Hyperspectral tailing pond dry beach line intelligent analysis method and system

Abstract

The invention provides a hyperspectral tailing pond dry beach line intelligent analysis method, and belongs to the technical field of image processing. Through multiband characteristics of a hyperspectral remote sensing image, principal component clustering analysis and characteristic wavelength selection are combined, so that the problems of insufficient spectral information of the image after dimension reduction and serious data redundancy caused by use of a full-band image are solved; single-band images corresponding to different characteristic bands emphasize different target information in the original image, and different combinations are beneficial to fully exerting the advantages of different target information of each band, so that a water area and a dry beach are distinguished more accurately; a double-branch network model is adopted, the difference between spectral dimensions and spatial dimensions is synchronously analyzed, spectral features and spatial information are fused, and the two branches have certain complementarity, so that rich spatial spectrum information of an input hyperspectral remote sensing image can be fully excavated, a dry beach line can be identified more accurately and effectively, the length of the dry beach is calculated, manual intervention is reduced, and the working efficiency is improved. The treatment efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to a hyperspectral tailing pond dry beach line intelligent analysis method and system. BACKGROUND

[0002] A tailing pond refers to a place formed by dam interception of a valley mouth or a surrounding land, used for storing tailings discharged after ore selection in a metal or non-metal mine or other industrial waste residues. The tailing pond is a man-made debris flow hazard source with high potential energy, and has the risk of dam break. Once an accident occurs, it is easy to cause a major accident. At present, there are many tailing ponds in China, which are widely distributed. Many tailing ponds have been in operation for many years, and the capacity is gradually decreasing, the ability to resist natural disasters is declining, and the safety hazards are increasing. The safety of the tailing pond is related to the safety of people's lives and property and the environment in its influence area, so the safety monitoring of the tailing pond is of great significance.

[0003] The measurement of the dry beach in front of the tailing pond dam body is a very important link in the safety monitoring of the tailing pond. The main parameters of the dry beach include the dry beach length, the dry beach area and the dry beach water surface boundary line. These parameters change with external factors, including the injection of tailing sand, the change of water quantity in tailings, rain and sunlight, which will cause changes in the dry beach, leading to a certain degree of danger of the tailings. These parameters of the dry beach directly determine the stability of the dam body. If the dry beach monitoring and maintenance are not in place, the dam body will lose stability and even collapse. Therefore, corresponding monitoring needs to be carried out in key areas and special positions, and the dry beach length, the dry beach area and the dry beach water surface boundary line are measured in real time as the basis for maintenance and care.

[0004] The traditional dry beach length measurement method has various limitations: the manual visual estimation method highly depends on the experience of workers, has strong subjectivity and difficult to guarantee the accuracy, especially in adverse weather conditions, which has safety hazards; the hardware collection device has high accuracy, but has high installation and maintenance cost, and is difficult to adapt to the dynamic changes of the tailing pond terrain; the method based on camera and marker has high accuracy in theory, but the marker is easy to be corroded, buried or damaged by the tailing sand, and the construction and maintenance are difficult, and manual intervention is still needed for data recording.

[0005] In recent years, a large number of image processing methods and video monitoring schemes have been proposed. There are also schemes for image processing through deep learning and neural networks.

[0006] A Chinese patent with publication number CN114120119A proposes a dry beach detection method, system and storage medium. The method relies heavily on the environment. In turbid water, heavy fog, rainy and snowy weather and light (such as reflection, shadow), it will lead to a decline in image quality, and the edge recognition accuracy is not enough. Moreover, the preset threshold of RGB difference is difficult to adapt to dynamic light changes, which may lead to boundary misjudgment (such as the reflection area being identified as the edge). Secondly: the trapezoidal area proposed by this method only exists in individual tailings pond areas and cannot adapt to all tailings pond situations under complex terrain. In addition, the wide-angle lens of the camera may introduce image distortion, which may not be true in the real three-dimensional terrain, resulting in distance calculation errors. Furthermore, the pre-calibrated trapezoidal area cannot automatically adapt to terrain changes (such as deposition, erosion), and manual intervention is required to update the parameters. Moreover, the camera and laser range finder need to be calibrated regularly, otherwise the cumulative error will affect the results. For example, the camera position offset will cause the calibration coordinate system to fail, and the deviation of the calibration coordinates will directly affect the subsequent calculation.

