Water supply network risk prediction method based on image recognition

Abstract

The invention discloses a water supply network risk prediction method based on image recognition, and belongs to the technical field of water supply network monitoring, and the method specifically comprises the steps: obtaining a multi-temporal remote sensing image and network geographic information data of a target region; dividing pipe section units based on pipeline coordinates and establishing corresponding banded analysis areas; extracting multispectral and thermal infrared features of pixels in the region, generating pixel feature vectors, and aggregating the pixel feature vectors into reference vectors of the pipe section units; constructing a pipe network characteristic graph according to a pipe network topology connection relation, and calculating spatial distribution parameters of the pipe network characteristic graph to learn a positive zone spatial gradient rule of environmental characteristics along the pipe network; identifying abnormal nodes by comparing the deviation between the current state of each pipe section unit and the state predicted based on the topology law; and calculating the dispersion of the pixel features in the region corresponding to the abnormal node for verification, and outputting a high-confidence risk pipe section position and early warning information. According to the invention, accurate identification of the leakage risk of the water supply network is realized.

Description

Technical Field

[0001] This invention relates to the field of water supply network monitoring technology, specifically to a method for predicting water supply network risks based on image recognition. Background Technology

[0002] Water supply networks are a core component of urban infrastructure, and their safe and stable operation is directly related to water resource security and public safety. Traditional network monitoring mainly relies on manual inspections and fixed-point sensors, which suffers from low efficiency, narrow coverage, and insensitivity to minor leaks. In recent years, with the popularization of remote sensing technology, using multispectral and thermal infrared images of the Earth's surface acquired by satellites or drones for indirect identification of network risks has become an important technological development direction due to its advantages such as wide coverage and non-contact operation.

[0003] In existing technologies, image recognition-based risk prediction methods primarily rely on analyzing surface anomalies in remote sensing images. Specifically, one common approach is to use thermal infrared bands to retrieve surface temperature and detect areas of temperature anomalies to infer potential underground leaks. Another approach involves analyzing visible and near-infrared bands to monitor spatial variations in vegetation growth or soil moisture along pipeline routes, thereby indirectly assessing the health of the pipeline network. Both methods aim to establish a correlation model between surface image features and the condition of underground pipeline networks.

[0004] However, these existing technical solutions still face significant challenges. First, features such as surface temperature and vegetation index are easily affected by various environmental factors, including meteorological conditions, seasonal changes, land use types, and irrigation activities. This makes it extremely difficult to effectively separate weak signals caused by pipeline leakage from complex background noise, resulting in a high false alarm rate. Second, existing methods often employ segmentation based on fixed thresholds or single statistical learning models. Their judgment logic often lacks consideration of underground physical processes (such as water transport and heat conduction), limiting the interpretability and generalization ability of the models across different geographical environments. More importantly, existing technologies typically treat image analysis units (such as pixels or regions) as independent entities, failing to fully utilize the inherent, continuous topological connections of the water supply network—a crucial spatial structural information. Consequently, they fail to construct a benchmark model that reflects the inherent spatial correlation and gradual changes in environmental characteristics along the pipeline network, limiting their ability to proactively identify localized minor anomalies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the risks of water supply networks based on image recognition, and to solve the following technical problems: Existing technologies typically treat image analysis units as independent entities, failing to fully utilize the important spatial structural information of the inherent and continuous topological connections of the water supply network. Consequently, they fail to construct a benchmark model that reflects the inherent spatial correlation and gradual change patterns of the environmental characteristics along the network, limiting their ability to proactively identify local minor anomalies.

[0006] The objective of this invention can be achieved through the following technical solutions: A risk prediction method for water supply networks based on image recognition includes the following steps: S1, acquire multi-temporal remote sensing images and pipeline geographic information data within the target area. The multi-temporal remote sensing images consist of pixels arranged in a regular grid. Each pixel contains measurement values ​​of thermal infrared band and multiple optical bands. The geographic information data includes the spatial coordinates and topological connection relationships of the pipelines. S2, based on the spatial coordinates, the pipeline is divided into multiple pipe segment units, and the corresponding strip-shaped analysis area is formed by expanding the projection of each pipe segment unit as the center line; S3, extract the thermal infrared radiation value and optical band reflectance of the pixels in each strip analysis area in the multi-temporal remote sensing image, and generate the feature vector of each pixel in the strip analysis area. S4, aggregate all pixel feature vectors corresponding to each pipe segment unit and calculate the statistical center value, and set the statistical center value as the reference vector of the pipe segment unit; S5. Based on the topological connection relationship, construct a pipeline network feature map with each pipe segment unit as a node, and calculate the spatial distribution parameters of each pipe segment unit on the pipeline network feature map. S6. Calculate the current reference vector of each pipe segment unit based on the current time-phase remote sensing image and update the attributes of the corresponding nodes in the pipe network feature map. Calculate the prediction deviation of the updated attribute values ​​of each node according to the spatial distribution parameters, and mark nodes whose prediction deviation exceeds a preset threshold as abnormal nodes. S7. Calculate the dispersion of the pixel feature vector in the strip analysis area corresponding to the abnormal node relative to the reference vector of the node. When the dispersion exceeds the set range, output the spatial coordinates of the pipe segment unit corresponding to the abnormal node and issue an early warning message.

[0007] As a further aspect of the present invention: the specific process for generating the banded analysis region in step S2 is as follows: Read the spatial coordinate sequence of pipelines in the geographic information data, wherein the spatial coordinate sequence is a series of coordinate points arranged in connection order; Connecting any two adjacent coordinate points forms a spatial line segment, and each of these spatial line segments is defined as a pipe segment unit. Using the spatial line segment corresponding to each pipe segment unit as the central reference line, the system extends to both sides along the vertical direction of the central reference line to form a strip-shaped polygonal region with a preset width. The strip polygon regions corresponding to pipe segments that belong to the same pipeline and are connected to each other are merged at the connection points of their spatial line segments to generate continuous polygon regions. The vector boundary of the continuous polygonal region is integrated with the pixel grid of the multi-temporal remote sensing image in a spatial coordinate system and overlaid. The image pixels whose center point is located within the continuous polygonal region are identified as the strip analysis area of ​​the pipe segment unit.

