Flood monitoring system and method for mountain hydrological environment
By using multi-scale feature extraction and dynamic weight adjustment, the problem of insufficient timeliness in mountainous hydrological environment monitoring systems has been solved, enabling accurate simulation and real-time monitoring of complex hydrological network structures in mountainous areas.
Patent Information
- Application Number
- CN202511190782.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-25
- Publication Date
- 2025-12-09
Smart Images
Figure CN121093263A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of water monitoring and detection, and particularly relates to a flood monitoring system and method for a mountain hydrological environment. BACKGROUND
[0002] Hydrological environment monitoring plays an important role in flood prevention, drought resistance, rational use of water resources and national defense. The monitoring of the hydrological environment is a complex process of interaction of numerous factors. The common monitoring system generalizes complex hydrological phenomena, ignores secondary and random factors, retains main factors and parts with basic rules, establishes a mathematical and physical model with certain physical significance, and is implemented on a computer.
[0003] With the rapid development of remote sensing technology, Internet of Things, big data analysis and artificial intelligence, new flood monitoring systems have emerged. These technologies can realize real-time monitoring and analysis of water resources, improve the accuracy and response speed of flood warning by integrating multi-modal data (such as rainfall, river water level, soil moisture, weather forecast, etc.).
[0004] However, the hydrological environment in mountainous areas presents a more complex topological structure. It is difficult to accurately simulate the complex network structure through local data collected by a multi-sensor network, especially the physical environment of vegetation and rocks in the mountains has a great adverse effect on remote data collection. Secondly, the water flow environment shows complex dynamic characteristics in space and time. The existing monitoring system usually uses rules or simple models for modeling, which cannot capture the complex spatio-temporal relationship and affect the accurate understanding of the dynamic changes of water flow. Finally, as the amount of fitting data increases during the model construction process, the learning and processing time of the monitoring system for data increases significantly, and the existing monitoring system has more serious risks in mountainous areas where monitoring timeliness is required. SUMMARY
[0005] The main purpose of the present application is to provide a flood monitoring system and method for a mountain hydrological environment, which aims to solve the problem of large hysteresis of the monitoring system in the prior art when monitoring the mountain hydrological environment.
[0006] To achieve the above purpose, the present application provides a flood monitoring system for a mountain hydrological environment, which comprises:
[0007] A data acquisition module, the acquisition module comprising a hydrological monitoring unit, a remote sensing unit and a view unit, the hydrological monitoring unit being arranged at a plurality of monitoring points to obtain hydrological data in the monitored area, the remote sensing unit being used to obtain remote sensing data in the monitored area, and the view unit being used to obtain view data in the monitored area;
[0008] The edge processing module comprises a convolutional encoder, a self-attention encoder and a decoding unit, the convolutional encoder is used for multi-scale extraction of image features of view data from low level to high level, the self-attention encoder is used for extracting global features of the image features, an augmented weight vector is output in a training stage of the convolutional encoder, and the mapping results of a plurality of feature spaces extracted by the convolutional encoder and the self-attention encoder are weighted and fused to generate a classification decision basis, and the decoding unit is further used for deciding risk data and non-risk data according to the classification decision basis.
[0009] The deep fusion module comprises a space-time alignment unit, a feature encoding unit, a dynamic fusion unit and a risk monitoring unit, the space-time alignment unit is used for unifying the space-time benchmarks of the risk data, the hydrological data and the remote sensing data, the feature encoding unit is used for extracting deep features of each data to generate feature vectors of a unified dimension, and the dynamic fusion unit is used for regulating weights through the feature vectors of the hydrological data, weighting and fusing the feature vectors of the remote sensing data, the feature vectors of the hydrological data and the feature vectors of the risk data, and generating a flood monitoring map.
[0010] Optionally, the convolutional encoder further comprises a training unit, the convolutional encoder is equipped with a ResNet50 network model, and the training unit is used for pre-training the ResNet50 network model through a multi-experience kernel learning machine.
[0011] Optionally, the decoding unit comprises a decoder.
[0012] The first layer of the decoder is a channel-by-channel convolutional layer to receive the classification decision basis.
[0013] The second layer of the decoder is an activation function.
[0014] The third layer of the decoder is a transposed convolutional layer to halve the number of channels and expand the spatial size by 2 times.
[0015] The fourth layer of the decoder is an activation function.
[0016] The fifth layer of the decoder is an up-sampling layer to expand the spatial size by 4 times.
[0017] The sixth layer of the decoder is a point-by-point convolutional layer to reduce the number of channels to the initial number of channels of the model to keep the spatial size unchanged.
[0018] To achieve the above object, the application further provides a flood monitoring method for a mountain hydrological environment, the method comprising the following steps:
[0019] Step 1, establishing a topological structure of a river basin in a monitored region to set a plurality of monitoring points and acquire hydrological data, remote sensing data and view data in the monitored region.
[0020] Step 2, multi-scale extraction of image features of view data from low to high level, and then extraction of global features of image features;
[0021] Step 3, weighted fusion of mapping results of multiple feature spaces extracted by convolutional encoder and self-attention encoder, and generation of classification decision basis;
[0022] Step 4, decision of risk data and non-risk data according to the classification decision basis;
[0023] Step 5, unification of spatio-temporal benchmarks of risk data, hydrological data and remote sensing data;
[0024] Step 6, extraction of deep features of each data to generate feature vectors of unified dimension;
[0025] Step 7, weighted fusion of feature vectors of remote sensing data, feature vectors of hydrological data and feature vectors of risk data by adjusting weights of feature vectors of hydrological data, and generation of flood monitoring map;
[0026] Wherein, the hydrological data includes water level data, rainfall data and flow data; the risk data is image features when the flood probability is greater than a first set threshold; the flood monitoring map has submergence depth matrix information and submergence loss thermal information.
[0027] Optionally, after obtaining the hydrological data, remote sensing data and view data in the monitored area, the multi-scale extraction of image features of the view data from low to high level further includes the following steps:
[0028] Unifying the resolution of the view data and normalizing the image data;
[0029] Using median filtering method to denoise the image data;
[0030] Obtaining historical image data, and summarizing the image data to form an image data set, dividing the image data set into a training set and a test set, and the training set is used for the training stage of the convolutional encoder.
[0031] Optionally, the training stage of the convolutional encoder includes the following steps:
[0032] Step 101, marking part of the samples of the training set centroid part;
[0033] Step 102, calculating the sample centroid that is not marked;
[0034] Step 103, calculating the distance array of the training set to the centroid and incrementally sorting;
[0035] Step 104, mark part of the samples in the distance array;
[0036] Step 105, repeat steps 102 to 104, and the ending condition is that the proportion of unmarked samples is less than a second set threshold;
[0037] Step 106, delete the marked samples and keep the unmarked samples to obtain the final training samples;
[0038] Step 107, select a plurality of subspaces, each of which is constructed by the final training samples obtained in step 106;
[0039] Step 108, calculate the experience kernel mapping corresponding to each subspace;
[0040] Step 109, calculate the augmented experience feature vector of each subspace;
[0041] Step 110, construct the target matrix of the training samples;
[0042] Step 111, output the augmented weight vector.