[0007] A Chinese patent with publication number CN110132200A proposes a tailings pond dry beach dynamic monitoring method and system based on Beidou and video recognition. The invention first obtains the sampling value of the tailings pond dry beach length and the corresponding water surface elevation information, and then calculates the tailings pond slope ratio and the beach top elevation. Then, by real-time collection of tailings pond water surface elevation information, combined with tailings pond slope ratio and beach top elevation, the real-time dry beach length of tailings pond is obtained, thereby realizing the dynamic monitoring of tailings pond dry beach. The method has the following problems: 1. Data synchronization and calibration problem: If the collection time of Beidou water level data and video stream data is not strictly synchronized (for example, when the water level changes rapidly), S1 / S2 and H1 / H2 will not match, the slope ratio calculation will be distorted, and the data calibration depends on historical data. The method of taking the average slope ratio by multiple measurements for calibration cannot adapt to sudden terrain changes (such as dam break or landslide), and calibration lag may cause monitoring failure; 2. The model assumption of this method has certain limitations: first, the dry beach slope ratio is assumed to be a constant, but the actual terrain may have nonlinear changes (such as local steep slope or depression), which may lead to cumulative errors in real-time calculation of S S At the same time, when the water level rises and falls sharply (such as heavy rain), the slope ratio calculation may fail due to insufficient data update frequency; 3. Challenge of device deployment and computing resources: This method requires the Beidou water level monitoring device to be fixed on the water surface. If the water surface fluctuates violently or freezes, the device is easy to shift or damage, and requires certain manpower and financial maintenance. Moreover, under the conditions of rain, snow, fog, strong light or night low light, the video monitoring device is difficult to accurately extract the dry beach boundary, leading to distorted length information. More importantly, real-time processing of video stream data requires high-performance hardware support.

[0008] Based on the above background, at present there are many tailings dry beach monitoring methods based on machine vision, but there are problems such as low segmentation accuracy, inability to accurately monitor in time, high requirements for the performance and installation position of hardware devices, high requirements for weather conditions and the like. In addition, the water line accuracy obtained by the existing monitoring system through image analysis method is not high, the image is not orthographic image, and the error is large during algorithm processing. The binocular vision technology can accurately measure the distance when the target is obvious and close, but the water line of the tailings is generally far away from the beach top, up to dozens of meters to hundreds of meters, so the three-dimensional model established basically has no reference significance and the error is large. Therefore, there is an urgent need for a dry beach monitoring method based on an unmanned aerial vehicle carrying a hyperspectral camera to solve the problems existing in the prior art. SUMMARY

[0009] The purpose of the present application is to overcome the problems of measurement error accumulation caused by insufficient spatial positioning accuracy of the dry beach boundary in the prior art, limited precision caused by single spectral or spatial feature representation capability, and misjudgment risk caused by low recognition of small-scale ground features under complex ground conditions, and to provide a hyperspectral tailings dry beach line intelligent analysis method and system.

[0010] In order to achieve the above purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a hyperspectral tailings dry beach line intelligent analysis method, comprising the following steps: Obtaining a hyperspectral remote sensing image of a tailings; Splicing the obtained hyperspectral remote sensing image through GPS, and pre-processing the image spliced through GPS; Performing principal component clustering analysis on different spectral ranges of the pre-processed hyperspectral remote sensing image, selecting characteristic wavelengths based on the load map, and obtaining a new hyperspectral remote sensing image based on the selected characteristic wavelengths; Calibrating the new hyperspectral remote sensing image to obtain image labels, constructing a double-branch network model based on GCN and space-spectral convolution dense network, inputting the new hyperspectral remote sensing image and its corresponding image labels into the constructed double-branch network model, and obtaining a prediction map; Based on the prediction map, extracting dry beach line information, and combining GPS information to obtain dry beach length data.

