A method and system for real-time early warning of watershed floods based on multi-source data fusion

Abstract

This invention relates to a method and system for real-time flood warning in watersheds based on multi-source data fusion, belonging to the field of flood warning technology. The method includes: acquiring multiple multi-source data from a target watershed and preprocessing the data; fusing the preprocessed multi-source data to obtain fused data; and inputting the fused data into a trained real-time flood warning model for processing to obtain flood warning results for the target watershed. The real-time flood warning model is trained based on a CNN-LSTM model. During the training process, a Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model so that the feature flattening layer in the real-time flood warning model flattens the input multi-dimensional features into a one-dimensional vector based on the optimized hyperparameters, thereby improving the accuracy and timeliness of flood warnings.

Description

Technical Field

[0001] This invention relates to the field of flood early warning technology, and in particular to a method and system for real-time early warning of watershed floods based on multi-source data fusion. Background Technology

[0002] Floods are among the most frequent and destructive natural disasters globally, posing a significant threat to people's lives and property, infrastructure, and socio-economic development. Accurate and timely flood warnings are a key non-engineering measure to minimize flood damage.

[0003] In recent years, with the development of artificial intelligence technology, data-driven machine learning methods (such as support vector machines and random forests) have been introduced into the field of flood forecasting. These methods have reduced the dependence on physical parameters to some extent and demonstrated the ability to handle nonlinear problems. However, existing machine learning methods still have the following problems: most models only use single or a few types of hydrological and meteorological data (such as rainfall and water level), failing to fully integrate multi-source data such as meteorological, hydrological, and geographic information, and ignoring the complex spatiotemporal interaction mechanism between "driving forces, states, and constraints" in the flood formation process, thus causing the problem of low accuracy in flood warnings. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method and system for real-time early warning of watershed floods based on multi-source data fusion, which solves the technical problem of low accuracy of flood early warning in the prior art.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] In a first aspect, embodiments of the present invention provide a method for real-time early warning of watershed floods based on multi-source data fusion, comprising: acquiring multiple multi-source data of a target watershed, and preprocessing the multiple multi-source data to obtain multiple preprocessed multi-source data; fusing the multiple preprocessed multi-source data to obtain fused data; and inputting the fused data into a trained real-time early warning model for watershed floods for processing to obtain flood warning results for the target watershed; wherein, the real-time early warning model for watershed floods is trained based on a CNN-LSTM model, and during the training of the CNN-LSTM model, a Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model, and the feature channel weights are adjusted by the Bayesian optimizer to enhance the contribution of key features related to the flood formation mechanism, thereby optimizing the feature quality of the feature flattening layer input to the CNN-LSTM model.

[0009] In one possible embodiment, preprocessing of multiple multi-source data includes at least one of data cleaning, missing value handling, outlier detection, and data standardization.

[0010] In one possible embodiment, the multiple multi-source data includes multi-source hydrological data, multi-source geographic information data, and multi-source meteorological data. The multiple preprocessed multi-source data are fused to obtain fused data, including: evaluating the prediction accuracy of multiple data sources corresponding to the multi-source meteorological data based on historical multi-source hydrological data and historical multi-source meteorological data; dynamically adjusting the weight coefficients of the meteorological data corresponding to each data source based on the prediction accuracy of the multiple data sources; performing a weighted summation of the meteorological data corresponding to each data source and its corresponding weight coefficient to obtain the final meteorological data; and fusing the final meteorological data, multi-source hydrological data, and multi-source geographic information data to obtain the fused data.

[0011] In one possible embodiment, the real-time flood warning model for a watershed includes an input layer, a CNN layer, a pooling layer, a feature flattening layer, an LSTM layer, and a fully connected layer connected in sequence.

[0012] In one possible embodiment, the real-time flood warning model for the watershed is obtained by training a CNN-LSTM model. The training process of the CNN-LSTM model includes: optimizing the hyperparameters of the CNN-LSTM model using a Bayesian optimizer to find the optimal combination of hyperparameters; and constructing and training the CNN-LSTM model based on the optimal combination of hyperparameters to generate a trained real-time flood warning model for the watershed.