[0008] As a further aspect of the present invention: in step S3, the specific process of generating the feature vector of each pixel within the strip analysis area is as follows: For each pixel within the strip analysis area, the thermal infrared radiance value of that pixel in each time phase and the reflectance value in each optical band are extracted from the multi-temporal remote sensing image. The mean value of the thermal infrared radiation value of the pixel in all time phases is calculated as the first characteristic value; the mean value of the reflectance value of the pixel in each optical band in all time phases is calculated to obtain multiple second characteristic values ​​corresponding to each optical band. The first feature value and all the second feature values ​​are combined into a one-dimensional array in a predetermined fixed order, wherein the first feature value occupies the first element position of the array, and the remaining element positions are filled with the second feature values ​​corresponding to each optical band in sequence. This one-dimensional array is defined as the feature vector of the pixel.

[0009] As a further aspect of the present invention: in step S4, the process of obtaining the reference vector of the pipe segment unit is as follows: Obtain the feature vectors of all pixels in the strip analysis area corresponding to the current pipe segment unit; arrange all the obtained feature vectors to form a two-dimensional data matrix, where each row of the matrix corresponds to the feature vector of a pixel, and each column corresponds to the value at a fixed sorting position in the feature vector. For each column of the two-dimensional data matrix, the arithmetic mean of all values ​​in that column is calculated. The arithmetic means of each column are arranged in the fixed order to obtain a one-dimensional array, and this one-dimensional array is defined as the reference vector of the pipe segment unit.

[0010] As a further aspect of the present invention: the specific calculation process of the spatial distribution parameters in S5 is as follows: Based on the topological connection relationship, each pipe segment unit is identified as an independent node; if two pipe segment units are recorded as directly connected in the topological connection relationship, an undirected edge is established between the nodes corresponding to these two units to generate a pipe network feature map. The baseline vector of each pipe segment unit is used as a set of numerical attributes and assigned to the corresponding node in the pipe network feature graph. For each node in the pipe network feature graph, all neighboring nodes directly connected to it through edges are retrieved. The difference vector between the numerical attributes of the node itself and the numerical attributes of each of its neighboring nodes is calculated. The difference vectors generated by the node and all its neighboring nodes are collected. Based on the arrangement order of nodes in the pipeline feature diagram, construct a square matrix with the number of rows and columns equal to the total number of nodes as a relation matrix. For each element in the relation matrix, if there is an edge between two nodes corresponding to its row and column, the value of the element is the modulus of the difference between the numerical attributes of the two nodes; if there is no edge between the two nodes, the value of the element is zero. Perform matrix spectral decomposition on the assigned relation matrix to obtain its eigenvalues ​​and corresponding eigenvectors. Select the smallest eigenvalue from all eigenvalues ​​greater than zero and obtain its corresponding eigenvector. Record the specific value of the selected eigenvalue and the order of the components in its corresponding eigenvector. Use the recorded value and the order of arrangement to define the spatial distribution parameters of each pipe segment unit.

[0011] As a further aspect of the present invention: in step S6, the specific process for determining abnormal nodes is as follows: Acquire remote sensing images of the current time phase. For each pipe segment unit, extract the thermal infrared radiation value and optical band reflectance of each pixel in its strip analysis area at the current time phase, generate the feature vector of each pixel at the current time phase, and aggregate to calculate the reference vector of the pipe segment unit at the current time phase. Replace the reference vector associated with each node in the pipeline feature graph with the reference vector calculated in the current time phase; for each node in the updated pipeline feature graph, obtain all neighboring nodes directly connected to it through connecting edges, and obtain the updated reference vector of the neighboring nodes. The feature vector in the spatial distribution parameters is obtained. According to the fixed order of the nodes in the pipeline feature map, the values ​​of each component of the feature vector are sequentially assigned to the neighboring nodes as weight coefficients. The order of the neighboring nodes is consistent with the allocation order of the feature vector components. Using the assigned weight coefficients, a weighted summation operation is performed on the updated reference vectors of the neighboring nodes to calculate a prediction vector for the node; the Euclidean distance between the node's current reference vector and the prediction vector is calculated and calibrated as the prediction deviation; if the prediction deviation of the node is greater than a preset threshold, the node is marked as an abnormal node.

[0012] As a further aspect of the present invention: in step S7, the specific calculation process of the discreteness is as follows: For each pipe segment unit marked as an abnormal node, its strip analysis area is determined, and the feature vectors of all pixels in the strip analysis area are obtained. At the same time, the reference vector corresponding to the abnormal node is obtained. Calculate the difference between the feature vector of each pixel and the reference vector in each dimension to obtain the difference vector corresponding to the pixel; calculate the square of the difference in all dimensions in each difference vector and sum them, then take the square root of the sum to obtain the distance scalar corresponding to the pixel; accumulate the distance scalars corresponding to all pixels in the strip analysis area and sum them, then divide the sum by the total number of pixels in the strip analysis area to calculate the reference distance scalar value, and define the reference distance scalar value as the dispersion corresponding to the anomaly node.

[0013] The beneficial effects of this invention are: 1) This invention establishes an environmental state benchmark based on long-term multi-temporal observations for each pipe segment unit. It is understood that, under conditions free from external disturbances (such as leakage), the thermal infrared and multispectral characteristics of any surface location, while fluctuating with natural conditions, will generally revolve around a long-term stable mean. This invention first delineates a strip-shaped analysis area strictly corresponding to the spatial location of each pipe segment unit. Then, it extracts the observation values ​​of all pixels within this area at different times from historical images and fuses them to calculate a feature vector representing the long-term average state of the area. This vector is the "environmental fingerprint" of the pipe segment, inherently containing the statistical characteristics of all inherent environmental noise at that location. During risk detection, the currently observed state is compared with this long-term benchmark. Since the benchmark has absorbed normal environmental fluctuation patterns, significant deviations between the current state and the benchmark are more likely to be attributed to abnormal disturbances such as leakage, rather than natural changes. This principle of "first establishing a personalized normal baseline, then detecting deviations" fundamentally improves the robustness and accuracy of identifying early, weak leakage signals in variable environments.

