Method for detecting dam abnormal seepage region based on thermal imaging module carrying deep learning by unmanned aerial vehicle

Abstract

The invention discloses a method for detecting an abnormal seepage region of a dam based on a thermal imaging module with deep learning carried by an unmanned aerial vehicle, and the method comprises the steps: S1, carrying out the information collection of the downstream of the dam through a hyperspectral imager carried by the unmanned aerial vehicle, and shooting a hyperspectral image; s2, processing the collected original hyperspectral image, removing redundant data, converting the original hyperspectral image into a low-dimensional hyperspectral image, and converting the hyperspectral features into low-dimensional features; s3, aiming at the data processed in the step S2, selecting a water body characteristic wave band, calculating a reflectivity ratio of each pixel in the spectral image in the selected wave band, and judging a leakage area based on the reflectivity ratio; and S4, calculating the leakage amount according to the area of the leakage area identified in the step S3, the length of the leakage path and the water head difference. Finally, the dam abnormal seepage area is accurately determined, and a reliable basis is provided for dam safety maintenance. The method effectively solves the problem of dam abnormal seepage area detection, improves the detection precision and efficiency, and guarantees the safety of a water conservancy project.

Description

Technical Field

[0001] The present invention belongs to the technical field of seepage detection of flood control dikes, and particularly relates to a method for detecting abnormal seepage areas of dikes based on a thermal imaging module carried by an unmanned aerial vehicle (UAV) and deep learning. Background Art

[0002] The problem of dike leakage seriously threatens the safety of water conservancy projects. The water permeability of the dike body and the surrounding rock and soil masses will cause corrosion of the rock and soil masses, resulting in subsidence and leakage of the dike section, and further damaging the dike and buildings. There are many limitations in traditional dike detection technologies. Manual detection has low efficiency and high cost. Although satellite remote sensing detection has a wide coverage range, its spatial and temporal resolutions and accuracy are limited. With the rapid development of UAV technology, its advantages of low cost and strong flexibility have attracted much attention in the field of water conservancy detection. However, existing UAV detections mostly rely on direct measurements by sensors, and it is difficult to achieve comprehensive evaluation and accurate prediction for the complex and changeable seepage conditions of earth-rock dikes. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method for detecting abnormal seepage areas of dikes based on a thermal imaging module carried by an unmanned aerial vehicle and deep learning.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A method for detecting abnormal seepage areas of dikes based on a thermal imaging module carried by an unmanned aerial vehicle and deep learning includes the following steps:

[0006] Step S1: Data acquisition step: Use an unmanned aerial vehicle carrying a hyperspectral imager to collect information on the downstream of the dike and capture a hyperspectral image.

[0007] Step S2: Data processing and feature extraction:

[0008] The data collected by the hyperspectral imaging system is huge, and there is a large amount of redundant information in it, which will interfere with the determination of leakage characteristics. Therefore, it is necessary to process the collected original hyperspectral image, eliminate redundant data, convert it into a low-dimensional hyperspectral image, and at the same time convert the hyperspectral features into low-dimensional features.

[0009] Step S3: Band selection and ratio calculation to judge the leakage area:

[0010] For the data processed in step S2, select the water body feature band, calculate the reflectance ratio of each pixel in the hyperspectral image in the selected band, and judge the leakage area based on this.

[0011] Step S4: Calculate the leakage amount of the leakage area:

[0012] Calculate the leakage volume based on the area of the identified leakage area and the leakage path length and the head difference in step S3.

[0013] For further optimization, in step S1, use a drone to focus on inspecting the parts suspected of having water bodies, and take pictures of the identified rainfall accumulation areas in the earth-rock dam, and deeply analyze the hyperspectral conditions of the precipitation areas; through this step, the original data of the relevant areas of the dam can be comprehensively obtained, laying a foundation for subsequent analysis.

[0014] For further optimization, in step S2, in order to improve the effect of feature extraction, a local adaptive dimensionality reduction measure learning method is adopted. This method can more accurately extract the low-dimensional feature data reflecting leakage characteristics by setting thresholds and analyzing the change of Mahalanobis distance before and after measure learning. The specific steps include:

[0015] Step S2.1: Preprocess the original hyperspectral image to remove noise and normalize the data;

[0016] Step S2.2: Use the locally linear embedding dimensionality reduction method to reduce the data dimension and retain the local structure information of the data;

[0017] Step S2.3: Extract spatial and spectral features through a two-dimensional Gabor filter and a spectral feature extraction tunnel respectively;

[0018] Step S2.4: Fuse the extracted features to form the final low-dimensional feature representation.