[0013] Secondly, embodiments of the present invention provide a real-time flood early warning system for river basins based on multi-source data fusion, comprising:

[0014] The acquisition module is used to acquire multiple sources of data from the target watershed and preprocess the multiple sources of data to obtain multiple preprocessed sources of data.

[0015] The fusion module is used to fuse multiple preprocessed multi-source data to obtain fused data;

[0016] The input module is used to input the fused data into the trained real-time flood warning model for processing, so as to obtain the flood warning results for the target watershed. The real-time flood warning model is trained based on the CNN-LSTM model. During the training of the CNN-LSTM model, the hyperparameters of the CNN-LSTM model are optimized using a Bayesian optimizer, and the feature channel weights are adjusted by the Bayesian optimizer to enhance the contribution of key features related to the flood formation mechanism, thereby optimizing the feature quality of the feature flattening layer input to the CNN-LSTM model.

[0017] In one possible embodiment, preprocessing of multiple multi-source data includes at least one of data cleaning, missing value handling, outlier detection, and data standardization.

[0018] In one possible embodiment, the multiple multi-source data includes multi-source hydrological data, multi-source geographic information data, and multi-source meteorological data; the fusion module is specifically used for: evaluating the prediction accuracy of multiple data sources corresponding to the multi-source meteorological data based on historical multi-source hydrological data and historical multi-source meteorological data; dynamically adjusting the weight coefficient of the meteorological data corresponding to each data source in the multiple data sources based on the prediction accuracy of the multiple data sources; performing a weighted summation calculation on the meteorological data corresponding to each data source and its corresponding weight coefficient to obtain the final meteorological data; and fusing the final meteorological data, multi-source hydrological data, and multi-source geographic information data to obtain the fused data.

[0019] In one possible embodiment, the real-time flood warning model for a watershed includes an input layer, a CNN layer, a pooling layer, a feature flattening layer, an LSTM layer, and a fully connected layer connected in sequence.

[0020] In one possible embodiment, the real-time flood warning model for the basin is obtained by training a CNN-LSTM model; the real-time flood warning system for the basin also includes: a CNN-LSTM model training module, used to: optimize the hyperparameters of the CNN-LSTM model using a Bayesian optimizer to find the optimal hyperparameter combination; and construct and train the CNN-LSTM model based on the optimal hyperparameter combination to generate a trained real-time flood warning model for the basin.

[0021] Thirdly, embodiments of this application also provide an electronic device, which includes: a memory for storing a computer program; and a processor for executing the computer program stored in the memory. When the computer program is executed, the processor is used to execute a method for real-time early warning of watershed floods based on multi-source data fusion as described above.

[0022] Fourthly, embodiments of this application also provide a storage medium including computer instructions that, when executed on an electronic device, cause the electronic device to perform a real-time early warning method for watershed floods based on multi-source data fusion as described above.

[0023] (III) Beneficial Effects

[0024] The beneficial effects of this invention are:

[0025] This application provides a method and system for real-time flood warning in watersheds based on multi-source data fusion. It acquires multiple multi-source data from a target watershed, preprocesses these data to obtain preprocessed multi-source data, fuses the preprocessed data to obtain fused data, and inputs the fused data into a trained real-time flood warning model for further processing to obtain flood warning results for the target watershed. This real-time flood warning model is trained based on a CNN-LSTM model. During the training process, a Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model and adjust the feature channel weights to enhance the contribution of key features related to flood formation mechanisms. This optimizes the feature quality of the feature flattening layer input to the CNN-LSTM model, thereby fully exploring the spatiotemporal characteristics of the multi-source data, improving the accuracy and timeliness of flood warnings, and providing strong support for watershed flood control and disaster reduction.

[0026] To make the above-mentioned objectives, features and advantages to be achieved by the embodiments of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1The flowchart shown is a method for real-time early warning of watershed floods based on multi-source data fusion provided in an embodiment of this application;

[0029] Figure 2 A flowchart illustrating a real-time early warning method for watershed floods based on multi-source data fusion, provided in an embodiment of this application, is shown.