[0011] In the step of splicing the obtained hyperspectral remote sensing image through GPS and pre-processing the image spliced through GPS, the following pre-processing process is included: Performing radiation calibration, atmospheric correction, orthographic correction and regional cropping basic preprocessing operations on the image spliced through GPS; Performing standard normal transformation, wavelet transformation and filter processing on the image after basic preprocessing to obtain the pre-processed hyperspectral remote sensing image.

[0012] The step of performing principal component cluster analysis on different spectral ranges of the preprocessed hyperspectral remote sensing image, selecting characteristic wavelengths based on the loading map, and obtaining a new hyperspectral remote sensing image based on the selected characteristic wavelengths includes the following specific methods: The preprocessed hyperspectral remote sensing image is divided into two different waveband ranges. The loading curve of the principal component pc1 of each waveband range is visualized, and the principal component cluster effects of tailings and water areas on different waveband ranges are compared and analyzed, respectively. The loading value of each waveband is obtained through the loading curve. The loading values of each waveband are compared, and the wavelengths corresponding to the peaks and troughs of the loading curve are selected as the characteristic wavelengths. A new hyperspectral remote sensing image is obtained based on the selected characteristic wavelengths.

[0013] The step of calibrating the new hyperspectral remote sensing image to obtain image labels, constructing a dual-branch network model based on GCN and spatial-spectral convolution dense network, and inputting the new hyperspectral remote sensing image and its corresponding image labels into the constructed dual-branch network model to obtain a predicted image includes calibrating the new hyperspectral remote sensing image using ENVI or GeoLabel.

[0014] The dual-branch network model is divided into two channels: The upper part is a graph convolution channel, and the lower part is a convolution network channel. The graph convolution channel mainly focuses on irregular and long-range spatial relationships between hyperspectral images, models the graph structure data, and the convolution network channel mainly extracts information for regular and short-range spatial relationships in the hyperspectral image.

[0015] The graph convolution channel is optimized as follows: Replace the original SLIC in the graph convolution channel with the improved ERS; and optimize the PCA-ERS based on the idea of boundary adjustment.

[0016] The step of calibrating the new hyperspectral remote sensing image to obtain image labels, constructing a dual-branch network model based on GCN and spatial-spectral convolution dense network, and inputting the new hyperspectral remote sensing image and its corresponding image labels into the constructed dual-branch network model to obtain a predicted image includes the following specific methods: Remove inter-spectral noise and reduce channel dimension of the new hyperspectral remote sensing image through a 1*1 convolution layer; Use the improved ERS superpixel segmentation to denoise the new hyperspectral remote sensing image to obtain a superpixel-level adjacency matrix; Send the superpixel nodes in the superpixel-level adjacency matrix to the graph convolution channel to capture global features in non-local and non-Euclidean space; The pixel-level node is sent into a convolution network channel to extract local pixel-level features; The extracted global features and local features are fused to obtain a final prediction map.

[0017] In the process of fusing the extracted global features and local features to obtain the final prediction map, the fusion is performed through weighted connection, and the specific formula is as follows:

[0018] Among them, represents the output result of the double-branch network model; represents a splicing function for splicing tensors along a specified dimension; and respectively represent learnable weight parameters, represents a feature representation processed by the GCN network, represents a feature representation processed by the spatial-spectral convolution dense network.

[0019] The difference between the predicted value and the actual value of the model is calculated through a cross-entropy loss function, and the parameters of the model are updated through a back propagation algorithm.

[0020] In a second aspect, the present application provides a hyperspectral tailing dam dry beach line intelligent analysis system, comprising: An acquisition module is configured to acquire a hyperspectral remote sensing image of a tailing dam. A preprocessing module is configured to perform GPS splicing on the acquired hyperspectral remote sensing image, and perform preprocessing on the image after GPS splicing. A feature selection module is configured to perform principal component cluster analysis on different spectral ranges of the preprocessed hyperspectral remote sensing image, select characteristic wavelengths based on a load diagram, and obtain a new hyperspectral remote sensing image based on the selected characteristic wavelengths. A model construction module is configured to calibrate the new hyperspectral remote sensing image to obtain an image label, construct a double-branch network model based on a GCN and a spatial-spectral convolution dense network, input the new hyperspectral remote sensing image and its corresponding image label into the constructed double-branch network model, and obtain a prediction map. A data extraction module is configured to extract dry beach line information based on the prediction map, and obtain dry beach length data in combination with GPS information.

