House leakage identification method and system based on hyperspectral and imaging technology

Abstract

The present invention relates to the technical field of leakage identification, and discloses a housing leakage identification method and system based on hyperspectral and imaging technology, which is used to realize the effective fusion of hyperspectral, thermal infrared and visible light multi-source data, so as to significantly improve the accuracy and quantitative evaluation ability of housing leakage detection. The method includes: performing spectral correction processing on hyperspectral image data, obtaining a reflectance data cube and performing spectral index calculation to obtain a moisture-related index set, performing smoothing filtering processing on each pixel in the reflectance data cube to obtain a denoised spectral curve and performing spectral derivative feature extraction to obtain spectral derivative features; performing multi-source data fusion on the moisture-related index set, spectral derivative features, thermal infrared images and visible light images to obtain multi-feature image data and perform multi-scale segmentation, obtain a spatial distribution map and perform temperature anomaly degree calculation and surface damage area measurement to obtain a leakage degree quantitative data set and a multi-dimensional risk assessment report.

Description

Technical Field

[0001] The present invention relates to the technical field of leakage identification, and in particular to a method and system for identifying house leakage based on hyperspectral and imaging technologies. Background Art

[0002] Existing methods for detecting house leakage mainly include traditional manual detection and emerging technical means. Traditional methods include visual inspection, percussion method and instrument testing. Visual inspection relies on the experience and professional knowledge of the inspectors, and can quickly find obvious signs of leakage, but it is easy to miss hidden leakage and cannot perform quantitative analysis. The percussion method judges whether there is hollowing or leakage by knocking on the wall and listening to the sound changes, but it can only detect surface leakage and cannot detect internal leakage. Instrument tests such as resistance method and ultrasonic method can provide certain quantitative data, but the detection range is limited, and the equipment needs to be moved frequently, which is inefficient. With the advancement of science and technology, emerging detection technologies such as infrared thermal imaging, microwave detection and geological radar have begun to be applied to house leakage detection. Among them, infrared thermal imaging technology identifies potential leakage locations by detecting abnormal temperature areas. It has the advantages of non-contact and large-area rapid detection, but it is easily affected by environmental factors such as sunlight and air conditioning use, which may lead to misjudgment.

[0003] In recent years, with the rapid development of remote sensing technology, hyperspectral imaging technology has begun to be applied to the field of building leakage detection. Hyperspectral imaging can obtain continuous narrowband spectral information in the range of visible light to short-wave infrared, and can capture subtle spectral changes that cannot be recognized by the naked eye. This technology can identify potential leakage areas by analyzing the spectral characteristics of materials, especially in detecting early leakage or hidden leakage. However, the existing hyperspectral leakage detection methods still have some significant shortcomings: 1) Although a single hyperspectral data contains rich spectral information, it is difficult to fully characterize all the characteristics of leakage, such as temperature anomalies and surface damage; 2) Hyperspectral images are easily affected by factors such as atmosphere and light, and there are large noise and distortion, which affects the detection accuracy; 3) There is a lack of effective multi-source data fusion methods, and it is difficult to fully utilize the complementary advantages of multimodal information such as hyperspectral, thermal infrared and visible light; 4) The existing leakage degree assessment methods are mostly based on empirical judgment or simple indicator calculation, lack of a systematic quantitative assessment system, and it is difficult to accurately reflect the severity and potential risks of leakage; 5) For large buildings or houses with complex structures, the existing methods are difficult to achieve rapid and comprehensive detection and evaluation, and the efficiency needs to be improved. These problems seriously limit the application effect and promotion of hyperspectral technology in house leakage detection. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a house leakage identification method and system based on hyperspectral and imaging technology, which is used to realize the effective fusion of hyperspectral, thermal infrared and visible light multi-source data, so as to significantly improve the accuracy and quantitative evaluation capability of house leakage detection.

[0005] The present invention provides a method for identifying house leakage based on hyperspectral and imaging technology, comprising: performing multi-band hyperspectral scanning on a building to obtain hyperspectral image data, wherein the spectral range of the hyperspectral image data is 400nm-2500nm, and performing thermal infrared imaging and high-resolution visible light imaging on the building to obtain thermal infrared images and visible light images; performing multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube; performing characteristic band selection and spectral index calculation on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index; The spectral data of each pixel in the reflectance data cube is smoothed and filtered to obtain a denoised spectral curve, and the spectral derivative feature of the denoised spectral curve is extracted to obtain a spectral derivative feature, and the spectral derivative feature is used to characterize the spectral change rate; the moisture-related index set, the spectral derivative feature, the thermal infrared image and the visible light image are multi-source data fused to obtain multi-feature image data; the multi-feature image data is multi-scale segmented to obtain a spatial distribution map of potential leakage areas; the spatial distribution map is quantitatively analyzed by spectral index, the temperature anomaly degree is calculated, and the surface damage area is measured to obtain a leakage degree quantitative data set and a multi-dimensional risk assessment report.

[0006] The present invention also provides a house leakage identification system based on hyperspectral and imaging technology, comprising:

[0007] A scanning module is used to perform multi-band hyperspectral scanning on the building to obtain hyperspectral image data, wherein the spectral range of the hyperspectral image data is 400nm-2500nm, and perform thermal infrared imaging and high-resolution visible light imaging on the building to obtain a thermal infrared image and a visible light image;

[0008] A correction module, used for performing multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube;

[0009] A calculation module, used for performing characteristic band selection and spectral index calculation on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index;

[0010] A filtering module, used for performing smoothing filtering on the spectral data of each pixel in the reflectance data cube to obtain a de-noised spectral curve, and performing spectral derivative feature extraction on the de-noised spectral curve to obtain a spectral derivative feature, wherein the spectral derivative feature is used to characterize the spectral change rate;

[0011] A fusion module, used for performing multi-source data fusion on the moisture-related index set, the spectral derivative features, the thermal infrared image and the visible light image to obtain multi-feature image data;

[0012] A segmentation module, used for performing multi-scale segmentation on the multi-feature image data to obtain a spatial distribution map of potential leakage areas;

[0013] The analysis module is used to perform spectral index quantitative analysis, temperature anomaly degree calculation and surface damage area measurement on the spatial distribution map to obtain a leakage degree quantitative data set and a multi-dimensional risk assessment report.

[0014] In the technical solution provided by the present invention, a multi-band hyperspectral scan is performed on the building to obtain hyperspectral image data in the range of 400nm-2500nm. Combined with thermal infrared imaging and high-resolution visible light imaging, a multi-source, multi-dimensional raw data basis is provided for leakage identification, which enhances the information richness and reliability of detection. The original hyperspectral image data is subjected to multi-factor correction and spectral correction processing to obtain a reflectance data cube, which effectively eliminates interference factors such as atmospheric influence and sensor error, and ensures the data quality of subsequent analysis. By selecting characteristic bands and calculating spectral indexes on the reflectance data cube, a set of moisture-related indexes including normalized difference water index and moisture stress index is obtained, which directly reflects the moisture status of the building surface and provides a key basis for accurately locating the leakage area. The spectral data of each pixel in the reflectance data cube is smoothed and filtered and the spectral derivative feature is extracted to obtain the spectral derivative feature that characterizes the spectral change rate. This processing can highlight the subtle changes in the spectral curve and help identify early or minor leakage problems. The moisture-related index set, spectral derivative features, thermal infrared images and visible light images are fused with multi-source data to obtain multi-feature image data. This fusion makes full use of the complementarity of different data sources and improves the recognition of leakage features. The multi-feature image data is segmented at multiple scales to obtain the spatial distribution map of potential leakage areas. This step can adapt to leakage areas of different sizes and shapes, improving the flexibility and applicability of detection. Finally, the spatial distribution map is quantitatively analyzed by spectral index, the degree of temperature anomaly is calculated, and the surface damage area is measured to obtain a quantitative data set of leakage degree and a multi-dimensional risk assessment report. This comprehensive assessment process can not only accurately determine the location and degree of leakage, but also assess the potential risk of leakage to building structures, realizing the automation of the entire process from data collection, processing, analysis to risk assessment, greatly improving the efficiency and accuracy of leakage detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 This is a flow chart of a method for identifying house leakage based on hyperspectral and imaging technology in an embodiment of the present invention;

[0017] Figure 2 Schematic diagram of a house leakage identification system based on hyperspectral and imaging technology in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0020] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0021] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , Figure 1 is a flow chart of a housing leakage identification method based on hyperspectral and imaging technology according to an embodiment of the present invention. Figure 1 As shown, the following steps are included:

[0022] S101, performing multi-band hyperspectral scanning on the building to obtain hyperspectral image data, wherein the spectral range of the hyperspectral image data is 400nm-2500nm, and performing thermal infrared imaging and high-resolution visible light imaging on the building to obtain a thermal infrared image and a visible light image;

[0023] Specifically, a push-broom hyperspectral imager is used to perform multi-band scanning of buildings. The device is usually equipped with a spectrometer with a spectral range of 400nm-2500nm and a high-sensitivity two-dimensional detector array. During the scanning process, the hyperspectral imager moves along a preset path, collecting spectral information on the surface of the building line by line, and records the spatial position information of each pixel in combination with a high-precision positioning system (such as RTK-GPS). To ensure data quality, equipment calibration is required before scanning, including dark current calibration, whiteboard reference calibration, etc., and spectral calibration is performed regularly during the scanning process to compensate for the impact of ambient light changes.

[0024] At the same time, long-wave infrared thermal imagers are used to perform thermal infrared imaging of buildings. Thermal imagers usually use equipment with high temperature resolution (such as 0.05°C) and wide temperature measurement range (such as -20°C to 120°C) to capture small temperature changes on the surface of buildings. Thermal infrared imaging needs to consider the influence of factors such as ambient temperature and relative humidity, so these environmental parameters need to be recorded synchronously during the shooting process to facilitate subsequent accurate temperature correction.

