Multi-modal feature fusion system and method for wireless sensor network facing flood disaster

Abstract

The present application relates to the technical field of data processing, especially to a wireless sensor network multi-modal feature fusion system and method for flood disasters, the present application proposes the following scheme, through the wireless sensor network obtains multidimensional sensing data, constructs the multi-modal feature matrix, and inputs the preset intelligent AI model to carry out disaster assessment. Through the adaptive adjustment time window and the sliding step, combining the LSTM prediction model, the dynamic capture precipitation intensity change trend, the optimization disaster prediction accuracy. Solve the problem of multi-source data fusion, poor real-time, low precision in the traditional flood disaster warning system, realize efficient, accurate flood disaster prediction and risk assessment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multimodal feature fusion system and method for wireless sensor networks for flood disasters. Background Technology

[0002] With the intensification of global climate change, the frequency and intensity of floods are constantly increasing, causing serious impacts on human society, economy, and ecological environment. Traditional flood monitoring methods rely on a single data source (such as satellite remote sensing, ground weather stations, or hydrological monitoring), but these methods have certain limitations in terms of spatiotemporal resolution, real-time performance, and accuracy. For example, although satellite remote sensing can cover a wide area, its data acquisition frequency is low and it is difficult to provide detailed ground information; although ground sensors can collect data in real time, their spatial coverage is limited and they cannot fully reflect changes in the development process of regional disasters.

[0003] For example, Chinese patent CN112949414B discloses an intelligent mapping method for surface water bodies using wide-field-of-view Gaofen-6 satellite imagery. Based on high-quality wide-field-of-view Gaofen-6 remote sensing imagery, it uses a grid-based geometric correction and image fusion method to obtain fused Gaofen-6 imagery data with 2m resolution across 8 bands. It proposes a method for rapid multi-scale segmentation of images based on a combination of principal component analysis and object-oriented approaches. At the segmentation object scale, feature classification vectors are constructed. Based on a spatial grid, training samples with high spatiotemporal representativeness of water / non-water bodies are selected to construct an object-oriented deep neural network-based intelligent mapping model for surface water bodies, achieving automated mapping of surface water bodies from wide-field-of-view Gaofen-6 remote sensing imagery. This method primarily improves the efficiency and accuracy of wide-field-of-view imagery geometric correction, multi-scale segmentation, and intelligent extraction of surface water bodies, showing good application potential in flood monitoring, river and lake management, and water ecological surveys.

[0004] All of the above patents suffer from the problems described in the background section: insufficient accuracy and poor real-time performance in multimodal data fusion, spatiotemporal alignment, and feature extraction, making it difficult to effectively support dynamic and refined disaster risk assessment and early warning. To address these issues, this application designs a wireless sensor network multimodal feature fusion system and method for flood disasters. Summary of the Invention

[0005] The technical problem this invention addresses is the shortcomings of existing technologies. It provides a multimodal feature fusion system and method for wireless sensor networks in flood disaster prevention. The system acquires multidimensional sensor data through a wireless sensor network, constructs a multimodal feature matrix, and inputs it into a pre-set intelligent AI model for disaster assessment. By adaptively adjusting the time window and sliding step size, combined with an LSTM prediction model, it dynamically captures the trend of precipitation intensity changes, optimizing disaster prediction accuracy. The system effectively improves the timeliness and accuracy of data processing through multi-physics coupling compensation, cross-modal data fusion, and dynamic feature distillation. A second processing network further enhances the accuracy of disaster assessment results through physical-driven nonlinear mapping, feature reconstruction, and a bidirectional collaborative optimization mechanism. This invention solves the problems of poor multi-source data fusion, poor real-time performance, and low accuracy in traditional flood disaster early warning systems, achieving efficient and accurate flood disaster prediction and risk assessment.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A multimodal feature fusion method for wireless sensor networks in response to flood disasters, comprising:

[0008] Multidimensional sensing data of the watershed is acquired through a time window to construct a multimodal feature matrix. The length of the time window is adaptively adjusted according to the rate of change of precipitation intensity, and the sliding step size of the window has a nonlinear mapping relationship with the watershed runoff propagation speed.

[0009] The multimodal feature matrix is ​​input into a preset intelligent AI model to output disaster assessment results. The intelligent AI model includes a first processing network and a second processing network. The first processing network extracts features from the multimodal feature matrix, and the second processing network compensates for environmental factors in the output of the first processing network through compensation factors in the multimodal feature matrix.

[0010] The disaster assessment results are fed back to the monitoring platform to generate a dynamic risk map of the watershed and send tiered early warning signals.

[0011] The length and window sliding step size are calculated based on the LSTM prediction model, specifically including:

[0012] Obtain historical precipitation intensity data, and predict the trend of precipitation intensity changes in the future period based on the historical precipitation intensity data;

[0013] Calculate the maximum variation range of precipitation intensity based on the aforementioned trend of precipitation intensity variation;

[0014] The length of the time window is set according to the maximum change amplitude;

[0015] The precipitation intensity change trend is input into a preset modeling model, and the output window sliding step size is generated. The modeling model is established based on the topography, precipitation pattern and runoff propagation characteristics of the watershed.

[0016] The construction of the multimodal feature matrix includes:

[0017] The multidimensional sensing data is mapped to the three-dimensional geographic coordinate system of the watershed to generate spatiotemporally unified gridded sensing data, wherein the multidimensional sensing data includes heterogeneous sensing data distributed in water bodies, soil and meteorological monitoring nodes.

[0018] Multiphysics coupling compensation is performed on the gridded sensing data;

[0019] The compensated sensor data is fused with weather radar and satellite remote sensing data across modes to generate a coupled tensor.

[0020] Dynamic feature distillation is performed on the coupled tensor to obtain a six-dimensional feature matrix after dimensional compression, wherein the dimensions of the six-dimensional feature matrix include timestamp, spatial grid, hydrological, meteorological, geological characteristics and compensation factor dimensions.

[0021] The compensated sensor data is fused with weather radar and satellite remote sensing data across modes, including:

[0022] Construct hydrological feature subtensors;

[0023] The meteorological radar reflectivity tensor and remote sensing deformation tensor are projected onto the hydrological feature space of the hydrological feature sub-tensor by a decomposition algorithm to generate a coupled tensor.