[0014] 2) This invention abstracts a physically continuous pipeline network into a mathematical graph structure and uses this structure to learn the spatial propagation laws that the surface features along a healthy pipeline network should follow. Specifically, each pipe segment unit and its environmental reference are mapped to a node and its attributes in the graph, while the actual physical connections between pipes constitute the edges in the graph. This mapping constructs a spatial topology model of the pipeline network. By analyzing the distribution of node attributes on the graph structure, we can quantify the spatial continuity pattern of "how the environmental characteristics of the upstream pipe segment should smoothly influence and transition to the downstream adjacent pipe segment under a leak-free state." It is understandable that when leakage occurs, the leak point will disrupt this continuous and gradual relationship between its location and the upstream and downstream, causing the measured attributes of the node to be unpredictable by the state of its neighboring nodes through the learned patterns, thus producing a significant prediction bias. The anomaly manifests as a disruption of the spatial continuity of the pipeline network. It is understandable that since the learning objective of this model is directly derived from the inherent physical connection relationship of the pipeline network, rather than the statistical characteristics of a specific dataset, its core logic does not rely on a large number of labeled samples. It can adapt to pipeline networks of different regions and configurations, showing excellent generalization ability. Furthermore, its "prediction based on connection relationship - comparison bias" decision-making process is easy to understand and verify.

[0015] 3) This invention transforms the absolute analysis of a single image into continuous state tracking: by establishing and continuously comparing long-term historical benchmarks with real-time observation data, it can capture subtle trends in surface environmental parameters deviating from their normal state. This dynamic comparison mechanism allows the method to perceive the early stages before leakage causes obvious surface traces, thus advancing the warning window; spatially, after identifying potentially abnormal pipe sections, it further implements local consistency verification. It is understandable that anomalies caused by different factors have fundamentally different spatial distribution patterns. Anomalies caused by local events such as point leakage will cause drastic changes in the characteristics of some pixels within the affected area, while other pixels remain relatively stable, resulting in high dispersion of feature values; conversely, uniform changes caused by widespread natural factors such as rainfall and seasonal changes will cause a consistent shift in feature values ​​across the entire area, with lower dispersion. By calculating and analyzing the degree of feature dispersion within the abnormal area, this method can effectively distinguish between local leakage signals and wide-area environmental noise, thereby performing secondary filtering and confirmation of the initial warning. This dual-criteria system, combining "spatiotemporal dynamic deviation monitoring" and "spatial local consistency verification," means that the output of this invention is no longer a single risk coordinate, but a multi-dimensional risk profile that includes anomaly intensity, spatial distribution pattern, and warning confidence level. It represents an evolution from simply reporting abnormal "locations" to accurately assessing the "nature" and "priority" of risks, providing intelligent support for inspection decision-making. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of a water supply network risk prediction method based on image recognition according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a method for predicting the risk of water supply networks based on image recognition, comprising the following steps: S1, acquire multi-temporal remote sensing images and pipeline geographic information data within the target area. The multi-temporal remote sensing images consist of pixels arranged in a regular grid. Each pixel contains measurement values ​​of thermal infrared band and multiple optical bands. The geographic information data includes the spatial coordinates and topological connection relationships of the pipelines. The multi-temporal remote sensing images are acquired via satellite or UAV platforms, covering the area traversed by the target water supply pipeline network. The image data must include thermal infrared bands and at least three optical bands: red, green, and near-infrared. Thermal infrared measurements are used to retrieve surface brightness temperature, and optical band reflectance is used to calculate spectral characteristics such as the normalized difference vegetation index (NDVI) and humidity index. Image pixels are arranged in a regular geographic coordinate grid, with each pixel corresponding to a fixed-size rectangular area on the surface (e.g., 10m x 10m). The geographic information system (GIS) data is in vector format, with the pipeline layer stored as polyline features. Each polyline is defined by a series of ordered vertex coordinates, containing a unique identifier. The connections between pipelines are defined through a node topology table or directly in the spatial data model, clearly recording upstream and downstream connections and intersections between pipeline segments to form a complete pipeline network topology. Remote sensing images and GIS data must be spatially registered in a unified coordinate system.

[0020] S2, based on the spatial coordinates, the pipeline is divided into multiple pipe segment units, and the corresponding strip-shaped analysis area is formed by expanding the projection of each pipe segment unit as the center line; The target pipeline's polyline coordinate sequence is retrieved from Geographic Information System (GIS) data. A straight line segment formed by two adjacent coordinate points is defined as a basic pipeline segment unit. A dedicated strip analysis area is established for each basic pipeline segment unit: using the vertical projection of the spatial line segment of that segment onto the Earth's surface as the center line, and expanding to both sides vertically along the center line according to a preset width rule (e.g., 10 meters on each side based on typical pipeline burial depth and diameter), a continuous polygonal area covering the area directly above and laterally influencing the pipeline is generated. If the pipeline is a continuous polyline, the polygonal strip analysis areas of adjacent pipeline segment units are merged at their connection points to ensure a seamless, continuous monitoring strip along the pipeline's direction. Finally, through spatial overlay analysis, all remote sensing image pixels whose center points fall within this continuous polygonal area are identified as the set of pixels belonging to the strip analysis area of ​​that pipeline segment unit.