[0043] Optionally, in step 2, the global feature of the extracted image feature includes the following steps:
[0044] Step 201, divide the obtained image feature into blocks of fixed size, flatten each block into a vector, and then map it to a high-dimensional space through linear projection to obtain a sequence;
[0045] Step 202, input the sequence into the multi-head self-attention layer of the self-attention encoder, calculate the single-head attention, and then calculate the multi-head attention in parallel to capture long-distance dependency;
[0046] Step 203, apply two fully connected layers to the output of the self-attention;
[0047] Step 204, repeat steps 202 to 203 to gradually abstract the features, and finally take the feature vector at the first position as the global feature.
[0048] Optionally, in step 3, after generating the classification decision basis, the step 4 further includes the following steps:
[0049] The mapping results of the plurality of feature spaces extracted by the convolutional encoder and the self-attention encoder are weighted and fused to obtain a fused feature, and the fused feature is mapped into a segmentation map with the same size as the original input image to obtain risk assessment information with flood disaster image semantic information and feature position information.
[0050] Optionally, in step 4, the channel-by-channel convolutional layer randomly inputs an image data for feature extraction, the number of convolution kernels corresponds to the depth of the input image data, and three 5*5 images are output after completion; the three 5*5 images are input into the point-by-point convolutional layer, the number of convolution kernels is selected as 1*1*3, and three 5*5 risk data are output after completion.
[0051] Optionally, the weight is a matrix product of a feature vector of hydrological data and an augmented weight vector.
[0052] The present application makes up for the lack of local information through multi-source data collaborative collection, and the existing system adopts a fixed fusion strategy which is difficult to adapt to complex scenes, and the model generalization ability is enhanced through dynamic weight adjustment, the common model is affected by the hierarchical feature fragmentation, and the decision accuracy is affected, the semantic understanding is improved through multi-scale and global feature complementation, in addition, the edge processing module completes feature extraction and classification locally, reduces data transmission delay, solves the timeliness problem, effectively integrates multi-modal data, accurately simulates the complex hydrological network structure of mountainous areas, through hierarchical feature extraction and dynamic fusion mechanism, the capture ability of spatio-temporal dynamic characteristics is enhanced, edge computing is used to reduce data processing delay and improve monitoring real-time performance, and finally the generated flood monitoring map combines the submerged depth and loss heat information, provides reliable basis for disaster warning and emergency response. BRIEF DESCRIPTION OF DRAWINGS
[0053] Figure 1 It is a structural schematic diagram of the monitoring system in embodiment 1 of the present application;
[0054] Figure 2 It is a flowchart of the monitoring method in embodiment 2 of the present application;
[0055] Figure 3 It is an interface prototype diagram of the training unit in embodiment 1 of the present application;
[0056] Figure 4 It is an interface prototype diagram of the existing training unit in embodiment 1 of the present application;
[0057] Figure 5 It is a comparison diagram of original image information and monitoring image information in embodiment 2 of the present application.
[0058] The implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION
[0059] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort are within the scope of the present application.
[0060] It should be noted that all the direction indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the direction indications also change accordingly.
[0061] In the present application, unless otherwise explicitly specified and limited, the terms "connection", "fixation", etc. should be understood in a broad sense, for example, "fixation" can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through an intermediate medium; can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.
[0062] In addition, if the present application has descriptions involving "first", "second", etc., the descriptions of "first", "second", etc. are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first" and "second" can explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel solutions. For example, "A and / or B" includes A solution, or B solution, or A and B solution. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the scope of protection claimed by the present application.
[0063] Embodiment 1:
[0064] As shown in the accompanying drawings, Figure 1 The present embodiment provides a flood monitoring system for mountain hydrological environment, which comprises:
[0065] A data acquisition module, the acquisition module comprising a hydrological monitoring unit, a remote sensing unit and a view unit, the hydrological monitoring unit being arranged at a plurality of monitoring points to obtain hydrological data in the monitored area, the remote sensing unit being used to obtain remote sensing data in the monitored area, and the view unit being used to obtain view data in the monitored area;
[0066] An edge processing module comprising a convolutional encoder, a self-attention encoder and a decoding unit, the convolutional encoder being used to perform multi-scale extraction of image features of view data from low level to high level, the self-attention encoder being used to extract global features of image features, the convolutional encoder being output with an augmented weight vector in a training stage to perform weighted fusion on mapping results of a plurality of feature spaces extracted by the convolutional encoder and the self-attention encoder, so as to generate a classification decision basis, and the decoding unit being further used to decide risk data and non-risk data according to the classification decision basis;
[0067] A deep fusion module comprising a space-time alignment unit, a feature encoding unit, a dynamic fusion unit and a risk monitoring unit, the space-time alignment unit being used to unify the space-time reference of risk data, hydrological data and remote sensing data; the feature encoding unit being used to extract deep features of each data to generate a feature vector of a unified dimension; the dynamic fusion unit being used to regulate weights through the feature vector of the hydrological data, to weightedly fuse the feature vector of the remote sensing data, the feature vector of the hydrological data and the feature vector of the risk data, and to generate a flood monitoring map.
[0068] It should be noted that in the prior art, hydrological environment monitoring plays an important role in flood prevention, drought resistance, rational use of water resources and national defense. The traditional monitoring system generalizes complex hydrological phenomena through mathematical and physical models, but ignores secondary factors and random factors, making it difficult to accurately reflect the actual environment. With the development of remote sensing technology, Internet of Things and big data analysis, a new type of flood monitoring system realizes real-time monitoring and early warning by integrating multi-modal data; however, the hydrological environment in mountainous areas has a complex topological structure, and the discrete data collected by the multi-sensor network is difficult to simulate the network structure. The physical environment such as vegetation and rock affects the data acquisition accuracy. The existing system uses rules or simple models for modeling, which cannot capture the spatio-temporal dynamic characteristics, and the data processing time is significantly prolonged with the increase of data volume, resulting in insufficient timeliness.