[0025] For high-resolution visible light imaging, a high-pixel digital camera is used, such as a professional-grade camera with a full-frame sensor and a high-quality lens. Visible light imaging requires attention to lighting conditions, and auxiliary lighting equipment may be required to ensure image clarity and uniformity. In order to ensure the geometric accuracy of the image, control points are usually placed on the surface of the building, and a survey-grade tripod and gimbal are used to ensure the stability of the camera.

[0026] In practical applications, these three imaging methods often need to be performed simultaneously or sequentially, and their spatial alignment must be ensured. For example, when conducting a leak detection on an office building, a hyperspectral imager can be used to scan the building's exterior wall to obtain spectral data in the range of 400nm-2500nm, which can reflect the chemical composition and moisture content changes of the wall material. Next, a thermal imager is used to thermally image the same area to capture temperature anomalies that may be caused by water seepage. Finally, a high-resolution camera is used to capture clear visible light images to record visible damage and discoloration on the surface.

[0027] S102, performing multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube;

[0028] Specifically, dark current correction is performed, which is achieved by collecting images under no illumination conditions and subtracting them from the raw data to eliminate the inherent electronic noise of the sensor. Next, flat field correction is performed, using images collected with a uniform light source or a standard white board to compensate for the non-uniform response of the imaging system to ensure the consistency of the response of each pixel in the image in each band. Subsequently, atmospheric correction is performed, which aims to eliminate the effects of atmospheric absorption, scattering and reflection on the spectrum of ground objects. Usually, a radiation transfer model (such as MODTRAN) is combined with measured atmospheric parameters (such as water vapor content, aerosol optical thickness, etc.) to calculate the influence of the atmosphere on the spectrum and remove these effects from the raw data. In some cases, alternative methods such as the empirical linear method or the dark pixel method can also be used for atmospheric correction.

[0029] Geometric correction is the next key step, which aims to eliminate image distortion caused by sensor field of view angle, terrain undulation and attitude changes during scanning. This usually involves using ground control points and digital elevation models (DEM), combined with sensor position and attitude data (such as GPS / IMU data), and reconstructing the correct geospatial relationship through complex geometric transformation algorithms. Spectral correction is a key step to ensure the spectral accuracy of hyperspectral data. This includes wavelength calibration and spectral response function correction. Wavelength calibration is usually performed using a standard light source with known spectral characteristics or atmospheric absorption characteristics, while spectral response function correction needs to take into account the sensitivity differences of the sensor at different wavelengths. Finally, radiometric calibration and reflectance conversion are performed. Radiometric calibration converts the original digital count value into an absolute radiance value, while reflectance conversion further converts the radiance value into the reflectance of the object. This is usually done by using a standard reflector plate arranged on site or an object with known reflectance to establish the relationship between radiance and reflectance.

[0030] For example, suppose a hyperspectral scan is performed on the exterior wall of a building. First, a dark current image is acquired in complete darkness, and then a flat field correction is performed using a standard white plate before scanning. During the scan, atmospheric parameters such as relative humidity, air temperature, and air pressure are measured simultaneously. After the scan is completed, these parameters and the MODTRAN model are used to perform atmospheric correction. Next, geometric correction is performed using control points pre-arranged on the building and RTK-GPS data. Then, wavelength correction is performed using the known emission lines of the xenon light source, and spectral correction is performed using the spectral response function measured in the laboratory. Finally, the grayscale gradient reflector arranged on site is used to establish the correspondence between radiance and reflectance, and the entire data set is converted into a reflectance data cube.

[0031] S103, performing characteristic band selection and spectral index calculation on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index;

[0032] Specifically, a band correlation analysis is performed on the reflectance data cube, which is usually achieved by calculating the correlation coefficient matrix between the bands. Based on the correlation analysis, highly correlated band groups can be identified, thereby reducing redundant information. Next, dimensionality reduction techniques such as principal component analysis (PCA) or independent component analysis (ICA) are used to further extract the main features in the data, which helps to identify the bands that are most sensitive to changes in moisture content. In the process of feature band selection, it is also necessary to combine prior knowledge, especially spectral knowledge about moisture absorption characteristics. Moisture has obvious absorption characteristics in the near-infrared and short-wave infrared regions, especially near 970nm, 1200nm, 1450nm and 1950nm. Therefore, bands near these wavelengths are usually given priority. In addition, the continuous removal method can be used to enhance these absorption features, so as to more accurately locate the position and depth of the absorption band.

[0033] In order to further optimize the band selection, intelligent optimization methods such as genetic algorithms or particle swarm optimization algorithms are used. These algorithms can search for the best combination among a large number of possible band combinations to maximize the accuracy of moisture content estimation. In this process, some sample data are usually used for cross-validation to ensure the universality and stability of the selected bands. Once the best characteristic bands are determined, the moisture-related index can be calculated. The normalized difference water index (NDWI) is one of the most commonly used moisture indices, which is usually calculated using the reflectivity difference between the near-infrared band (about 860nm) and the short-wave infrared band (about 1240nm). The calculation formula of NDWI is: NDWI = (R860-R1240) / (R860+R1240), where R represents the reflectivity of the corresponding wavelength. The higher the NDWI value, the higher the moisture content.

[0034] The moisture stress index (MSI) is another important moisture-related index, which is calculated using the reflectance near 1600nm and 820nm. The calculation formula of MSI is: MSI = R1600 / R820. Unlike NDWI, the higher the MSI value, the greater the moisture stress, that is, the lower the moisture content.

[0035] In addition to NDWI and MSI, other water-related indices can be calculated, such as the Water Absorption Index (WAI) or the Shortwave Infrared Water Stress Index (SWIR Water Stress Index), to provide more comprehensive water information. The selection and calculation of these indices need to be determined according to the specific application scenario and research objectives.

[0036] For example, suppose a building’s exterior wall is being analyzed for potential water seepage. First, a band correlation analysis is performed on the acquired hyperspectral data, and it is found that the bands in the 850-870nm and 1230-1250nm ranges are highly correlated with changes in moisture content. PCA analysis further confirms that these two band ranges are most sensitive to moisture changes. Combined with prior knowledge, 860nm and 1240nm are selected as bands for calculating NDWI. At the same time, 1600nm and 820nm bands are also selected to calculate MSI. When calculating these indices, the entire reflectance data cube is processed to obtain two new image layers: NDWI map and MSI map. In the NDWI map, you may find that some areas have abnormally high values, which may indicate potential areas of water seepage. In the MSI map, you may notice that some areas have abnormally low values, which may also suggest areas with higher moisture content.

[0037] S104, performing smoothing filtering on the spectral data of each pixel in the reflectance data cube to obtain a de-noised spectral curve, and extracting spectral derivative features from the de-noised spectral curve to obtain spectral derivative features, where the spectral derivative features are used to characterize the spectral change rate;

[0038] Specifically, for smoothing filtering, commonly used methods include moving average filtering, Savitzky-Golay filtering, and wavelet transform filtering. Among them, Savitzky-Golay filtering is widely used because it can retain the high-order moments of the spectral curve while smoothing noise. This filtering method is implemented by fitting a polynomial to the data points within a sliding window, where the window size and the polynomial order are key parameters that need to be adjusted according to the specific data characteristics. When applying Savitzky-Golay filtering, the entire spectral range is usually filtered first, and then different parameter settings are used for refinement processing for different spectral regions (such as visible light, near infrared, and short-wave infrared). This is because the noise characteristics and signal strength of different wavelength ranges may be different. For example, in the short-wave infrared region, due to the influence of atmospheric absorption, a larger window size may be required to effectively suppress noise.

[0039] After the filtering process is completed, the next step is to extract the spectral derivative features. The spectral derivative can enhance the subtle changes in the spectral curve and help identify the absorption features that are not easy to detect in the original spectrum. The first-order derivative is usually used to enhance the change of the spectral slope, while the second-order derivative can highlight the change of the spectral curvature. There are many methods to calculate the spectral derivative, including the finite difference method, the polynomial fitting method and the convolution method. The finite difference method is the most direct method, which approximates the derivative by calculating the difference between adjacent wavelength points. However, this method is sensitive to noise. The polynomial fitting rule is to perform a polynomial fitting on the spectrum within a certain wavelength range and then derive the fitting function. This method can suppress the influence of noise to a certain extent. The convolution method is to convolve the spectral data with a predefined derivative kernel. This method is computationally efficient and the smoothness of the derivative calculation can be controlled by adjusting the kernel function.

[0040] In practical applications, the first-order and second-order derivatives are usually calculated at the same time because they provide complementary information. The first-order derivative mainly reflects the change in the slope of the spectral curve and can be used to locate the boundary of the absorption band; while the second-order derivative can accurately locate the center position and depth of the absorption band. In addition, higher-order derivatives (such as third-order and fourth-order derivatives) can also be calculated, but higher-order derivatives are more sensitive to noise and require extra caution when used.

[0041] For example, suppose that hyperspectral data of building materials are being analyzed to detect potential moisture infiltration. First, a Savitzky-Golay filter is applied to the spectral data of each pixel, using a window of 11 points and a 3rd order polynomial fit. This setting strikes a good balance between smoothing noise and preserving spectral details. Next, the first and second order derivatives are calculated using the convolution method, and a Gaussian shaped derivative kernel is selected with a kernel width of 5 wavelength units. By analyzing the first order derivative, it can be found that there are obvious slope changes around 970nm and 1450nm, which correspond to the absorption features of moisture. The second order derivative further enhances these features, making it possible to pinpoint the center of the absorption bands, such as the strong negative peak at 1450nm indicating significant moisture presence. By comparing the derivative features of different regions, areas with abnormally high moisture content can be identified, which may indicate a leakage problem. In addition, by monitoring the changes in these derivative features over time, the dynamic changes in moisture content can also be tracked.