[0024] The dynamic feature distillation includes a forward propagation path and a backward propagation path, specifically including:

[0025] In the forward propagation path, hydrological features, meteorological features, and geological features are used as query vectors, key vectors, and value vectors, respectively, to calculate the dynamic correlation between different feature channels and generate a weight distribution map.

[0026] The weight distribution map is subjected to a Hadamard product operation with the coupling tensor to output a preliminarily optimized feature subspace.

[0027] In the backpropagation path, a gradient-sensitive feature selector is constructed to filter the feature subspace. Hard truncation is applied to feature dimensions with gradient magnitudes lower than the specified feature dimensions to generate a sparse mask matrix.

[0028] The sparse mask matrix is ​​multiplied bitwise with the feature sub-control to output a six-dimensional feature matrix.

[0029] The first processing network includes:

[0030] The spatiotemporal feature extraction layer is used to perform spatiotemporal joint modeling on the multimodal feature matrix and output a feature mapping matrix. The spatiotemporal joint modeling includes a spatiotemporal convolutional kernel group that is dynamically adjusted based on timestamps and spatial grids. The spatiotemporal convolutional kernel group slides along the time dimension to capture the evolution trend of hydrological parameters and expands the modeling of flood propagation correlation between watershed nodes along the spatial dimension.

[0031] The branch feature extraction layer is used to process the feature mapping matrix through a neural network model and output branch features;

[0032] The feature output layer is used to adaptively allocate the weights of each feature dimension through an attention mechanism, which includes a channel attention mechanism and a spatial attention mechanism. The branch features after dynamic channel and spatial weighting are output through an activation function.

[0033] The branch feature extraction layer includes hydrological branches, meteorological branches, and geological branches, wherein:

[0034] The hydrological branch extracts multi-scale runoff features through a dilated convolutional network.

[0035] The meteorological branch captures the spatiotemporal propagation pattern of precipitation through a two-way gated circulation unit;

[0036] The geological branch is modeled using graph convolutional networks to model the spatial topological relationships of watershed soil permeability.

[0037] The second processing network includes:

[0038] The data mapping layer is used to model the dynamic relationship between environmental data and the output of the first processing network through a physical-driven nonlinear mapping mechanism, and generate an influence weight matrix.

[0039] A feature reconstruction layer is used to reconstruct the output of the first processing network dimension by dimension based on the influence weight matrix.

[0040] The output layer is used to perform consistency correction on the reconstructed features through a bidirectional collaborative optimization mechanism. Based on the corrected features, the features are nonlinearly coupled through the structural domain and the spatial domain to calculate the disaster assessment results.

[0041] A wireless sensor network multimodal feature fusion system for flood disaster prevention includes a data acquisition module, a data processing module, and a result feedback module, wherein:

[0042] The data acquisition module is used to acquire multidimensional sensing data of the watershed through a time window, wherein the length of the time window is adaptively adjusted according to the rate of change of precipitation intensity, and the window sliding step size has a non-linear mapping relationship with the watershed runoff propagation speed.

[0043] The data processing module is used to construct a multimodal feature matrix based on the output of the data acquisition module, and input the multimodal feature matrix into a preset intelligent AI model to output disaster assessment results;

[0044] The result feedback module is used to feed back the output of the data processing module to the monitoring platform to generate a dynamic risk map of the watershed and send graded early warning signals.

[0045] The data processing module includes:

[0046] The matrix generation unit is used to map and organize multidimensional sensing data into a multimodal feature matrix according to spatiotemporal relationships;

[0047] The feature extraction unit is equipped with a first processing network to extract features from the multimodal feature matrix;

[0048] The output compensation unit is equipped with a second processing network, which compensates for environmental factors in the output of the first processing network using compensation factors in the multimodal feature matrix.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] This invention constructs a multimodal feature matrix, which not only realizes the spatiotemporal mapping of multimodal heterogeneous sensor data such as hydrology, meteorology, and geology to a unified three-dimensional geographic coordinate system, but also overcomes the problems of scale difference, heterogeneity and spatiotemporal asynchrony during data fusion through multi-physics coupling compensation and cross-modal fusion mechanism. Attached Figure Description

[0051] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0052] Figure 1 This is a flowchart illustrating the multimodal feature fusion method for wireless sensor networks for flood disasters according to Embodiment 1 of the present invention.

[0053] Figure 2 This is a network structure diagram of the intelligent AI model in Embodiment 1 of the present invention;

[0054] Figure 3 This is a diagram of the first processing network structure according to an embodiment of the present invention;

[0055] Figure 4 This is a structural diagram of the branch feature extraction layer in an embodiment of the present invention;

[0056] Figure 5 This is a diagram of the second processing network structure according to an embodiment of the present invention;

[0057] Figure 6This is a block diagram of a wireless sensor network multimodal feature fusion system for flood disasters, as described in Embodiment 2 of the present invention. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0059] Example 1:

[0060] Please see Figure 1 The present invention provides an embodiment of a multimodal feature fusion method for wireless sensor networks for flood disasters, the specific steps of which are as follows:

[0061] S1: Acquire multidimensional sensing data of the watershed through a time window;

[0062] In this embodiment, multi-dimensional sensor data (such as soil moisture, precipitation intensity, water level, and wind speed) from different monitoring nodes within the watershed are first acquired via a wireless sensor network. A time window mechanism is used to aggregate data from different time periods for subsequent processing. Specifically, the length of the time window is adaptively adjusted based on the rate of change in precipitation intensity, and the data acquisition time is dynamically optimized according to the trend of precipitation changes to ensure that critical moments of flooding are captured. Simultaneously, the data update frequency is precisely adjusted based on the topography and hydrological characteristics of the watershed. This achieves accurate disaster prediction and improves the timeliness and relevance of the data.

[0063] S2: Construct a multimodal feature matrix;

[0064] In this embodiment, the acquired multidimensional sensor data is preprocessed (e.g., spatiotemporal alignment, standardization), and then a multimodal feature matrix is ​​constructed based on the spatiotemporal relationships between the data. Specifically, the data is first mapped to the three-dimensional geographic coordinate system of the watershed and then subjected to spatiotemporally unified gridding. This gridded sensor data undergoes multiphysics coupling compensation to unify and fuse data collected by different sensors (e.g., soil moisture, precipitation, meteorological data, etc.). Then, cross-modal data fusion is performed, combining the compensated sensor data with meteorological radar and satellite remote sensing data to generate a coupling tensor. Finally, through dynamic feature distillation, a six-dimensional feature matrix is ​​obtained, covering timestamp, spatial grid, hydrological, meteorological, geological characteristics, and compensation factor dimensions.

