Hyperspectral remote sensing detection method and system for low-altitude unmanned aerial vehicle

Abstract

The invention discloses a low-altitude unmanned aerial vehicle hyperspectral remote sensing detection method and system. The detection method comprises the following steps: acquiring at least two low-altitude hyperspectral remote sensing sample images; acquiring a sub-pixel sample image of the hyperspectral remote sensing sample image; extracting one or more sample regions of interest in the sub-pixel sample image; configuring a region label of each region of interest; selecting a neural network model; obtaining the optimal neural network model at different low altitudes; obtaining the optimal hyperspectral remote sensing to-be-measured data of the low altitude; acquiring a sub-pixel to-be-detected image of the hyperspectral remote sensing to-be-detected image; extracting one or more regions of interest to be detected in the sub-pixel image to be detected; predicting a to-be-measured label of the regional to-be-measured spectral curve through the prediction model; the to-be-tested label represents mineral resources existing in the to-be-tested region of interest.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing detection, and in particular to a low-altitude unmanned aerial vehicle hyperspectral remote sensing detection method and system for detecting mineral resources using a hyperspectral spectrum and a neural network. Background Art

[0002] At present, satellite hyperspectral remote sensing is the trend in geological exploration. Hyperspectral remote sensing is generally carried out by photographing the ground through aircraft. The shooting distance is far and the accuracy of ground observation is insufficient.

[0003] In recent years, geological discrimination based on deep learning is a key technology to replace manual analysis of hyperspectral images. However, hyperspectral prospecting-related technologies have always lacked a large number of learning samples that can be used for training, as well as hyperspectral images taken at different low altitudes. There are situations where sample requirements are not met, which results in the hyperspectral remote sensing exploration technology based on deep learning being unable to meet the needs of metal mineral resource detection. Summary of the invention

[0004] Based on this, the first aspect of an embodiment of the present invention discloses a low-altitude UAV hyperspectral remote sensing detection method.

[0005] Wherein, the detection method includes:

[0006] Acquire hyperspectral remote sensing sample images of the low-altitude UAV at at least two low-altitude altitudes;

[0007] Acquire a sub-pixel sample image of the hyperspectral remote sensing sample image;

[0008] Extracting one or more sample regions of interest from the sub-pixel sample image; wherein the pixel values ​​of each pixel point in the sample region of interest are continuous within at least one error range;

[0009] Extract the average reflectance of the sample region of interest as a regional sample spectrum curve, and configure a regional label for each region of interest;

[0010] Select a neural network model;

[0011] Obtaining sample combinations of a plurality of sample spectral curves of the regions and their regional labels at the same low altitude, selecting a part of the sample combinations as a training combination, and another part of the sample combinations as a test combination;

[0012] Training the neural network model corresponding to the low altitude by the training combination;

[0013] Evaluating the neural network model corresponding to the low altitude by the test combination;

[0014] Select the neural network model with the best evaluation as the prediction model, and record the low altitude of the prediction model as the prediction altitude;

[0015] Controlling the UAV to obtain the hyperspectral remote sensing data to be measured at the predicted height;

[0016] Acquire a sub-pixel image to be tested of the hyperspectral remote sensing image to be tested;

[0017] Extracting one or more regions of interest to be tested in the sub-pixel image to be tested; wherein the pixel values ​​of each pixel point in the region of interest to be tested are continuous within the error range;

[0018] Extracting the average reflectivity of the region of interest to be measured as the regional spectral curve to be measured;

[0019] The prediction model is used to predict the labels to be tested of the spectral curve to be tested in the area; wherein the labels to be tested represent the mineral resources existing in the area of ​​interest to be tested.

[0020] In some embodiments disclosed in the present invention,

[0021] Acquiring the sub-pixel sample image and / or the sub-pixel image to be tested includes:

[0022] One or more interpolation points are respectively set between all adjacent pixel points of the hyperspectral remote sensing sample image and / or the hyperspectral remote sensing image to be tested, and sub-pixels are randomly filled in the interpolation points; wherein the value of the sub-pixel filled in the interpolation point is between the pixel values ​​of two adjacent pixel points.