[0042] S105, performing multi-source data fusion on the moisture-related index set, the spectral derivative features, the thermal infrared image, and the visible light image to obtain multi-feature image data;

[0043] It should be noted that in this step, it is first necessary to ensure that all data sources are aligned in space and time, which is usually achieved through image registration technology. Image registration can use control point method, feature matching method or mutual information method, etc. The choice of which method depends on the nature of the data and the accuracy requirements. For example, the corners of the building or other obvious features can be used as control points, and images from different sources can be aligned to the same coordinate system through affine transformation or polynomial transformation. Next, the data from different sources are normalized to eliminate the scale and unit differences between different data types. This can be achieved through methods such as linear normalization, Z-score standardization or histogram equalization. Normalization ensures that different features are comparable during the fusion process, avoiding some features from dominating the fusion results due to their large numerical range.

[0044] The core step of data fusion can be achieved by a variety of methods, such as simple weighted averaging, principal component analysis (PCA), high-pass filter fusion, wavelet transform fusion, or deep learning methods. Among them, wavelet transform fusion is a common and effective method that can fuse low-frequency information while maintaining high-frequency detail information. This method first performs wavelet decomposition on each data source, then fuses them at different scales, and finally reconstructs the fused image through inverse wavelet transform. Another advanced method is to use deep learning techniques such as convolutional neural networks (CNN) or generative adversarial networks (GAN). These methods can automatically learn the complex relationships between different data sources and produce high-quality fusion results. For example, a multi-input CNN architecture can be designed to process moisture index, spectral derivative, thermal infrared, and visible light data separately, and then generate the final multi-feature image through multiple convolution and fusion layers.

[0045] During the fusion process, the reliability and importance of different data sources also need to be considered. An adaptive weight mechanism can be introduced to dynamically adjust the weight of each data source based on the quality of the data, the signal-to-noise ratio, or the contribution to the target task. For example, in shadowed areas, the reliability of visible light images may be low, and the weights of thermal infrared and hyperspectral data can be increased. Finally, the fused multi-feature image data needs to be quality assessed and post-processed. The fusion quality can be evaluated using indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and visual inspection can be used to ensure that the fusion result has no obvious artifacts or distortion. Post-processing steps may include edge enhancement, contrast adjustment, or noise suppression to further improve the quality and interpretability of the fused image.

[0046] For example, suppose a large office building is being analyzed for potential leakage. There is a set of moisture-related indices consisting of the Normalized Difference Water Index (NDWI) and the Moisture Stress Index (MSI), first-order and second-order spectral derivative features, thermal infrared images in the 8-14 μm band, and high-resolution RGB visible light images. First, all images are precisely aligned to the same coordinate system through affine transformation using the corners and window edges of the building as control points. Then, all data are Min-Max normalized to unify the value range to between 0 and 1. Next, a multi-scale fusion method based on wavelet transform is used. Each data source is subjected to 3-layer discrete wavelet decomposition, and at each layer, weights are assigned according to the importance of different features. For example, in the low-frequency part, NDWI and MSI are given higher weights because they directly reflect the moisture content; in the high-frequency part, the weight of the visible light image is increased to retain the building details. For the mid-frequency part, the contribution of all data sources is balanced. During the fusion process, the weight is also dynamically adjusted according to the thermal infrared data at each pixel location. When a temperature anomaly is detected, the weight of the thermal infrared data in the corresponding area is increased. After fusion, a comprehensive multi-feature image containing moisture information, temperature distribution and visible light details is obtained. In this image, potential leakage areas can be clearly seen: these areas show high values ​​at the moisture index level, may appear as temperature anomalies at the thermal infrared level, and may appear as color changes or surface damage in visible light images.

[0047] S106, performing multi-scale segmentation on the multi-feature image data to obtain a spatial distribution map of potential leakage areas;

[0048] Specifically, the multi-feature image data is first preprocessed, including noise removal and edge enhancement. This can be done by applying a Gaussian filter to smooth the image, and then using the Sobel or Canny operator to enhance the edge features. Next, a multi-scale analysis method, such as a Gaussian pyramid or wavelet transform, is used to decompose the image into representations of different scales. This multi-scale representation is able to capture potential leakage areas of different sizes and shapes. Based on the multi-scale representation, a region growing algorithm or a watershed algorithm can be applied for preliminary segmentation. The region growing algorithm starts from a pre-selected seed point and gradually expands the region based on pixel similarity; while the watershed algorithm regards the image as a topographic map and divides the region by simulating water flow. Both methods can effectively consider the spatial continuity of multi-feature data. However, these methods may lead to over-segmentation, so subsequent region merging is required.

[0049] Region merging can be based on a variety of criteria, such as spectral similarity, texture features, or shape features between regions. Graph theory methods can be used to treat each initial segmented region as a node of the graph, with the similarity between regions as the weight of the edge, and then merged through algorithms such as minimum spanning tree or spectral clustering. In addition, morphological operations such as opening and closing operations can be introduced to further optimize the segmentation results, eliminate small isolated regions, and fill holes. In order to more accurately identify potential leakage areas, prior knowledge and machine learning techniques can be combined. For example, a support vector machine (SVM) or random forest classifier can be trained to learn the characteristics of leakage areas using known leakage samples as training data. This classifier can be applied to the segmentation results to further filter and mark possible leakage areas. Finally, the segmentation results are post-processed and evaluated. This includes boundary refinement, small area filtering, and smoothing. Active contour models (such as the snake algorithm) can be used to refine the region boundaries so that they better fit the actual leakage boundaries. At the same time, an area threshold filter is applied to remove too small areas to reduce false detections. The final segmentation results should be verified for accuracy and effectiveness through visual inspection and quantitative evaluation.

[0050] For example, suppose that a large sports stadium is being analyzed for roof leakage. The multi-feature image data contains moisture index, temperature distribution, and visible light information. First, a Gaussian filter is applied to smooth the data, and then the Canny edge detector is used to enhance the edge features. Next, a three-layer Gaussian pyramid is constructed to capture leakage features at different scales. On each layer of the pyramid, the watershed algorithm is applied for preliminary segmentation. Due to the complexity of the roof structure, the initial segmentation result contains a large number of small areas. Then a spectral clustering algorithm is used to merge regions to merge adjacent regions with high spectral similarity. During the merging process, special attention is paid to retaining those regions that show anomalies in moisture index and temperature distribution. Next, a pre-trained random forest classifier is applied, which has learned the characteristics of typical leakage areas. This classifier helps further identify and mark possible leakage areas. For areas marked as potential leakage by the classifier, the snake algorithm is used to refine their boundaries so that they more accurately fit the actual leakage contour.

[0051] Finally, an area threshold filter is applied to remove areas with an area of ​​less than 5 square meters, because these small areas may be false detections or minor leaks that are not worthy of special attention. After this series of processing, a clear spatial distribution map of potential leakage areas is obtained. In this map, several major leakage areas can be clearly seen: one is located in the northwest corner of the roof, with an area of ​​about 20 square meters; the other extends along the drainage ditch in the center of the roof, about 30 meters long.

[0052] S107. Perform spectral index quantitative analysis on the spatial distribution map, calculate the degree of temperature anomaly, and measure the surface damage area to obtain a quantitative data set of leakage degree and a multi-dimensional risk assessment report.

[0053] Specifically, for the quantitative analysis of spectral indices, it is necessary to calculate the average normalized difference water index (NDWI) and moisture stress index (MSI) values ​​for each potential leakage area. This can be achieved by performing statistical analysis on the pixel values ​​in each area, calculating statistics such as mean, median, and standard deviation to comprehensively describe the moisture status of the area. At the same time, the degree of difference between these index values ​​and the normal area of ​​the entire building can also be calculated to quantify the severity of the leakage. Next, the degree of temperature anomaly is calculated. This step first requires determining the normal temperature range of the building surface, which can be obtained by analyzing the temperature distribution of non-leakage areas. Then, for each potential leakage area, the deviation of its average temperature from the normal temperature range is calculated. Statistical methods such as z-score or Mahalanobis distance can be introduced to quantify this degree of temperature anomaly, which can take into account the overall characteristics of the temperature distribution. In addition, the temperature gradient can also be analyzed because leakage areas often show a large temperature change rate.

[0054] Measuring the surface damage area is the next key step. This involves the combined analysis of visible light images and hyperspectral images. Image segmentation techniques, such as threshold segmentation or edge detection, can be used to identify visible damage such as surface discoloration, cracks or peeling. For each potential leakage area, the proportion of its damaged area to the total area is calculated, which can be used as an important indicator of the severity of the leakage. At the same time, the morphological characteristics of the damage, such as the length, width and direction of the cracks, can also be analyzed. This information is crucial for assessing structural risks. Combining the above analysis, a multidimensional leakage degree quantitative data set can be constructed. This data set should include indicators such as spectral index statistics, temperature anomaly, and surface damage ratio for each potential leakage area. In order to more comprehensively assess the risk, the structural characteristics and usage of the building also need to be considered. For example, the risk of leakage occurring on load-bearing walls is higher than that of non-load-bearing areas, and leakage close to electrical equipment is more dangerous than leakage far away from the equipment. Therefore, a location weight factor can be introduced to adjust the risk assessment according to the importance of the leakage location.

[0055] Finally, based on this quantitative data set, a multidimensional risk assessment report can be generated. This report should include the following aspects: first, a detailed description of each leakage area, including basic information such as its location, area, severity, etc.; second, an assessment of the impact of the leakage on the overall safety of the building, which needs to consider the location, scope and duration of the leakage; then an estimate of potential economic losses, including repair costs and possible collateral losses; and finally, repair recommendations and priority rankings to provide guidance for maintenance work based on the degree of risk and urgency.