[0065] S3: Input the multimodal feature matrix into the preset intelligent AI model to output disaster assessment results;

[0066] In this embodiment, the constructed multimodal feature matrix is ​​input into a preset intelligent AI model. Please refer to [link to relevant documentation]. Figure 2The present invention provides an intelligent AI model network structure diagram, comprising a first processing network and a second processing network. The first processing network extracts spatiotemporal features from the feature matrix using a neural network to capture key features of disaster occurrence. The second processing network compensates for environmental factors in the output of the first processing network based on compensation factors in the multimodal feature matrix, achieving a more refined disaster assessment by considering the specific meteorological, geological, and hydrological environment of the watershed.

[0067] S4: Feed back the disaster assessment results to the monitoring platform to generate a dynamic risk map of the watershed and send tiered early warning signals;

[0068] In this embodiment, the disaster assessment results output by the intelligent AI model are fed back to the monitoring platform. The platform uses these results to generate a dynamic risk map of the watershed and sends corresponding early warning signals according to different disaster risk levels. The risk map can dynamically display the predicted results of disaster occurrence, including the possible impact range and time scale, helping relevant decision-makers to take timely emergency measures.

[0069] In this embodiment, the disaster assessment results output by the intelligent AI model are fed back to the monitoring platform. The platform uses these results to generate a dynamic risk map of the watershed and sends corresponding early warning signals according to different disaster risk levels. The risk map can dynamically display the predicted results of disaster occurrence, including the possible impact range and time scale, helping relevant decision-makers to take timely emergency measures.

[0070] In the field of practical flood disaster monitoring, although existing technologies can acquire local hydrological, meteorological, or geological data using single types of sensors (such as hydrological sensors, meteorological stations, or remote sensing satellites), the spatiotemporal heterogeneity of the data and differences in monitoring methods make it difficult for existing technologies to achieve a refined, dynamic, and global description of the disaster occurrence and development process. For example, when traditional technologies use wireless sensor networks to acquire soil moisture and water level data from different monitoring points in a watershed, the lack of a unified mapping and fusion of the spatiotemporal relationships between heterogeneous data makes it difficult to determine the accurate correlation between the local disaster characteristics at a specific moment and the entire watershed flood evolution process. Furthermore, when independent technologies such as meteorological radar or satellite remote sensing provide large-scale regional data, insufficient spatial resolution or timeliness makes it difficult for traditional technologies to accurately project this large-scale information into the data space of micro-scale monitoring nodes. This gap between data scale and model greatly limits the accuracy and real-time performance of existing disaster early warning technologies.

[0071] In this embodiment, by constructing a multimodal feature matrix, not only is the spatiotemporal mapping of multimodal heterogeneous sensing data such as hydrology, meteorology, and geology to a unified three-dimensional geographic coordinate system realized, but also the scale difference, heterogeneity, and spatiotemporal asynchrony problems in data fusion in the prior art are overcome through multi-physics coupling compensation and cross-modal fusion mechanism.

[0072] Specifically, the method proposed in this application first maps the data to a unified three-dimensional spatial coordinate system through gridding. Then, it utilizes cross-modal tensor fusion technology to efficiently fuse large-scale monitoring data from meteorological radar and satellite remote sensing with local node data, forming a coupled tensor with precise fusion scale and multi-dimensional information. Furthermore, through feature distillation technology, the resulting six-dimensional feature matrix efficiently preserves the spatiotemporal characteristics of key data, thereby significantly improving the accuracy of disaster prediction.

[0073] Furthermore, by utilizing the forward and backward propagation paths in dynamic feature distillation, adaptive optimization of the weight distribution among multi-source data is achieved. This approach spontaneously uncovers and strengthens the implicit dynamic correlations between data, avoiding errors caused by manually setting data weights in traditional methods. First, a weight distribution map is constructed in the forward propagation path based on the dynamic correlation between hydrological, meteorological, and geological features. Then, in the backward propagation path, feature selection is performed based on gradient sensitivity, and data channels with weak correlations are dynamically sparsified. This creatively reduces the interference caused by redundant data, ensuring that the final feature matrix retains only the effective features that actually contribute to disaster prediction, thus achieving optimal effectiveness in data feature fusion.

[0074] In this embodiment, the length and window sliding step size are calculated based on the LSTM prediction model, and the specific steps are as follows:

[0075] S1.1: Obtain historical precipitation intensity data, and predict the trend of precipitation intensity changes in the future period based on the historical precipitation intensity data;

[0076] In this embodiment, historical precipitation intensity data is first acquired through a wireless sensor network or meteorological monitoring equipment. This data may include precipitation intensity information from the past several days or hours. This data, after spatiotemporal alignment, is used to train an LSTM model. The LSTM model excels in time-series data modeling, capturing long-term dependencies and short-term fluctuations in precipitation intensity data. To improve prediction accuracy, historical precipitation intensity data is normalized using a sliding window method to remove outliers and weighted according to factors such as watershed geographical characteristics and seasonal variations. After training, the LSTM model can accurately predict the trend of precipitation intensity changes over a future period. The output of the LSTM model provides the trajectory of precipitation intensity changes over the future, predicting potential future precipitation extremes and the magnitude of precipitation intensity changes, providing a basis for subsequent calculations of the time window length and sliding step size.

[0077] S1.2: Calculate the maximum variation range of precipitation intensity based on the aforementioned precipitation intensity variation trend;

[0078] Specifically, the range of precipitation intensity variation within a future time period is quantified to obtain the maximum value of the variation. This maximum variation reflects the magnitude of potential drastic changes in precipitation intensity, providing a more accurate basis for setting the time window. For example, if the predicted precipitation intensity change is drastic, a longer time window needs to be set to ensure that the entire process of change is fully captured. Conversely, if the precipitation intensity change is small, a shorter time window can be selected. By calculating the maximum variation, the problem of time windows being too long or too short can be avoided, ensuring that the length of the time window is more in line with actual meteorological conditions.

[0079] S1.3: Set the length of the time window according to the maximum change amplitude;

[0080] Specifically, the maximum magnitude of change serves as a key parameter for the length of the time window, ensuring that the time window can encompass all possible changes in precipitation intensity.