[0021] S3, extract the thermal infrared radiation value and optical band reflectance of the pixels in each strip analysis area in the multi-temporal remote sensing image, and generate the feature vector of each pixel in the strip analysis area. For each pixel within the defined analysis zone, the thermal infrared radiance value and reflectance values ​​of each optical band (red, green, near-infrared, etc.) for all historical time phases are extracted from its corresponding multi-temporal image sequence. Statistical analysis is performed on the time series data for each band: the arithmetic mean of the thermal infrared radiance value of the pixel across all time phases is calculated and denoted as the thermal infrared characteristic value of the pixel; the arithmetic mean of the reflectance of each optical band for the pixel across all time phases is calculated to obtain a series of optical characteristic values. The thermal infrared characteristic value and all optical characteristic values ​​are concatenated in a predefined, fixed order (e.g., [thermal infrared mean, red light mean, green light mean, near-infrared mean]) to form a one-dimensional numerical array. This one-dimensional array is defined as the feature vector of the pixel, which comprehensively characterizes the stable spectral and thermal radiation state of the pixel location during long-term observation.

[0022] S4, aggregate all pixel feature vectors corresponding to each pipe segment unit and calculate the statistical center value, and set the statistical center value as the reference vector of the pipe segment unit; For a given pipe segment, feature vectors of all pixels within its strip analysis area are collected. These feature vectors are then arranged as row vectors in pixel order to construct a two-dimensional data matrix. Each column of this matrix corresponds to a feature value at a fixed position in the feature vectors (e.g., the first column consists entirely of the average thermal infrared values ​​of all pixels). The arithmetic mean of all values ​​in each column of this two-dimensional data matrix is ​​calculated to obtain a new feature vector. This new vector has the same dimension as the pixel feature vector, and the value of each dimension represents the average level of the corresponding feature within the area covered by the pipe segment under long-term historical conditions. This new vector is formally defined as the baseline vector for the pipe segment, characterizing the inherent, average "health" of the surface environment above the pipeline during periods of no leakage disturbance.

[0023] S5. Based on the topological connection relationship, construct a pipeline network feature map with each pipe segment unit as a node, and calculate the spatial distribution parameters of each pipe segment unit on the pipeline network feature map. Each pipe segment unit is abstracted as a graph node, and the reference vector of each pipe segment unit is used as the attribute of that node. Based on the pipe network topology connections defined in the Geographic Information System (GIS) data, if two pipe segments are physically directly connected, an undirected edge is established between their corresponding two graph nodes. All nodes and edges constitute an abstract pipe network feature graph, which accurately reflects the topology of the real pipe network. To quantify the spatial association and smooth variation of node attributes (i.e., reference vectors) in this graph structure, a graph Laplacian matrix is ​​constructed based on this graph: the rows and columns of the matrix correspond to the node order; if there is an edge connecting two nodes, the corresponding matrix element value is the reciprocal of the Euclidean distance between the reference vectors of these two nodes (or a negative constant); if there is no connection, it is 0; the diagonal elements of the matrix are the negatives of the sum of the absolute values ​​of all off-diagonal elements in the row. Eigenvalue decomposition is performed on this graph Laplacian matrix to extract its smallest non-zero eigenvalue and its corresponding eigenvector. The minimum eigenvalue quantifies the overall smoothness of attribute changes on the graph, while its corresponding eigenvector indicates the smoothest "intrinsic" association pattern of each node under the constraints of the graph structure. This set of eigenvalues ​​and eigenvectors are recorded together as the spatial distribution parameters, which encode the prior knowledge of spatial continuity that the surface environmental characteristics along the healthy pipeline network should follow.

[0024] S6. Calculate the current reference vector of each pipe segment unit based on the current time-phase remote sensing image and update the attributes of the corresponding nodes in the pipe network feature map. Calculate the prediction deviation of the updated attribute values ​​of each node according to the spatial distribution parameters, and mark nodes whose prediction deviation exceeds a preset threshold as abnormal nodes. Acquire a latest (current time phase) remote sensing image. For each pipe segment unit, repeat steps S3 and S4: extract the band values ​​of all pixels within its strip analysis area for the current time phase, generate the pixel feature vector for the current time phase, and aggregate to calculate the current reference vector for that pipe segment unit. Then, replace the original attribute values ​​of the corresponding nodes in the network feature map with the current reference vectors of each pipe segment unit, completing the node attribute update. For any node in the updated map, use the spatial distribution parameters obtained in S5 (especially the feature vector), combined with the updated reference vectors of all its neighboring nodes, to perform a weighted sum to calculate the predicted value of the node's attribute: the weights are taken from the components in the feature vector corresponding to the positions of each neighboring node. Calculate the Euclidean distance between the node's current reference vector and this predicted value; this distance is the prediction deviation of the node. This deviation quantifies the degree to which the node's current state deviates from the "expected normal state" determined by the network topology and neighboring states. After calculating the prediction deviation for all nodes, set a global deviation threshold. Nodes whose prediction deviation consistently exceeds this threshold (e.g., whose deviation value is consistently significantly higher than the network average throughout the analysis period) are marked as anomalous nodes. The core of this step is to use spatial correlation under network topology constraints for anomaly detection: a real leak point will disrupt its environmental continuity with upstream and downstream pipe segments, causing its state to be unpredictable by its neighbors, thus generating high deviation.

[0025] S7. Calculate the dispersion of the pixel feature vector in the strip analysis area corresponding to the abnormal node relative to the reference vector of the node. When the dispersion exceeds the set range, output the spatial coordinates of the pipe segment unit corresponding to the abnormal node and issue an early warning message.