[0056] For example, suppose you are analyzing a historical building for leakage. Through the previous steps, you have obtained a spatial distribution map of potential leakage areas. First, calculate the NDWI and MSI values ​​for each area. It is found that there is an area in the northwest corner of the roof with an NDWI value 50% higher than the normal area and an MSI value 20% lower, indicating serious moisture accumulation.

[0057] Temperature analysis shows that the average temperature in this area is 2.5℃ lower than the surrounding area, with a z-score of -3.2, indicating significant temperature anomalies. Surface damage analysis reveals that about 30% of the area has obvious discoloration and slight peeling. Considering that this leakage area is located near the main load-bearing structure, it is given a higher location weight factor of 1.5. Combining these factors, the leakage risk of this area is rated as "high". In the generated risk assessment report, the various indicators of this leakage area are described in detail, and the structural safety hazards that may result if not handled in time are warned. The repair cost is estimated, and it is recommended to be repaired before the start of the next dry season. In addition, the report includes analysis of several other medium and low risk leakage areas and provides a comprehensive assessment of the building's waterproof performance.

[0058] By executing the above steps, a multi-band hyperspectral scan of the building is performed to obtain hyperspectral image data in the range of 400nm-2500nm. Combined with thermal infrared imaging and high-resolution visible light imaging, it provides a multi-source, multi-dimensional raw data basis for leakage identification, enhancing the information richness and reliability of detection. The original hyperspectral image data is processed by multi-factor correction and spectral correction to obtain a reflectance data cube, which effectively eliminates interference factors such as atmospheric influence and sensor error, and ensures the data quality of subsequent analysis. By selecting characteristic bands and calculating spectral indices for the reflectance data cube, a set of moisture-related indices including normalized difference water index and moisture stress index are obtained, which directly reflects the moisture status of the building surface and provides a key basis for accurately locating the leakage area. The spectral data of each pixel in the reflectance data cube is smoothed and filtered, and the spectral derivative feature is extracted to obtain the spectral derivative feature that characterizes the spectral change rate. This processing can highlight subtle changes in the spectral curve and help identify early or minor leakage problems. The moisture-related index set, spectral derivative features, thermal infrared images and visible light images are fused with multi-source data to obtain multi-feature image data. This fusion makes full use of the complementarity of different data sources and improves the recognition of leakage features. The multi-feature image data is segmented at multiple scales to obtain the spatial distribution map of potential leakage areas. This step can adapt to leakage areas of different sizes and shapes, improving the flexibility and applicability of detection. Finally, the spatial distribution map is quantitatively analyzed by spectral index, the degree of temperature anomaly is calculated, and the surface damage area is measured to obtain a quantitative data set of leakage degree and a multi-dimensional risk assessment report. This comprehensive assessment process can not only accurately determine the location and degree of leakage, but also assess the potential risk of leakage to building structures, realizing the automation of the entire process from data collection, processing, analysis to risk assessment, greatly improving the efficiency and accuracy of leakage detection.

[0059] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0060] (1) performing dark current correction on the hyperspectral image data, collecting a dark current image without illumination, and performing dark current data analysis on the dark current image without illumination to obtain dark current reference data, and subtracting the dark current value of each pixel in each band of the hyperspectral image data based on the dark current reference data to obtain dark current corrected image data;

[0061] (2) calculating the response coefficient of each pixel in the dark current correction image data in each band to obtain a response coefficient set, and performing pixel response inconsistency correction on the dark current correction image data based on the response coefficient set to obtain flat field correction image data;

[0062] (3) Calculating the atmospheric transmittance and path radiance based on the pre-collected measured atmospheric parameters, and correcting the spectral value of each pixel of the flat-field corrected image data by the atmospheric transmittance and path radiance to obtain the atmospheric corrected image data;

[0063] (4) performing geometric distortion elimination processing on the atmospheric correction image data to obtain geometric correction image data, and performing spectral resampling and calibration on the geometric correction image data to obtain spectral correction image data;

[0064] (5) performing radiometric calibration on the spectrally corrected image data, converting the digital count values ​​into radiometric brightness values, and obtaining radiometric calibration image data;

[0065] (6) Perform reflectivity conversion on the radiation calibration image data, convert the radiation brightness value into surface reflectivity, and obtain the reflectivity data cube.

[0066] Specifically, dark current correction of hyperspectral image data is a key step in hyperspectral data processing. First, a non-illuminated image is acquired, which is achieved by imaging in complete darkness. The non-illuminated image reflects the response of the sensor when there is no incident light, mainly including dark current and readout noise. The non-illuminated image is statistically analyzed, and the average dark current value of each pixel in each band is calculated to obtain the dark current benchmark data. This process can be expressed as:

[0067]

[0068] Among them, D ij is the dark current reference value of pixel (i, j), D ijk is the dark current value collected for the kth time, and N is the number of collections.

[0069] Then, the dark current value of each pixel in each band in the hyperspectral image data is subtracted based on the dark current reference data. This process can be expressed as:

[0070] I corr (i,j,λ)=I raw (i,j,λ)-D ij ;

[0071] Among them, I corr (i, j, λ) is the corrected image intensity, I raw ,j,λ) is the original image intensity, λ represents the wavelength.

[0072] Next, flat field correction is performed to calculate the response coefficient of each pixel in the dark current correction image data in each band. This is obtained by imaging a uniform light source (such as an integrating sphere). The calculation formula for the response coefficient is:

[0073]

[0074] Among them, R ijλ is the response coefficient of pixel (i, j) at wavelength λ, I flat (i,j,λ) is the flat-field image intensity, is the average intensity of all pixels at that wavelength.

[0075] Based on the response coefficient set, the dark current correction image data is corrected for pixel response inconsistency to obtain flat field correction image data. The correction formula is:

[0076]

[0077] Atmospheric correction is the next key step. Atmospheric transmittance and path radiance are calculated based on pre-collected measured atmospheric parameters (such as air temperature, humidity, air pressure, etc.). This is usually done using an atmospheric radiation transfer model. Then, the spectral value of each pixel of the flat-field corrected image data is corrected and calculated using the atmospheric transmittance and path radiance to obtain the atmospheric corrected image data. Geometric correction aims to eliminate image distortion caused by factors such as the sensor field of view and terrain undulation. This is achieved by establishing a correspondence between the image coordinate system and the geographic coordinate system. Spectral resampling and calibration are then performed to ensure that the spectral data of each pixel corresponds to the correct wavelength. Radiometric calibration converts digital count values ​​into absolute radiance values. This requires calibration using a standard light source with known radiant intensity. Finally, reflectance conversion is performed to convert the radiance value into surface reflectance to obtain a reflectance data cube. This is usually achieved by using a standard reflector or a target with known reflectance within the site.

[0078] For example, when leak detection was performed on an office building, 100 unilluminated images were first collected in a completely dark environment. By performing statistical analysis on these images, the average dark current value of each pixel was obtained. For example, at a wavelength of 1000nm, the average dark current value of a certain pixel was 10 digital counts. Next, the actual acquired hyperspectral image was dark current corrected by subtracting the corresponding dark current value from the actual reading of each pixel. Flat field correction uses an integrating sphere as a uniform light source. By imaging the integrating sphere, the response coefficient of each pixel in each band is calculated. For example, at a wavelength of 1000nm, the response coefficient of a certain pixel is 1.05, which means that the response of this pixel is 5% higher than the average. Based on these response coefficients, the image after dark current correction was corrected for non-uniformity.

[0079] Atmospheric correction uses atmospheric parameters measured on site, including temperature 25°C, relative humidity 60%, air pressure 101.3 kPa, etc. Through the atmospheric radiation transmission model, it is calculated that at a wavelength of 1000 nm, the atmospheric transmittance is 0.85 and the path radiance is 0.02 W / (m2 ·sr·μm). These parameters were used to correct the atmospheric effects on the image. Geometric correction used the building corners as control points to establish the correspondence between the image coordinates and the actual geographic coordinates. Spectral correction ensured that the central wavelength accuracy of each band reached ±1nm. Radiometric calibration used a spectral radiance of 0.1W / (m 2 Finally, a standard reflector (reflectivity 30%) placed on the roof was used for reflectivity conversion.

[0080] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0081] (1) Perform band correlation analysis on the reflectance data cube, calculate the correlation coefficient matrix between each band, and obtain band correlation data;

[0082] (2) Perform eigenvalue decomposition on the band correlation data, extract the main eigenvectors, and obtain the band contribution ranking;

[0083] (3) Select the first N bands as candidate feature bands according to the band contribution ranking to obtain a candidate feature band set, where N is a positive integer greater than 1;

[0084] (4) Perform spectral resolution analysis on the candidate feature band set, calculate the spectral angular distance between adjacent bands, and obtain the band discrimination index;

[0085] (5) Screening the candidate feature band sets according to the band discrimination index, selecting the band combination with the highest discrimination, and obtaining the target feature band set;

[0086] (6) performing normalized difference water index calculation on the bands in the target characteristic band set to obtain difference water index data, and performing water stress index calculation on the bands in the target characteristic band set to obtain water stress index data;

[0087] (7) The difference water index data and the water stress index data are combined to obtain a set of water-related indices.

[0088] Specifically, the correlation coefficient matrix between each band is calculated, which is achieved by the Pearson correlation coefficient.

[0089] For bands i and j, the correlation coefficient is calculated as:

[0090]

[0091] Among them, x qi and x qj are the reflectance values ​​of the qth pixel in bands i and j, respectively. and is the average reflectivity of the corresponding band. After obtaining the correlation coefficient matrix, perform eigenvalue decomposition to extract the main eigenvectors. The basic equation of eigenvalue decomposition is:

[0092] Av=ωv;

[0093] Where A is the correlation coefficient matrix, v is the eigenvector, and ω is the corresponding eigenvalue.