[0023] In some embodiments disclosed in the present invention,

[0024] Selecting the sample region of interest and / or the region of interest to be tested comprises:

[0025] extracting one or more regions of interest in the sub-pixel sample image using a region growing method, and / or

[0026] A region growing method is used to extract one or more regions of interest in the sub-pixel image to be tested.

[0027] In some embodiments disclosed in the present invention,

[0028] The prediction model configuration is selected as follows:

[0029] A plurality of the neural network models are obtained by an optimal search method, and the neural network model that meets the evaluation requirements is selected as the prediction model.

[0030] In some embodiments disclosed in the present invention,

[0031] The neural network models at multiple different low-altitude altitudes are obtained through a search method of Levy flight or simulated annealing.

[0032] In some embodiments disclosed in the present invention,

[0033] The optimal neural network model configuration is obtained as follows:

[0034] Configure convergence conditions for simulated annealing;

[0035] Initialize the number of iterations and temperature T;

[0036] Obtain a training combination and a test combination corresponding to the low altitude h, train the neural network model m through the training combination, and obtain the accuracy p of the trained neural network model m through a plurality of the test combinations;

[0037] Update the low altitude h to the low altitude h',

[0038] Obtaining a training combination and a test combination corresponding to the low altitude h', training the neural network model m' through the training combination, and obtaining the accuracy p' of the trained neural network model m' through a plurality of the test combinations;

[0039] Calculate the probability difference Δp between the low altitude h and the low altitude h', Δp = p'-p,

[0040] When Δp≥0, the neural network model m is used as the prediction model.

[0041] When Δp<0, whether to use the neural network model m' as the prediction model is selected according to an acceptance probability; wherein the acceptance probability is related to the temperature T and the probability difference Δp;

[0042] Increase the current iteration number;

[0043] When the current number of iterations is less than an iteration number threshold, the low altitude h' is used as the current low altitude h and the update of the low altitude h is returned.

[0044] When the current number of iterations is greater than or equal to the iteration number threshold, determine whether the convergence condition is met;

[0045] When the convergence conditions are met, select the current prediction model.

[0046] When the convergence condition is not met, the current temperature is reduced to the value of the current temperature, the number of iterations is reset, and the process returns to the update of the low altitude h.

[0047] In some embodiments disclosed in the present invention,

[0048] The convergence conditions include:

[0049] The accuracy of the current prediction model is greater than the accuracy threshold.

[0050] In some embodiments disclosed in the present invention,

[0051] Obtaining the accuracy rate includes:

[0052] Acquire test labels of regional sample spectral curves of a plurality of the test combinations through the neural network model;

[0053] Comparing whether the test label of each test combination is consistent with the corresponding area label;

[0054] The accuracy is calculated based on the proportion of test combinations with the same label.

[0055] In some embodiments disclosed in the present invention,

[0056] The neural network model is trained to be configured as follows:

[0057] Acquiring frequency domain waveform data of the sample spectrum curve of the region;

[0058] Acquiring amplitude points of different frequency components according to the frequency domain waveform data;

[0059] Constructing an original feature matrix according to a plurality of amplitude points in the same region of interest at the same low altitude;

[0060] The neural network model is trained according to the multiple original feature matrices X and their region labels.

[0061] And, the second aspect of the embodiment of the present invention discloses at least a low-altitude UAV hyperspectral remote sensing detection system.