[0026] For each marked anomalous node (i.e., its corresponding pipe segment unit), internal consistency verification is performed. The feature vectors of all pixels within the strip analysis area of ​​that pipe segment unit in the current time phase are obtained, along with the node's own (long-term historical) baseline vector. For each pixel within the area, the Euclidean distance between its current feature vector and the baseline vector is calculated. This distance measures the degree to which the pixel's current position deviates from the long-term average state of the area. Next, the average of the Euclidean distances from all pixels to the baseline vector is calculated; this average is defined as the "dispersion" corresponding to the anomalous node. This dispersion index reveals the heterogeneity within the anomalous area: if the leakage is a local point-source event, the features of a small number of pixels near the leak point will change drastically (large distance), while other pixels will change little, resulting in high overall dispersion; if it is a large-scale uniform environmental change (such as rainfall), the features of all pixels will consistently deviate from the baseline, resulting in a lower calculated dispersion. A dispersion threshold is set. If the dispersion of an anomaly node exceeds this threshold, it indicates the existence of a highly heterogeneous, localized anomaly pattern within it, consistent with the characteristics of point source leakage, thus confirming the anomaly as a high-confidence risk. Finally, the system outputs the geographic coordinates, prediction deviation, and dispersion of the pipe segment units corresponding to all anomaly nodes that pass this verification, and generates tiered early warning information.

[0027] In a preferred embodiment of the present invention, the specific process of generating the banded analysis region in step S2 is as follows: Read the spatial coordinate sequence of pipelines in the geographic information data, wherein the spatial coordinate sequence is a series of coordinate points arranged in connection order; Connecting any two adjacent coordinate points forms a spatial line segment, and each of these spatial line segments is defined as a pipe segment unit. Using the spatial line segment corresponding to each pipe segment unit as the central reference line, the system extends to both sides along the vertical direction of the central reference line to form a strip-shaped polygonal region with a preset width. The strip polygon regions corresponding to pipe segments that belong to the same pipeline and are connected to each other are merged at the connection points of their spatial line segments to generate continuous polygon regions. The vector boundary of the continuous polygonal region is integrated with the pixel grid of the multi-temporal remote sensing image in a spatial coordinate system and overlaid. The image pixels whose center point is located within the continuous polygonal region are identified as the strip analysis area of ​​the pipe segment unit.

[0028] In another preferred embodiment of the present invention, the specific process of generating the feature vector of each pixel in the strip analysis area in step S3 is as follows: For each pixel within the strip analysis area, the thermal infrared radiance value of that pixel in each time phase and the reflectance value in each optical band are extracted from the multi-temporal remote sensing image. The mean value of the thermal infrared radiation value of the pixel in all time phases is calculated as the first characteristic value; the mean value of the reflectance value of the pixel in each optical band in all time phases is calculated to obtain multiple second characteristic values ​​corresponding to each optical band. The first feature value and all the second feature values ​​are combined into a one-dimensional array in a predetermined fixed order, wherein the first feature value occupies the first element position of the array, and the remaining element positions are filled with the second feature values ​​corresponding to each optical band in sequence. This one-dimensional array is defined as the feature vector of the pixel.

[0029] First, the thermal infrared radiance value of the pixel in each historical time phase and the reflectance value in each optical band (e.g., visible red, green, blue bands and near-infrared band) are extracted from the multi-temporal remote sensing images. Each pixel is an independent and fixed-position observation unit in each image, and its corresponding band measurement value is stored in a data cube arranged in time series. Therefore, the complete time series observation record of the point can be obtained by indexing its geographic coordinates and tracing back to the image data of all time phases. Subsequently, the arithmetic mean of the thermal infrared radiation values ​​of the pixel across all time phases is calculated, and this result is used as the first feature value. The arithmetic mean of the reflectance values ​​of the pixel in each optical band across all time phases is calculated to obtain multiple second feature values ​​corresponding to each optical band. The principle of calculating the mean is that for a point with a fixed geographical location, its long-term environmental state (such as average temperature and average vegetation cover) will fluctuate around a stable value without severe external interference. The arithmetic mean can effectively estimate and represent this long-term stable background state. By aggregating multi-temporal data into a mean, random noise such as seasonal changes and short-term weather fluctuations can be smoothed out, thereby extracting the most essential and stable environmental characteristics of the point. Finally, the first feature value and all the second feature values ​​are combined into a one-dimensional array in a predetermined and fixed order. The first feature value is placed in the first element position of the array, and the remaining elements are filled with the corresponding second feature values ​​in the order of the bands (e.g., red, green, near-infrared). This one-dimensional array is defined as the feature vector of the pixel. The principle of this fixed-order combination is that it creates a digital feature representation with a unified structure and fixed dimensions for each pixel. This vector integrates information from thermal infrared and multispectral dimensions, and the value of each dimension points to a specific physical or environmental attribute (such as heat, specific vegetation response). This allows subsequent calculations to consistently measure and compare all pixels in a multidimensional feature space with clear physical meaning.

[0030] This provides a robust and computationally feasible framework for extracting key information characterizing surface stability from massive, multi-temporal remote sensing imagery. By using multi-temporal fusion to "de-noise" and establish a "baseline"—that is, by calculating the average over time series—it effectively filters out instantaneous fluctuations caused by differences in solar angle, short-term meteorological events (such as single rainfall events or cloud shadows), and sensor random noise, thereby extracting stable signals that reflect the inherent environmental characteristics of the location over several months or years. This allows for the identification of weak anomalies caused by underground pipe network leakage, as leakage signals are often hidden within these long-term stable background features.

[0031] In another preferred embodiment of the present invention, the process of obtaining the reference vector of the pipe segment unit in step S4 is as follows: Obtain the feature vectors of all pixels in the strip analysis area corresponding to the current pipe segment unit; arrange all the obtained feature vectors to form a two-dimensional data matrix, where each row of the matrix corresponds to the feature vector of a pixel, and each column corresponds to the value at a fixed sorting position in the feature vector. For each column of the two-dimensional data matrix, the arithmetic mean of all values ​​in that column is calculated. The arithmetic means of each column are arranged in the fixed order to obtain a one-dimensional array, and this one-dimensional array is defined as the reference vector of the pipe segment unit.