[0030] Figure 3 This paper shows a schematic diagram of the structure of a real-time early warning model for watershed floods provided in an embodiment of this application;

[0031] Figure 4 This illustration shows a parameterizable dynamic adaptive pooling layer provided in an embodiment of this application;

[0032] Figure 5 This illustration shows a training diagram of a real-time flood early warning model for a watershed provided in an embodiment of this application;

[0033] Figure 6 The diagram shows a structural block diagram of a real-time early warning system for watershed floods based on multi-source data fusion, as provided in an embodiment of this application. Detailed Implementation

[0034] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Currently, in addition to the problem of low accuracy in flood early warning, existing technologies also have the following problems:

[0036] The model structure is too simple: ordinary feedforward neural networks or recurrent neural networks are difficult to capture spatial features (such as the spatial distribution of rainfall and topography) and temporal features (such as the time series changes of water level).

[0037] Hyperparameter tuning is difficult: Model performance is highly sensitive to the setting of hyperparameters (such as the number of network layers and the learning rate), and relying on manual experience or grid search for tuning is extremely inefficient, computationally expensive, and difficult to obtain the optimal configuration, thus limiting the upper limit of model performance.

[0038] Based on this, this application provides a method and system for real-time flood warning in watersheds based on multi-source data fusion. The method involves acquiring multiple multi-source data from a target watershed, preprocessing these data to obtain preprocessed multi-source data, fusing the preprocessed data to obtain fused data, and inputting the fused data into a trained real-time flood warning model for further processing to obtain flood warning results for the target watershed. This real-time flood warning model is trained based on a CNN-LSTM model. During the training process, a Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model and adjust the feature channel weights to enhance the contribution of key features related to flood formation mechanisms. This optimizes the feature quality of the feature flattening layer input to the CNN-LSTM model, thereby fully exploring the spatiotemporal characteristics of the multi-source data, improving the accuracy and timeliness of flood warnings, and providing strong support for watershed flood control and disaster reduction.

[0039] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0040] Please see Figure 1 , Figure 1 A flowchart of a real-time watershed flood warning method based on multi-source data fusion, provided in an embodiment of this application, is shown. It should be understood that this real-time watershed flood warning method can be executed by a real-time watershed flood warning system based on multi-source data fusion, and the specific apparatus of this system can be configured according to actual needs; this embodiment is not limited thereto. For example, the real-time watershed flood warning system can be a computer or a server, etc. Specifically, the real-time watershed flood warning method includes:

[0041] Step S110: Obtain multiple sources of data from the target watershed and preprocess the multiple sources of data to obtain multiple preprocessed sources of data.

[0042] It should be understood that the specific multi-source data can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0043] Optionally, the multiple sources of data may include multi-source hydrological data, multi-source geographic information data, and multi-source meteorological data. Among them, the multi-source hydrological data may include water level and flow rate, etc.; the multi-source geographic information data may include watershed topography, soil type, and vegetation cover, etc.; and the multi-source meteorological data may include precipitation, temperature, humidity, and wind speed, etc.

[0044] It should also be understood that the specific process of this preprocessing can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0045] Optionally, preprocessing for each type of multi-source data may include at least one of data cleaning, missing value handling, outlier detection, and data standardization.

[0046] Step S120: The multi-source data after multiple preprocessing are fused to obtain the fused data.

[0047] Specifically, based on historical multi-source hydrological and meteorological data, the prediction accuracy of multiple data sources corresponding to the multi-source meteorological data is evaluated; based on the prediction accuracy of multiple data sources, the weighting coefficients of the meteorological data corresponding to each data source are dynamically adjusted; the meteorological data corresponding to each data source and its corresponding weighting coefficient are weighted and summed to obtain the final meteorological data; the final meteorological data, multi-source hydrological data, and multi-source geographic information data are then fused to obtain the fused data. The data sources for the meteorological data may include different satellites, etc.

[0048] For example, based on the predicted precipitation data from multiple meteorological data sources, the water level data of the basin can be predicted. Furthermore, considering the differences in the predicted hydrological data from various data sources, the predicted hydrological data of the basin can be compared with the actual water level data of the basin from historical multi-source hydrological data to determine the prediction accuracy of each data source. Additionally, each data source's meteorological data has a corresponding initial weight, which can be dynamically adjusted based on the prediction accuracy of each data source (for example, if the prediction accuracy is high, the current data source is considered reliable, and its weight can be adjusted upwards; if the prediction accuracy is low, the current data source is considered unreliable, and its weight can be adjusted downwards; this step must be performed in every prediction process). The final meteorological data is obtained by weighted summation of the meteorological data corresponding to each data source and its corresponding weight coefficients. The aforementioned content describes... Figure 2 The dynamic weight allocation part is then used. Subsequently, the final meteorological data, multi-source hydrological data, and multi-source geographic information data can be fused to obtain the fused data.