[0062] Wherein, the detection system includes a sample module, a training module and a detection module;

[0063] The sample module is used to obtain the hyperspectral remote sensing sample images of the low-altitude UAV at at least two low altitudes; obtain the sub-pixel sample images of the hyperspectral remote sensing sample images; extract one or more sample regions of interest in the sub-pixel sample images; wherein the pixel values ​​of each pixel point in the sample region of interest are continuous within at least one error range; extract the average reflectivity of the sample region of interest as a regional sample spectral curve, and configure the regional label of each of the regions of interest;

[0064] The training module is used to select a neural network model; obtain a sample combination of several regional sample spectral curves and their regional labels at the same low altitude, select a part of the sample combination as a training combination, and another part of the sample combination as a test combination; train the neural network model corresponding to the low altitude through the training combination; evaluate the neural network model corresponding to the low altitude through the test combination; select the neural network model with the best evaluation as the prediction model, and record the low altitude of the prediction model as the prediction altitude;

[0065] The detection module is used to control the UAV to obtain the hyperspectral remote sensing test data of the predicted height; obtain the sub-pixel test image of the hyperspectral remote sensing test image; extract one or more test interest regions in the sub-pixel test image; wherein the pixel values ​​of each pixel point in the test interest region are continuous within the error range; extract the average reflectivity of the test interest region as the regional test spectral curve; predict the test label of the regional test spectral curve through the prediction model; wherein the test label represents the mineral resources existing in the test interest region.

[0066] Compared with the prior art, the embodiments of the present invention first enhance the features of the hyperspectral remote sensing sample image by sub-pixel interpolation, then find the optimal prediction model at different low altitudes, and then use the prediction model to identify the mineral resources in the hyperspectral remote sensing image to be tested at its predicted altitude, thereby comprehensively avoiding the influence of drones on image shooting altitude, image resolution and other aspects.

[0067] In view of the above-mentioned solutions, the present invention will describe the disclosed exemplary embodiments in detail below with reference to the accompanying drawings, so as to make other features and advantages of the embodiments of the present invention clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0069] Figure 1 A schematic diagram showing the process of a low-altitude UAV hyperspectral remote sensing detection method;

[0070] Figure 2 The schematic diagram of the structure of the low-altitude UAV hyperspectral remote sensing detection system is shown. DETAILED DESCRIPTION

[0071] In order to facilitate understanding of the present application, the present application will be described more fully below with reference to the relevant drawings. Embodiments of the present application are provided in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0072] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0073] The embodiment of the present invention discloses a low-altitude unmanned aerial vehicle hyperspectral remote sensing detection method. The detection method can realize geological exploration based on hyperspectral remote sensing images, especially the exploration of mineral resources.

[0074] Figure 1 A schematic diagram showing the process of low-altitude UAV hyperspectral remote sensing detection method. Figure 1 The detection method is shown to include steps 10 to 110 .

[0075] 10. Obtain hyperspectral remote sensing sample images taken by low-altitude UAVs at multiple low altitudes.

[0076] 20 The hyperspectral remote sensing sample image is updated to a sub-pixel sample image through pixel interpolation logic.

[0077] The pixel interpolation logic is to set one or more sub-pixel interpolation points between two adjacent pixels of the hyperspectral remote sensing sample image, and randomly fill the pixel values ​​in the interpolation points.

[0078] Preferably, the pixel value filled in each interpolation point is based on the pixel values ​​of two adjacent pixel points and the difference between them.

[0079] Furthermore, when the pixel value difference between two adjacent pixel points is less than or equal to a threshold, each interpolation point is filled with the maximum pixel value of the two adjacent pixel points. When the pixel value difference between two adjacent pixel points is greater than the threshold, a finite linear function is selected and the two adjacent pixel points are used as the two end points of the linear function. Each interpolation point is evenly distributed between the two end points, and the pixel value filled in each interpolation point can be calculated.

[0080] For example, let the linear function be y=2x, where x represents the lateral distance from the origin and y represents the pixel value; at the same time, assume that one of the two adjacent pixels is the origin (0, 0) and the other is the end point (5, 10), then when there is one interpolation point, the interpolation point is (2.5, 5), and when there are four interpolation points, the two interpolation points are (1, 2), (2, 4), (3, 6), and (4, 8) respectively.