[0069] Based on the above problems, the flood monitoring system for mountain hydrological environment is provided in the embodiment. Through the combination of multi-scale feature extraction and dynamic weight adjustment, a hierarchical processing framework is constructed to realize data complementation and efficient fusion. Specifically, multi-modal data is synchronously acquired by distributed hydrological sensors, remote sensing equipment and camera devices. The edge processing module first performs multi-scale convolution operation on view data to extract local features such as edges and textures, and then models global spatial relationships through a self-attention mechanism. The augmented weight vector generated in the training stage dynamically weights different levels of features, and the decoding unit maps the fused features to risk classification results. The deep fusion module performs spatio-temporal alignment of risk data and hydrological and remote sensing data. The feature encoding unit extracts uniform dimension features through a fully connected network, and the dynamic fusion unit assigns weights according to the importance of hydrological features to finally generate a monitoring map containing flood depth and loss information.
[0070] It can be understood that the embodiment makes up for the lack of local information through multi-source data collaborative acquisition. The existing system adopts a fixed fusion strategy, which is difficult to adapt to complex scenes. The embodiment enhances the model generalization ability through dynamic weight adjustment. The conventional model is affected by the fragmentation of hierarchical features, which affects the decision accuracy. The embodiment improves semantic understanding through complementary multi-scale and global features. In addition, the edge processing module completes feature extraction and classification locally, reduces data transmission delay, and solves the timeliness problem. The multi-modal data is effectively integrated, and the complex hydrological network structure in the mountainous area is accurately simulated. Through hierarchical feature extraction and dynamic fusion mechanism, the ability to capture spatio-temporal dynamic characteristics is enhanced. The edge computing is used to reduce data processing delay and improve monitoring real-time performance. The final generated flood monitoring map combines flood depth and loss heat information to provide reliable basis for disaster warning and emergency response.
[0071] It can also be understood that the core logic of the embodiment is to obtain risk data (a set of water boundary images with time series, which can represent the development trend of flood) and non-risk data (relatively low flood disaster occurrence situation) by fusing and classifying the corresponding water boundary images in remote sensing data (remote sensing image information) and view data (image information or video information); and then assist in representing the risk data through hydrological data to make it have more detailed flood information, including flood depth matrix information and flood loss heat information, etc.
[0072] In some embodiments, the first set threshold is the probability of occurrence of a flood disaster, or the degree of danger.
[0073] In some embodiments, the hydrological monitoring unit refers to a sensor network deployed at the monitoring point, which can be implemented by a water level gauge, a rain gauge and a flow meter to collect water level, rainfall and flow data in a local area.
[0074] In some embodiments, the remote sensing unit refers to a satellite or unmanned aerial vehicle device carrying optical or radar sensors, which can specifically be implemented by using multispectral imaging technology to obtain land cover and terrain change information.
[0075] In some embodiments, the view unit refers to a camera installed at a fixed position, which can specifically be implemented by using infrared or visible light cameras to capture real-time ground images.
[0076] In some embodiments, the convolutional encoder refers to a feature extraction network based on deep learning, which can specifically be implemented by using a ResNet50 model to extract low-level to high-level features of images through convolutional layer stacking.
[0077] In some embodiments, the self-attention encoder refers to a feature modeling module based on attention mechanism, which can specifically be implemented by using a multi-head self-attention layer, preferably a visual transformer, to establish long-distance dependency between pixels.
[0078] In some embodiments, the augmented weight vector refers to a dynamic adjustment parameter generated during model training, which can specifically be obtained by calculating subspace mapping relationship through a multi-experience kernel learning machine to optimize feature fusion weights.
[0079] In some embodiments, the dynamic fusion unit refers to a weight regulation module based on feature vectors, which can specifically be implemented by using matrix multiplication operations to adjust the fusion ratio of remote sensing and risk features through hydrological feature vectors.
[0080] In this embodiment, the convolutional encoder further includes a training unit, the convolutional encoder is equipped with a ResNet50 network model, and the training unit is used to pre-train the ResNet50 network model through a multi-experience kernel learning machine.
[0081] It can be understood that the ResNet50 network model refers to a deep residual network with a structure of 50 layers, which can specifically be implemented by using a skip connection structure to establish an identity mapping path between adjacent layers, alleviate the gradient vanishing problem in the training process of deep network, and enhance the extraction ability of multi-scale features of images. It can also be understood that the multi-experience kernel learning machine refers to a multi-kernel collaborative training mechanism, which can specifically be implemented by constructing multiple heterogeneous subspace kernel mappings, calculating the experience kernel mapping and augmented feature vector in different subspaces, forming a composite kernel function matrix, and improving the adaptability of the model to complex data distribution.
[0082] It should be noted that the residual connection structure of the ResNet50 network model is used for the convolutional encoder, shallow features are directly transmitted to the deep network through the skip connection, low-level texture information is retained while high-level semantic features are extracted, and the problem of multi-scale feature loss in mountain hydrological images is solved; when the training unit is pre-trained by the multiple experience kernel learning machine, firstly, the sample centroid is marked by iteration and the subspace is constructed, multiple heterogeneous kernel mapping spaces are generated, then the augmented experience feature vectors of each subspace are calculated to form a composite kernel matrix as the basis for weight optimization, and finally the augmented weight vector is output for feature fusion; this training method expands the expression dimension of the feature space through the multi-kernel collaborative mechanism, so that the model can adapt to the feature distribution deviation caused by complex interference factors such as light changes and vegetation shielding in the mountain hydrological environment.
[0083] It should also be noted that through the combination of the residual network structure and the multiple experience kernel learning machine, the extraction ability of the deep network to high-level semantic features is retained, and the adaptability of the model to heterogeneous data is enhanced through the multi-kernel mapping mechanism, effectively solving the limitation problem of single kernel function in complex hydrological scenes. Hierarchical extraction and robust expression of mountain hydrological image features are realized, the integrity of feature extraction and the efficiency of model training are improved, and the generalization ability of the model to complex hydrological data distribution is enhanced through the multi-kernel collaborative mechanism, providing a more reliable feature basis for subsequent flood monitoring.
[0084] The core steps of network training in this embodiment are: marking the centroid samples of the training set; iteratively calculating the centroid of the unmarked samples and marking the sample closest to the centroid; deleting the marked samples and retaining the boundary samples to construct a subspace; calculating the experience kernel mapping and augmented feature vector of the subspace.
[0085] The experience kernel mapping expression satisfies:
[0086]
[0087] Wherein, K is an N*N kernel matrix, and K i，j =k(x i , x j );
[0088] L is a Laplace matrix, and L=D-W;
[0089] W is an intra-class weight matrix, satisfying
[0090] D is a diagonal matrix, satisfying D i，i =∑ j W i，j ;
[0091] K i is the kernel vector of sample x i .