[0081] In this embodiment, the initial time window is weighted by the maximum variation range to provide a suitable time scale for monitoring flood disasters. For example, when the precipitation intensity varies greatly, the time window is extended to fully capture large fluctuations in precipitation intensity; conversely, when the precipitation variation range is small, the window length is shortened to reduce interference from irrelevant data. The specific weighting method can be determined by those skilled in the art through numerous repeated experiments.

[0082] S1.4: Input the precipitation intensity change trend into the preset modeling model and output the window sliding step size, wherein the modeling model is established based on the topography, precipitation pattern and runoff propagation characteristics of the watershed;

[0083] Specifically, the modeling system simulates how precipitation intensity propagates within a watershed under different sliding step sizes, based on the watershed's topographic features (such as slope, river distribution, and vegetation cover) and the spatiotemporal patterns of precipitation. The watershed's runoff propagation characteristics, the spatial distribution of precipitation, and the drainage capacity of different regions all influence the setting of the sliding step size. By simulating actual conditions at different step sizes, the modeling system can output the most suitable window sliding step size, ensuring the continuity and real-time nature of monitoring data and adapting to different geographical conditions and climatic characteristics.

[0084] The specific steps of S2 are as follows:

[0085] S2.1: Map the multidimensional sensing data to the three-dimensional geographic coordinate system of the watershed to generate spatiotemporally unified gridded sensing data, wherein the multidimensional sensing data includes heterogeneous sensing data distributed in water bodies, soil and meteorological monitoring nodes.

[0086] Specifically, the spatiotemporal distribution of the sensor data exhibits heterogeneity, and due to differences in monitoring equipment, the acquired data may involve different spatial scales and temporal granularities. Therefore, to ensure that various types of data can be fused and processed within a unified framework, these data must be converted into point data in a three-dimensional coordinate system through spatial mapping. For example, a soil moisture sensor may be located at one location in the watershed, while a weather station may be located at another, and a water level sensor may be located elsewhere. By using GIS (Geographic Information System) technology, these heterogeneous data are uniformly mapped to the three-dimensional geographic coordinate system of the watershed, ensuring that data collected by different sensors can be processed within the same spatiotemporal framework. This mapping not only guarantees the spatial consistency of the data but also ensures spatiotemporal uniformity in the subsequent fusion process.

[0087] S2.2: Perform multiphysics coupling compensation on the gridded sensing data;

[0088] Specifically, gridded sensor data, especially hydrological, meteorological, and geological data, are typically influenced by multiple physical fields. Natural factors such as meteorological changes, geological characteristics, and watershed topography affect the data, potentially causing systematic biases. For example, within the same watershed, hydrological data from different monitoring nodes may differ due to local climate change or topographical variations. A water level sensor located upstream of a high-altitude area will be affected by precipitation differently than a sensor located in a low-lying area. Furthermore, soil moisture variations may be influenced by seasonal climate, groundwater levels, and soil type. To eliminate these influences and improve data accuracy, multi-physics coupling compensation is essential.

[0089] In this embodiment, a multi-physics mathematical model is first established using the geographical characteristics, meteorological data, and hydrological features of the watershed. This model includes meteorological fields (such as precipitation, temperature, and humidity), geological fields (such as soil type, stratigraphic structure, and groundwater level), and hydrological fields (such as water level changes and flow velocity). Spatial distribution data of these physical fields are collected using Geographic Information System (GIS) and remote sensing technology.

[0090] Furthermore, after obtaining the basic data for each physical field, a preliminary analysis is performed on the gridded sensor data to identify the sources of error. For example, by comparing hydrological sensor data from different geographical regions, it can be found that some data have systematic errors due to differences in climatic factors or geological conditions.

[0091] Furthermore, by comparing the results obtained from physical field modeling with sensor data, errors are identified, and numerical optimization methods are used to compensate for the data. For example, if there are deviations in the hydrological sensor data, the compensation method will adjust the data to a range that better reflects the actual situation, taking into account the topography and meteorological conditions of the watershed. This compensation process considers the interactive effects between physical fields, such as the effects of topography on precipitation accumulation and drainage, and the soil's infiltration characteristics for precipitation.

[0092] Furthermore, the optimal compensation value is found by minimizing the difference between the actual sensor data and the data predicted by the physical field model. For hydrological data, the error between the water level sensor reading and the water level at the corresponding location in the model can be calculated using the least squares method, and then adjustments can be made.

[0093] Preferably, the compensated data is updated in real time and dynamically adjusted through a feedback mechanism. For example, in practical applications, weather changes are dynamic, so the compensation model continuously adjusts the compensation strategy based on the latest meteorological data and real-time monitoring results. This dynamic compensation method ensures that data errors in long-term monitoring are corrected in real time.

[0094] S2.3: Construct a hydrological feature sub-tensor. By using a decomposition algorithm, project the meteorological radar reflectivity tensor and the remote sensing deformation tensor onto the hydrological feature space of the hydrological feature sub-tensor to generate a coupled tensor.

[0095] Specifically, the key to constructing a hydrological feature subtensor lies in extracting a subspace related to hydrological features from multi-source data. Hydrological features typically include data related to hydrological dynamics, such as precipitation intensity, flow rate, water level changes, and soil moisture. To extract these hydrological features from sensor data, it is first necessary to extract key feature values ​​representing watershed hydrological changes from existing hydrological data (such as water level sensors and precipitation intensity sensors). Then, data dimensionality reduction methods are used to reduce the dimensionality of the data while retaining the features that best reflect hydrological changes. By projecting the original multidimensional data into a lower-dimensional space, a hydrological feature subtensor containing the main hydrological features is generated.

[0096] Furthermore, hydrological feature sub-tensors are fused with meteorological radar reflectivity tensors and remote sensing deformation tensors through nonnegative matrix factorization. This decomposition algorithm decomposes the meteorological radar reflectivity tensor and remote sensing deformation tensor into several factors across multiple dimensions, including time and space. These factors represent the underlying structure and relationships between these dimensions. The goal of the decomposition is to represent each tensor as a product of several low-rank matrices, thus extracting the low-dimensional representations that best reflect the relationships between various data types. For example, the meteorological radar reflectivity tensor can be decomposed into a set of spatially relevant factors, representing the meteorological reflectivity characteristics of different geographical regions; the remote sensing deformation tensor can be decomposed to extract factors representing surface deformation over different time periods. This factor extraction process ensures that the inherent structure of each type of data is fully reflected.