[0019] For further optimization, traditional hyperspectral image feature extraction methods are usually based on statistical hypothesis models, but these models generally have poor universality problems. In addition, the samples are not distributed in a simple linear Euclidean space, which makes feature extraction more complex. To solve these problems, in step S2 of the present invention, a measure learning method is adopted to extract hyperspectral features. Specifically, the Mahalanobis distance between samples is calculated through the following formula to measure the similarity between samples:

[0020]

[0021] where x i and x j are two sample vectors, M is a symmetric positive definite matrix, M ∈ R D×D is a symmetric semi-positive definite matrix, M = WW T , W ∈ R D×d , d ≤ D; W is a dimensionality reduction matrix.

[0022] For further optimization, step S3 specifically includes the following steps:

[0023] Step S3.1: Select bands:

[0024] By analyzing the spectral characteristics of the hyperspectral image, two characteristic bands λ1 and λ2 are selected, and the reflectivities of these two bands and usually have significant differences between the leakage area and the non-leakage area. Among them, in the leakage area, is significantly higher than In the non-leakage area, and the difference is small. Usually, λ1 and λ2 can be selected according to the spectral characteristics of water bodies. For example, λ1 may be located in the visible light band, while λ2 is located in the near-infrared band.

[0025] Step S3.2: Calculate the reflectivity:

[0026] For each pixel (i, j) in the hyperspectral image, calculate its reflectivities in the two characteristic bands and

[0027]

[0028] where, and are the reflected light intensities of the pixel (i, j) in bands λ1 and λ2 respectively, and L 入射 (λ1) and L 入射 (λ2) are the incident light intensities of bands λ1 and λ2 respectively;

[0029] Step S3.3: Calculate the reflectivity ratio:

[0030] For each pixel (i, j) in the hyperspectral image, calculate the reflectivity ratio of its bands λ1 and λ2

[0031]

[0032] Step S3.4: Classify the pixels according to the reflectivity ratio:

[0033] First, determine a reasonable threshold range [K min , K max according to the previous experimental data and analysis;

[0034] Secondly, for each pixel (i, j) in the hyperspectral image, compare its reflectivity ratio with the preset threshold:

[0035] When then mark this pixel as a suspected leakage area;

[0036] When then mark this pixel as a non-leakage area.

[0037] Step S3.4: Generate the Shen Lu area map:

[0038] By classifying all pixels in the hyperspectral image in Step S3.3, a preliminary leakage area map is generated. All pixels marked as suspected leakage areas are highlighted, while other pixels remain unchanged or are marked as non-leakage areas.

[0039] The generation formula of the leakage area map can be expressed as: \(\begin{cases}\text{Suspected leakage area},&\text{if}B_{\lambda_1 / \lambda_2}(i,j)\in[T_{\min},T_{\max}]\\\text{Non-leakage area},&\text{otherwise}\end{cases}\). Through these steps, the leakage areas in the hyperspectral image can be initially identified, providing basic data for further leakage area analysis and confirmation. This process can be implemented through programming, for example, by writing a traversal algorithm using programming languages such as Python or MATLAB.

[0040] Taking into comprehensive consideration factors such as the dam type, surrounding environment, and water body characteristics, determine appropriate leakage area extraction thresholds, further screen and confirm the marked suspected leakage areas, and finally accurately determine the abnormal seepage areas of the dam, providing a reliable basis for dam safety maintenance.