[0092] The principle can be understood as: by the intra-class scatter matrix Constraint projection space, retain the same sample distribution structure, so that the classifier focus on the difference between different classes, and then save a lot of training time, such as Figure 3 And the interface prototype diagram of the training unit Figure 4 As shown in the accompanying Figure 4 ) and the training unit Figure 3 ) in the present embodiment, respectively, the accuracy is the proportion of the number of samples predicted correctly by the model to the total number of samples, reflecting the overall prediction accuracy of the model; The precision is the proportion of the number of actual positive samples in the samples predicted by the model as positive, reflecting the accuracy of the model in predicting positive samples; The recall rate is the proportion of the number of samples predicted by the model as positive in the actual positive samples, which measures the capture ability of the model to positive samples; The F1 score is used to balance the performance of the model in precision and recall.
[0093] In the present embodiment, the decoding unit comprises a decoder;
[0094] The first layer of the decoder is a channel-by-channel convolution layer to receive classification decision basis;
[0095] The second layer of the decoder is an activation function;
[0096] The third layer of the decoder is a transpose convolution layer to reduce the number of channels by half and expand the spatial size by 2 times;
[0097] The fourth layer of the decoder is an activation function;
[0098] The fifth layer of the decoder is an up-sampling layer to expand the spatial size by 4 times;
[0099] The sixth layer of the decoder is a point-by-point convolution layer to reduce the number of channels to the initial number of channels of the model to keep the spatial size unchanged.
[0100] It can be understood that the channel-by-channel convolutional layer receives the classification decision basis and independently extracts the feature information of each channel to avoid interference between different channel features. The activation function performs a nonlinear transformation on the output to improve the model expression ability. The transposed convolutional layer reduces the number of channels by half and expands the spatial size through inverse convolution kernel operation, which reduces the consumption of computing resources while restoring the feature resolution. The activation function is applied again to strengthen the feature discrimination. The upsampling layer further expands the feature map to 4 times the original input through the interpolation algorithm to compensate for the loss of position information caused by downsampling. The point-by-point convolutional layer adjusts the number of channels to the initial number through a 1x1 convolution kernel to ensure that the output structure is compatible with the model input. Each level processes the feature map in turn through the alternating operation of channel number adjustment and spatial size expansion to realize the mapping from abstract classification basis to high-resolution risk data, reducing information loss while improving computing efficiency.
[0101] It can also be understood that traditional decoders usually use a single upsampling method, which leads to insufficient detail recovery, or directly stack fully connected layers, which causes a sharp increase in computational complexity. In this embodiment, the alternating use of channel-by-channel convolution and transposed convolution reduces the number of channels while gradually expanding the spatial size, controlling the consumption of computing resources and achieving multi-scale feature fusion; the combination of the upsampling layer and the point-by-point convolutional layer maintains the channel dimension consistent with the original data when restoring spatial information, avoiding feature distortion caused by mismatched channel numbers.
[0102] In some embodiments, the channel-by-channel convolutional layer refers to a layer that independently performs convolutional operations on each channel of the input feature map, which can be implemented using the depth convolution part in the depth separable convolution. By independently processing the channel features, interference between channels is avoided.
[0103] In some embodiments, the activation function refers to an operation unit that introduces a nonlinear transformation, which can be implemented using the ReLU function to enhance the model's fitting ability for complex feature distribution.
[0104] In some embodiments, the transposed convolutional layer refers to a layer that realizes feature map upsampling through inverse convolution operation, which can be implemented using an inverse convolution kernel with a step to reduce the number of channels and reduce computational complexity.
[0105] In some embodiments, the upsampling layer refers to a layer that expands the spatial size of the feature map through interpolation methods, which can be implemented using the bilinear interpolation algorithm to restore compressed detail information.
[0106] In some embodiments, the point-by-point convolutional layer refers to a layer that performs cross-channel linear combination for each spatial position, which can be implemented using a 1x1 convolution kernel to adjust the number of channels and maintain the spatial dimension.
[0107] In some embodiments, the decoder satisfies the expression:
[0108] Z = D k × D k × M + D k × D k × M × N
[0109] where D k is the size of the convolution kernel
[0110] M is the number of input channels
[0111] N is the number of output channels
[0112] The core principle is:
[0113] Decompose the standard convolution into two steps:
[0114] 1. Channel-wise convolution: each input channel is independently spatially convolved, outputting M D k × D k feature maps, with a parameter amount of
[0115] 2. Point-wise convolution: 1x1 convolution fuses channels, with a parameter amount of MxN.
[0116] Embodiment 2:
[0117] As shown in the accompanying Figure 2 , the present embodiment provides a flood monitoring method for mountain hydrological environment, which comprises the following steps:
[0118] Step 1, establish the topography of the monitored area, set up several monitoring points, and obtain the hydrological data, remote sensing data and view data in the monitored area;
[0119] Step 2, multi-scale extraction of image features of view data from low to high, and then extraction of global features of image features;
[0120] Step 3, weighted fusion of the mapping results of the multiple feature spaces extracted by the convolution encoder and the self-attention encoder, and generation of classification decision basis;
[0121] Step 4, decision of risk data and non-risk data according to the classification decision basis;
[0122] Step 5, unify the space-time reference of risk data, hydrological data and remote sensing data;
[0123] Step 6, extract deep features of each data to generate a unified dimension feature vector;
[0124] Step 7, regulating weights by eigenvectors of hydrological data to weightedly fuse eigenvectors of remote sensing data, eigenvectors of hydrological data and eigenvectors of risk data, and generating a flood monitoring map;
[0125] The hydrological data includes water level data, rainfall data and flow data; the risk data is image features when the flood probability is greater than a first set threshold; the flood monitoring map is provided with submergence depth matrix information and submergence loss thermal information.
[0126] It should be noted that the conventional method relies on fixed rules to lay out monitoring points, resulting in data missing in complex mountainous terrain. It should also be noted that the embodiment dynamically optimizes the monitoring network layout through basin topology analysis, improves data collection integrity, and the existing system uses a single convolutional network to extract image features, which easily ignores global context relationships. The embodiment combines multi-scale convolution and self-attention mechanism to simultaneously capture local details and long-range dependencies, enhancing feature expression capability. Conventional fusion methods directly splice multi-source features, resulting in information redundancy. The embodiment dynamically allocates feature contribution through hydrological data weight regulation, focusing on strengthening the impact of key parameters on the fusion result. Traditional spatio-temporal alignment uses simple interpolation, causing data distortion. The embodiment realizes accurate benchmark unification based on geographic registration and time series analysis, ensuring the spatial consistency of multi-modal data fusion.