[0081] Alternatively, when the pixel value difference between two adjacent pixel points is greater than the threshold, a finite linear function is selected and the two adjacent pixel points are used as two end points of the linear function, and each interpolation point is randomly distributed between the two end points.

[0082] 30 extracts one or more sample regions of interest in the sub-pixel sample image.

[0083] The sample region of interest (ROI) is extracted by using a region growing method within the range of a sub-pixel sample image to obtain a plurality of sample regions of interest.

[0084] The region growing method is used to extract the region of interest (ROI). It is a ROI extraction method based on pixel similarity. It starts with a pixel point as a seed and adds other pixels adjacent to the pixel point into the target region one by one until a preset first similarity threshold is reached. This target region can be used as a sample region of interest separated from the sub-pixel sample image.

[0085] Preferably, the sample region of interest may be obtained by interpolating and combining multiple target regions using a region growing method.

[0086] Specifically, when there are two adjacent target areas with an interval less than a distance threshold, first calculate the pixel means of the two target areas, then compare the pixel means of the two target areas, and if the pixel means of the two target areas are less than a mean threshold, merge the two target areas. When merging two target areas, first traverse the two pixel points closest to each other in the two target areas, obtain the shortest line between the two pixel points, and then cover all the pixel points on the shortest line. The covered pixel value is the average of the pixel means of the two target areas. Then, using each pixel point on the shortest line as a seed, the regional growing method is used to add and cover the pixels of the attachment one by one until a pre-set second similarity threshold is reached. The second similarity threshold will be much larger than the first similarity threshold.

[0087] Then, when there are multiple mineral resources in the pixel sample image, the sample area of ​​interest can use the concentration of mineral resources to classify different mineral resources, so that the trained neural network model can be used to distinguish multiple different mineral resources at the same time. At the same time, multiple target areas in the sample area of ​​interest are combined by interpolation, which is conducive to clustering the same mineral resources that are scattered and close in distance in the remote sensing image, and is conducive to improving the robustness of the trained neural network model.

[0088] 40 extracts the average reflectance of the sample interest region as a regional sample spectrum curve, and configures a regional label for the interest region.

[0089] Among them, the regional sample spectral curve can reflect the average spectral information in the region, and the regional label is an artificial mark of the main mineral resources in the region.

[0090] 50Select the neural network model.

[0091] Among them, the neural network model can be selected as a convolutional neural network or a BP neural network.

[0092] 60 Obtain a sample combination of multiple regional sample spectral curves and their regional labels at each low altitude, select a part of the sample combination as a training combination, and another part of the sample combination as a test combination, train the neural network model corresponding to the low altitude through the training combination, evaluate the neural network model corresponding to the low altitude through the test combination, select the neural network model with the best evaluation as the prediction model, and record the low altitude of the prediction model as the prediction altitude.

[0093] Among them, step 60 updates different low-altitude altitudes through a simulated annealing search method, and obtains multiple training combinations and test combinations of different low-altitude altitudes to train and obtain a prediction model.

[0094] Furthermore, obtaining the prediction model includes,

[0095] 61. Configure the convergence conditions of simulated annealing and initialize the number of iterations L and the initial temperature T.

[0096] 62 obtains the sample combination of the current low altitude h, configures a part of the sample combination as a training combination, and another part of the sample combination as a test combination, and the number of training combinations and test combinations is controlled at 4:1.

[0097] 63 The neural network model m is trained through the training combination, and the accuracy p of the trained neural network model m is obtained through the test combination.

[0098] 64 Update the low altitude h to the low altitude h'.

[0099] Preferably, the hyperspectral remote sensing sample images at different low altitudes can be configured to be specially taken by the same drone at the geological site. Considering the uncontrollable pitch angle changes of the drone when shooting at different low altitudes, a certain distance of tilt or rotation in the horizontal direction is allowed for different low altitudes.