[0097] Furthermore, the projection process essentially maps meteorological radar and remote sensing data into a hydrological feature subspace, transforming their spatiotemporal characteristics into a feature space representation closely related to hydrological changes. To achieve projection, the similarity between the meteorological radar reflectivity tensor and the remote sensing deformation tensor and the hydrological feature subspace tensor must first be calculated. Similarity can be calculated using inner product, correlation, or other metrics. Then, using these calculated similarities, the meteorological radar and remote sensing deformation data are mapped to the hydrological feature subspace through linear or nonlinear mapping. The core objective of the projection process is to fuse large-scale meteorological and remote sensing data with intra-basin hydrological data in a common space, thereby effectively reflecting the correlation between different data sources and generating a unified coupling tensor. The generated coupling tensor will simultaneously contain multi-dimensional information from hydrology, meteorology, and remote sensing, comprehensively and accurately describing disaster processes within the basin.

[0098] S2.4: Perform dynamic feature distillation on the coupled tensor to obtain a six-dimensional feature matrix after dimensional compression, wherein the dimensions of the six-dimensional feature matrix include timestamp, spatial grid, hydrological, meteorological, geological characteristics and compensation factor dimensions;

[0099] Specifically, the coupled tensor after cross-modal fusion typically contains multi-dimensional data, potentially exhibiting high dimensionality and redundant information, leading to high computational and storage costs. To effectively reduce computational complexity while retaining useful information, dynamic feature distillation is employed to compress the tensor's dimensionality. Dynamic feature distillation automatically identifies the feature dimensions most influential on disaster prediction through forward and backward propagation processes, removing redundant or irrelevant information, thereby compressing the tensor into a six-dimensional feature matrix.

[0100] The dynamic feature distillation includes a forward propagation path and a backward propagation path. The specific steps of S2.4 are as follows:

[0101] S2.4.1: In the forward propagation path, hydrological features, meteorological features, and geological features are used as query vectors, key vectors, and value vectors, respectively. The dynamic correlation between different feature channels is calculated, and a weight distribution map is generated.

[0102] Specifically, hydrological, meteorological, and geological features are first input into a self-attention model as query vectors, key vectors, and value vectors, respectively. Each feature corresponds to a different feature channel. The feature vector of each feature channel is mapped into a high-dimensional vector space. These high-dimensional vectors represent an abstract representation of each feature and preserve their potential contribution to the final decision in disaster prediction through mapping. The self-attention mechanism learns the dynamic dependencies between different feature channels by calculating the correlation between these feature channels.

[0103] Furthermore, query vectors, key vectors, and value vectors are extracted from different feature spaces. First, the similarity between each pair of query vectors and key vectors is calculated using an inner product, yielding a relevance score for each pair of features. These similarity measures represent the degree of correlation between different features, reflecting their influence on the same disaster early warning target. For example, some meteorological features may have a high similarity to hydrological features, indicating a close correlation in their impact on flood disasters; while other geological features may have a low similarity to meteorological features, indicating a smaller predictive impact on the disaster.

[0104] Furthermore, based on the calculated similarity scores, the model uses these similarities to generate a weight distribution map. The weight distribution map is a matrix where the value at each position represents the strength of the association between a specific feature and other features. These weight values ​​can be seen as the relative importance of features in disaster prediction. For example, if hydrological features and meteorological features have a high similarity, they will have a larger weight, indicating a stronger influence on disaster prediction; while relatively weak features (such as the low correlation between certain geological features and hydrological features) will be assigned smaller weights.

[0105] S2.4.2: Perform a Hadamard product operation on the weight distribution map and the coupling tensor to output a preliminarily optimized feature subspace;

[0106] Specifically, the resulting weight distribution map is subjected to a Hadamard product operation with the coupling tensor, which involves multiplying the weight of each feature channel element-wise by the corresponding element of the coupling tensor. This weighting of each dimension in the coupling tensor optimizes the representation of the feature space. Through the Hadamard product operation, the dynamic weights of the feature channels can be effectively applied to the coupling tensor, strengthening features most relevant to disaster prediction while weakening less relevant features. As an element-wise multiplication operation, the Hadamard product ensures that the contribution of each feature matches its importance in disaster prediction, resulting in a feature subspace that better meets the needs of actual disaster prediction.

[0107] S2.4.3: In the backpropagation path, a gradient-sensitive feature selector is constructed. The feature selector is used to filter the feature subspace and apply a hard truncation to the feature dimension whose gradient magnitude is lower than the specified feature dimension to generate a sparse mask matrix.

[0108] Specifically, the feature selector measures feature importance by calculating the rate of change of each feature during gradient descent. If a feature contributes little throughout training, resulting in a low gradient magnitude, it is hard-truncated by the feature selector and removed from the feature space. This operation generates a sparse mask matrix indicating which feature dimensions are not important in disaster prediction and can be discarded. This process is similar to feature selection, but here it is based on gradient sensitivity to ensure that features that contribute little to disaster prediction do not affect the model's learning performance.

[0109] S2.4.4: Multiply the sparse mask matrix bitwise with the feature sub-control to output a six-dimensional feature matrix;

[0110] Specifically, by multiplying by position, the sparse mask matrix sets the previously truncated feature dimensions to zero, thus outputting a simplified six-dimensional feature matrix. This ensures that only those features highly correlated with disaster prediction are retained in the feature matrix, thereby reducing the burden of computation and storage and improving the efficiency of the model when processing large-scale data.

[0111] Please see Figure 3 The first processing network structure diagram of this embodiment is shown. The design of the first processing network is to achieve efficient processing of multimodal feature matrices, especially by constructing spatiotemporal feature extraction, branch feature extraction, and feature output layers to improve the accuracy and timeliness of flood disaster prediction. The first processing network includes:

[0112] The spatiotemporal feature extraction layer is used to perform spatiotemporal joint modeling on the multimodal feature matrix and output a feature mapping matrix. The spatiotemporal joint modeling includes a spatiotemporal convolutional kernel group that is dynamically adjusted based on timestamps and spatial grids. The spatiotemporal convolutional kernel group slides along the time dimension to capture the evolution trend of hydrological parameters and expands the modeling of flood propagation correlation between watershed nodes along the spatial dimension.