[0127] Specifically, the construction of the basin topology first identifies the main river and tributary distribution based on digital water level data, deploys monitoring equipment at key nodes in the catchment area, and ensures that hydrological data collection covers runoff paths and prone areas. After the view data is normalized and denoised, it is input into the convolutional encoder, which extracts local features at different levels through the residual network, while the self-attention encoder captures the global association of mountain shape and water distribution. The augmented weight vector generated during the training phase dynamically weights the multi-scale features, giving higher weight to features in flood-sensitive areas in the classification decision. After the multi-source data is unified to the same spatio-temporal benchmark by the spatio-temporal alignment unit, the feature encoding unit extracts abstract features of uniform dimensions through a deep neural network, eliminating the impact of data heterogeneity on fusion. During the dynamic fusion process, the water level and rainfall eigenvectors generate weight coefficients through matrix operations, guiding the fusion ratio of remote sensing image features and risk area features. The final monitoring map superimposes submergence depth and loss thermal information, providing spatialized decision-making basis for emergency response.
[0128] Based on the above, the problem of incomplete coverage of monitoring data in complex mountainous terrain is effectively solved, and the data collection density in key areas is improved through topological optimization of the basin layout; the defects of traditional image feature extraction methods in capturing global correlation information are overcome, and the feature discrimination ability is improved by combining multi-scale convolution and self-attention mechanism; the problem of low fusion efficiency caused by simple superposition of multi-source data is improved, and the dominant role of key hydrological parameters is strengthened by using a dynamic weight distribution mechanism; the final flood monitoring map accurately reflects the flood depth distribution and disaster loss hotspots, providing high-precision spatial support for flood control decision-making in mountainous areas.
[0129] In some embodiments, the basin topology refers to a basin structure model constructed by a geographic information system, which can be implemented by extracting the river network and watershed boundary using a digital elevation model, and is used to optimize the layout of monitoring points to cover key areas.
[0130] In some embodiments, multi-scale extraction refers to the use of convolution kernels with different receptive fields to capture image details and semantic features in layers, which can be implemented through a multi-level convolution structure of a residual network to solve the problem of insufficient feature expression ability of traditional single-scale features.
[0131] In some embodiments, global features refer to correlation information across the entire image, which can be implemented by calculating the long-range dependency between pixels using a self-attention mechanism to overcome the limitations of local receptive fields of convolution operations.
[0132] In some embodiments, the spatio-temporal reference is unified, which means that multi-source data is mapped to the same timestamp and coordinate system, and can be implemented by using geographic registration and time interpolation algorithms to eliminate the impact of spatio-temporal misalignment on fusion.
[0133] In some embodiments, the dynamic fusion weight refers to the feature importance coefficient calculated based on the hydrological feature vector, which can be used to adjust the contribution ratio of remote sensing and risk features through matrix multiplication to realize a fusion process dominated by key hydrological parameters.
[0134] In this embodiment, after obtaining the hydrological data, remote sensing data and view data in the monitored area, the multi-scale extraction of the image features of the view data from low to high levels further includes the following steps:
[0135] Unify the resolution of the view data and normalize the image data;
[0136] Use the median filter method to denoise the image data;
[0137] Obtain historical image data, and aggregate the image data to form an image data set, divide the image data set into a training set and a test set, and the training set is used for the training stage of the convolutional encoder.
[0138] Specifically, the multi-source view data such as unmanned aerial vehicle aerial images and satellite remote sensing images are unified to 0.5-meter resolution by a bicubic interpolation algorithm to eliminate spatial scale differences, and then the pixel values are compressed to the interval [0, 1] by using maximum and minimum value normalization to avoid the influence of light intensity difference on feature extraction. Subsequently, the image is processed using a 5x5 window median filter to effectively suppress the salt and pepper noise caused by vegetation shielding or sensor failure in mountainous environments, while retaining key features such as river edges and terrain abruptness. Further, image data of the flood season and non-flood season in the past five years is extracted from historical monitoring records, and is randomly divided into a training set and a test set in a 7:3 ratio, wherein the training set contains submerged area samples under different water level conditions, enabling the convolutional encoder to learn dynamic features such as vegetation coverage changes and water turbidity differences during flood evolution.
[0139] Compared with the prior art, the conventional method usually directly uses the original resolution image for feature extraction, resulting in scale differences between different sensor data affecting the stability of model convergence. The existing noise reduction techniques mostly use mean filtering or Gaussian filtering, which can suppress noise but cause edge blurring, while median filtering can better preserve terrain details in complex texture scenes in mountainous areas. The conventional dataset construction only uses current monitoring period data, lacks historical multi-working condition sample support, and is difficult to cover extreme scenarios during mountain flood emergencies. The present embodiment unifies the resolution and enhances the data through strategies, making the model input have spatio-temporal consistency; through noise suppression and sample expansion, the robustness of the feature extraction network to interference factors is improved.
[0140] In some embodiments, unifying the resolution of view data refers to adjusting images of different sources or collected at different times to the same spatial resolution, which can be realized by using a bilinear interpolation or bicubic interpolation algorithm to eliminate scale deviations caused by sensor differences or changes in acquisition conditions.
[0141] In some embodiments, normalization processing refers to mapping image pixel values to a fixed numerical range, which can be realized by using the maximum and minimum value normalization method to make image data under different light conditions comparable.
[0142] In some embodiments, the median filtering method refers to using a sliding window to sort image pixels and replacing the original value with the median value, which can be realized by using a 3x3 or 5x5 filter window to remove impulse noise while preserving edge features.
[0143] In some embodiments, the historical image data collection refers to collecting monitoring area images under different seasons and weather conditions in the past, which can be stored by timestamp classification through a database management system to construct a dataset covering multiple scenarios to enhance the model generalization ability.
[0144] In the present embodiment, the training phase of the convolutional encoder includes the following steps:
[0145] Step 101, mark part of the samples in the centroid part of the training set;
[0146] Step 102, calculate the centroid of the unmarked samples;
[0147] Step 103, calculate the distance array of the training set to the centroid and incrementally sort;
[0148] Step 104, mark part of the samples in the distance array;
[0149] Step 105, repeat steps 102 to 104, and the ending condition is that the proportion of unmarked samples is less than a second set threshold;
[0150] Step 106, delete the marked samples and keep the unmarked samples to obtain the final training samples;
[0151] Step 107, select multiple subspaces, each of which is constructed from the final training samples obtained in step 106;
[0152] Step 108, calculate the experience kernel mapping corresponding to each subspace;
[0153] Step 109, calculate the augmented experience feature vector of each subspace;
[0154] Step 110, construct the target matrix of the training samples;
[0155] Step 111, output the augmented weight vector.
[0156] It should be noted that the key samples are selected by the iterative marking mechanism, and in each marking process, the samples far from the current centroid are preferentially selected, the redundant data is gradually eliminated, and when the proportion of unmarked samples is lower than a threshold, the selection is terminated, and the samples with high distinguishability are retained to construct multiple subspaces; each subspace is converted to generate a unique feature expression through kernel mapping, and finally the dynamic fusion of multiple space features is realized through the augmented weight vector. This process effectively reduces the data marking amount, and at the same time, the expression ability of the model to complex hydrological features is enhanced through multi-view feature learning.
