Weak target rapid remote sensing detection method and system

Abstract

The invention provides a weak target rapid remote sensing detection method and system. The method comprises the following steps: determining that each pixel belongs to a weak target pure pixel, a strong target pure pixel or a mixed target non-pure pixel according to abundance coefficients of a weak target and a strong target of each pixel in a remote sensing image of an area to be detected; dividing the remote sensing image of the to-be-detected area at equal intervals to form a plurality of weak target pure pixel sub-grids, strong target pure pixel sub-grids and mixed target non-pure pixel sub-grids; extracting a weak target pure pixel sample, a strong target pure pixel sample and a mixed target non-pure pixel sample to construct a training set; training a detection model by adopting the training set; and inputting a remote sensing image of a to-be-detected area into the trained detection model to obtain weak target distribution in the to-be-detected area. The weak target remote sensing detection performance and the result precision are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and remote sensing image technology, and specifically relates to a method and system for rapid remote sensing detection of weak targets. Background Art

[0002] In practical applications, such as geological disaster monitoring, resource exploration, and transportation, it is often necessary to quickly and accurately detect and identify weak targets in a large area. However, traditional remote sensing image processing methods often face multiple challenges when dealing with weak targets in complex scenes, including background interference, spectral overlap, and ambiguity of target features.

[0003] In recent years, with the rapid development of remote sensing and computer technology, especially the rise of deep learning, new opportunities have been brought for large-scale weak target extraction. However, deep learning target recognition and classification rely on a large amount of target sample data. The establishment of traditional sample libraries often selects appropriate samples through human eye interpretation, and the accuracy is generally high, but there are also problems such as low efficiency, subjectivity and data imbalance. In terms of feature extraction, convolutional neural networks have shown excellent performance and can learn different dimensional features from large-scale data, greatly improving the accuracy of target detection. In previous studies, most of them focused on the single feature collected in remote sensing images for identifying remote sensing targets, such as spectrum, geometry, texture, etc. However, since weak targets on remote sensing images are often interfered by shadows, noise and complex backgrounds, the use of a single feature and a single network structure often leads to the inability to maximize feature utilization and low target extraction accuracy. Therefore, for the task of weak target recognition in high-resolution remote sensing images, how to construct a high-quality target sample set and effectively utilize spectral and spatial information to achieve fast and accurate detection of weak targets is still an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the above-mentioned background technology and to provide a method and system for rapid remote sensing detection of weak targets. Target sample points are uniformly generated based on mixed pixel decomposition and classification, and boundary range points are automatically searched to generate a deep learning sample library with balanced data types and representativeness. A cascade structure model of a multidimensional convolutional neural network is constructed to realize spectral and spatial multidimensional feature fusion to improve the remote sensing detection performance and result accuracy of weak targets.

[0005] The technical solution adopted by the present invention is: a method for rapid remote sensing detection of weak targets, comprising the following steps:

[0006] Determine the ground object categories corresponding to weak targets and strong targets in the area to be detected;

[0007] According to the abundance coefficients of weak targets and strong targets of each pixel in the remote sensing image of the area to be detected, it is determined whether each pixel is a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel;

[0008] The remote sensing image of the area to be detected is divided into equal intervals to form a number of weak target pure pixel subgrids, strong target pure pixel subgrids and mixed target impure pixel subgrids;

[0009] In each weak target pure pixel subgrid, a single pixel belonging to a weak target pure pixel is selected as a weak target pure pixel sample;

[0010] Select a single pixel belonging to a strong target pure pixel in each strong target pure pixel subgrid as a strong target pure pixel sample;

[0011] In each mixed target impure pixel subgrid, a single pixel and its neighborhood belonging to the mixed target impure pixel are selected as mixed target impure pixel samples;

[0012] Constructing a training set: taking weak target pure pixel samples and strong target pure pixel samples as model inputs, and labeling them as weak targets or strong targets accordingly; taking mixed target impure pixel samples as model inputs, and further labeling each pixel in the mixed target impure pixel samples as weak targets or strong targets;

[0013] Use the training set to train the detection model;

[0014] The remote sensing image of the area to be detected is input into the trained detection model to obtain the distribution of weak targets in the area to be detected.