[0041] For further optimization, in Step S4, it specifically includes the following steps:

[0042] Step S4.1: Calculate the leakage area A:

[0043] A = ∑ (i,j)∈渗漏区域 ΔA;

[0044] Where: A represents the total area of the leakage area, (i, j) represents the pixel coordinates in the hyperspectral image; ΔA represents the area corresponding to each pixel (unit: square meters, m 2 ), which is calculated through the image resolution and the actual ground resolution;

[0045] Step S4.2: Calculate the hydraulic gradient q:

[0046]

[0047] Where, h1 and h2 respectively represent the water head heights on both sides of the leakage area, and S represents the length of the leakage path; Step S4.2: Calculate the leakage volume Q:

[0048] Q = G.A.q;

[0049] Among them, G represents the permeability coefficient, reflecting the seepage capacity of the soil; q is the hydraulic gradient; A represents the total area of the leakage area.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] The present invention discloses that the method uses an unmanned aerial vehicle (UAV) equipped with a hyperspectral imager to collect information on the downstream of the dam and capture a hyperspectral image; the collected original hyperspectral image is processed to eliminate redundant data, convert it into a low-dimensional hyperspectral image, and at the same time transform the hyperspectral features into low-dimensional features; for the processed data, the water body characteristic bands are selected, the reflectance ratio of each pixel in the selected bands in the spectral image is calculated, and based on this, the leakage area is judged; and considering factors such as the dam type, the surrounding environment, and the water body characteristics, a suitable leakage area extraction threshold is determined, and the marked suspected leakage areas are further screened and confirmed, and finally the abnormal seepage area of the dam is accurately determined. The leakage volume is calculated according to the area of the identified leakage area and the leakage path length and the head difference, and the leakage volume of the leakage area at different time periods is calculated, so as to provide a scientific basis for the dynamic monitoring and evaluation of the dam leakage and ensure the safety of water conservancy projects. Description of the Drawings

[0052] Figure 1 It is a flow chart of a method for detecting an abnormal seepage area of a dam based on a thermal imaging module carried by an unmanned aerial vehicle with deep learning. Detailed Embodiments

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] As Figure 1 shown, a method for detecting an abnormal seepage area of a dam based on a thermal imaging module carried by an unmanned aerial vehicle with deep learning includes the following steps:

[0055] Step S1: Data acquisition step: Use an unmanned aerial vehicle equipped with a hyperspectral imager to collect information on the downstream of the dam and capture a hyperspectral image. The unmanned aerial vehicle focuses on inspecting the parts suspected of having water bodies and captures the determined rainfall accumulation areas in the earth-rock dam, and deeply analyzes the hyperspectral conditions of the precipitation areas; through this step, the original data of the relevant areas of the dam can be comprehensively obtained, laying a foundation for subsequent analysis.

[0056] Step S2: Data processing and feature extraction:

[0057] The data collected by the hyperspectral imaging system is huge, and there is a large amount of redundant information in it, which will interfere with the determination of leakage characteristics. Therefore, it is necessary to process the collected original hyperspectral images, remove redundant data, convert them into low-dimensional hyperspectral images, and at the same time transform the hyperspectral features into low-dimensional features.

[0058] In order to improve the effect of feature extraction, a local adaptive dimensionality reduction measure learning method is adopted. This method can more accurately extract the low-dimensional feature data reflecting leakage characteristics by setting thresholds and analyzing the change of Mahalanobis distance before and after measure learning. The specific steps include:

[0059] Step S2.1: Preprocess the original hyperspectral image to remove noise and normalize the data;

[0060] Step S2.2: Use the locally linear embedding dimensionality reduction method to reduce the data dimension and retain the local structure information of the data;

[0061] Step S2.3: Extract spatial and spectral features through a two-dimensional Gabor filter and a spectral feature extraction tunnel respectively;

[0062] Step S2.4: Fuse the extracted features to form the final low-dimensional feature representation.

[0063] Traditional hyperspectral image feature extraction methods are usually based on statistical hypothesis models, but these models generally have poor universality. In addition, the samples are not distributed in a simple linear Euclidean space, which makes feature extraction more complex. To solve these problems, in step S2 of the present invention, a measure learning method is adopted to extract hyperspectral features. Specifically, the Mahalanobis distance between samples is calculated through the following formula to measure the similarity between samples:

[0064]

[0065] where, x i and x j are two sample vectors, M is a symmetric positive definite matrix, M ∈ R D×D is a symmetric semi-positive definite matrix, M = WW T , W ∈ R D×d , d ≤ D; W is a dimensionality reduction matrix.