[0021] Compared with the prior art, the present application has the following beneficial effects: The application provides a hyperspectral tailing pond dry beach line intelligent analysis method, including the following steps: acquiring a hyperspectral remote sensing image of a tailing pond; performing GPS splicing on the acquired hyperspectral remote sensing image, and performing pretreatment on the image after GPS splicing; performing principal component cluster analysis on different spectral ranges of the pretreated hyperspectral remote sensing image, selecting characteristic wavelengths based on a load diagram, and obtaining a new hyperspectral remote sensing image based on the selected characteristic wavelengths; calibrating the new hyperspectral remote sensing image to obtain an image label, constructing a double-branch network model based on GCN and a space-spectrum convolution dense network, inputting the new hyperspectral remote sensing image and the corresponding image label into the constructed double-branch network model to obtain a prediction image; based on the prediction image, extracting dry beach line information, and combining GPS information to obtain dry beach length data. Through the multi-band characteristics of the hyperspectral remote sensing image, combined with principal component cluster analysis and characteristic wavelength selection, the problem of insufficient spectral information of the image after dimension reduction and the serious data redundancy phenomenon existing in the use of full-band images is solved, and different characteristic bands correspond to single-band images that emphasize different target information in the original image. Different combinations are conducive to giving full play to the advantages of different target information of each band, so as to more accurately distinguish water areas and dry beaches; the double-branch network model is adopted to simultaneously analyze the differences between the spectral dimension and the spatial dimension, fuse the spectral characteristics and the spatial information, and the two branches have a certain complementarity, which is more conducive to fully mining the rich space-spectrum information of the input hyperspectral remote sensing image, so as to more accurately and effectively identify the dry beach line, thereby calculating the dry beach length, realizing end-to-end intelligent identification of the dry beach line, reducing manual intervention, improving processing efficiency, and being suitable for large-scale tailing pond monitoring requirements.

[0022] Further, through the GPS splicing and calibration technology, the spatial consistency of the remote sensing image is ensured, the dry beach line information in the prediction image and the GPS coordinates are combined, and the dry beach length data is directly output, thereby improving the geographical positioning accuracy of the results and providing a reliable basis for tailing pond safety evaluation.

[0023] Further, the neural network is used to identify the dry beach length slope, which has high automation, simple and convenient operation, reduces the safety risk and labor intensity of the inspection personnel, and provides help for mine enterprises to realize labor reduction. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The method flowchart of the application; Figure 2 The visualization diagram of the load curve in the application; Figure 3 The structure diagram of the double-branch network model in the application. DETAILED DESCRIPTION

[0025] For further understanding of the present application, the following will make a detailed description of the present application in combination with the drawings and specific embodiments. It should be understood that the embodiments are only to explain the present application but not to limit it.

[0026] Embodiment 1 As Figure 1 shown, a hyperspectral tailings dry beach line intelligent analysis method, comprising the following steps: S1: obtaining a hyperspectral remote sensing image of a tailings pond; S2: the obtained hyperspectral remote sensing image is spliced through GPS, and the image after GPS splicing is preprocessed; S3: principal component clustering analysis is performed on different spectral ranges of the preprocessed hyperspectral remote sensing image, feature wavelengths are selected based on the load diagram, and a new hyperspectral remote sensing image is obtained based on the selected feature wavelengths; S4: calibrating the new hyperspectral remote sensing image to obtain image labels, constructing a double-branch network model based on GCN and space-spectral convolution dense network, inputting the new hyperspectral remote sensing image and its corresponding image labels into the constructed double-branch network model to obtain a prediction map; S5: based on the prediction map, extracting dry beach line information, combining GPS information to obtain dry beach length data.