[0100] 65 Obtain sample combinations of low altitude h', configure a part of the sample combinations as training combinations, and another part of the sample combinations as test combinations, and the number of training combinations and test combinations is controlled at 4:1.

[0101] 66 A neural network model m' is trained through the training combination, and the accuracy p' of the trained neural network model m' is obtained through the test combination.

[0102] 67 Calculate the probability difference Δp between the low altitude h and the low altitude h', Δp = p'-p,

[0103] When Δp≥0, the neural network model m is used as the prediction model. When Δp<0, whether to use the neural network model m' as the prediction model is selected according to an acceptance probability.

[0104] 68 increases the current number of iterations.

[0105] When the current iteration number L is less than an iteration threshold, the low altitude h' is used as the current low altitude h and the process returns to step 64;

[0106] When the current number of iterations is greater than or equal to the iteration threshold, step 69 is entered.

[0107] Preferably, the current number of iterations increases by 1, ie, L=L+1.

[0108] 69 Determine whether the convergence conditions are met.

[0109] When the convergence condition is met, the current prediction model is accepted and the process goes to step 70.

[0110] When the convergence condition is not satisfied, the current temperature is reduced to the value of the current temperature, the number of iterations is reset, and the process returns to step 64 .

[0111] Preferably, the current temperature value is lowered based on the annealing coefficient k, and the value range of k is between 0 and 1, that is, T=Tk.

[0112] The convergence condition is that the accuracy of the current prediction model is greater than an accuracy threshold.

[0113] Preferably, obtaining the accuracy rate includes obtaining test labels of the regional sample spectral curves of multiple test combinations through the neural network model, comparing whether the test label of each test combination is consistent with the corresponding regional label, and then calculating the accuracy rate based on the proportion of test combinations with the same labels.

[0114] Preferably, simulated annealing is a method used in this embodiment to traverse the training neural network models corresponding to different low altitudes, and other random search methods (such as Levy flight, greedy search, etc.) can also be used to select different low altitudes. Even, it can be considered to traverse several low altitude neural network models in sequence at certain intervals within a limited altitude range and find the best.

[0115] Preferably, training the neural network model by the training combination includes,

[0116] Acquire multiple frames of hyperspectral remote sensing sample data at the same low altitude.

[0117] Among them, multiple frames of hyperspectral remote sensing sample data are taken at different times or under different lighting conditions.

[0118] Each frame of hyperspectral remote sensing sample image is iterated into a sub-pixel sample image through pixel interpolation logic.

[0119] Extract sample regions of interest from sub-pixel sampled images per frame.

[0120] The average reflectance of the sample area of ​​interest is extracted as the regional sample spectral curve.

[0121] Get the frequency domain waveform data of the regional sample spectrum curve.

[0122] Extract M amplitude points of different frequency components in the frequency domain waveform data.

[0123] The original feature matrix X is constructed based on the M amplitude points corresponding to the same region of interest of N sub-pixel sample images at the same low altitude:

[0124] The covariance formula C for calculating the original feature matrix X is:

[0125] in, is the characteristic mean. When the covariance formula of X and Y is positive, it means that X and Y are positively correlated; when the covariance formula is negative, it means that X and Y are negatively correlated. When the covariance is 0, X and Y are independent of each other.

[0126] The simplified covariance formula C is,

[0127] Calculate the covariance matrix C as And perform singular value decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors.

[0128] Construct a dimension reduction matrix P based on eigenvalues ​​and eigenvectors;

[0129] Arrange the eigenvectors corresponding to the eigenvalues ​​in order according to the size of the eigenvalues;

[0130] Establish an eigenvector matrix Z based on the arranged eigenvectors;

[0131] The first k rows of the eigenvector matrix Z are selected to construct the dimension reduction matrix P, where k is less than N and is a positive integer.

[0132] Calculate the product of the original feature matrix X and the dimension reduction matrix P, that is, PX is the data set T after the original data is reduced in dimension.

[0133] A neural network model is trained based on multiple datasets T and their corresponding region labels.