[0157] In some embodiments, the centroid marking refers to a dynamic screening mechanism based on the sample distribution center point, and specifically, the k-means clustering algorithm can be used to determine the initial centroid, and the centroid position is adjusted through iterative calculation, which is used to identify representative edge samples.
[0158] In some embodiments, the distance array sorting refers to arranging the Euclidean distances of samples to the centroid in ascending order, and specifically, the quicksort algorithm can be used to realize it, which is used to preferentially select abnormal samples far from the distribution center.
[0159] In some embodiments, the subspace construction refers to dividing the screened samples into multiple heterogeneous feature spaces. Specifically, a random sampling method can be used to generate different sample subsets, and each subspace corresponds to a unique feature distribution pattern.
[0160] In this embodiment, in step 2, the global feature of the re-extracted image feature includes the following steps:
[0161] In step 201, the acquired image feature is divided into fixed-size blocks, each block is flattened into a vector, and then linear projection is used to map it to a high-dimensional space to obtain a sequence.
[0162] In step 202, the sequence is input into the multi-head self-attention layer of the self-attention encoder, and the single-head attention is calculated, and then the multi-head attention is calculated in parallel to capture the long-distance dependency.
[0163] In step 203, two fully connected layers are applied to the output of the self-attention.
[0164] In step 204, steps 202 to 203 are repeated to gradually abstract the features, and finally the feature vector at the first position is taken as the global feature.
[0165] It should be noted that the image feature is first cut into multiple local blocks, and each block is vectorized to obtain an initial feature representation. The low-dimensional vector is lifted to a high-dimensional latent space through linear projection operation to establish a more rich feature representation basis. In the parallel calculation process of the multi-head attention mechanism, each attention head independently learns the correlation weight between different positions to capture the dependency relationship with a larger spatial span. The two fully connected layers perform nonlinear transformation on the attention output to eliminate redundant information and strengthen key features. The attention calculation and full connection processing are repeated to form a feature abstraction level, and finally the feature vector at the starting position of the sequence is selected as the global representation. This position usually corresponds to the overall semantic information of the image.
[0166] In some embodiments, the fixed-size block refers to dividing the image feature map into local areas with equal area. Specifically, a division method of 16x16 pixels can be used to realize it, and the spatial block is used to reduce the complexity of feature processing.
[0167] In some embodiments, the linear projection refers to mapping the flattened vector to a higher dimension through a fully connected layer. Specifically, a 512-dimensional embedding space can be used to realize it to enhance the feature expression ability.
[0168] In some embodiments, the multi-head self-attention layer refers to decomposing the attention mechanism into multiple parallel computing paths, which can be implemented by 8 independent computing heads, each of which focuses on different dimensional feature association relationships. The two-layer fully connected layer refers to stacking a nonlinear transformation layer after the attention output, which can be implemented by combining the ReLU activation function and layer normalization to promote feature deep fusion.
[0169] In some embodiments, the multi-head attention satisfies the expression:
[0170] MultiHeadAttention(X p )=Concat(head1,...,head h )W o ;
[0171] Where X p is the input sequence;
[0172] head i is the single-head attention output, and satisfies
[0173] Q i , K i , V i are query, key, and value matrices generated by linear projection, respectively;
[0174] W o is the output weight matrix.
[0175] By parallel computing multiple attention heads, different dimensional long-range dependencies (such as the association between the flood area and the surrounding terrain) are captured, making up for the limitation of the local receptive field in the prior art.
[0176] In the present embodiment,
[0177] After the classification decision basis is generated in step 3, the following step is further included before step 4:
[0178] The mapping results of the multiple feature spaces extracted by the convolutional encoder and the self-attention encoder are weighted and fused to obtain fused features, and the fused features are mapped to a segmentation map of the same size as the original input image to obtain risk assessment information with flood disaster image semantic information and feature position information.
[0179] It can be understood that after the convolutional encoder extracts the multi-scale local features, the self-attention encoder models the global context, dynamically weights and fuses the two types of features through the augmented weight vector, and generates the fused features containing both the detailed texture and the overall structure. The fused features are upsampled to the original image size layer by layer through the transpose convolution operation to form a segmentation map with the same spatial resolution as the input. The probability value of each pixel point in the segmentation map reflects the risk level of the position occurring flood, and the pixel coordinates correspond to the actual geographical space position, so that the output risk assessment information contains both disaster semantic categories and accurate geographical coordinate mapping relationship.
[0180] It can also be understood that compared with the prior art, the traditional method usually directly classifies the down-sampled feature map, resulting in loss of spatial information and inability to accurately delineate the flood boundary. In the prior art, feature fusion usually uses fixed weights or simple splicing, which is difficult to adaptively adjust the contribution of multi-scale features. The present scheme integrates local and global features through a dynamic weighting mechanism, realizes the synchronous preservation of semantic information and position information under the premise of maintaining the consistency of input and output spaces, and overcomes the defect of insufficient risk assessment accuracy in the traditional method.
[0181] Through the above technical solutions, the present embodiment solves the problem of semantic information loss caused by the lack of spatial correlation of multi-scale features, and at the same time avoids the positioning deviation of geographical coordinates caused by the reduction of feature map size. The risk assessment information generated thereby can accurately reflect the probability of flood occurrence and its spatial distribution, providing a data basis with both semantic abstraction ability and spatial positioning accuracy for subsequent flood depth calculation and loss assessment.
[0182] In some embodiments, weighted fusion refers to integrating the outputs of different feature spaces through a dynamic weight distribution mechanism, which can be implemented by multiplying the augmented weight vector with the feature vector matrix, and is used to solve the problem of semantic information loss caused by the lack of spatial correlation of multi-scale features. The fused features refer to the combination of multi-scale features after weight adjustment, which can be generated through cross-layer feature splicing and channel attention mechanism, and are used to enhance the feature distinction between flood disaster areas and background environment.
[0183] In some embodiments, the segmentation map refers to a probability distribution map with the same spatial resolution as the input image, which can be implemented by using transpose convolution layers and up-sampling layers to recover the spatial size step by step, and is used to preserve the position information of the flood inundation area boundary.
[0184] In some embodiments, semantic information refers to the probability of occurrence of flood disaster and the disaster level, which can be mapped to the category probability space through the fully connected layer and the softmax function, and is used to identify the flood risk degree of different regions in the image.
[0185] In some embodiments, the feature position information refers to the geographical coordinates of the flood inundation area in the original image, and can be specifically implemented by keeping the spatial size of the encoder consistent with that of the decoder feature map, so as to accurately locate the latitude and longitude range of the disaster area.