[0027] Specifically, in S1, the research area of the tailings pond is circled, and the flight route of the unmanned aerial vehicle is planned. The hyperspectral camera is carried on the unmanned aerial vehicle, the relevant flight parameters are set, the unmanned aerial vehicle is controlled to perform the flight task, and the hyperspectral remote sensing image (Hyperspectral Imaging, HSI) of the tailings pond is obtained.

[0028] Considering that the shape and position of the object in the image will change due to the problems such as lens distortion, shooting angle tilt and the like of the traditional RGB image, but the accuracy is limited in the correction process and depends on the control points, and multiple shooting angles, the present application uses the hyperspectral camera carried by the unmanned aerial vehicle to collect data in the tailings pond area. Compared with the ordinary camera, the hyperspectral camera contains rich spectral information and high spatial resolution, can capture the ground object categories in detail, is helpful to identify small-scale ground object changes, and can calculate the texture features in the image, such as entropy, correlation and the like to reveal the spatial structure and distribution pattern of the ground object.

[0029] Specifically, in S2, the obtained HSI is firstly spliced through GPS to restore the original scene layout of the tailings pond, and then subjected to preprocessing operations such as radiation calibration, atmospheric correction, orthorectification, and interested region clipping, so that the spectral curve in the HSI eliminates the influence of components such as aerosols and water vapor on the reflectivity of the ground object, ensures that the data of the sensor truly reflects the spectral differences of various ground objects, and gives the image accurate geographic coordinates, ensuring that the image after basic preprocessing is aligned with the real geographic space.

[0030] In the image acquisition process of the tailings pond, due to the differences in the acquisition environment, hyperspectral equipment, and sample individuals, the acquired spectral information will be affected to a certain extent. Even through basic preprocessing, these influences cannot be completely removed, and some interference causes the actual shape and characteristic wavelength of the spectral curve to change, such as various noise, spectral line drift, and other interference phenomena. Therefore, the influence of scattering effects and optical path changes on the spectrum in the spectral data needs to be eliminated through standard normal variate (SNV), the noise components in the spectral signal need to be removed through wavelet transform (WT), the differences under different measurement conditions need to be eliminated, and the high-frequency characteristics of the signal need to be retained through the Savitzky-Golay filter, so as to retain important spectral information while improving the signal-to-noise ratio.

[0031] Specifically, in S3, principal component cluster analysis is performed on different spectral ranges of the preprocessed hyperspectral remote sensing image, feature wavelengths are selected based on the load map, and a new hyperspectral remote sensing image is obtained based on the selected feature wavelengths.

[0032] Traditional RGB images only contain three channels of information, while hyperspectral images usually contain hundreds of channel data, and can collect spectral information that traditional cameras lack. The spectral characteristics of tailings and water areas in hyperspectral images are significantly different, and these differences mainly exist in the wavelength bands of spectral absorption and reflection. Tailings usually contain specific minerals, which will show absorption peaks at specific wavelengths, while the spectral characteristics of water areas are mainly determined by the absorption and scattering of water molecules. Water bodies have high reflectivity in blue and green light bands, but the reflectivity rapidly decreases in red and near-infrared bands. Therefore, tailings and water areas can be identified by analyzing the specific absorption characteristics in the spectral curve. The addition of spectral information breaks the single-dimensional information bottleneck, synchronously analyzes the differences in spectral and spatial dimensions, and helps to improve the accuracy of dry beach line recognition in bad weather.

[0033] The specific method is as follows: The pretreated HSI is divided into two different waveband ranges 400-1000 nm and 900-1700 nm, the load curve of the principal component pc1 of each waveband range is visualized, and the principal component clustering effect of tailings and water area in different waveband ranges is compared and analyzed respectively, and the HSI corresponding to the feature waveband range with the best clustering effect is used as the input data for subsequent deep learning.

[0034] As shown in Figure 2 The feature waveband range with the best clustering effect refers to that in the two waveband ranges 400-1000 nm and 900-1700 nm, water and tailings have different clustering effects. As shown in Fig. b, the substances corresponding to the red part can be better clustered together compared with Fig. a, and have better clustering effect, so the waveband range corresponding to Fig. b is more suitable for distinguishing tailings and water.