[0066] Step S3: Band selection and ratio calculation to judge the leakage area:

[0067] For the data processed in step S2, select the water body feature bands, calculate the reflectance ratio of each pixel in the hyperspectral image in the selected bands, and judge the leakage area based on this; specifically, it includes the following steps:

[0068] Step S3.1: Select bands:

[0069] By analyzing the spectral characteristics of the spectral image, two characteristic bands λ1 and λ2 are selected, and the reflectivities of these two bands and usually have significant differences between the leakage area and the non-leakage area. Among them, in the leakage area, is significantly higher than In the non-leakage area, and the difference is small. Usually, λ1 and λ2 can be selected according to the spectral characteristics of water bodies. For example, λ1 may be located in the visible light band, while λ2 is located in the near-infrared band.

[0070] Step S3.2: Calculate the reflectivity:

[0071] For each pixel (i, j) in the hyperspectral image, calculate its reflectivities in the two characteristic bands and

[0072]

[0073] where and are the reflected light intensities of the pixel (i, j) in bands λ1 and λ2 respectively, and L 入射 (λ1) and L 入射 (λ2) are the incident light intensities of bands λ1 and λ2 respectively.

[0074] Step S3.3: Calculate the reflectivity ratio:

[0075] For each pixel (i, j) in the hyperspectral image, calculate the reflectivity ratio of its bands λ1 and λ2

[0076]

[0077] Step S3.4: Classify the pixels according to the reflectivity ratio:

[0078] First, according to the previous experimental data and analysis, determine a reasonable threshold range [K min , K max ;

[0079] Secondly, for each pixel (i, j) in the hyperspectral image, compare its reflectivity ratio with the preset threshold:

[0080] When then mark this pixel as a suspected leakage area;

[0081] When then mark this pixel as a non-leakage area.

[0082] Step S3.4: Generate the Shen Lu area map:

[0083] By classifying all pixels in the hyperspectral image in Step S3.3, a preliminary leakage area map is generated, and all pixels marked as suspected leakage areas are highlighted, while other pixels remain unchanged or are marked as non-leakage areas.

[0084] The generation formula of the leakage area map can be expressed as: \(\begin{cases}\text{Suspected leakage area},&\text{if}B_{\lambda_1 / \lambda_2}(i,j)\in[T_{\min},T_{\max}]\\\text{Non-leakage area},&\text{otherwise}\end{cases}\). Through these steps, the leakage areas in the hyperspectral image can be initially identified, providing basic data for further leakage area analysis and confirmation. This process can be implemented through programming, for example, by writing a traversal algorithm using programming languages such as Python or MATLAB.

[0085] Step S4: Calculate the leakage volume of the leakage area:

[0086] Calculate the leakage volume based on the area of the leakage area identified in Step S3 and the leakage path length and head difference. The specific steps are as follows:

[0087] Step S4.1: Calculate the leakage area A:

[0088] A = ∑ (i,j)∈渗漏区域 ΔA;

[0089] Where: A represents the total area of the leakage area, (i, j) represents the pixel coordinates in the hyperspectral image; ΔA represents the area corresponding to each pixel (unit: square meters, m 2 ), which is calculated through the image resolution and the actual ground resolution.

[0090] Step S4.2: Calculate the hydraulic gradient q:

[0091]

[0092] Where h1 and h2 respectively represent the head heights on both sides of the leakage area, and S represents the length of the leakage path;

[0093] Step S4.2: Calculate the leakage volume Q:

[0094] Q = G.A.q;

[0095] Where G represents the permeability coefficient, reflecting the permeability of the soil; q is the hydraulic gradient; A represents the total area of the leakage area.

[0096] By calculating the leakage volume of the leakage area in different time periods, it provides a scientific basis for the dynamic monitoring and evaluation of dam leakage, ensuring the safety of water conservancy projects.

[0097] Inspired by the ideal embodiments of the present invention as described above, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of this invention. The technical scope of this invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for detecting abnormal seepage areas of dams based on a thermal imaging module equipped with deep learning by a drone, characterized in that: The steps include: Step S1: Data collection step: using a drone equipped with a hyperspectral imager to collect information downstream of the dam and take hyperspectral images; Step S2: Data processing and feature extraction: Process the collected original hyperspectral images, remove redundant data, convert them into low-dimensional hyperspectral images, and convert hyperspectral features into low-dimensional features; Step S3: Band selection and ratio calculation to determine the leakage area: For the data processed in step S2, the characteristic bands of the water body are selected, and the reflectance ratio of each pixel in the hyperspectral image in the selected band is calculated, and the leakage area is determined based on the ratio; Step S4: Calculate the leakage amount of the leakage area: The leakage amount is calculated according to the area of ​​the leakage region identified in step S3 and the length of the leakage path and the water head difference.