[0186] In the present embodiment,
[0187] In step 4, the channel-by-channel convolutional layer randomly inputs an image data for feature extraction, and the number of convolution kernels corresponds to the depth of the input image data. After completion, three 5x5 images are output. The three 5x5 image input values are input into the point-by-point convolutional layer, and the number of convolution kernels is selected as 1x1x3. After completion, three 5x5 risk data are output.
[0188] In the present embodiment,
[0189] In step 7, the weight is the matrix product of the feature vector of the hydrological data and the augmented weight vector.
[0190] It can be understood that when the dynamic fusion unit performs the weighted fusion operation, the feature vector of the hydrological data and the augmented weight vector are calculated interactively through matrix multiplication. The feature vector of the hydrological data contains real-time water level changes, rainfall intensity and other physical quantity information of the current monitoring point, which is converted into a unified dimensional vector expression through the feature encoding unit. The augmented weight vector is derived from the subspace feature learning result of the convolutional encoder on the historical image data set in the training stage, and contains the association mode of image features and flood risks under different space-time conditions. The matrix multiplication operation of the two constructs a weight generation mechanism based on data characteristics, so that the hydrological monitoring data can dynamically adjust the contribution of remote sensing data and risk data in the fusion process according to the current environmental state, while inheriting the image feature space distribution rule established in the training stage. This fusion method not only retains the core decision-making value of hydrological data for flood monitoring, but also introduces the feature association rules formed by historical data training through the augmented weight vector, effectively overcoming the adaptability problem of traditional static weight allocation method in complex space-time dynamic environment.
[0191] As shown in the accompanying Figure 5 , it is a comparison chart of original image information and monitoring image information, wherein Figure 5 the upper part of the image information in the accompanying Figure 5 is the original image information of Sichuan Liangshan area, and the lower part of the image information in the accompanying Figure 3 is the corresponding monitoring image information, and the upper left image is the downstream position of the upper right image, and the positions correspond to each other, and in combination with the accompanying Figure 4It can be seen that the monitoring method / system in the embodiment can accurately simulate the complex hydrological network structure in mountainous areas with lower calculation time; and through hierarchical feature extraction and dynamic fusion mechanism, the ability to capture the spatio-temporal dynamic characteristics is enhanced; and edge computing is used to reduce data processing delay and improve monitoring real-time performance, and finally the generated flood monitoring image information provides a reliable basis for disaster warning and emergency response.
[0192] Embodiment 3:
[0193] In order to make the technical solutions in the embodiment more clear, the monitoring process of the water level data in the hydrological data is described in detail, specifically:
[0194] The water level data monitoring includes the following steps:
[0195] The initial water level data is called,
[0196] The water level anomaly data is obtained through neural network fitting;
[0197] The water level anomaly data is received;
[0198] The water level anomaly data obtained through neural network fitting includes the following steps:
[0199] The water level difference between the monitoring points is obtained;
[0200] The average normal water level value between the monitoring points is obtained iteratively;
[0201] The water level high is obtained, and the difference between the water level high and the normal high difference is taken as the water level anomaly data.
[0202] It should be noted that during the system installation and debugging stage, the spatial position coordinates of the reference point and the monitoring point are obtained through the GNSS receiver to form the initial layout of the sensor network. During the water level data monitoring process, the liquid level data of each monitoring point is converted into an electrical signal by the sensor and transmitted to the data processing unit. The neural network fitting unit calls the water level field model to calculate the water level difference between the monitoring points, and iteratively optimizes the normal water level reference value of the ellipsoid surface, and finally outputs the water level anomaly data.
[0203] In order to make the technical solutions in the embodiment more clear, the specific calculation process is explained in detail, wherein,
[0204] The expression of the water level anomaly data represented as the difference between the water level high and the normal high difference satisfies:
[0205]
[0206] Wherein, W AB represents the water level difference between two monitoring points on the section;
[0207] represents the average normal water level value between two monitoring points on the section;
[0208] h AB Output normal high difference.
[0209] The water level difference between the monitoring points comprises the following steps:
[0210] The geocentric coordinates of the two monitoring points are obtained, and the expression satisfies:
[0211]
[0212] wherein (X, Y, Z) is the obtained geocentric coordinate;
[0213] is the radius of curvature of the quadrant;
[0214] B, L, H are the geodetic latitude, longitude and water level height, respectively.
[0215] The geocentric coordinates are converted into spherical coordinates, and the expression satisfies:
[0216]
[0217] wherein (ρ, θ, λ) is the spherical coordinate, representing the geocentric radial, geocentric latitude and geocentric longitude, respectively;
[0218] The water level field model is called to output the water level position, and the expression satisfies:
[0219]
[0220] wherein GM is the earth's gravitational constant;
[0221] α is the reference ellipsoid long axis;
[0222] ω is the earth's rotation angular velocity;
[0223] N is the maximum order of the water level field model, and in this embodiment, it is preferably 2190;
[0224] is the completely normalized position coefficient (cosine term);
[0225] is the completely normalized position coefficient (sine term);
[0226] is the completely normalized associated Legendre function;
[0227] n, m are the corresponding order and degree, respectively.
[0228] The water level difference between the two monitoring points is obtained, and the expression satisfies:
[0229] W AB =W A -W B ;
[0230] wherein, W A , W B are water level values of two monitoring points respectively.
[0231] The iterative acquisition of the average normal water level value between the monitoring points comprises the following steps:
[0232] The normal water level reference value on the ellipsoid is calculated, and the expression satisfies:
[0233] γ0=γ e (1+a2sin 2 B+a4sin 4 B+a6sin 6 B+…);
[0234] wherein, a2, a4… are taken from the ellipsoid parameters.
[0235] The water level correction term is introduced, and the expression satisfies:
[0236]
[0237] wherein, k1, k2 control the core coefficient of the water level correction term.
[0238] The initial difference is iteratively optimized to obtain the average normal water level value satisfying the termination condition
[0239] wherein, the termination condition is preferably: < water level data threshold, and the water level data threshold is preferably 0.001 m.
[0240] Based on the above content, the geocentric coordinates are converted into spherical coordinates by the coordinate conversion module, input into the preset water level field model, and the water level values of two monitoring points are directly output, and the water level difference is obtained through difference operation; in the average normal water level value calculation stage, first, the reference value is generated based on the ellipsoid water level formula, and then the water level correction term is superimposed to compensate the actual water level, forming the initial difference parameter; through the cyclic iteration process, the initial difference is substituted into the BP neural network for weight optimization, and when the output result satisfies the preset convergence condition, the calculation is terminated, and finally the accurate average normal water level value is obtained; the above process is through automatic data conversion and model calculation, replacing manual measurement and visual calibration, eliminating subjective error sources.