[0134] Corresponding to the neural network model and the data set T, the neural network model is a convolutional neural network model. Configure the number of convolutional layers and the number and size of feature maps applied by all convolutional layers; convolutional layer A has a total of 16 feature vectors, each feature vector is deconvolved using an 11×1 convolution kernel, the moving step is set to 2, and this layer outputs 32 feature vectors of size 93×1; convolutional layer B uses a 9×1 convolution kernel and outputs 32 feature vectors of size 85×1; pooling layer a uses a 3×1 pooling kernel and a moving step of 2 to produce 32 feature vectors of size 42×1; convolutional layer C uses a 7×1 convolution kernel and a moving step of 2 to output 32 convolution kernels of size 18×1; convolutional layer D uses a 5×1 convolution kernel and outputs 32 feature vectors of size 14×1; pooling layer b uses a 3×1 pooling kernel and a moving step of 2 to produce 32 feature vectors of size 6×1; convolutional layer E uses a 6×1 convolution kernel and outputs 32 feature vectors of size 1×1.

[0135] Then, the amount of spectral data in hyperspectral remote sensing images is large, and there is redundant information, which can easily cause resolution errors in deep learning models and reduce resolution efficiency and accuracy. Then, reducing the dimension of the original feature matrix, that is, reducing the dimension of the spectral information, is conducive to reflecting the wavelength data that is highly correlated with mineral resources.

[0136] 70 Obtain the hyperspectral remote sensing data to be measured at low altitudes corresponding to the prediction model.

[0137] Among them, the UAV is controlled to shoot at a low altitude corresponding to the prediction model to obtain multiple frames of hyperspectral remote sensing test data.

[0138] 80 updates the hyperspectral remote sensing image to be tested into a sub-pixel image to be tested through pixel interpolation logic.

[0139] 90 extracts one or more regions of interest to be tested in the sub-pixel image to be tested, wherein the pixel values ​​of each pixel point in the region of interest to be tested are continuous within an error range.

[0140] 100 The average reflectivity of the region of interest to be measured is extracted as the regional spectral curve to be measured.

[0141] Obtain the frequency domain waveform data of the sample spectrum curve of each frame area. Extract M amplitude points of different frequency components in each frame of frequency domain waveform data. Construct the original feature matrix X according to the M amplitude points corresponding to the same region of interest of N frames of sub-pixel sample images. Obtain the dataset T after the dimension reduction of the original feature matrix X by principal component analysis method.

[0142] 110 predicts the tags to be tested corresponding to the spectral curve of the area to be tested through the prediction model; wherein the tags to be tested represent the mineral resources existing in the area of ​​interest to be tested.

[0143] The data set T is input into the prediction model to predict the mineral resources of the corresponding area of ​​interest to be tested.

[0144] In summary, this embodiment enhances the features of hyperspectral remote sensing images by sub-pixel interpolation, compensates for the low resolution of hyperspectral remote sensing images caused by aerial photography, and helps to serve as training samples; and, by traversing the neural network models trained at different low altitudes, to find the prediction model corresponding to the optimal low altitude, and finally realize the resolution of mineral resources in the hyperspectral remote sensing images to be tested at low altitudes through the prediction model.

[0145] The embodiment of the present invention discloses a low-altitude unmanned aerial vehicle hyperspectral remote sensing detection system.

[0146] Figure 2 The schematic diagram of the structure of the low-altitude UAV hyperspectral remote sensing detection system is shown. Figure 2 It is shown that the detection system includes a sample module, a training module and a detection module.