[0094] The eigenvectors are sorted according to the eigenvalues ​​to obtain the band contribution ranking. The bands corresponding to the first N eigenvectors are selected as candidate feature bands to form a candidate feature band set. The selection of N is based on the cumulative contribution rate, and the eigenvectors with a cumulative contribution rate of more than 95% are usually selected.

[0095] The spectral resolution analysis is performed on the candidate feature band set, and the spectral angular distance between adjacent bands is calculated.

[0096] The calculation formula of SpectralAngleMapper (SAM) is:

[0097]

[0098] Among them, x i and i are the reflectance values ​​of two adjacent bands.

[0099] The candidate feature band sets are screened according to the band discrimination index, and the band combination with the highest discrimination is selected to obtain the target feature band set. This step can be achieved by setting a threshold or selecting a local maximum value.

[0100] Next, the Normalized Difference Water Index (NDWI) and Moisture Stress Index (MSI) are calculated for the bands in the target characteristic band set. The calculation formula of NDWI is:

[0101]

[0102] Among them, R 860 and R 1240 They are the reflectivity of 860nm and 1240nm bands respectively.

[0103] The calculation formula of MSI is:

[0104]

[0105] Among them, R 1600 and R 820 They are the reflectivity in the 1600nm and 820nm bands respectively.

[0106] Finally, the NDWI and MSI data are combined to obtain a set of moisture-related indices.

[0107] For example, when leak detection is performed on a large office building, the acquired hyperspectral data is first subjected to band correlation analysis. Assuming that the original data contains 200 bands, a 200x200 correlation coefficient matrix is ​​obtained by calculation. The matrix is ​​subjected to eigenvalue decomposition, and it is found that the cumulative contribution rate of the first 10 eigenvalues ​​reaches 97.5%. Therefore, the eigenvectors corresponding to these 10 eigenvalues ​​are selected to obtain the candidate feature band set. In the spectral resolution analysis, the spectral angular distances between these 10 bands are calculated. Assume that the spectral angular distance between bands 4 and 5 is the largest, which is 0.15 radians, while the distance between bands 7 and 8 is the smallest, which is 0.02 radians. Based on these results, 6 bands are selected as the final target feature band set, including bands with central wavelengths of 860nm, 1240nm, 1600nm, and 820nm.

[0108] The NDWI is calculated using the 860nm and 1240nm bands to obtain an NDWI image. In this image, the average NDWI value of the normal area is 0.2, while the NDWI value of the suspected leakage area is as high as 0.6. At the same time, the MSI is calculated using the 1600nm and 820nm bands to obtain an MSI image. The average MSI value of the normal area is 1.1, while the MSI value of the suspected leakage area drops to 0.8. The NDWI and MSI data are combined into a moisture-related index set, which provides key input data for subsequent leakage area identification. This method not only greatly reduces the amount of data that needs to be processed, but also extracts the spectral features that best reflect the moisture status of the building surface.

[0109] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0110] (1) Gaussian filtering is performed on the spectral data of each pixel in the reflectance data cube to obtain denoised spectral data;

[0111] (2) performing wavelet transform decomposition on the denoised spectral data to obtain multi-scale wavelet coefficients, and performing threshold processing on the multi-scale wavelet coefficients to remove the noise coefficients to obtain denoised wavelet coefficients;

[0112] (3) Perform wavelet reconstruction on the denoised wavelet coefficients to obtain the denoised spectrum curve;

[0113] (4) performing a first-order difference calculation on the denoised spectrum curve to obtain a first-order derivative spectrum, and performing a second-order difference calculation on the first-order derivative spectrum to obtain a second-order derivative spectrum;

[0114] (5) extracting feature points from the first-order derivative spectrum and the second-order derivative spectrum to obtain a derivative feature point set, wherein the derivative feature point set includes an inflection point, an extreme value point, and a zero crossing point;

[0115] (6) calculating the spectral slope, curvature and area characteristics according to the derivative feature point set to obtain the spectral derivative feature parameters;

[0116] (7) The first-order derivative spectrum, the second-order derivative spectrum and the spectral derivative characteristic parameters are combined to obtain the spectral derivative characteristics.

[0117] Specifically, Gaussian filtering of the spectral data of each pixel in the reflectance data cube is the first step in the noise reduction process. Gaussian filtering uses the Gaussian function as a weight to perform weighted averaging on the spectral data. The kernel function of the one-dimensional Gaussian filter is defined as:

[0118]

[0119] Among them, σ is the standard deviation of the Gaussian distribution, and x is the distance from the center point. By adjusting the σ value, the intensity of the filter is controlled. Next, the denoised spectral data is decomposed by wavelet transform. Wavelet transform can analyze the signal in both time and frequency domains, and is suitable for processing non-stationary signals. Commonly used wavelet basis functions include Haar, Daubechies, etc. Wavelet transform decomposes the signal into approximate coefficients and detail coefficients. Threshold processing is performed on the multi-scale wavelet coefficients to remove the noise coefficient. Threshold processing usually uses soft threshold or hard threshold methods, and the threshold can be determined by methods such as VisuShrink or BayesShrink.

[0120] The denoised wavelet coefficients are reconstructed by wavelet to obtain the denoised spectrum curve. The reconstruction process is the inverse process of decomposition, and the processed wavelet coefficients are recombined into a time domain signal.

[0121] Perform the first-order difference calculation on the noise reduction spectrum curve to obtain the first-order derivative spectrum. The first-order derivative calculation formula is:

[0122]

[0123] Among them, f(λ i ) is the wavelength λ i is the reflectivity at , Δλ is the wavelength interval.

[0124] The second-order derivative spectrum is obtained by performing a differential calculation on the first-order derivative spectrum. The second-order derivative can better highlight the changing characteristics of the spectral curve.

[0125] Feature points are extracted from the first-order derivative spectrum and the second-order derivative spectrum to obtain a set of derivative feature points. The inflection point is the extreme point of the first-order derivative, the extreme point is the zero point of the first-order derivative, and the zero crossing point is the zero point of the second-order derivative. These feature points reflect the key change positions of the spectral curve.

[0126] The spectral slope, curvature and area characteristics are calculated based on the derivative feature point set. The spectral slope reflects the rate of spectral change, the curvature indicates the degree of bending of the spectral curve, and the area characteristic reflects the spectral integrated intensity within a specific band. Finally, the first-order derivative spectrum, the second-order derivative spectrum and the spectral derivative characteristic parameters are combined to obtain the spectral derivative characteristics. These characteristics comprehensively reflect the changing characteristics of the spectral curve and play an important role in identifying subtle differences in materials.

[0127] For example, when leak detection is performed on the roof of a large industrial building, the acquired hyperspectral data is first processed. A pixel of a suspected leaking area is selected, and its original spectral data contains 1000 bands in the range of 400-2500nm. A Gaussian filter is applied to this spectral curve, and σ=2 is selected, with a window size of 5 bands. After filtering, the sharp noise peak originally near 1450nm is smoothed, and the signal-to-noise ratio is improved from the original 15dB to 22dB. Then a wavelet transform is performed, and the db4 wavelet basis function is selected for 5-layer decomposition. The obtained wavelet coefficients are soft-thresholded, and the threshold is set to 3 times the standard deviation. Through this step, high-frequency noise is further removed. The reconstructed spectral curve retains the main absorption features while significantly reducing the noise level. When calculating the first-order derivative, the difference between adjacent bands is used. Near 1450nm, the first-order derivative reaches a maximum value of 0.015 / nm, indicating strong moisture absorption. The second-order derivative calculation further enhances this feature, and a clear negative peak appears at 1450nm with a peak value of -0.0005 / nm^2.

[0128] Feature point extraction found that there are obvious inflection points at 1450nm and 1940nm, which correspond to the two main absorption bands of water. The integral area in the range of 1400-1500nm is calculated to be 25 (normalized unit), which is much higher than the average value of 10 in the non-leakage area.

[0129] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0130] (1) The moisture-related index set is processed by nonlinear adaptive histogram specification to obtain enhanced moisture index data, and the spectral derivative characteristics are processed by dual-tree complex wavelet packet transform to obtain multi-scale and multi-directional spectral feature data;

[0131] (2) Perform anisotropic diffusion filtering and local binary pattern feature extraction on the thermal infrared image to obtain temperature texture feature data;

[0132] (3) Perform multi-scale and multi-directional Shearlet transform on the visible light image to obtain geometric structure feature data;

[0133] (4) Performing high-order singular value decomposition on the enhanced moisture index data, multi-scale and multi-directional spectral feature data, temperature texture feature data, and geometric structure feature data to obtain multi-modal tensor features;

[0134] (5) Performing a tensor kernel principal component analysis on the multimodal tensor features to obtain nonlinear dimensionality reduction feature data, and performing a hypergraph clustering process on the nonlinear dimensionality reduction feature data to obtain a multi-scale feature association graph;

[0135] (6) Perform graph attention analysis on the multi-scale feature association graph to obtain adaptive weight fusion features;

[0136] (7) Perform inverse Hadamard transform on the adaptive weight fusion features to obtain multi-feature image data.

[0137] Specifically, the moisture-related index set is processed by nonlinear adaptive histogram normalization to enhance the contrast and details of the moisture index. This processing method first calculates the cumulative distribution function of the input image and then maps it to the desired output distribution. This process can effectively highlight the slight changes in the moisture index, making the leakage area more obvious. The enhanced moisture index data obtained after processing has a higher dynamic range and clarity. The spectral derivative features are processed by dual-tree complex wavelet packet transform. Dual-tree complex wavelet packet transform is an advanced multi-resolution analysis method that can not only analyze signals at multiple scales, but also provide directional information. This transform decomposes the spectral derivative features into sub-bands of different scales and directions, thereby obtaining multi-scale and multi-directional spectral feature data. These data can more comprehensively capture the local and global variation characteristics of the spectrum.