[0049] It should be noted here that... Figure 2The CNN-LSTM model in this application is the real-time early warning model for watershed floods.

[0050] It should be noted that since the hydrological data collected from each data source is essentially identical, in the process of obtaining the fused data, the hydrological data from one of the multiple sources can be selected for fusion, or the multiple sources of hydrological data can be fused first and then fused with other data (e.g., the final meteorological data). The embodiments of this application are not limited to this.

[0051] Correspondingly, multi-source geographic information data is similar; for details, please refer to the relevant descriptions of the hydrological data mentioned above, which will not be repeated here.

[0052] It should also be understood that the specific process of integrating the final meteorological data, multi-source hydrological data and multi-source geographic information data can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0053] Optionally, when selecting specific hydrological data from multi-source hydrological data and specific geographic information from multi-source geographic information data, for the final meteorological data, a spatial interpolator is used to convert the discrete point data or coarse-grid data into a meteorological element grid field consistent with the spatial resolution of the target watershed. The spatial interpolator includes at least one of Kriging interpolation, inverse distance weighted interpolation, or radial basis function interpolation. Furthermore, for the specified hydrological data, a hydrological state grid field covering the entire watershed can be generated using regression or machine learning methods based on the correlation between hydrological station observations and watershed spatial characteristics.

[0054] In other words, meteorological data, hydrological station time-series data, and geographic information static data are uniformly converted into spatiotemporal grid data covering the entire watershed through spatial interpolation and station-grid association methods.

[0055] Subsequently, by unifying the geographic projection coordinate system and grid resolution, all data layers are strictly registered in space, and data with different time frequencies are unified to the same sampling interval and timestamps are aligned through temporal resampling.

[0056] In other words, spatial benchmark unification and time step alignment are performed on spatiotemporal grid data to ensure that all data have consistent spatial resolution and time series.

[0057] Subsequently, by multi-channel tensor splicing, the three types of data are integrated into an input that can be processed by the real-time flood early warning model of the basin (i.e., the fused data).

[0058] Step S130: The fused data is input into the trained real-time flood warning model for processing to obtain the flood warning results for the target watershed. The real-time flood warning model is trained based on a CNN-LSTM model. During the training process, a Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model and adjust the feature channel weights to enhance the contribution of key features related to the flood formation mechanism, thereby optimizing the feature quality of the feature flattening layer input to the CNN-LSTM model.

[0059] It should be understood that the specific structure of the real-time flood early warning model for the basin can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0060] Optionally, the real-time flood warning model for the watershed includes an input layer, a CNN layer, a pooling layer, a feature flattening layer, an LSTM layer, and a fully connected layer connected in sequence. In other words, the real-time flood warning model for the watershed is obtained by training a CNN-LSTM model.

[0061] It should also be understood that the specific layer structure of each layer in the model and the specific pooling layer can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0062] Optionally, such as Figure 3 As shown, this pooling layer can be a max pooling layer, allowing the model input data to be obtained through the input layer and the CNN layer to slide and scan the model input data to extract local spatiotemporal features. Furthermore, the max pooling layer is used to downsample the feature map output by the CNN layer to reduce feature dimensionality, lower computational complexity, and retain key information within the features. Additionally, a feature flattening layer converts the multidimensional features output by the max pooling layer into a one-dimensional vector (after replacing it with a parameterizable dynamic adaptive pooling layer, this means converting the multidimensional features output by the parameterizable dynamic adaptive pooling layer into a one-dimensional vector), making it compatible with the input requirements of the LSTM layer. Finally, the LSTM layer can model long-term temporal dependencies of the spatiotemporal features. The output of the LSTM layer is then passed through a fully connected layer to generate flood warning results.