[0147] The sample module is used to obtain the hyperspectral remote sensing sample images of the low-altitude UAV at at least two low altitudes; obtain the sub-pixel sample images of the hyperspectral remote sensing sample images; extract one or more sample regions of interest in the sub-pixel sample images; wherein the pixel values ​​of each pixel point in the sample region of interest are continuous within at least one error range; extract the average reflectivity of the sample region of interest as a regional sample spectral curve, and configure the regional label of each of the regions of interest;

[0148] The training module is used to select a neural network model; obtain a sample combination of several regional sample spectral curves and their regional labels at the same low altitude, select a part of the sample combination as a training combination, and another part of the sample combination as a test combination; train the neural network model corresponding to the low altitude through the training combination; evaluate the neural network model corresponding to the low altitude through the test combination; select the neural network model with the best evaluation as the prediction model, and record the low altitude of the prediction model as the prediction altitude;

[0149] The detection module is used to control the UAV to obtain the hyperspectral remote sensing test data of the predicted height; obtain the sub-pixel test image of the hyperspectral remote sensing test image; extract one or more test interest regions in the sub-pixel test image; wherein the pixel values ​​of each pixel point in the test interest region are continuous within the error range; extract the average reflectivity of the test interest region as the regional test spectral curve; predict the test label of the regional test spectral curve through the prediction model; wherein the test label represents the mineral resources existing in the test interest region.

[0150] Based on this understanding, the technical solution of the present invention can essentially or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of various embodiments of the present invention.

[0151] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A low-altitude UAV hyperspectral remote sensing detection method, characterized in that: The detection method comprises: Acquire hyperspectral remote sensing sample images of the low-altitude UAV at at least two low-altitude altitudes; Acquire a sub-pixel sample image of the hyperspectral remote sensing sample image; Extracting one or more sample regions of interest from the sub-pixel sample image; wherein the pixel values ​​of each pixel point in the sample region of interest are continuous within at least one error range; Extract the average reflectance of the sample region of interest as a regional sample spectrum curve, and configure a regional label for each region of interest; Select a neural network model; Obtaining sample combinations of a plurality of sample spectral curves of the regions and their regional labels at the same low altitude, selecting a part of the sample combinations as a training combination, and another part of the sample combinations as a test combination; Training the neural network model corresponding to the low altitude by the training combination; Evaluating the neural network model corresponding to the low altitude by the test combination; Select the neural network model with the best evaluation as the prediction model, and record the low altitude of the prediction model as the prediction altitude; Controlling the UAV to obtain the hyperspectral remote sensing data to be measured at the predicted height; Acquire a sub-pixel image to be tested of the hyperspectral remote sensing image to be tested; Extracting one or more regions of interest to be tested in the sub-pixel image to be tested; wherein the pixel values ​​of each pixel point in the region of interest to be tested are continuous within the error range; Extracting the average reflectivity of the region of interest to be measured as the regional spectral curve to be measured; The prediction model is used to predict the labels to be tested of the spectral curve to be tested in the area; wherein the labels to be tested represent the mineral resources existing in the area of ​​interest to be tested.

2. The low-altitude UAV hyperspectral remote sensing detection method according to claim 1 is characterized in that: Acquiring the sub-pixel sample image and / or the sub-pixel image to be tested includes: One or more interpolation points are respectively set between all adjacent pixel points of the hyperspectral remote sensing sample image and / or the hyperspectral remote sensing image to be tested, and sub-pixels are randomly filled in the interpolation points; wherein the value of the sub-pixel filled in the interpolation point is between the pixel values ​​of two adjacent pixel points.

3. The low-altitude UAV hyperspectral remote sensing detection method according to claim 1 is characterized in that: Selecting the sample region of interest and / or the region of interest to be tested comprises: extracting one or more regions of interest in the sub-pixel sample image using a region growing method, and / or A region growing method is used to extract one or more regions of interest in the sub-pixel image to be tested.

4. The low-altitude UAV hyperspectral remote sensing detection method according to claim 1 is characterized in that: The prediction model configuration is selected as follows: A plurality of the neural network models are obtained by an optimal search method, and the neural network model that meets the evaluation requirements is selected as the prediction model.

5. The low-altitude UAV hyperspectral remote sensing detection method according to claim 4 is characterized in that: The neural network models at multiple different low-altitude altitudes are obtained through a search method of Levy flight or simulated annealing.