[0035] Further, considering that on the principal component load curve, the wavelength points corresponding to the peaks and troughs have more obvious contribution (i.e., the wavelengths corresponding to these points account for a larger proportion in this principal component and have more priority), therefore, the load values of each waveband are visualized by the load curve, and the wavebands corresponding to these positions are used as feature wavelengths.

[0036] If the single waveband images corresponding to each feature wavelength are directly applied to distinguish different ground objects, the spectral feature information cannot be fully utilized. However, directly using the full waveband spectrum will cause serious data redundancy. By selecting multiple feature wavebands to synthesize a new hyperspectral image, the key information of each feature wavelength can be fully utilized, and different ground object categories in complex images can be effectively detected and recognized.

[0037] Specifically, in S4, the new hyperspectral remote sensing image is calibrated to obtain an image label, a double-branch network model is constructed based on GCN and spatial-spectral convolution dense network, the new hyperspectral remote sensing image and the corresponding image label are input into the constructed double-branch network model to obtain a prediction map, and the specific method is as follows: The HSI obtained after the above processing is calibrated by ENVI (The Environment for Visualizing Images, remote sensing image processing environment) or GeoLabel (geographic labeling) to obtain a water area and dry beach area label image, which is named as GT.

[0038] A deep learning environment is built, the new HSI image obtained after the above steps and the label image GT are input into the double-branch network model based on GCN (Graph Convolutional Network, graph convolutional network) and spatial-spectral convolution dense network, and the output result is a prediction map.

[0039] The optimized double-branch network based on GCN and spectral-spatial convolution dense network is as follows: the whole model is divided into two channels, the upper channel is a graph convolution channel, and the lower channel is a convolution network channel. The graph convolution channel mainly focuses on irregular and long-range spatial relationships between hyperspectral images, models the spatial relationships to effectively process graph structure data, and constructs an adjacency matrix of spectral information and an adjacency matrix of spatial information when extracting spectral and spatial information simultaneously. The convolution network channel mainly extracts information about regular and short-range spatial relationships in the hyperspectral image.

[0040] The original network model is optimized as follows: The original SLIC (Simple Linear Iterative Clustering, superpixel segmentation algorithm) in the graph convolution channel is replaced with the improved ERS (Edge-Reweighting Strategy, edge weighting strategy), and the PCA-ERS (Principal Component Analysis-Edge-Reweighting Strategy, optimization algorithm combining principal component analysis and edge weighting strategy) is optimized according to the boundary adjustment idea. After the initial superpixel segmentation is completed, the boundary pixels of the extracted superpixels are fine-tuned using full-band data. First, update the boundary pixels to form a candidate superpixel set, then eliminate isolated pixels, repeat the above two steps until a threshold is reached, so that the spectral similarity within the superpixel is higher.

[0041] When the double-branch network is fused, in order to retain more original features and complementary information of the GCN and spectral-spatial convolution dense network branches, the subsequent network (such as 1×1 convolution or attention mechanism) is used to adaptively learn the weight, and the robustness of classification is improved. The original addition fusion is improved to weighted connection fusion, and the cross-entropy loss function is used for backpropagation training parameters at the end of the model.

[0042] The weighted fusion method is as follows:

[0043] wherein, represents the output result of the double-branch network model; represents a concatenation function used to concatenate tensors along a specified dimension; and respectively represent learnable weight parameters, represents the feature representation processed by the GCN network, represents the feature representation processed by the spectral-spatial convolution dense network.

[0044] The cross entropy loss function is expressed as follows:

[0045] in, Represents the loss function, which is used to measure the difference between the predicted result P and the true result Y. represents the number of image samples, represents the number of image categories, represents the one-hot encoding of the true label of the i-th sample in category c, Indicates the model i samples belong to the category c The predicted probability of .

[0046] Taking into account the complexity of the dual-branch network itself, in order to prevent overfitting and considering that the piecewise linear characteristics of the original LeakyReLU activation function limit the modeling ability of complex nonlinear relationships and the derivative of the function is a piecewise constant, it is easy to cause gradient discontinuity in deep networks and affect the back propagation efficiency. Therefore, the original LeakyReLU activation function is replaced by the Mish activation function to better balance the input information and network sparsity.