[0063] It should also be understood that the content included in the flood warning result can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0064] For example, the flood warning results may include information such as the flood risk level, the time and location of the flood, the water level, and the flow rate.

[0065] Furthermore, traditional pooling operations use fixed, predefined sampling windows (e.g., 2x2) and step sizes, and their functions (max or average) are also fixed. This "one-size-fits-all" approach cannot adapt to the varying importance of different regions and feature maps in the input data. For example, in flood prediction, topographic features around the main channel of a river are far more important than those in plains far from the river. Max pooling may retain maxima in irrelevant regions while losing sequence information within important regions that has predictive value but whose values ​​are not the largest. Therefore, it is necessary to consider alternative approaches. Figure 3 The max pooling layer in the code is replaced with a parameterized dynamic adaptive pooling layer (PDAP layer), specifically, as follows: Figure 4 As shown, the parameterizable dynamic adaptive pooling layer includes:

[0066] The Global Context (GC) module is used to first calculate the global average pooling (GAP) of the entire input feature map to obtain a global context information, and this global context information encodes the overall state and information of the current feature map.

[0067] The control module includes a first fully connected layer, a second fully connected layer, a reshaping module, and a Softmax module. The first fully connected layer contains H neurons (H is a positive integer much smaller than C, such as 32, and C represents the number of channels in the feature map), and uses the ReLU activation function to perform nonlinear transformation and feature compression on global context information. The second fully connected layer contains C×K neurons (K is the number of pooling operation types) and outputs the original weights of each channel for each pooling operation. The reshaping module reshapes the output of the second fully connected layer into a C-row, K-column weight matrix, where each row of the weight matrix is ​​the specific weight coefficient for the corresponding channel, which determines how multiple pooling operations are mixed for that channel. The Softmax module performs Softmax normalization on each row of the weight matrix to obtain the final pooling weight coefficients for each channel. And satisfy .in, i K represents the channel index; K represents the pooling operation type index.

[0068] The dynamic pooling module is used to pool the input feature map using K different pooling operations to obtain K pooling result feature maps. For each channel, the data of the corresponding channel in the K pooling result feature maps are weighted and fused according to the weight coefficient of that channel (i.e. the weight coefficient of the corresponding row) to obtain the final output feature map. The final output feature maps of all channels are combined into the final output feature map.

[0069] Therefore, this application's embodiments overcome the "fixed" nature of traditional pooling layers by enabling the downsampling process to dynamically adjust its behavior based on the global context of the input data, thus achieving dynamic weight allocation. Furthermore, by integrating the advantages of multiple pooling strategies, it retains both the ability of max pooling to highlight salient features and the advantage of average pooling in preserving overall background information, thereby reducing data dimensionality while maximizing the preservation of key spatial information relevant to flood prediction. Moreover, richer and more accurate feature representations inevitably enable subsequent LSTM and fully connected layers to make more accurate predictions, directly improving the accuracy of flood warnings.

[0070] Based on the structure of the above-mentioned real-time flood early warning model for river basins, the training process of this CNN-LSTM model can also be set according to actual needs, and the embodiments of this application are not limited thereto.

[0071] Optionally, such as Figure 5 As shown, the Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model to find the optimal hyperparameter combination. Simultaneously, the Bayesian optimizer is also used to adjust the feature channel weights to enhance the contribution of key features related to the flood formation mechanism. Here, feature channel weights refer to the weight coefficients applied to different channels of the input features; each weight value corresponds to a feature channel, used to adjust the relative importance of that channel in the model. The specific adjustment process is as follows: the feature channel weights are included as part of the hyperparameters, and the Bayesian optimizer searches the hyperparameter space for the optimal hyperparameter combination, including the feature channel weights. The feature channel weights are vectors with the same dimension as the number of input feature channels, each weight value is between 0 and 1, and the sum of all weights is 1. The Bayesian optimizer models the objective function (such as the validation set loss function) using Gaussian process regression and selects the next evaluation point using a collection function (such as the desired improvement in EI). By iteratively evaluating the impact of different weight vectors on model performance, it finds the feature channel weight vector that optimizes the validation set performance. In this way, key feature channels related to flood formation mechanisms (such as feature channels from multi-source hydrological data, geographic information data, or meteorological data) are assigned higher weights, thereby enhancing their contribution and optimizing the feature quality of the feature flattening layer input to the CNN-LSTM model. Based on the optimal hyperparameter combination and feature channel weight vector, the CNN-LSTM model is constructed and trained to generate a trained real-time flood early warning model for the watershed. The Bayesian optimizer simultaneously optimizes the basic hyperparameters of the CNN-LSTM (such as the number of filters in the CNN and the number of units in the LSTM) and the feature channel weight vector in the hyperparameter space; both constitute the optimization object.