[0138] The processing of thermal infrared images includes two steps: anisotropic diffusion filtering and local binary pattern feature extraction. Anisotropic diffusion filtering can maintain edge information while smoothing the image, effectively reducing noise without blurring important temperature boundaries. Local binary pattern feature extraction can capture the texture information of temperature distribution. The combination of these two steps produces temperature texture feature data, which contains both the spatial distribution information of temperature and the texture characteristics of temperature changes. Visible light images are processed by multi-scale and multi-directional Shearlet transform to obtain geometric structure feature data. Shearlet transform is a multi-scale method that can effectively represent anisotropic features in high-dimensional data. It can accurately capture edges and other geometric structures in images, and is particularly suitable for analyzing subtle structural changes on the surface of buildings.

[0139] The enhanced moisture index data, multi-scale and multi-directional spectral feature data, temperature texture feature data and geometric structure feature data are processed by high-order singular value decomposition to obtain multimodal tensor features. High-order singular value decomposition is a method of tensor decomposition that can effectively process multidimensional data. This step integrates feature data from different sources into a unified multimodal representation, retaining the relationships between various features. The multimodal tensor features are processed by tensor kernel principal component analysis to achieve nonlinear dimensionality reduction. Tensor kernel principal component analysis is an extension of traditional principal component analysis on high-dimensional tensor data, which can capture nonlinear relationships in the data. This step greatly reduces the dimensionality of the data while retaining the most important feature information.

[0140] The feature data after dimensionality reduction is processed by hypergraph clustering to obtain a multi-scale feature association graph. Hypergraph clustering is an advanced clustering method that can handle complex high-dimensional data structures. It clusters similar features together to form a multi-scale feature association network. Graph attention analysis is performed on the multi-scale feature association graph to obtain adaptive weight fusion features. The graph attention mechanism can automatically learn the importance relationship between different feature nodes, thereby assigning appropriate weights to different features. This step makes the final feature representation more focused on key information related to leakage detection.

[0141] Finally, the adaptive weight fusion features are subjected to inverse Hadamard transform processing to obtain multi-feature image data. Hadamard transform is an orthogonal transform, and its inverse transform can remap the data in feature space back to image space. This step reconstructs the fused features into a visual image form, which is convenient for subsequent analysis and interpretation.

[0142] For example, when leak detection is performed on the exterior wall of a large commercial building, the moisture index image is first processed by nonlinear adaptive histogram normalization. Before processing, the moisture index ranges from 0.1 to 0.4, and after processing, it is expanded to -0.2 to 0.8, making the originally inconspicuous moisture changes clearer. The spectral derivative features are transformed by a 5-level double-tree complex wavelet packet transform to obtain 32 directional subbands, each of which contains spectral change information of different scales and directions. After anisotropic diffusion filtering, the noise level of the thermal infrared image is reduced by 60%, while the clarity of the temperature boundary is retained. By extracting the local binary pattern features, a 16-dimensional feature vector reflecting the temperature distribution texture is obtained. Visible light images are transformed by 4-scale and 8-directional Shearlet transform to generate multidimensional data that can accurately describe geometric features such as cracks and shedding on the building surface.

[0143] These features are represented by a 5th-order tensor, and the first 100 principal components are retained after high-order singular value decomposition, explaining 95% of the data variance. Tensor kernel principal component analysis reduces the feature dimension from 100 to 20 while retaining 90% of the information. Hypergraph clustering divides these 20-dimensional features into 5 main categories, forming a feature association graph containing 500 nodes. Graph attention analysis assigns importance weights to each node, ranging from 0.01 to 0.95. Finally, a comprehensive image containing multiple feature information is generated through inverse Hadamard transformation. In this image, the leakage area appears as a high-brightness area, which is in sharp contrast with the surrounding area. For example, in the northwest corner of the building's exterior wall, a highlight area of ​​about 2 square meters appears, with an average pixel value of 220 (full scale 255), while the average value of the normal area is only 100, clearly identifying a serious leakage point.

[0144] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0145] (1) Perform Gaussian pyramid decomposition on multi-feature image data to obtain image pyramids at three scales;

[0146] (2) Perform Sobel operator edge detection on the image pyramids at three scales to obtain a multi-scale edge map;

[0147] (3) Perform OTSU adaptive threshold segmentation on the multi-scale edge map to obtain the initial binary segmentation map;

[0148] (4) Perform morphological opening and closing operations on the initial binary segmentation image to obtain a denoised binary image;

[0149] (5) marking the connected regions of the denoised binary image to obtain an initial region marking map, and merging the regions of the initial region marking map to merge the regions with an area less than 50 square pixels to obtain a merged region map;

[0150] (6) Calculating the spectral similarity and texture similarity of adjacent regions in the merged region map to obtain a region similarity matrix;

[0151] (7) Based on the regional similarity matrix, adjacent regions with similarity greater than 0.85 are merged to obtain a candidate leakage region map;

[0152] (8) Refine the boundaries of the candidate leakage area map to obtain the spatial distribution map of the potential leakage area.

[0153] Specifically, in this method, three-scale image pyramids are generated, and the image resolution of each scale is reduced by half successively. This multi-scale representation can capture the characteristics of leakage areas of different sizes. The Sobel operator edge detection is performed on the three-scale image pyramids respectively to obtain a multi-scale edge map. The Sobel operator is a classic edge detection operator that detects edges by calculating the gradients of the image in the horizontal and vertical directions. In leakage detection, edge information is crucial for locating potential leakage areas because leakage areas usually show obvious feature changes at the edges.

[0154] OTSU adaptive threshold segmentation is a method to automatically determine the optimal threshold. It determines the optimal threshold by maximizing the inter-class variance and divides the image into foreground and background. Perform OTSU adaptive threshold segmentation on the multi-scale edge map to obtain an initial binary segmentation map. This step converts the continuous edge intensity map into a discrete binary image and preliminarily distinguishes potential leakage areas. Morphological opening and closing operations are effective methods for removing noise and smoothing boundaries. The opening operation corrodes first and then expands, which can remove small protrusions; the closing operation expands first and then corrodes, which can fill small holes. Performing these two operations on the initial binary segmentation map will obtain a denoised binary image, which helps to eliminate the fine noise and irregular edges generated during the segmentation process.

[0155] Connected region labeling is the process of grouping connected pixels in a binary image into a region. Connected region labeling is performed on the denoised binary image to obtain an initial region labeling map. Regions are then merged to merge regions with an area of ​​less than 50 square pixels into adjacent larger regions. This step helps to eliminate regions that are too small and reduce the complexity of subsequent processing. The spectral similarity and texture similarity of adjacent regions in the merged region map are calculated to obtain a region similarity matrix. Spectral similarity can be measured by calculating the Euclidean distance of the average spectral features of the two regions, while texture similarity can be determined by comparing the grayscale co-occurrence matrix features of the regions. This step provides a basis for subsequent region merging.

[0156] According to the region similarity matrix, adjacent regions with similarity greater than 0.85 are further merged to obtain a candidate leakage region map. The selection of this threshold is based on empirical values ​​and aims to balance the trade-off between over-segmentation and under-segmentation. Regions with high similarity are more likely to belong to the same leakage region, so merging them can obtain a more accurate leakage region boundary.

[0157] Finally, the boundary of the candidate leakage region map is refined to obtain the spatial distribution map of the potential leakage region. Boundary refinement can be achieved by active contour model or level set method, which can make the segmentation boundary better fit the actual leakage region edge.

[0158] For example, when leak detection is performed on the exterior wall of a large office building, the multi-feature image data is first decomposed by Gaussian pyramid. The original image resolution is 2048×2048 pixels, and two additional scale images of 1024×1024 and 512×512 are obtained through Gaussian pyramid decomposition. The Sobel operator is applied to the images of these three scales to obtain edge intensity maps. At the original scale, the edge intensity value ranges from 0 to 255, where the average edge intensity of the leaking area reaches 180, while that of the normal area is only 50. The optimal threshold automatically calculated by the OTSU algorithm is 125, based on which the initial binary segmentation map is obtained. Morphological processing uses 3×3 structural elements for opening and closing operations, effectively removing noise with an area less than 9 square pixels. After the connected regions are marked, the initial number of regions is 500, which is reduced to 200 after small regions are merged. When calculating the regional similarity, the contrast features of the spectral angular distance and the gray-level co-occurrence matrix are used. In the similarity matrix, the average similarity between normal areas is 0.95, while the average similarity between leakage areas and normal areas is only 0.6. After setting the threshold of 0.85 for region merging, 15 candidate leakage areas were finally obtained. The boundary refinement used the level set method. After 50 iterations, the consistency between the boundary of the leakage area and the actual water seepage trace reached 95%. The final spatial distribution map clearly shows 5 major leakage areas, the largest of which is located in the southwest corner of the building, with an area of ​​about 4 square meters and an irregular elliptical shape, which is highly consistent with the field survey results.

[0159] In a specific embodiment, the process of executing step S107 may specifically include the following steps:

[0160] (1) Perform spectral index statistical analysis on each potential leakage area in the spatial distribution map, calculate the mean, variance and distribution characteristics of the normalized difference water index and water stress index, and obtain the regional spectral index feature set;

[0161] (2) Threshold segmentation is performed on the regional spectral index feature set to divide it into three leakage levels: mild, moderate, and severe, and the preliminary leakage degree classification results are obtained;

[0162] (3) Perform thermal infrared temperature extraction on the potential leakage area, calculate the temperature difference and temperature gradient inside and outside the area, and obtain the temperature anomaly measurement value;

[0163] (4) Correct the preliminary leakage degree classification result according to the temperature anomaly measurement value to obtain the comprehensive leakage degree classification result;

[0164] (5) Perform morphological processing on the potential leakage area, extract the regional boundary and area information, and obtain surface damage characteristic data;

[0165] (6) Fusing the comprehensive leakage degree classification results with the surface damage feature data to obtain a leakage degree quantitative data set;

[0166] (7) Perform spatial cluster analysis on the leakage degree quantitative data set to identify high-risk leakage clusters and obtain a risk distribution map;

[0167] (8) Based on the risk distribution map and the building structure information, conduct leakage impact analysis on the building and obtain the structural risk assessment results;

[0168] (9) Data fusion is performed on the leakage degree quantitative data set, risk distribution map and structural risk assessment results to generate a multidimensional risk assessment report.