[0113] Specifically, the spatiotemporal convolutional kernel group is a key component of the spatiotemporal feature extraction layer. In order to achieve spatiotemporal joint modeling, the spatiotemporal convolutional kernel not only slides to capture the evolution trend of hydrological parameters in the time dimension, but also expands the modeling of the flood propagation correlation between watershed nodes in the spatial dimension.

[0114] In this embodiment, the spatiotemporal convolution kernel group consists of multiple convolution kernels, each of which has two dimensions: time and space, and is responsible for capturing temporal features and spatial features, respectively.

[0115] Furthermore, in the time dimension, the convolutional kernel slides along the time axis to extract the evolution trend of hydrological parameters over time. To capture hydrological changes over long periods, the size and stride of the convolutional kernel can be dynamically adjusted based on the actual data. For example, hydrological data may have long variation cycles (such as seasonal precipitation variations), so the sliding stride of the convolutional kernel can be set to a longer time span to better reflect long-term trends. In this way, the convolutional kernel can automatically learn the temporal patterns of hydrological data, thus providing a basis for flood prediction.

[0116] Furthermore, in the spatial dimension, the convolution kernel slides along the spatial grid to establish the flood propagation correlations between monitoring nodes within the watershed. The mutual influences and propagation relationships between various monitoring nodes in the watershed (such as water level, soil moisture, and meteorological data) need to be modeled using spatial convolution kernels. For example, precipitation in one area may affect soil moisture or water level changes in neighboring areas, and the interrelationships between these data are captured by spatial convolution kernels. To accurately model spatial propagation patterns within the watershed, the size and shape of the convolution kernel can be adjusted according to the geographical characteristics of the watershed. Through spatial convolution, the model can learn the dependencies between different spatial locations, thereby better predicting flood propagation paths.

[0117] Furthermore, the spatiotemporal convolutional kernel group performs convolution operations on the input multimodal feature matrix, outputting a feature mapping matrix. This matrix is ​​the final output of the spatiotemporal feature extraction layer, containing temporal and spatial information from the multimodal data. Each element in the feature mapping matrix represents a data feature at a specific time and spatial location, such as precipitation, water level, or soil moisture at a specific time and spatial location.

[0118] The branch feature extraction layer is used to process the feature mapping matrix through a neural network model and output branch features;

[0119] In this embodiment, the input feature mapping matrix is ​​further processed by a neural network model to extract representative branch features. This branch feature extraction layer typically includes multiple neural network modules, each focused on mining different types of useful information from the feature matrix.

[0120] The feature output layer is used to adaptively allocate the weights of each feature dimension through an attention mechanism, which includes a channel attention mechanism and a spatial attention mechanism. The branch features after dynamic channel and spatial weighting are output through an activation function.

[0121] Specifically, the channel attention mechanism is mainly used to adjust the weights of different feature channels. Through this mechanism, the model can adaptively assign different weights to different features, thereby highlighting those features that have a greater impact on disaster prediction. For example, hydrological features (such as water level and precipitation) are usually more important than meteorological or geological features in flood disaster prediction, so the system will assign higher weights to hydrological features based on the actual situation of the data.

[0122] Furthermore, the spatial attention mechanism is used to adjust weights based on the importance of different geographical regions. For example, in areas where there may have been significant rainfall or water level changes, features from those regions should receive greater attention, while features from other regions can have their weights reduced. The spatial attention mechanism adaptively adjusts feature importance based on spatial changes, ensuring the system focuses on data from key areas.

[0123] Furthermore, after dynamic channel and spatial weighting, all features are nonlinearly mapped through an activation function to generate the final output features. These output features will serve as inputs to subsequent disaster assessment and prediction models, ensuring the system can provide timely disaster warnings with high accuracy.

[0124] Please see Figure 4 This invention presents a branch feature extraction layer structure diagram. The design of the branch feature extraction layer aims to efficiently process the input multimodal feature matrix through multiple feature branches, extracting key hydrological, meteorological, and geological features. Each branch focuses on extracting features from a different domain, thereby providing rich and multidimensional information for subsequent disaster prediction. Specifically, the hydrological branch, meteorological branch, and geological branch employ three different network architectures—Dilated Convolutional Network, Bidirectional Gated Recurrent Unit (BiGRU), and Graph Convolutional Network (GCN)—to process features from their respective domains. The branch feature extraction layer includes hydrological, meteorological, and geological branches, wherein:

[0125] The hydrological branch extracts multi-scale runoff features through a dilated convolutional network.

[0126] Specifically, the main task of the hydrological branch is to extract multi-scale runoff features from the input hydrological data. Runoff features refer to the trends in hydrological data such as precipitation, flow, and water level over time and space. To capture these changes, the hydrological branch employs a dilated convolutional network. Unlike standard convolution, dilated convolution can skip some input values ​​during the convolution operation, thereby expanding the receptive field without increasing the number of parameters. In this way, dilated convolution can capture long-term dependencies in the data without increasing computational complexity. It also handles features at different scales. For example, in hydrological data, runoff features may change rapidly in a short period (such as sudden heavy rainfall) or evolve slowly over a long period (such as seasonal variations). Through dilated convolution, the network can automatically capture these multi-scale changing features, providing more detailed hydrological information for subsequent flood disaster prediction.

[0127] The meteorological branch captures the spatiotemporal propagation pattern of precipitation through a two-way gated circulation unit;

[0128] Specifically, the primary task of the meteorological branch is to capture the spatiotemporal propagation patterns in precipitation data. To this end, the meteorological branch uses bidirectional gated cyclic units (BiGRUs) to model the spatiotemporal evolution of precipitation intensity. BiGRU is a variant of the gated cyclic unit that can simultaneously consider both forward and backward dependencies in sequence data. In meteorological data, the spatiotemporal patterns of precipitation are influenced not only by past precipitation events but also by the inverse influence of future precipitation patterns. BiGRU can effectively capture these spatiotemporal dependencies and provide more comprehensive forecast information.

[0129] Furthermore, the input to each time step includes not only the precipitation data for the current time step, but also historical information from both the previous and next time steps. This enables the network to more accurately predict future precipitation trends. Especially in flood forecasting, the temporal relationship of precipitation is crucial, and the use of a bidirectional GRU allows the system to better capture temporal fluctuations and potential anomalies in precipitation.

[0130] The geological branch is modeled using graph convolutional networks to model the spatial topological relationships of watershed soil permeability.