[0072] It should be noted here that... Figure 5 The 1D convolutional layer in the paper is the CNN layer of this application.

[0073] In addition, this application may also include steps for visualizing and alerting flood warning results, so as to promptly convey flood warning information to relevant personnel and provide strong support for flood control and disaster reduction in the basin.

[0074] For example, when a flood warning includes multiple flood risk levels, a specific notification method can be set for each of these levels. For instance, the flashing frequency could increase as the flood risk level rises.

[0075] In summary, by utilizing the above technical solutions, this application embodiment acquires multiple multi-source data from the target watershed, preprocesses these data to obtain multiple preprocessed multi-source data, fuses the preprocessed multi-source data to obtain fused data, and inputs the fused data into a trained real-time flood warning model for processing to obtain flood warning results for the target watershed. This real-time flood warning model is trained based on a CNN-LSTM model. During the training process of the CNN-LSTM model, a Bayesian optimizer is used to optimize the hyperparameters of the CNN-LSTM model and adjust the feature channel weights to enhance the contribution of key features related to the flood formation mechanism. This optimizes the feature quality of the feature flattening layer input to the CNN-LSTM model, thereby fully exploring the spatiotemporal features of the multi-source data, improving the accuracy and timeliness of flood warnings, and providing strong support for watershed flood control and disaster reduction.

[0076] It should be understood that the above-described method for real-time early warning of watershed floods based on multi-source data fusion is merely exemplary. Those skilled in the art can make various modifications to the above method, and the modified solutions also fall within the protection scope of this application.

[0077] Please see Figure 6 , Figure 6 This diagram illustrates a structural block diagram of a real-time flood warning system 600 based on multi-source data fusion, according to an embodiment of this application. It should be understood that the real-time flood warning system 600 is capable of executing the steps described in the above method embodiments. The specific functions of the real-time flood warning system 600 can be found in the description above; detailed descriptions are omitted here to avoid repetition. The real-time flood warning system 600 includes at least one software functional module that can be stored in memory or embedded in the operating system (OS) of the real-time flood warning system 600 in the form of software or firmware. Specifically, the real-time flood warning system 600 includes:

[0078] The acquisition module 610 is used to acquire multiple types of multi-source data and preprocess the multiple types of multi-source data to obtain multiple types of preprocessed multi-source data.

[0079] The fusion module 620 is used to fuse multiple preprocessed multi-source data to obtain fused data;

[0080] The input module 630 is used to input the fused data into the trained real-time flood warning model for processing to obtain flood warning results. The real-time flood warning model is trained based on a CNN-LSTM model. During the training of the CNN-LSTM model, the hyperparameters of the CNN-LSTM model are optimized using a Bayesian optimizer, and the feature channel weights are adjusted by the Bayesian optimizer to enhance the contribution of key features related to the flood formation mechanism, thereby optimizing the feature quality of the feature flattening layer input to the CNN-LSTM model.

[0081] In one possible embodiment, preprocessing includes at least one of data cleaning, missing value handling, outlier detection, and data standardization.

[0082] In one possible embodiment, the multiple multi-source data includes multi-source hydrological data, multi-source geographic information data, and multi-source meteorological data; the fusion module 620 is specifically used for: evaluating the prediction accuracy of multiple data sources corresponding to the multi-source meteorological data based on historical multi-source hydrological data and historical multi-source meteorological data; dynamically adjusting the weight coefficient of the meteorological data corresponding to each data source in the multiple data sources based on the prediction accuracy of the multiple data sources; performing a weighted summation calculation on the meteorological data corresponding to each data source and its corresponding weight coefficient to obtain the final meteorological data; and fusing the final meteorological data, multi-source hydrological data, and multi-source geographic information data to obtain the fused data.