[0169] Specifically, statistical analysis of the spectral index for each potential leakage area in the spatial distribution map is a key step in quantifying the degree of leakage. First, the mean, variance, and distribution characteristics of the normalized difference water index (NDWI) and water stress index (MSI) in each region are calculated. NDWI reflects the difference in reflectance between the near-infrared and short-wave infrared bands, while MSI represents the reflectance ratio between the short-wave infrared band and the near-infrared band. The statistical characteristics of these indices constitute the regional spectral index feature set, which provides basic data for the subsequent classification of leakage degree. Next, the regional spectral index feature set is threshold segmented to divide the leakage degree into three levels: mild, moderate, and severe. The selection of thresholds is based on a large amount of experimental data and expert experience. Generally, the higher the NDWI value and the lower the MSI value, the more severe the leakage degree. This step obtains the preliminary leakage degree classification results, laying the foundation for the subsequent comprehensive evaluation.

[0170] Thermal infrared temperature extraction is another important dimension of leak identification. Thermal infrared temperature extraction is performed on potential leak areas to calculate the temperature difference and temperature gradient inside and outside the area. Leak areas usually show temperature anomalies, which may be lower (due to heat absorption due to water evaporation) or higher (due to changes in heat conduction caused by humidity) than the surrounding areas. The temperature anomaly metric takes into account the temperature difference and gradient information, providing an important reference for further assessment of the degree of leakage. The preliminary leakage degree classification results are corrected according to the temperature anomaly metric to obtain the comprehensive leakage degree classification results. This step combines spectral information with thermal information to improve the accuracy and reliability of classification. For example, if the spectral index of an area indicates that the leakage is mild, but the temperature anomaly is significant, it may be reclassified as moderate or severe leakage.

[0171] Morphological processing is performed on the potential leakage area to extract the regional boundary and area information and obtain surface damage characteristic data. Morphological processing includes operations such as corrosion and expansion, which can smooth the boundaries and eliminate small holes or protrusions. Boundary and area information reflects the spatial distribution characteristics of leakage and is an important indicator for assessing the severity of leakage. The comprehensive leakage degree classification results are fused with the surface damage characteristic data to obtain a leakage degree quantitative data set. This fusion process takes into account information in multiple dimensions such as spectrum, temperature and spatial morphology, and comprehensively reflects the characteristics and degree of leakage.

[0172] Spatial cluster analysis is performed on the quantitative leakage data set to identify high-risk leakage clusters and obtain a risk distribution map. Spatial cluster analysis can use the K-means algorithm or DBSCAN algorithm to aggregate similar leakage areas together, thereby discovering large-scale leakage problems or leakage diffusion trends. According to the risk distribution map and the building structure information of the building, the building is analyzed for leakage impact to obtain the structural risk assessment results. This step takes into account the relative position of the leakage location and the key structure of the building (such as load-bearing walls, electrical equipment, etc.) to evaluate the potential impact of leakage on building safety and function.

[0173] Finally, the leakage degree quantitative data set, risk distribution map and structural risk assessment results are fused to generate a multi-dimensional risk assessment report. This report integrates all analysis results and provides building managers with a comprehensive leakage risk assessment and repair recommendations.

[0174] For example, when conducting leakage detection on a large commercial complex, the spectral index analysis is first performed on the identified potential leakage area. In a suspected leakage area of ​​10 square meters, the average value of NDWI is 0.3 (normal areas are usually <0.1), the standard deviation is 0.05, and it shows a positive skewed distribution; the average value of MSI is 0.8 (normal areas are usually >1.2), the standard deviation is 0.1, and it shows a negative skewed distribution. Based on these data, the area was initially classified as moderate leakage. Thermal infrared analysis showed that the average temperature in the area was 2.5℃ lower than the surrounding area, and the temperature gradient reached 0.5℃ / m. Considering the significance of the temperature anomaly, the leakage degree was revised to severe. Morphological analysis showed that the leakage area was an irregular ellipse with a circumference of about 14 meters and an area that increased by 20% compared with the initial identification, indicating that the leakage had a tendency to spread. Spatial clustering analysis found that there was a high-risk leakage cluster in the southwest corner of the building, including 5 adjacent severe leakage areas with a total area of ​​50 square meters. Structural analysis showed that this gathering area was adjacent to the main load-bearing wall and the electrical equipment room, and the risk level was rated as "extremely high". The final risk assessment report pointed out that the leakage problem in the southwest corner needed to be addressed immediately, and it was recommended to carry out comprehensive waterproofing and structural reinforcement within 30 days. The report also included detailed information and repair priority rankings for the other 12 medium and low risk leakage points, providing comprehensive and precise guidance for building maintenance.

[0175] The embodiment of the present invention also provides a house leakage identification system based on hyperspectral and imaging technology, such as Figure 2 As shown, the house leakage identification system based on hyperspectral and imaging technology specifically includes:

[0176] The scanning module 201 is used to perform multi-band hyperspectral scanning on the building to obtain hyperspectral image data, wherein the spectral range of the hyperspectral image data is 400nm-2500nm, and perform thermal infrared imaging and high-resolution visible light imaging on the building to obtain a thermal infrared image and a visible light image;

[0177] A correction module 202 is used to perform multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube;

[0178] A calculation module 203 is used to perform characteristic band selection and spectral index calculation on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index;

[0179] The filtering module 204 is used to perform smoothing filtering on the spectral data of each pixel in the reflectance data cube to obtain a de-noised spectral curve, and to extract spectral derivative features on the de-noised spectral curve to obtain spectral derivative features, wherein the spectral derivative features are used to characterize the spectral change rate;

[0180] A fusion module 205 is used to perform multi-source data fusion on the moisture-related index set, the spectral derivative features, the thermal infrared image and the visible light image to obtain multi-feature image data;

[0181] A segmentation module 206 is used to perform multi-scale segmentation on the multi-feature image data to obtain a spatial distribution map of potential leakage areas;

[0182] The analysis module 207 is used to perform spectral index quantitative analysis, temperature anomaly degree calculation and surface damage area measurement on the spatial distribution map to obtain a leakage degree quantitative data set and a multi-dimensional risk assessment report.

[0183] Through the collaborative work of the above modules, a multi-band hyperspectral scan of the building is performed to obtain hyperspectral image data in the range of 400nm-2500nm. Combined with thermal infrared imaging and high-resolution visible light imaging, it provides a multi-source, multi-dimensional raw data basis for leakage identification, enhancing the information richness and reliability of detection. The original hyperspectral image data is subjected to multi-factor correction and spectral correction processing to obtain a reflectance data cube, which effectively eliminates interference factors such as atmospheric influence and sensor error, and ensures the data quality of subsequent analysis. By selecting characteristic bands and calculating spectral indices for the reflectance data cube, a set of moisture-related indices including the normalized difference water index and moisture stress index are obtained, which directly reflects the moisture status of the building surface and provides a key basis for accurately locating the leakage area. The spectral data of each pixel in the reflectance data cube is smoothed and filtered, and the spectral derivative feature is extracted to obtain the spectral derivative feature that characterizes the spectral change rate. This processing can highlight subtle changes in the spectral curve and help identify early or minor leakage problems. The moisture-related index set, spectral derivative features, thermal infrared images and visible light images are fused with multi-source data to obtain multi-feature image data. This fusion makes full use of the complementarity of different data sources and improves the recognition of leakage features. The multi-feature image data is segmented at multiple scales to obtain the spatial distribution map of potential leakage areas. This step can adapt to leakage areas of different sizes and shapes, improving the flexibility and applicability of detection. Finally, the spatial distribution map is quantitatively analyzed by spectral index, the degree of temperature anomaly is calculated, and the surface damage area is measured to obtain a quantitative data set of leakage degree and a multi-dimensional risk assessment report. This comprehensive assessment process can not only accurately determine the location and degree of leakage, but also assess the potential risk of leakage to building structures, realizing the automation of the entire process from data collection, processing, analysis to risk assessment, greatly improving the efficiency and accuracy of leakage detection.