[0015] In the above technical scheme, the process of calculating the abundance coefficients of weak targets and strong targets for each pixel in the remote sensing image of the area to be detected includes: extracting the spectral measurement values ​​of weak targets and strong targets from the remote sensing image of the area to be detected respectively through the coordinate positions of the weak targets and strong targets in the area to be detected; and calculating the abundance coefficients of weak targets and strong targets for each pixel in the remote sensing image of the area to be detected based on the spectral information of the remote sensing image of the area to be detected and the spectral measurement values ​​of the weak targets and strong targets.

[0016] In the above technical solution, the abundance coefficient is calculated by fitting and solving the following formula:

[0017]

[0018] Among them, y i represents the spectral parameter of any pixel in the i-th band; p ij Refers to the spectral measurement value of the jth end member in the i-th band; j∈{1,2,…,p}; α jRefers to the abundance coefficient of the jth end member of the pixel; β i Refers to the participating error term of the pixel in the i-th band; p represents the total number of endmembers; among all endmembers, at least one is a weak target and at least one is a strong target.

[0019] In the above technical solution, the criteria for judging whether each pixel is a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel are:

[0020]

[0021] Among them, P represents the pixel type of a single pixel in the remote sensing image; P a Refers to weak target pure pixel; P b Refers to mixed target non-pure pixels; P c Refers to strong target pure pixels; α A refers to the abundance coefficient of weak targets, α B refers to the abundance coefficient of strong targets; T refers to the set threshold.

[0022] In the above technical solution, the process of equally-spacedly dividing the remote sensing image of the area to be detected includes: obtaining the endpoint positions of the pixels of the specified target category in the area to be detected, including the minimum and maximum values ​​of the horizontal range, and the minimum and maximum values ​​of the vertical range, so as to determine the total pixel range of the target category to be divided in the remote sensing image of the area to be detected;

[0023] According to the preset sub-grid size and row and column spacing, the area to be detected is divided into a number of sub-target category grids in an equally spaced manner until the entire total pixel range to be divided is covered; the preset sub-grid size and row and column spacing ensure that the target category sub-grid obtained by division covers all pixels belonging to the target category;

[0024] The target category is a weak target pure pixel, a strong target pure pixel, or a mixed target impure pixel.

[0025] In the above technical solution, in the process of generating mixed target impure pixel samples, pixels whose correlation strength with a single mixed target impure pixel is greater than a set value are combined with the mixed target impure pixel to form a mixed target impure pixel sample.

[0026] In the above technical solution, the detection model includes a 1D-CNN model, a 2D-CNN model, a weighted sum feature fusion module, a fully connected layer and a normalized exponential function module;

[0027] The 1D-CNN model is used to extract the deep features of weak target pure pixel and strong target pure pixel samples in the spectral dimension when the detection model is trained, and to extract the deep features of the input remote sensing image in the spectral dimension when the detection model is applied;

[0028] The 2D-CNN model is used to extract the depth features of the mixed target impure pixel samples in the spatial dimension when the detection model is trained, and to extract the depth features of the input remote sensing image in the spectral dimension when the detection model is applied;

[0029] Extracting the weighted sum feature fusion module to fuse the depth features in the spectral dimension and the spatial dimension;

[0030] The fully connected layer is used to flatten the fused feature vector and then perform a linear transformation; the normalized exponential function module is used to convert the output of the fully connected layer into the probability distribution of the model input belonging to weak target pure pixels and strong target pure pixels.

[0031] In the above technical solution, after the remote sensing images of the area to be detected are subjected to radiation calibration, atmospheric correction, geometric correction, cloud and snow removal, fusion, mosaicking and cropping preprocessing operations, the steps of the weak target rapid remote sensing detection method are executed.

[0032] In the above technical scheme, the process of generating mixed target non-pure pixel training samples includes: using a manual labeling method, labeling each pixel inside the mixed target non-pure pixel sample one by one, and clarifying its corresponding category label as "weak target" or "strong target"; the manually labeled label matrix is ​​used to generate pixel-by-pixel training labels for the mixed target samples.