[0131] Specifically, the main task of the Geological Branch is to extract flood-related features from geological data, particularly soil permeability and the spatial topology of the watershed. To effectively process this spatial data, the Geological Branch employs graph convolutional networks (GCNs). GCNs can handle the relationships between nodes and edges in graph data, and watershed soil permeability data can be naturally represented as a graph structure, where different regions within the watershed are the nodes, and the connections between nodes represent the relationships between these regions (e.g., water propagation paths). By learning these connections between nodes, the GCN can model the impact of soil permeability within the watershed on water infiltration and flood propagation.

[0132] Furthermore, the characteristics of each node (such as soil permeability) are updated based on information from neighboring nodes, thereby capturing the impact of spatial topology on flood propagation. GCN can effectively handle complex watershed geological structures and help predict water flow dynamics in different regions.

[0133] Please see Figure 5 The second processing network structure diagram of this embodiment of the invention further optimizes the output features of the first processing network, combines environmental data for modeling and adjustment, and ultimately outputs accurate disaster assessment results. It employs technologies such as a physics-driven nonlinear mapping mechanism, dimension-by-dimensional reconstruction, and a bidirectional collaborative optimization mechanism. The second processing network includes:

[0134] The data mapping layer is used to model the dynamic relationship between environmental data and the output of the first processing network through a physical-driven nonlinear mapping mechanism, and generate an influence weight matrix.

[0135] In this embodiment, the main task of the data mapping layer is to model the dynamic relationship between the influence of environmental data and the output of the first processing network through a physical-driven nonlinear mapping mechanism, and to generate an influence weight matrix. First, environmental data such as meteorological data (wind speed, temperature, precipitation, etc.), hydrological data (flow rate, water level changes, etc.), and geological data (soil permeability, topography, etc.) are multidimensional data collected by sensors. These environmental data have a certain spatiotemporal correlation with the output features of the first processing network. Therefore, through a physical-driven nonlinear mapping mechanism, the influence of these environmental factors can be effectively mapped to the feature space output by the first processing network.

[0136] Specifically, the physics-driven nonlinear mapping mechanism considers the interaction of multiple physical factors and employs nonlinear functions to capture the nonlinear impact of environmental changes on disaster assessment results. The aim of this process is to combine environmental data with the spatiotemporal features of the network output to generate an influence weight matrix, which reflects the importance of each feature under different environmental factors. In this way, the model can more accurately identify which environmental factors have a greater impact on disaster occurrence, and by adjusting the weights, the model's predictions become more accurate.

[0137] A feature reconstruction layer is used to reconstruct the output of the first processing network dimension by dimension based on the influence weight matrix.

[0138] In this embodiment, the task of the feature reconstruction layer is to reconstruct the output of the first processing network dimension by dimension based on the generated influence weight matrix. The influence weight matrix provides the importance of each feature dimension under different environmental conditions, and the feature reconstruction layer uses this weight information to adjust the feature matrix. Specifically, feature reconstruction involves weighting the output features of the first processing network to ensure that important features receive higher weights in disaster assessment, while secondary features are appropriately weakened. This process essentially optimizes the model's focus on key features by dynamically adjusting each feature dimension.

[0139] Specifically, during feature reconstruction, the feature matrix is ​​weighted dimension-wise according to the influence weight matrix, enhancing the impact of key features in disaster prediction. For example, if rainfall in a certain region has a significant impact on flood occurrence, then the feature dimension corresponding to that region will be assigned a higher weight during reconstruction, thus playing a greater role in subsequent disaster assessment. Through this dimension-wise reconstruction, the feature reconstruction layer can effectively optimize input features, ensuring that the model fully considers environmental impacts and generates more accurate disaster prediction results.

[0140] The output layer is used to perform consistency correction on the reconstructed features through a bidirectional collaborative optimization mechanism. Based on the corrected features, the features are nonlinearly coupled through the structural domain and the spatial domain to calculate the disaster assessment results.

[0141] In this embodiment, the main function of the output layer is to perform consistency correction on the reconstructed features through a bidirectional collaborative optimization mechanism, and finally generate disaster assessment results. The core idea of ​​the bidirectional collaborative optimization mechanism is to optimize the model's prediction results in two directions simultaneously: on the one hand, the accuracy of the output features is optimized through forward propagation; on the other hand, the consistency correction of the output features is performed using backpropagation to ensure a high degree of consistency between the prediction results and the actual environmental data.

[0142] Specifically, the reconstructed features are first corrected to ensure consistency across all features in disaster assessment. This correction considers not only the interdependencies between features but also the impact of environmental data on disaster occurrence, ensuring the accuracy and reliability of the final results. The corrected features are then nonlinearly coupled through structural and spatial domains to further refine the spatial distribution characteristics of disaster prediction. Specifically, structural domain coupling considers the logical relationships between features, while spatial domain coupling considers the mutual influences of geographical locations. These two coupling methods help the model accurately predict the occurrence, development, and impact range of disasters in multi-dimensional, multi-scale spaces.

[0143] Example 2:

[0144] Please see Figure 6 This invention provides an embodiment: a multimodal feature fusion system for wireless sensor networks for flood disasters, the system comprising a data acquisition module, a data processing module, and a result feedback module, wherein:

[0145] The data acquisition module is used to acquire multidimensional sensing data of the watershed through a time window, wherein the length of the time window is adaptively adjusted according to the rate of change of precipitation intensity, and the window sliding step size has a non-linear mapping relationship with the watershed runoff propagation speed.

[0146] The data processing module is used to construct a multimodal feature matrix based on the output of the data acquisition module, and input the multimodal feature matrix into a preset intelligent AI model to output disaster assessment results;

[0147] The result feedback module is used to feed back the output of the data processing module to the monitoring platform to generate a dynamic risk map of the watershed and send graded early warning signals.

[0148] The data processing module includes:

[0149] The matrix generation unit is used to map and organize multidimensional sensing data into a multimodal feature matrix according to spatiotemporal relationships;

[0150] The feature extraction unit is equipped with a first processing network to extract features from the multimodal feature matrix;

[0151] The output compensation unit is equipped with a second processing network, which compensates for environmental factors in the output of the first processing network using compensation factors in the multimodal feature matrix.