[0241] The above merely describes the preferred embodiments of the present application, and is not intended to limit the patent scope of the present application, and any equivalent structure or equivalent process conversion, or direct or indirect application in other related technical fields, which are made by using the content of the present application specification and drawings, are also included in the patent protection scope of the present application.
Claims
1. A flood monitoring system for mountainous hydrological environments, characterized in that, The system includes: The data acquisition module includes a hydrological monitoring unit, a remote sensing unit, and a viewing unit. The hydrological monitoring unit is set at several monitoring points to acquire hydrological data within the monitored area. The remote sensing unit is used to acquire remote sensing data within the monitored area, and the viewing unit is used to acquire view data within the monitored area. The edge processing module includes a convolutional encoder, a self-attention encoder, and a decoding unit. The convolutional encoder is used to extract image features from low to high scales of the view data. The self-attention encoder is used to extract global features of the image. The training phase output of the convolutional encoder has an augmented weight vector to perform weighted fusion of the mapping results of multiple feature spaces extracted by the convolutional encoder and the self-attention encoder to generate a classification decision basis. The decoding unit is also used to determine risky data and non-risky data based on the classification decision basis. The deep fusion module includes a spatiotemporal alignment unit, a feature encoding unit, a dynamic fusion unit, and a risk monitoring unit. The spatiotemporal alignment unit is used to unify the spatiotemporal reference of risk data, hydrological data, and remote sensing data. The feature encoding unit is used to extract deep features from each data to generate a feature vector of a unified dimension. The dynamic fusion unit is used to adjust the weights through the feature vectors of hydrological data to weightedly fuse the feature vectors of remote sensing data, hydrological data, and risk data, and generate a flood monitoring map.
2. The flood monitoring system for mountainous hydrological environment as described in claim 1, characterized in that, The convolutional encoder also includes a training unit equipped with a ResNet50 network model. The training unit is used to pre-train the ResNet50 network model using a multi-empirical kernel learning machine.
3. A flood monitoring system for mountainous hydrological environments as described in claim 2, characterized in that, The decoding unit includes a decoder; The first layer of the decoder is a channel-wise convolutional layer to receive the classification decision basis; The second layer of the decoder is an activation function; The third layer of the decoder is a transposed convolutional layer, which halves the number of channels and doubles the spatial size; The fourth layer of the decoder is an activation function; The fifth layer of the decoder is an upsampling layer to increase the spatial size by 4 times; The sixth layer of the decoder is a pointwise convolutional layer, which reduces the number of channels to the initial number of channels in the model to keep the spatial size unchanged.
4. A flood monitoring method for the hydrological environment of mountainous areas, characterized in that, The flood monitoring system for mountainous hydrological environment according to claim 3 includes the following steps: Step 1: Establish the watershed topology within the monitored area to set up several monitoring points and acquire hydrological data, remote sensing data, and visual data within the monitored area. Step 2: Extract the image features of the view data from low to high scale, and then extract the global features of the image features. Step 3: Weighted fusion of the mapping results of multiple feature spaces extracted by the convolutional encoder and the self-attention encoder, and generation of classification decision basis; Step 4: Based on the classification criteria, determine whether the data is risky or non-risky. Step 5: Unify the spatiotemporal benchmarks for risk data, hydrological data, and remote sensing data; Step 6: Extract deep features from each data point to generate a feature vector with a unified dimension; Step 7: Adjust the weights of the feature vectors of hydrological data to weightedly fuse the feature vectors of remote sensing data, hydrological data, and risk data, and generate a flood monitoring map. The hydrological data includes water level data, rainfall data, and flow data; the risk data is the image features when the probability of flooding is greater than a first set threshold; the flood monitoring map includes inundation depth matrix information and inundation loss thermal information.
5. A flood monitoring method for mountainous hydrological environment as described in claim 4, characterized in that, After acquiring hydrological data, remote sensing data, and view data within the monitored area, and before performing multi-scale extraction of image features from low to high levels from the view data, the following steps are also included: Unify the resolution of view data and normalize the image data; The image data is denoised using median filtering. Historical image data is acquired and aggregated to form an image dataset. The image dataset is then divided into a training set and a test set, with the training set used for the training phase of the convolutional encoder.
6. A flood monitoring method for mountainous hydrological environments as described in claim 5, characterized in that, The training phase of the convolutional encoder includes the following steps: Step 101: Label a portion of the samples in the centroid region of the training set; Step 102: Calculate the centroids of the unlabeled samples; Step 103: Calculate the distance array from the training set to the centroid and sort it in ascending order; Step 104: Label a subset of samples in the distance array; Step 105: Repeat steps 102 to 104. The termination condition is: the proportion of unlabeled samples is less than the second set threshold. Step 106: Delete labeled samples and keep unlabeled samples to obtain the final training samples; Step 107: Select multiple subspaces, each subspace being constructed from the final training samples obtained in step 106; Step 108: Calculate and generate the empirical kernel mapping corresponding to each subspace; Step 109: Calculate the augmented empirical feature vector for each subspace; Step 110: Construct the target matrix of the training samples; Step 111: Output the augmented weight vector.
7. A flood monitoring method for mountainous hydrological environment as described in claim 4, characterized in that, Step 2, the global feature extraction of the image features includes the following steps: Step 201: Divide the acquired image features into blocks of fixed size, flatten each block into a vector, and then map it to a high-dimensional space through linear projection to obtain a sequence; Step 202: Input the sequence into the multi-head self-attention layer of the self-attention encoder, calculate single-head attention, and then calculate multi-head attention in parallel to capture long-distance dependencies; Step 203: Apply two fully connected layers to the output of self-attention; Step 204: Repeat steps 202 to 203 to gradually abstract features, and finally take the feature vector of the first position as the global feature.
8. A flood monitoring method for mountainous hydrological environment as described in claim 4, characterized in that, In step 3, after generating the classification decision criteria, and before step 4, the following steps are included: The mapping results of multiple feature spaces extracted by the convolutional encoder and the self-attention encoder are weighted and fused to obtain fused features. The fused features are then mapped onto a segmentation map of the same size as the original input image to obtain risk assessment information with semantic information and feature location information of flood disaster images.
9. A flood monitoring method for mountainous hydrological environment as described in claim 4, characterized in that, In step 4, a random image data is input into the channel-wise convolutional layer for feature extraction. The number of convolutional kernels corresponds to the depth of the input image data. After completion, three 5×5 images are output. In the pointwise convolutional layer with the input values of the three 5×5 images, the number of convolutional kernels is 1×1×3. After completion, three 5×5 risk data are output.
10. A flood monitoring method for mountainous hydrological environment as described in claim 6, characterized in that, In step 7, the weight is the matrix multiplication of the feature vector of the hydrological data and the augmented weight vector.