[0032] By querying the spatial range of the strip analysis area under the jurisdiction of the pipe segment unit, the feature vectors corresponding to all pixels within it are completely retrieved. Then, all the acquired pixel feature vectors are arranged to form a two-dimensional data matrix. The arrangement rule is that each row of the matrix strictly corresponds to the complete feature vector of one pixel, while each column of the matrix corresponds to the value in the same fixed sorting position among all pixel feature vectors. For example, if the first element of each feature vector is the thermal infrared mean, then the first column of the matrix contains the thermal infrared mean data of all pixels. This construction principle ensures that data from different pixels but representing the same physical or environmental attribute are organized in the same column, establishing a regular data structure for subsequent column-oriented statistical analysis. Next, for each column of the two-dimensional data matrix, the arithmetic mean of all values ​​in that column is calculated. The principle of calculating the column average is that the values ​​in the same column come from the observation of the same environmental attribute of all pixels within the coverage area of ​​the pipe segment unit. Averaging these spatially distributed point data can effectively integrate regional information to obtain a single value that can represent the general level or central trend of the pipe segment unit as a whole in this attribute. Finally, following the inherent fixed order of the feature vectors, the arithmetic mean values ​​calculated from each column are arranged sequentially and combined to obtain a new one-dimensional array with the same dimension as the original pixel feature vector. This one-dimensional array is defined as the baseline vector of the pipe segment unit. This step is a natural integration of the column average values ​​from the previous step. It reassembles the average values ​​of each attribute obtained separately according to their original physical meaning order, thereby generating a completely new and comprehensive feature vector. This vector no longer represents a single pixel, but rather the average state of the entire pipe segment unit's strip region under long-term observation, i.e., the "environmental fingerprint" or health baseline of the pipe segment.

[0033] Understandably, while the features of a single pixel can reflect a point-like state, they are easily affected by local accidental factors and are too scattered, lacking engineering guidance significance. By spatially aggregating the feature vectors of all pixels within a pipe segment unit and calculating the arithmetic mean of each feature dimension, a spatial domain data fusion and smoothing is essentially performed. The direct benefit of this is that it can minimize the impact of measurement errors or outliers that may exist in individual pixels, while condensing the point information of hundreds or thousands of pixels into a statistically significant and spatially representative overall feature descriptor. The resulting baseline vector robustly depicts the historically inherent comprehensive surface environment of the pipeline segment under its specific geographical location, soil conditions, and cover type. By comparing and analyzing the real-time observed pipe segment state (current baseline vector) with the historically aggregated baseline vector in the topology network, it becomes clear that without this scientifically aggregated baseline that represents a "healthy normal state," any anomaly detection would be isolated and absolute, unable to distinguish between a genuine leakage signal and a natural feature inherent to the area and different from other places. Therefore, this process establishes a regionally specific normal baseline through spatial averaging, which changes the detection target from "finding absolute anomalies" to "detecting deviations from its own historical normality", greatly improving the method's adaptability to different geographical environments and the accuracy of identifying weak leakage signals.

[0034] In another preferred embodiment of the present invention, the specific calculation process of the spatial distribution parameter in step S5 is as follows: Based on the topological connection relationship, each pipe segment unit is identified as an independent node; if two pipe segment units are recorded as directly connected in the topological connection relationship, an undirected edge is established between the nodes corresponding to these two units to generate a pipe network feature map. The baseline vector of each pipe segment unit is used as a set of numerical attributes and assigned to the corresponding node in the pipe network feature graph. For each node in the pipe network feature graph, all neighboring nodes directly connected to it through edges are retrieved. The difference vector between the numerical attributes of the node itself and the numerical attributes of each of its neighboring nodes is calculated. The difference vectors generated by the node and all its neighboring nodes are collected. Based on the arrangement order of nodes in the pipeline feature diagram, construct a square matrix with the number of rows and columns equal to the total number of nodes as a relation matrix. For each element in the relation matrix, if there is an edge between two nodes corresponding to its row and column, the value of the element is the modulus of the difference between the numerical attributes of the two nodes; if there is no edge between the two nodes, the value of the element is zero. Perform matrix spectral decomposition on the assigned relation matrix to obtain its eigenvalues ​​and corresponding eigenvectors. Select the smallest eigenvalue from all eigenvalues ​​greater than zero and obtain its corresponding eigenvector. Record the specific value of the selected eigenvalue and the order of the components in its corresponding eigenvector. Use the recorded value and the order of arrangement to define the spatial distribution parameters of each pipe segment unit.

[0035] By abstracting the physical pipeline network and its environmental characteristics into a mathematical graph model and conducting in-depth analysis, this approach aims to automatically learn and quantify the inherent spatial continuity patterns that a healthy pipeline network should follow from complex data. This is the core theoretical foundation for achieving accurate and interpretable anomaly detection. Specifically, treating each pipe segment as a node and constructing a graph based on actual connection relationships essentially encodes the spatial topology of the pipeline network into a computable mathematical form. Using the reference vector of each pipe segment as a node attribute "anchors" the local environmental state of each segment to its corresponding position in the graph. Subsequently, calculating the differences between the attributes of nodes and their neighbors and constructing a relation matrix systematically measures the local differences or gradients of environmental characteristics along each connection edge in the graph. Spectral decomposition of the relation matrix and extraction of the minimum non-zero eigenvalue and its eigenvector are key operations in graph signal processing. Their physical significance lies in the fact that the minimum eigenvalue quantifies the lower bound of the overall smoothness achievable by changes in all node attributes (i.e., environmental characteristics) across the entire pipeline network graph structure; while its corresponding eigenvector reveals an optimal relative relationship pattern between node attributes that maximizes global smoothness. This set of parameters collectively captures the inherent patterns of how environmental characteristics should naturally transition from one segment of pipeline to the next under network connectivity constraints. The core advantage of this approach is that it defines a "healthy" or "normal" state as a computationally achievable smoothness criterion based on network structure, rather than a series of isolated, fixed numerical thresholds. In a leaking pipeline segment, the surface environmental characteristics become incompatible with those of its upstream and downstream neighbors, leading to a sharp increase in local differences at that node, significantly deviating from the smoothness described by the feature vectors, and thus reliably being marked as an anomaly. Therefore, this process represents a leap from static, absolute feature value comparisons to dynamic, network-structure-based diagnosis of relative change patterns, greatly enhancing the method's ability to detect weak, early-stage leak signals and its generalization ability to different pipeline layouts.