6. The low-altitude UAV hyperspectral remote sensing detection method according to claim 1 is characterized in that: The optimal neural network model configuration is obtained as follows: Configure convergence conditions for simulated annealing; Initialize the number of iterations and temperature T; Obtain a training combination and a test combination corresponding to the low altitude h, train the neural network model m through the training combination, and obtain the accuracy p of the trained neural network model m through a plurality of the test combinations; Update the low altitude h to the low altitude h', Obtaining a training combination and a test combination corresponding to the low altitude h', training the neural network model m' through the training combination, and obtaining the accuracy p' of the trained neural network model m' through a plurality of the test combinations; Calculate the probability difference Δp between the low altitude h and the low altitude h', Δp = p'-p, When Δp≥0, the neural network model m is used as the prediction model. When Δp<0, whether to use the neural network model m' as the prediction model is selected according to an acceptance probability; wherein the acceptance probability is related to the temperature T and the probability difference Δp; Increase the current iteration count; When the current number of iterations is less than an iteration number threshold, the low altitude h' is used as the current low altitude h and the update of the low altitude h is returned. When the current number of iterations is greater than or equal to the iteration number threshold, determine whether the convergence condition is met; When the convergence conditions are met, select the current prediction model. When the convergence condition is not met, the current temperature is reduced to the value of the current temperature, the number of iterations is reset, and the process returns to the update of the low altitude h.

7. The low-altitude UAV hyperspectral remote sensing detection method according to claim 5 is characterized in that: The convergence conditions include: The accuracy of the current prediction model is greater than the accuracy threshold.

8. The low-altitude UAV hyperspectral remote sensing detection method according to claim 5 is characterized in that: Obtaining the accuracy rate includes: Acquire test labels of regional sample spectral curves of a plurality of the test combinations through the neural network model; Comparing whether the test label of each test combination is consistent with the corresponding area label; The accuracy is calculated based on the proportion of test combinations with the same label.

9. The low-altitude UAV hyperspectral remote sensing detection method according to claim 1 is characterized in that: The neural network model is trained to be configured as follows: Acquiring frequency domain waveform data of the sample spectrum curve of the region; Acquiring amplitude points of different frequency components according to the frequency domain waveform data; Constructing an original feature matrix according to a plurality of amplitude points in the same region of interest at the same low altitude; The neural network model is trained according to the multiple original feature matrices X and their region labels.

10. A low-altitude UAV hyperspectral remote sensing detection system, characterized in that: The detection system includes a sample module, a training module and a detection module; The sample module is used to obtain the hyperspectral remote sensing sample images of the low-altitude UAV at at least two low altitudes; obtain the sub-pixel sample images of the hyperspectral remote sensing sample images; extract one or more sample regions of interest in the sub-pixel sample images; wherein the pixel values ​​of each pixel point in the sample region of interest are continuous within at least one error range; extract the average reflectivity of the sample region of interest as a regional sample spectral curve, and configure the regional label of each of the regions of interest; The training module is used to select a neural network model; obtain a sample combination of several regional sample spectral curves and their regional labels at the same low altitude, select a part of the sample combination as a training combination, and another part of the sample combination as a test combination; train the neural network model corresponding to the low altitude through the training combination; evaluate the neural network model corresponding to the low altitude through the test combination; select the neural network model with the best evaluation as the prediction model, and record the low altitude of the prediction model as the prediction altitude; The detection module is used to control the UAV to obtain the hyperspectral remote sensing test data of the predicted height; obtain the sub-pixel test image of the hyperspectral remote sensing test image; extract one or more test interest regions in the sub-pixel test image; wherein the pixel values ​​of each pixel point in the test interest region are continuous within the error range; extract the average reflectivity of the test interest region as the regional test spectral curve; predict the test label of the regional test spectral curve through the prediction model; wherein the test label represents the mineral resources existing in the test interest region.

Images