[0083] In one possible embodiment, the real-time flood warning model for the watershed is obtained by training a CNN-LSTM model; the real-time flood warning system for the watershed also includes:

[0084] The CNN-LSTM model training module (not shown) is used to: optimize the hyperparameters of the CNN-LSTM model using a Bayesian optimizer to find the optimal hyperparameter combination; and build and train the CNN-LSTM model based on the optimal hyperparameter combination to generate a trained real-time early warning model for watershed floods.

[0085] In one possible embodiment, the real-time flood warning model for a watershed includes an input layer, a CNN layer, a pooling layer, a feature flattening layer, an LSTM layer, and a fully connected layer connected in sequence.

[0086] Since the system described in the above embodiments of the present invention is a system used to implement the method of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the system based on the method described in the above embodiments of the present invention, and therefore will not be described again here. All systems used in the method of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0087] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0089] It should be noted that the word "a" or "an" preceding a component does not preclude the existence of multiple such components. This invention can be implemented using hardware comprising several different components and using a suitably programmed computer. Among the listed devices, several of these devices may be embodied by the same hardware. The use of terms such as "first," "second," "third," etc., is merely for convenience and does not indicate any order. These terms can be understood as part of the component names.

[0090] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the technical solution should be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention's technical solutions and their equivalents, then the invention should also include these modifications and variations.

Claims

1. A method for real-time flood warning of a river basin based on multi-source data fusion, characterized in that, The method comprises the following steps: obtaining multiple types of multi-source data of a target basin, and preprocessing the multiple types of multi-source data to obtain multiple types of preprocessed multi-source data; fusing the multiple types of preprocessed multi-source data to obtain fused data; inputting the fused data into a trained basin flood real-time early warning model for processing to obtain a flood early warning result of the target basin; wherein the basin flood real-time early warning model is trained based on a CNN-LSTM model, and in the process of training the CNN-LSTM model, the hyperparameters of the CNN-LSTM model are optimized by using a Bayesian optimizer, and the feature channel weight is adjusted by the Bayesian optimizer to enhance the contribution degree of key features related to the flood formation mechanism, thereby optimizing the feature quality of the feature flattening layer input into the CNN-LSTM model; wherein the basin flood real-time early warning model comprises a parameterizable dynamic adaptive pooling layer, and the parameterizable dynamic adaptive pooling layer comprises: a global context module for calculating the global average pooling of the entire input feature map to obtain global context information encoding the state and information of the current feature map as a whole; The control module comprises a first full connection layer, a second full connection layer, a reshaping module and a Softmax module, the first full connection layer is used for containing H neurons, uses a ReLU activation function, is used for nonlinear transformation and feature compression on the global context information; the second full connection layer contains CxK neurons, is used for outputting the original weight proportion of each channel corresponding to each pooling operation, C represents the channel number of a feature map, and K represents the number of pooling operation types; the reshaping module is used for reshaping the output of the second full connection layer into a weight matrix with C rows and K columns, and each row of the weight matrix is a dedicated weight coefficient of a corresponding channel, which determines how the channel mixes multiple pooling operations; the Softmax module is used for performing Softmax normalization processing on each row of the weight matrix to obtain the final pooling weight coefficient of each channel , and satisfies , wherein i represents a channel index; K represents a pooling operation type index; a dynamic pooling module for pooling the input feature map using K types of pooling operations to obtain K pooled result feature maps; for each channel, the data of the corresponding channel in the K pooled result feature maps is weighted and fused according to the weight coefficient of the channel to obtain a final output feature map; and all the final output feature maps of all channels are combined into the final output feature map of the parameterizable dynamic adaptive pooling layer. 2.The method according to claim 1, wherein, The preprocessing of the multiple types of multi-source data comprises at least one of data cleaning, missing value processing, outlier detection and data standardization. 3.The method of Claim 1, wherein, The multiple types of multi-source data comprise multi-source hydrological data, multi-source geographic information data and multi-source meteorological data; and the fusing of the multiple types of preprocessed multi-source data to obtain fused data comprises: evaluating the prediction accuracy of multiple data sources corresponding to the multi-source meteorological data according to historical multi-source hydrological data and historical multi-source meteorological data; dynamically adjusting the weight coefficient of the meteorological data corresponding to each data source in the multiple data sources according to the prediction accuracy of the multiple data sources; performing weighted summation calculation on the meteorological data corresponding to each data source and the weight coefficient corresponding thereto to obtain final meteorological data; fusing the final meteorological data, the multi-source hydrological data and the multi-source geographic information data to obtain fused data. 4.The method according to claim 1, wherein, The basin flood real-time early warning model comprises an input layer, a CNN layer, a pooling layer, the feature flattening layer, an LSTM layer and a fully connected layer connected in sequence.