[0047] like Figure 3 As shown in the figure, the HSI processed by the above steps first passes through a 1*1 convolution layer to remove inter-spectral noise and reduce channel dimensionality. Then, the improved ERS superpixel segmentation is used to obtain a superpixel-level adjacency matrix. The HSI is converted into a graph structure, and the converted superpixel nodes are fed into a graph convolution channel to capture global features in non-local and non-Euclidean space. After removing noise and reducing channel dimensionality, another channel feeds the pixel-level nodes into a spectral-spatial convolution channel to extract local and fine pixel-level features. In the spectral-spatial convolution, each 3D convolution kernel is split into a 1-D kernel for extracting spectral features and a 2D kernel for utilizing spatial features. The two pixel-level features are then fed into a feature fusion module to obtain the final prediction image.

[0048] Feature extraction is performed simultaneously from both regional-level features and pixel-level features. Different features are weightedly connected and fused from different perspectives to reduce the occurrence of misjudgment of ground objects.

[0049] Specifically, in S5, based on the prediction map, the dry beach line information is extracted and combined with GPS information to obtain the dry beach length data. The specific method is as follows: Based on the output results of the dual-branch network model, the dry beach line information is extracted. Combined with the GPS information carried by the drone during flight, accurate dry beach length data can be obtained.

[0050] Example 2 A hyperspectral tailings pond dry beach line intelligent analysis system, comprising: An acquisition module is configured to acquire a hyperspectral remote sensing image of a tailing pond. A preprocessing module is configured to perform GPS splicing on the acquired hyperspectral remote sensing image, and perform preprocessing on the image after GPS splicing. A feature selection module is configured to perform principal component cluster analysis on different spectral ranges of the preprocessed hyperspectral remote sensing image, select characteristic wavelengths based on a load chart, and obtain a new hyperspectral remote sensing image based on the selected characteristic wavelengths. A model construction module is configured to calibrate the new hyperspectral remote sensing image to obtain an image label, and construct a double-branch network model based on GCN and a space-spectrum convolution dense network, input the new hyperspectral remote sensing image and the corresponding image label into the constructed double-branch network model, and obtain a prediction map. A data extraction module is configured to extract dry beach line information based on the prediction map, and obtain dry beach length data in combination with GPS information.

[0051] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: the specific embodiments of the present application can still be modified or replaced by the equivalent, without departing from the spirit and scope of the present application, any modification or equivalent replacement should be covered within the protection scope of the claims of the present application.

Claims

1. A hyperspectral tailings pond dry beach line intelligent analysis method, characterized in that: The steps include: Obtain hyperspectral remote sensing images of tailings ponds; The acquired hyperspectral remote sensing images are stitched by GPS, and the stitched images are preprocessed; Perform principal component cluster analysis on different spectral ranges of the preprocessed hyperspectral remote sensing image, select characteristic wavelengths based on the load diagram, and obtain new hyperspectral remote sensing images based on the selected characteristic wavelengths; The new hyperspectral remote sensing image is calibrated to obtain the image label. A dual-branch network model is constructed based on GCN and spatial-spectral convolutional dense network. The new hyperspectral remote sensing image and its corresponding image label are input into the constructed dual-branch network model to obtain the prediction map. Based on the prediction map, the dry beach line information is extracted and combined with GPS information to obtain the dry beach length data.

2. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 1, characterized in that: The steps of stitching the acquired hyperspectral remote sensing images by GPS and preprocessing the stitched GPS images include the following preprocessing steps: Perform basic pre-processing operations such as radiometric calibration, atmospheric correction, orthorectification and regional cropping on the GPS stitched images; The image after basic preprocessing is subjected to standard normal transformation, wavelet transformation and filter processing to obtain the preprocessed hyperspectral remote sensing image.

3. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 1, characterized in that: In the steps of performing principal component cluster analysis on different spectral ranges of the pre-processed hyperspectral remote sensing image, selecting characteristic wavelengths based on the load diagram, and obtaining a new hyperspectral remote sensing image based on the selected characteristic wavelengths, the specific method is as follows: The pre-processed hyperspectral remote sensing image is divided into two different band ranges; Visualize the load curve of the principal component pc1 in each band range, compare and analyze the principal component clustering effects of tailings and water areas in different band ranges, and obtain the load value of each band through the load curve; Compare the load values ​​of each band and select the wavelength corresponding to the peak and trough of the load curve as the characteristic wavelength; New hyperspectral remote sensing images are obtained based on the selected characteristic wavelengths.

4. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 1, characterized in that: The new hyperspectral remote sensing image is calibrated to obtain image labels, a dual-branch network model is constructed based on GCN and spatial-spectral convolutional dense network, the new hyperspectral remote sensing image and its corresponding image labels are input into the constructed dual-branch network model, and in the step of obtaining a prediction map, the new hyperspectral remote sensing image is calibrated with ENVI or GeoLabel.

5. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 4, characterized in that: The dual-branch network model is divided into two channels: The upper part is the graph convolution channel and the lower part is the convolutional network channel. The graph convolution channel mainly focuses on the irregular, medium- and long-range spatial relationships between hyperspectral images, and models and processes graph structure data. The convolutional network channel mainly extracts information from regular, short-range spatial relationships in hyperspectral images.

6. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 5, characterized in that: The graph convolution channel is optimized as follows: The SLIC in the original graph convolution channel is replaced by the improved ERS; PCA-ERS is optimized based on the idea of ​​boundary adjustment.

7. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 6, characterized in that: The new hyperspectral remote sensing image is calibrated to obtain image labels, a dual-branch network model is constructed based on GCN and spatial-spectral convolutional dense network, and the new hyperspectral remote sensing image and its corresponding image labels are input into the constructed dual-branch network model to obtain the prediction map. The specific method is as follows: The new hyperspectral remote sensing image is passed through a 1*1 convolutional layer to remove inter-spectral noise and reduce channel dimensions; The new hyperspectral remote sensing image after denoising is segmented using the improved ERS superpixel to obtain the superpixel-level adjacency matrix; The superpixel nodes in the superpixel-level adjacency matrix are fed into the graph convolution channel to capture global features in non-local and non-Euclidean spaces; Send pixel-level nodes into the convolutional network channel to extract local pixel-level features; The extracted global features and local features are fused to obtain the final prediction map.

8. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 7, characterized in that: In the process of fusing the extracted global features and local features to obtain the final prediction map, weighted connection fusion is performed. The specific formula is as follows: in, Represents the output of the two-branch network model; Represents a concatenation function, used to concatenate tensors along a specified dimension; and Represent the learnable weight parameters, Represents the feature representation after GCN network processing, Represents the feature representation after processing by the spatial-spectral convolutional dense network.

9. A hyperspectral tailings pond dry beach line intelligent analysis method according to claim 8, characterized in that: The difference between the model's predicted value and the actual value is calculated through the cross-entropy loss function, and the model parameters are updated through the back-propagation algorithm.

10. A hyperspectral tailings pond dry beach line intelligent analysis system according to any one of claims 1 to 9, characterized in that: include: Acquisition module, used to obtain hyperspectral remote sensing images of tailings ponds; A preprocessing module is used to stitch the acquired hyperspectral remote sensing images through GPS and preprocess the GPS stitched images; The feature selection module is used to perform principal component cluster analysis on different spectral ranges of the preprocessed hyperspectral remote sensing image, select characteristic wavelengths based on the load diagram, and obtain a new hyperspectral remote sensing image based on the selected characteristic wavelengths; The model construction module is used to calibrate the new hyperspectral remote sensing image to obtain the image label. A two-branch network model is constructed based on GCN and spatial-spectral convolutional dense network. The new hyperspectral remote sensing image and its corresponding image label are input into the constructed two-branch network model to obtain the prediction map. The data extraction module extracts the dry beach line information based on the prediction map and obtains the dry beach length data by combining it with GPS information.

Images