[0036] In another preferred embodiment of the present invention, the specific process for determining the abnormal node in step S6 is as follows: Acquire remote sensing images of the current time phase. For each pipe segment unit, extract the thermal infrared radiation value and optical band reflectance of each pixel in its strip analysis area at the current time phase, generate the feature vector of each pixel at the current time phase, and aggregate to calculate the reference vector of the pipe segment unit at the current time phase. Replace the reference vector associated with each node in the pipeline feature graph with the reference vector calculated in the current time phase; for each node in the updated pipeline feature graph, obtain all neighboring nodes directly connected to it through connecting edges, and obtain the updated reference vector of the neighboring nodes. The feature vector in the spatial distribution parameters is obtained. According to the fixed order of the nodes in the pipeline feature map, the values ​​of each component of the feature vector are sequentially assigned to the neighboring nodes as weight coefficients. The order of the neighboring nodes is consistent with the allocation order of the feature vector components. Using the assigned weight coefficients, a weighted summation operation is performed on the updated reference vectors of the neighboring nodes to calculate a prediction vector for the node; the Euclidean distance between the node's current reference vector and the prediction vector is calculated and calibrated as the prediction deviation; if the prediction deviation of the node is greater than a preset threshold, the node is marked as an abnormal node.

[0037] In another preferred embodiment of the present invention, the specific calculation process of the dispersion in step S7 is as follows: For each pipe segment unit marked as an abnormal node, determine its strip analysis area, obtain the feature vector of all pixels in the strip analysis area, and obtain the reference vector corresponding to the abnormal node. Calculate the difference between the feature vector of each pixel and the reference vector in each dimension to obtain the difference vector corresponding to the pixel; calculate the square of the difference in all dimensions in each difference vector and sum them, then take the square root of the sum to obtain the distance scalar corresponding to the pixel; accumulate the distance scalars corresponding to all pixels in the strip analysis area and sum them, then divide the sum by the total number of pixels in the strip analysis area to calculate the reference distance scalar value, and define the reference distance scalar value as the dispersion corresponding to the anomaly node.

[0038] The core objective is to distinguish between two different scenarios: one is anomalies caused by real faults such as localized point-source leaks in pipelines, and the other is overall state deviations caused by uniform environmental changes such as widespread rainfall or regional irrigation. Real localized leaks have an uneven impact on the surface environment, with a limited range of influence. Typically, they only cause drastic changes in the characteristics of a small number of pixels directly above and near the leak point, while other pixels in the area that are not directly affected remain close to their historical baseline. This "partial drastic change, partial normality" pattern leads to a high average distance (i.e., dispersion) within the calculated area. Conversely, large-scale uniform environmental events cause almost all pixels in the area to deviate from the historical baseline synchronously and to a similar degree. In this case, the distance scalar values ​​of each pixel are close to each other, and the dispersion obtained after averaging is relatively low. Therefore, calculating and evaluating this dispersion index is equivalent to performing a "secondary diagnosis" of the anomaly signal. The benefit of this is a significant improvement in the accuracy and reliability of the system's early warning system. It can effectively filter out false positive alarms triggered by widespread changes in the natural environment that are not due to faults in the pipeline network itself, thus focusing maintenance attention on high-confidence risk pipeline sections with high dispersion, indicating that there are indeed local anomalies. This achieves accurate, efficient, and low-false-reporting pipeline risk prediction, ensuring that the issued early warning information includes not only location, thereby enhancing its practical value and credibility in actual water affairs operation and maintenance management.

[0039] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting the risk of water supply networks based on image recognition, characterized in that, Includes the following steps: S1, acquire multi-temporal remote sensing images and pipeline geographic information data within the target area. The multi-temporal remote sensing images consist of pixels arranged in a regular grid. Each pixel contains measurement values ​​of thermal infrared band and multiple optical bands. The geographic information data includes the spatial coordinates and topological connection relationships of the pipelines. S2, based on the spatial coordinates, the pipeline is divided into multiple pipe segment units, and the corresponding strip-shaped analysis area is formed by expanding the projection of each pipe segment unit as the center line; S3, extract the thermal infrared radiation value and optical band reflectance of the pixels in each strip analysis area in the multi-temporal remote sensing image, and generate the feature vector of each pixel in the strip analysis area. S4, aggregate all pixel feature vectors corresponding to each pipe segment unit and calculate the statistical center value, and set the statistical center value as the reference vector of the pipe segment unit; S5. Based on the topological connection relationship, construct a pipeline network feature map with each pipe segment unit as a node, and calculate the spatial distribution parameters of each pipe segment unit on the pipeline network feature map. S6. Calculate the current reference vector of each pipe segment unit based on the current time-phase remote sensing image and update the attributes of the corresponding nodes in the pipe network feature map. Calculate the prediction deviation of the updated attribute values ​​of each node according to the spatial distribution parameters, and mark nodes whose prediction deviation exceeds a preset threshold as abnormal nodes. S7. Calculate the dispersion of the pixel feature vector in the strip analysis area corresponding to the abnormal node relative to the reference vector of the node. When the dispersion exceeds the set range, output the spatial coordinates of the pipe segment unit corresponding to the abnormal node and issue an early warning message.