5. A real-time flood warning system for a river basin based on multi-source data fusion, characterized in that, The method comprises the following steps: an acquisition module for acquiring multiple types of multi-source data of a target basin, and preprocessing the multiple types of multi-source data to obtain multiple types of preprocessed multi-source data; a fusion module for fusing the multiple types of preprocessed multi-source data to obtain fused data; The input module is configured to input the fused data into a trained basin flood real-time warning model for processing to obtain a flood warning result of the target basin. The basin flood real-time warning model comprises a parameterizable dynamic adaptive pooling layer, and the parameterizable dynamic adaptive pooling layer comprises: The global context module is configured to calculate a global average pooling of an entire input feature map to obtain global context information encoding a state and information of the entire input feature map. The control module comprises a first full connection layer, a second full connection layer, a reshaping module and a Softmax module, the first full connection layer is used for containing H neurons, uses a ReLU activation function, is used for nonlinear transformation and feature compression on the global context information; the second full connection layer contains CxK neurons, is used for outputting the original weight proportion of each channel corresponding to each pooling operation, C represents the channel number of a feature map, and K represents the number of pooling operation types; the reshaping module is used for reshaping the output of the second full connection layer into a weight matrix with C rows and K columns, and each row of the weight matrix is a dedicated weight coefficient of a corresponding channel, which determines how the channel mixes multiple pooling operations; the Softmax module is used for performing Softmax normalization processing on each row of the weight matrix to obtain the final pooling weight coefficient of each channel , and satisfies , wherein i represents a channel index; K represents a pooling operation type index; The dynamic pooling module is configured to pool the input feature map using K types of pooling operations to obtain K pooled result feature maps; for each channel, the data of the corresponding channel in the K pooled result feature maps is weighted and fused according to a weight coefficient of the channel to obtain a final output feature map; and the final output feature maps of all channels are combined into a final output feature map of the parameterizable dynamic adaptive pooling layer.

6. The real-time flood warning system for a river basin according to claim 5, wherein The preprocessing of the multiple types of multi-source data comprises at least one of data cleaning, missing value processing, outlier detection, and data standardization.

7. The real-time flood warning system for a river basin according to claim 5, wherein The multiple types of multi-source data comprise multi-source hydrological data, multi-source geographic information data, and multi-source meteorological data. The fusion module is specifically configured to: according to historical multi-source hydrological data and historical multi-source meteorological data, evaluate prediction accuracies of a plurality of data sources corresponding to the multi-source meteorological data; according to the prediction accuracies of the plurality of data sources, dynamically adjust weight coefficients of meteorological data corresponding to each data source in the plurality of data sources; and perform weighted summation calculation on the meteorological data corresponding to each data source and the weight coefficients corresponding to each data source to obtain final meteorological data.

8. The real-time flood warning system for a river basin according to claim 5, wherein The final meteorological data, the multi-source hydrological data, and the multi-source geographic information data are fused to obtain fused data.

9. A storage medium having stored thereon a computer program, characterized in that The basin flood real-time warning model comprises an input layer, a CNN layer, a pooling layer, the feature flattening layer, an LSTM layer, and a fully connected layer connected in sequence.

10. An electronic device comprising a processor, a memory, and a computer program stored on the memory, wherein the computer program, when executed by the processor, is arranged to perform the method of any one of claims 1 to 9. The computer program is configured to, when executed by a processor, perform the basin flood real-time warning method based on multi-source data fusion. The processor is configured to execute the computer program to implement the basin flood real-time warning method based on multi-source data fusion.

Images