[0184] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the embodiments, a person skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for identifying house leakage based on hyperspectral and imaging technology, characterized in that: include: Performing multi-band hyperspectral scanning on the building to obtain hyperspectral image data, wherein the spectral range of the hyperspectral image data is 400nm-2500nm, and performing thermal infrared imaging and high-resolution visible light imaging on the building to obtain a thermal infrared image and a visible light image; Performing multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube; Selecting characteristic bands and calculating spectral indices on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index; Gaussian filtering is performed on the spectral data of each pixel in the reflectance data cube to obtain denoised spectral data; wavelet transform is performed on the denoised spectral data to obtain multi-scale wavelet coefficients, and threshold processing is performed on the multi-scale wavelet coefficients to remove noise coefficients and obtain denoised wavelet coefficients; wavelet reconstruction is performed on the denoised wavelet coefficients to obtain a denoised spectral curve; first-order difference calculation is performed on the denoised spectral curve to obtain a first-order derivative spectrum, and second-order difference calculation is performed on the first-order derivative spectrum to obtain a second-order derivative spectrum; feature point extraction is performed on the first-order derivative spectrum and the second-order derivative spectrum to obtain a derivative feature point set, and the derivative feature point set includes an inflection point, an extreme value point and a zero crossing point; spectrum slope, curvature and area features are calculated according to the derivative feature point set to obtain spectrum derivative feature parameters; Combining the first-order derivative spectrum, the second-order derivative spectrum and the spectrum derivative characteristic parameter to obtain a spectrum derivative characteristic; Performing multi-source data fusion on the moisture-related index set, the spectral derivative features, the thermal infrared image, and the visible light image to obtain multi-feature image data; Performing multi-scale segmentation on the multi-feature image data to obtain a spatial distribution map of potential leakage areas; A spectral index statistical analysis is performed on each potential leakage area in the spatial distribution map, and the mean, variance and distribution characteristics of the normalized difference water index and the water stress index are calculated to obtain a regional spectral index feature set; the regional spectral index feature set is threshold segmented and divided into three leakage levels of mild, moderate and severe to obtain a preliminary leakage degree classification result; thermal infrared temperature extraction is performed on the potential leakage area, and the temperature difference and temperature gradient inside and outside the area are calculated to obtain a temperature anomaly measurement value; the preliminary leakage degree classification result is corrected according to the temperature anomaly measurement value to obtain a comprehensive leakage degree classification result; the potential leakage area is classified into a comprehensive leakage degree classification result according to the temperature anomaly measurement value. The leakage area is morphologically processed to extract the regional boundary and area information to obtain surface damage feature data; the comprehensive leakage degree classification result is data-fused with the surface damage feature data to obtain a leakage degree quantitative data set; spatial clustering analysis is performed on the leakage degree quantitative data set to identify high-risk leakage clustering areas to obtain a risk distribution map; based on the risk distribution map and the building structure information of the building, the building is analyzed for leakage impact to obtain a structural risk assessment result; the leakage degree quantitative data set, risk distribution map and structural risk assessment result are data-fused to generate a multidimensional risk assessment report.

2. The method for identifying house leakage based on hyperspectral and imaging technology according to claim 1, characterized in that: The multi-factor correction includes dark current correction, flat field correction, atmospheric correction, geometric correction and spectral correction. The step of performing multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube includes: Performing dark current correction on the hyperspectral image data, collecting a no-light image, and performing dark current data analysis on the no-light image to obtain dark current reference data, and subtracting a dark current value from each pixel in each band of the hyperspectral image data based on the dark current reference data to obtain dark current corrected image data; Calculating the response coefficient of each pixel in the dark current correction image data in each band to obtain a response coefficient set, and performing pixel response inconsistency correction on the dark current correction image data based on the response coefficient set to obtain flat field correction image data; Calculating atmospheric transmittance and path radiance based on pre-collected measured atmospheric parameters, and performing correction calculation on the spectral value of each pixel of the flat-field corrected image data by using the atmospheric transmittance and the path radiance to obtain atmospheric corrected image data; Performing geometric distortion elimination processing on the atmospheric correction image data to obtain geometric correction image data, and performing spectral resampling and calibration on the geometric correction image data to obtain spectral correction image data; Performing radiation calibration on the spectrally corrected image data, converting the digital count value into a radiation brightness value, and obtaining radiation calibration image data; The radiation calibration image data is subjected to reflectivity conversion, the radiation brightness value is converted into the surface reflectivity, and the reflectivity data cube is obtained.

3. The housing leakage identification method based on hyperspectral and imaging technology according to claim 1 is characterized in that: The step of selecting characteristic bands and calculating spectral indices on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index, comprises: Performing band correlation analysis on the reflectivity data cube, calculating the correlation coefficient matrix between each band, and obtaining band correlation data; Performing eigenvalue decomposition on the band correlation data, extracting main eigenvectors, and obtaining band contribution ranking; According to the band contribution ranking, the first N bands are selected as candidate feature bands to obtain a candidate feature band set, where N is a positive integer greater than 1; Performing spectral resolution analysis on the candidate feature band set, calculating the spectral angular distance between adjacent bands, and obtaining a band discrimination index; The candidate feature band sets are screened according to the band discrimination index, and the band combination with the highest discrimination is selected to obtain the target feature band set; Performing normalized difference water index calculation on the bands in the target characteristic band set to obtain difference water index data, and performing water stress index calculation on the bands in the target characteristic band set to obtain water stress index data; The differential water index data and the water stress index data are combined to obtain a water-related index set.

4. The method for identifying house leakage based on hyperspectral and imaging technology according to claim 1, characterized in that: The step of performing multi-source data fusion on the moisture-related index set, the spectral derivative features, the thermal infrared image and the visible light image to obtain multi-feature image data includes: The moisture-related index set is subjected to nonlinear adaptive histogram regulation processing to obtain enhanced moisture index data, and the spectral derivative characteristics are subjected to dual-tree complex wavelet packet transformation processing to obtain multi-scale and multi-directional spectral characteristic data; Performing anisotropic diffusion filtering and local binary pattern feature extraction processing on the thermal infrared image to obtain temperature texture feature data; Performing multi-scale and multi-directional Shearlet transform processing on the visible light image to obtain geometric structure feature data; Performing high-order singular value decomposition processing on the enhanced moisture index data, the multi-scale multi-directional spectral feature data, the temperature texture feature data, and the geometric structure feature data to obtain multi-modal tensor features; Performing a tensor kernel principal component analysis process on the multimodal tensor features to obtain nonlinear dimensionality reduction feature data, and performing a hypergraph clustering process on the nonlinear dimensionality reduction feature data to obtain a multi-scale feature association graph; Performing graph attention analysis on the multi-scale feature association graph to obtain an adaptive weight fusion feature; The adaptive weight fusion feature is subjected to inverse Hadamard transform processing to obtain multi-feature image data.

5. The method for identifying house leakage based on hyperspectral and imaging technology according to claim 1, characterized in that: The step of performing multi-scale segmentation on the multi-feature image data to obtain a spatial distribution map of potential leakage areas includes: Performing Gaussian pyramid decomposition on the multi-feature image data to obtain image pyramids of three scales; Performing Sobel operator edge detection on the image pyramids of the three scales respectively to obtain a multi-scale edge map; Performing OTSU adaptive threshold segmentation on the multi-scale edge map to obtain an initial binary segmentation map; Performing morphological opening and closing operations on the initial binary segmentation image to obtain a denoised binary image; Marking connected regions on the de-noised binary image to obtain an initial region marking map, and merging regions on the initial region marking map to merge regions with an area less than 50 square pixels to obtain a merged region map; Calculate the spectral similarity and texture similarity of adjacent regions in the merged region map to obtain a region similarity matrix; According to the regional similarity matrix, adjacent regions with similarities greater than 0.85 are merged to obtain a candidate leakage region map; The boundary of the candidate leakage area map is refined to obtain a spatial distribution map of potential leakage areas.

6. A house leakage identification system based on hyperspectral and imaging technology, used to execute the house leakage identification method based on hyperspectral and imaging technology as claimed in any one of claims 1 to 5, characterized in that: include: A scanning module is used to perform multi-band hyperspectral scanning on the building to obtain hyperspectral image data, wherein the spectral range of the hyperspectral image data is 400nm-2500nm, and perform thermal infrared imaging and high-resolution visible light imaging on the building to obtain a thermal infrared image and a visible light image; A correction module, used for performing multi-factor correction and spectral correction processing on the hyperspectral image data to obtain a reflectance data cube; A calculation module, used for performing characteristic band selection and spectral index calculation on the reflectance data cube to obtain a moisture-related index set, wherein the moisture-related index set includes a normalized difference water index and a moisture stress index; A filtering module, used to perform Gaussian filtering on the spectral data of each pixel in the reflectance data cube to obtain denoised spectral data; perform wavelet transform decomposition on the denoised spectral data to obtain multi-scale wavelet coefficients, and perform threshold processing on the multi-scale wavelet coefficients to remove noise coefficients to obtain denoised wavelet coefficients; perform wavelet reconstruction on the denoised wavelet coefficients to obtain a denoised spectral curve; perform first-order difference calculation on the denoised spectral curve to obtain a first-order derivative spectrum, and perform second-order difference calculation on the first-order derivative spectrum to obtain a second-order derivative spectrum; perform feature point extraction on the first-order derivative spectrum and the second-order derivative spectrum to obtain a derivative feature point set, the derivative feature point set includes inflection points, extreme points and zero crossing points; calculate spectral slope, curvature and area features according to the derivative feature point set to obtain spectral derivative feature parameters; combine the first-order derivative spectrum, the second-order derivative spectrum and the spectral derivative feature parameters to obtain spectral derivative features; A fusion module, used for performing multi-source data fusion on the moisture-related index set, the spectral derivative features, the thermal infrared image and the visible light image to obtain multi-feature image data; A segmentation module, used for performing multi-scale segmentation on the multi-feature image data to obtain a spatial distribution map of potential leakage areas; The analysis module is used to perform spectral index statistical analysis on each potential leakage area in the spatial distribution map, calculate the mean, variance and distribution characteristics of the normalized difference water index and the water stress index, and obtain a regional spectral index feature set; perform threshold segmentation on the regional spectral index feature set to divide it into three leakage levels of light, moderate and severe to obtain a preliminary leakage degree classification result; perform thermal infrared temperature extraction on the potential leakage area, calculate the temperature difference and temperature gradient inside and outside the area, and obtain a temperature anomaly measurement value; modify the preliminary leakage degree classification result according to the temperature anomaly measurement value to obtain a comprehensive leakage degree classification result; The potential leakage area is morphologically processed to extract the regional boundary and area information to obtain surface damage feature data; the comprehensive leakage degree classification result is data-fused with the surface damage feature data to obtain a leakage degree quantification data set; spatial clustering analysis is performed on the leakage degree quantification data set to identify high-risk leakage clustering areas to obtain a risk distribution map; according to the risk distribution map and the building structure information of the building, the building is analyzed for leakage impact to obtain a structural risk assessment result; the leakage degree quantification data set, the risk distribution map and the structural risk assessment result are data-fused to generate a multidimensional risk assessment report.

Images