[0152] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multimodal feature fusion method for wireless sensor networks for flood disasters, characterized in that, The wireless sensor network multimodal feature fusion method includes: Multidimensional sensing data of the watershed is acquired through a time window to construct a multimodal feature matrix. The length of the time window is adaptively adjusted according to the rate of change of precipitation intensity, and the sliding step size of the window has a nonlinear mapping relationship with the watershed runoff propagation speed. The multimodal feature matrix is ​​input into a preset intelligent AI model to output disaster assessment results. The intelligent AI model includes a first processing network and a second processing network. The first processing network extracts features from the multimodal feature matrix, and the second processing network compensates for environmental factors in the output of the first processing network through compensation factors in the multimodal feature matrix. The disaster assessment results are fed back to the monitoring platform to generate a dynamic risk map of the watershed and send tiered early warning signals. The construction of the multimodal feature matrix includes: The multidimensional sensing data is mapped to the three-dimensional geographic coordinate system of the watershed to generate spatiotemporally unified gridded sensing data, wherein the multidimensional sensing data includes heterogeneous sensing data distributed in water bodies, soil and meteorological monitoring nodes. Multiphysics coupling compensation is performed on the gridded sensing data; The compensated sensor data is fused with weather radar and satellite remote sensing data across modes to generate a coupled tensor. Dynamic feature distillation is performed on the coupled tensor to obtain a six-dimensional feature matrix after dimensional compression, wherein the dimensions of the six-dimensional feature matrix include timestamp, spatial grid, hydrological, meteorological, geological characteristics and compensation factor dimensions.

2. The multimodal feature fusion method for wireless sensor networks for flood disasters according to claim 1, characterized in that, The length and window sliding step size are calculated based on the LSTM prediction model, specifically including: Obtain historical precipitation intensity data, and predict the trend of precipitation intensity changes in the future period based on the historical precipitation intensity data; Calculate the maximum variation range of precipitation intensity based on the aforementioned trend of precipitation intensity variation; The length of the time window is set according to the maximum change amplitude; The precipitation intensity change trend is input into a preset modeling model, and the output window sliding step size is output. The modeling model is established based on the topography, precipitation pattern and runoff propagation characteristics of the watershed. The first processing network includes: The spatiotemporal feature extraction layer is used to perform spatiotemporal joint modeling on the multimodal feature matrix and output a feature mapping matrix. The spatiotemporal joint modeling includes a spatiotemporal convolutional kernel group that is dynamically adjusted based on timestamps and spatial grids. The spatiotemporal convolutional kernel group slides along the time dimension to capture the evolution trend of hydrological parameters and expands the modeling of flood propagation correlation between watershed nodes along the spatial dimension. The branch feature extraction layer is used to process the feature mapping matrix through a neural network model and output branch features; The feature output layer is used to adaptively allocate the weights of each feature dimension through an attention mechanism, which includes a channel attention mechanism and a spatial attention mechanism. The branch features after dynamic channel and spatial weighting are output through an activation function. The branch feature extraction layer includes hydrological branches, meteorological branches, and geological branches, wherein: The hydrological branch extracts multi-scale runoff features through a dilated convolutional network. The meteorological branch captures the spatiotemporal propagation pattern of precipitation through a two-way gated circulation unit; The geological branch is modeled using graph convolutional networks to model the spatial topological relationships of watershed soil permeability.

3. The multimodal feature fusion method for wireless sensor networks for flood disasters according to claim 1, characterized in that, The compensated sensor data is fused with weather radar and satellite remote sensing data across modes, including: Construct hydrological feature subtensors; The meteorological radar reflectivity tensor and remote sensing deformation tensor are projected onto the hydrological feature space of the hydrological feature sub-tensor by a decomposition algorithm to generate a coupled tensor.

4. The multimodal feature fusion method for wireless sensor networks for flood disasters according to claim 1, characterized in that, The dynamic feature distillation includes a forward propagation path and a backward propagation path, specifically including: In the forward propagation path, hydrological features, meteorological features, and geological features are used as query vectors, key vectors, and value vectors, respectively, to calculate the dynamic correlation between different feature channels and generate a weight distribution map. The weight distribution map is subjected to a Hadamard product operation with the coupling tensor to output a preliminarily optimized feature subspace. In the backpropagation path, a gradient-sensitive feature selector is constructed. The feature selector filters the feature subspace and applies hard truncation to features in the feature subspace whose feature dimension is lower than the gradient magnitude, generating a sparse mask matrix. The sparse mask matrix is ​​multiplied bitwise with the feature subspace to output a six-dimensional feature matrix.

5. The multimodal feature fusion method for wireless sensor networks for flood disasters according to claim 1, characterized in that, The second processing network includes: The data mapping layer is used to model the dynamic relationship between environmental data and the output of the first processing network through a physical-driven nonlinear mapping mechanism, and generate an influence weight matrix. A feature reconstruction layer is used to reconstruct the output of the first processing network dimension by dimension based on the influence weight matrix. The output layer is used to perform consistency correction on the reconstructed features through a bidirectional collaborative optimization mechanism. Based on the corrected features, the features are nonlinearly coupled through the structural domain and the spatial domain to calculate the disaster assessment results.

6. A wireless sensor network multimodal feature fusion system for flood disasters, used to implement the wireless sensor network multimodal feature fusion method for flood disasters as described in any one of claims 1-5, characterized in that, The system includes a data acquisition module, a data processing module, and a result feedback module, wherein: The data acquisition module is used to acquire multidimensional sensing data of the watershed through a time window, wherein the length of the time window is adaptively adjusted according to the rate of change of precipitation intensity, and the window sliding step size has a non-linear mapping relationship with the watershed runoff propagation speed. The data processing module is used to construct a multimodal feature matrix based on the output of the data acquisition module, and input the multimodal feature matrix into a preset intelligent AI model to output disaster assessment results; The result feedback module is used to feed back the output of the data processing module to the monitoring platform to generate a dynamic risk map of the watershed and send graded early warning signals.

7. The wireless sensor network multimodal feature fusion system for flood disaster relief according to claim 6, characterized in that, The data processing module includes: The matrix generation unit is used to map and organize multidimensional sensing data into a multimodal feature matrix according to spatiotemporal relationships; The feature extraction unit is equipped with a first processing network to extract features from the multimodal feature matrix; The output compensation unit is equipped with a second processing network, which compensates for environmental factors in the output of the first processing network using compensation factors in the multimodal feature matrix.

Images