6: establishing a hierarchical early warning triggering mechanism based on the third hierarchical reference and feeding back to a hierarchical decision terminal.

2. The data-driven flood staging method of claim 1, wherein: In the S5, the first reference feature is fused through a federated learning framework, and a three-dimensional flood evolution is constructed by combining a water power model to obtain a flood classification feature set, which specifically comprises: aggregating the first reference feature and a high-precision digital elevation model of each edge computing node based on a differential privacy federated learning framework; coupling a water power model and multi-spectral remote sensing data to perform three-dimensional dynamic flood evolution; generating a time and space labeled training set through historical data backtracking in a system running database.

3. The data-driven flood staging method of claim 1, wherein: The S5 adopts a double-channel neural network to learn the flood classification feature set, and adjusts the flood grade classification model parameters through a dynamic weight average algorithm, and specifically includes: A plurality of double-channel feature extractors are constructed to process multiple types of flood feature data in parallel; A dynamic weight average algorithm is used to aggregate the plurality of double-channel feature extractors, and the aggregation weight is adjusted according to the KL divergence.

4. The data-driven flood staging method of claim 3, wherein: The S5 adjusts the aggregation weight according to the KL divergence, which is specifically represented as: ， wherein, denotes the aggregated weights of the th dual-channel feature extractor, denotes a preset sensitivity parameter, denotes a KL divergence, and denotes the model parameters of the th dual-channel feature extractor and the model parameters of the th dual-channel feature extractor, respectively, denotes the average parameters of all feature extractors, denotes the total number of dual-channel feature extractors, and denotes the index of the dual-channel feature extractor, is used to measure the distribution difference between the model parameters of the th dual-channel feature extractor and the average parameters of all feature extractors.

5. The data-driven flood staging method of claim 1, wherein: The S5, the disaster chain topological relationship of the target basin includes disaster nodes and edge weights calculated based on the Pearson correlation coefficient.

6. A data-driven flood staging system comprising a system operation database, a system central processing module and a staging decision terminal, which are suitable for the data-driven flood staging method according to any one of claims 1-5, characterized in that, Also includes: A multi-source data acquisition module is used to obtain a first hierarchical reference through a multi-source heterogeneous data interface, the first hierarchical reference including satellite remote sensing data, ground sensor monitoring data and social media public opinion data; A feature extraction module is used to perform a feature extraction task on the first hierarchical reference through a convolutional neural network to generate a first reference feature that is spatiotemporally synchronized; A computing load acquisition module is used to acquire a second hierarchical reference of each edge computing node in real time, the second hierarchical reference being a load pressure value of each edge computing node predicted based on a deep learning model; A load balancing module is used to dynamically distribute the extraction task of the first reference feature to each edge computing node through a deep reinforcement learning algorithm according to the second hierarchical reference; A flood grade analysis module is used to obtain a third hierarchical reference by passing the first reference feature through a flood grade classification model; A hierarchical decision module establishes a hierarchical early warning triggering mechanism based on the third hierarchical reference and feeds back to a hierarchical decision terminal; The system operation database includes all data texts of the data-driven flood grading system, and real-time collection of 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 hierarchical decision terminal is an information output device receiving the data-driven flood grading system.

7. A data-driven flood staging system according to claim 6, wherein: When the third hierarchical reference reaches a preset early warning triggering threshold, the hierarchical decision module triggers the corresponding early warning level; The preset early warning triggering threshold is based on historical flood events in the system operation database to construct a threshold matrix using a quantile regression algorithm, and the feature value distribution interval corresponding to each early warning level is calculated as the early warning triggering threshold.

Images