2. The method for detecting abnormal seepage areas of dams based on a thermal imaging module equipped with deep learning by an unmanned aerial vehicle according to claim 1 is characterized in that: In step S1, the drone is used to focus on inspecting the areas where water bodies are suspected to exist, and to photograph the areas of rainfall accumulation in the earth-rock dam that have been determined.

3. The method for detecting abnormal seepage areas of dams based on a thermal imaging module equipped with deep learning by an unmanned aerial vehicle according to claim 2 is characterized in that: In step S2, the specific steps include: Step S2.1: preprocess the original hyperspectral image to remove noise and normalize the data; Step S2.2: Use the local preserving dimensionality reduction method to reduce the data dimension and preserve the local structural information of the data; Step S2.3: extracting spatial and spectral features respectively through two-dimensional Gabor filter and spectral feature extraction tunnel; Step S2.4: Fuse the extracted features to form the final low-dimensional feature representation.

4. The method for detecting abnormal seepage areas of dams based on a thermal imaging module equipped with deep learning by an unmanned aerial vehicle according to claim 3 is characterized in that: In step S2, the hyperspectral features are extracted using a measurement learning method, and the similarity between samples is measured by calculating the Mahalanobis distance between samples using the following formula: Among them, x i and x j are two sample vectors, M is a symmetric positive definite matrix, M∈R D×D is a symmetric semi-positive definite matrix, M = WW T ,W∈R D×d ,d≤D; W is a dimensionality reduction matrix. In view of the poor universality of the statistical hypothesis model for extracting features from hyperspectral images and the fact that samples are not distributed in a simple linear Euclidean space, the present invention adopts a measure learning method to extract features.

5. The method for detecting abnormal seepage areas of dams based on a thermal imaging module equipped with deep learning by an unmanned aerial vehicle according to claim 4 is characterized in that: The step S3 specifically includes the following steps: Step S3.1: Select the band: By analyzing the spectral characteristics of the hyperspectral image, two characteristic bands λ1 and λ2 are selected. The reflectance of these two bands and There are usually significant differences between leaky and non-leaky areas; Step S3.2: Calculate reflectivity: For each pixel (i, j) in the hyperspectral image, calculate its reflectance in two characteristic bands and in, and are the reflected light intensity of pixel (i, j) in bands λ1 and λ2, respectively, 入射 (λ1) and L 入射 (λ2) are the incident light intensities of the wavelength bands λ1 and λ2 respectively; Step S3.3: Calculate the reflectivity ratio: For each pixel (i, j) in the hyperspectral image, calculate the reflectance ratio of its bands λ1 and λ2 Step S3.4: Classify pixels according to reflectivity ratio: First, based on the previous experimental data and analysis, determine the reasonable threshold range [K min , K max ]; Secondly, for each pixel (i, j) in the spectral image, its reflectance ratio Compare with preset threshold: when Then the pixel is marked as a suspected leakage area; when Then the pixel is marked as a non-leakage area; Step S3.4: Generate leakage area map: By classifying all pixels in the hyperspectral image in step S3.3, a preliminary leakage area map is generated, and all pixels marked as suspected leakage areas are highlighted, while other pixels remain as they are or are marked as flying leakage areas.

6. The method for detecting abnormal seepage areas of dams based on a thermal imaging module equipped with deep learning by an unmanned aerial vehicle according to claim 5 is characterized in that: The step S4 specifically includes the following steps: Step S4.1: Calculation of leakage area A: A=∑ (i,j)∈渗漏区域 ΔA; Where: A represents the total area of ​​the leakage area, (i, j) represents the pixel coordinates in the hyperspectral image; ΔA represents the area corresponding to each pixel (unit: square meters, m 2 ), calculated from the image resolution and the actual ground resolution; Step S4.2: Calculate the hydraulic slope q: Among them, h1 and h2 represent the water head heights on both sides of the leakage area, and S represents the length of the leakage path; Step S4.2: Calculate the leakage amount Q: Q = G Aq; Among them, G is the permeability coefficient, which reflects the permeability of the soil; q is the hydraulic gradient; and A is the total area of ​​the leakage area.

Images