2. The method for predicting water supply network risks based on image recognition according to claim 1, characterized in that, In S2, the specific process of generating the banded analysis region is as follows: Read the spatial coordinate sequence of pipelines in the geographic information data, wherein the spatial coordinate sequence is a series of coordinate points arranged in connection order; Connecting any two adjacent coordinate points forms a spatial line segment, and each of these spatial line segments is defined as a pipe segment unit. Using the spatial line segment corresponding to each pipe segment unit as the central reference line, the system extends to both sides along the vertical direction of the central reference line to form a strip-shaped polygonal region with a preset width. The strip polygon regions corresponding to pipe segments that belong to the same pipeline and are connected to each other are merged at the connection points of their spatial line segments to generate continuous polygon regions. The vector boundary of the continuous polygonal region is integrated with the pixel grid of the multi-temporal remote sensing image in a spatial coordinate system and overlaid. The image pixels whose center point is located within the continuous polygonal region are identified as the strip analysis area of ​​the pipe segment unit.

3. The method for predicting water supply network risks based on image recognition according to claim 1, characterized in that, In step S3, the specific process of generating the feature vector of each pixel within the strip analysis area is as follows: For each pixel within the strip analysis area, the thermal infrared radiance value of that pixel in each time phase and the reflectance value in each optical band are extracted from the multi-temporal remote sensing image. Calculate the mean value of the thermal infrared radiation value of the pixel in all time phases, and use it as the first characteristic value; Calculate the mean value of the reflectance of the pixel in each optical band in all time phases to obtain multiple second characteristic values ​​corresponding to each optical band; The first feature value and all the second feature values ​​are combined into a one-dimensional array in a predetermined fixed order, wherein the first feature value occupies the first element position of the array, and the remaining element positions are filled with the second feature values ​​corresponding to each optical band in sequence. This one-dimensional array is defined as the feature vector of the pixel.

4. The method for predicting water supply network risks based on image recognition according to claim 1, characterized in that, In step S4, the process of obtaining the reference vector of the pipe segment unit is as follows: Obtain the feature vectors of all pixels in the strip analysis area corresponding to the current pipe segment unit; arrange all the obtained feature vectors to form a two-dimensional data matrix, where each row of the matrix corresponds to the feature vector of a pixel, and each column corresponds to the value at a fixed sorting position in the feature vector. For each column of the two-dimensional data matrix, the arithmetic mean of all values ​​in that column is calculated. The arithmetic means of each column are arranged in the fixed order to obtain a one-dimensional array, and this one-dimensional array is defined as the reference vector of the pipe segment unit.

5. The method for predicting water supply network risks based on image recognition according to claim 1, characterized in that, In S5, the specific calculation process for the spatial distribution parameters is as follows: Based on the topological connection relationship, each pipe segment unit is identified as an independent node; if two pipe segment units are recorded as directly connected in the topological connection relationship, an undirected edge is established between the nodes corresponding to these two units to generate a pipe network feature map. The baseline vector of each pipe segment unit is used as a set of numerical attributes and assigned to the corresponding node in the pipe network feature graph. For each node in the pipe network feature graph, all neighboring nodes directly connected to it through edges are retrieved. The difference vector between the numerical attributes of the node itself and the numerical attributes of each of its neighboring nodes is calculated. The difference vectors generated by the node and all its neighboring nodes are collected. Based on the arrangement order of nodes in the pipeline feature diagram, construct a square matrix with the number of rows and columns equal to the total number of nodes as a relation matrix. For each element in the relation matrix, if there is an edge between two nodes corresponding to its row and column, the value of the element is the modulus of the difference between the numerical attributes of the two nodes; if there is no edge between the two nodes, the value of the element is zero. Perform matrix spectral decomposition on the assigned relation matrix to obtain its eigenvalues ​​and corresponding eigenvectors. Select the smallest eigenvalue from all eigenvalues ​​greater than zero and obtain its corresponding eigenvector. Record the specific value of the selected eigenvalue and the order of the components in its corresponding eigenvector. Use the recorded value and the order of arrangement to define the spatial distribution parameters of each pipe segment unit.

6. The method for predicting water supply network risks based on image recognition according to claim 1, characterized in that, In step S6, the specific process for determining abnormal nodes is as follows: Acquire remote sensing images of the current time phase. For each pipe segment unit, extract the thermal infrared radiation value and optical band reflectance of each pixel in its strip analysis area at the current time phase, generate the feature vector of each pixel at the current time phase, and aggregate to calculate the reference vector of the pipe segment unit at the current time phase. Replace the reference vector associated with each node in the pipeline feature graph with the reference vector calculated in the current time phase; for each node in the updated pipeline feature graph, obtain all neighboring nodes directly connected to it through connecting edges, and obtain the updated reference vector of the neighboring nodes. The feature vector in the spatial distribution parameters is obtained. According to the fixed order of the nodes in the pipeline feature map, the values ​​of each component of the feature vector are sequentially assigned to the neighboring nodes as weight coefficients. The order of the neighboring nodes is consistent with the allocation order of the feature vector components. Using the assigned weight coefficients, a weighted summation operation is performed on the updated reference vectors of the neighboring nodes to calculate a prediction vector for the node; the Euclidean distance between the node's current reference vector and the prediction vector is calculated and calibrated as the prediction deviation; if the prediction deviation of the node is greater than a preset threshold, the node is marked as an abnormal node.

7. The method for predicting water supply network risks based on image recognition according to claim 1, characterized in that, In S7, the specific calculation process for the dispersion is as follows: For each pipe segment unit marked as an abnormal node, determine its strip analysis area, obtain the feature vector of all pixels in the strip analysis area, and obtain the reference vector corresponding to the abnormal node. Calculate the difference between the feature vector of each pixel and the reference vector in each dimension to obtain the difference vector corresponding to the pixel; calculate the square of the difference in all dimensions in each difference vector and sum them, then take the square root of the sum to obtain the distance scalar corresponding to the pixel; accumulate the distance scalars corresponding to all pixels in the strip analysis area and sum them, then divide the sum by the total number of pixels in the strip analysis area to calculate the reference distance scalar value, and define the reference distance scalar value as the dispersion corresponding to the anomaly node.

Images