[0033] The present invention also provides a weak target rapid remote sensing detection system, which is used to implement the weak target rapid remote sensing detection method described in the above technical solution; comprising:

[0034] A ground object category recognition module is used to determine the ground object categories corresponding to weak targets and strong targets in the area to be detected;

[0035] A pixel classification module is used to determine whether each pixel belongs to a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel according to the abundance coefficient of weak targets and strong targets of each pixel in the remote sensing image of the area to be detected;

[0036] A subgrid division module is used to divide the remote sensing image of the area to be detected at equal intervals to form a number of weak target pure pixel subgrids, strong target pure pixel subgrids and mixed target impure pixel subgrids;

[0037] A sample acquisition module is used to select a single pixel belonging to a weak target pure pixel in each weak target pure pixel subgrid as a weak target pure pixel sample; select a single pixel belonging to a strong target pure pixel in each strong target pure pixel subgrid as a strong target pure pixel sample; select a single pixel belonging to a mixed target non-pure pixel and its neighborhood in each mixed target non-pure pixel subgrid as a mixed target non-pure pixel sample;

[0038] The training set construction module is used to take weak target pure pixel samples and strong target pure pixel samples as model inputs, and label weak targets or strong targets accordingly; take mixed target impure pixel samples as model inputs, and further label each pixel in the mixed target impure pixel samples as weak targets or strong targets;

[0039] A model training module is used to train the detection model using a training set;

[0040] The detection area weak target recognition module is used to input the remote sensing image of the area to be detected into the trained detection model to obtain the distribution of weak targets in the area to be detected.

[0041] The present invention also provides a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the method for rapid remote sensing detection of weak targets described in the above technical solution is implemented.

[0042] The beneficial effects of the present invention are as follows: by combining the establishment of a spectral library and the construction of a multi-terminal mixed model, the purity of pixels can be accurately determined, the representativeness of samples can be improved, the false detection rate can be reduced, and a reliable basis and data foundation can be provided for the selection of subsequent deep learning samples; object-level samples are generated for mixed targets, the spatial correlation of samples is improved, and the model is helped to better capture the spatial distribution characteristics of weak targets. By combining the abundance map and statistical rules of mixed pixel decomposition, a multi-level "point-surface" sample set is constructed, the balance and diversity of samples are ensured, and the model's ability to recognize targets of different scales and complexities is improved. By extracting spectral features and spatial geometric features, feature fusion is used to identify weak targets in remote sensing images, the model's ability to characterize weak targets is enhanced, and the multi-dimensional information of high-resolution remote sensing images is fully utilized.

[0043] Furthermore, the present invention proposes a method for calculating the abundance coefficient of weak targets and strong targets for each pixel. The method can enhance the accuracy of target recognition. By extracting the spectral measurement values ​​of weak targets and strong targets and calculating the abundance coefficient using the spectral information, the mixing ratio of different target categories in each pixel can be quantitatively reflected, thereby improving the accuracy of target classification. The method also supports the accuracy of subsequent classification. The calculated abundance coefficient gives a specific target feature weight to each pixel, making the subsequent classification of weak targets, strong targets and mixed targets more reliable.

[0044] Furthermore, the present invention calculates the abundance coefficient by formula fitting, and the method can improve the calculation accuracy of the abundance coefficient. By combining the multi-band spectral information with the end-member spectral measurement value through the formula-based fitting calculation method, the abundance coefficients of weak targets and strong targets can be separated more accurately, and the influence of noise and model error can be reduced. The consideration of noise and error terms is included in the formula, making the method suitable for complex spectral environments (such as mixed pixels) actually existing in remote sensing images.

[0045] Furthermore, the present invention proposes a standard for judging pixel types. The present invention provides clear classification standards, and through the classification standards based on the abundance coefficients of weak targets and strong targets, and threshold setting, the pixel classification method is unified and easy to implement. The present invention uses clear threshold judgment rules to quickly distinguish weak target pure pixels, strong target pure pixels and mixed target impure pixels, providing efficient pre-processing for subsequent operations (such as subgrid division and sample selection).

[0046] Furthermore, the present invention divides the remote sensing image of the area to be detected at equal intervals to ensure the uniform distribution of samples. By dividing the grid at equal intervals, the samples of the target category in the subgrid are evenly distributed, avoiding the deviation of sample selection. The present invention obtains the endpoint position of the target category and divides it based on the subgrid size to ensure that the pixels of all target categories are covered, thereby providing protection for the integrity of the sample. The adjustability of the subgrid size and interval allows the division process to adapt to the distribution density of different targets, thereby improving the flexibility of the method.

[0047] Furthermore, the present invention can more comprehensively capture the spatial distribution characteristics of mixed targets by merging pixels with a strong correlation with a single mixed target impure pixel into one sample. Generating more complete and accurate mixed target samples helps the classifier better learn the differences between mixed targets and other target categories, thereby improving classification accuracy.

[0048] Furthermore, the module design of the detection model adopted by the present invention, the 1D-CNN model and the 2D-CNN model effectively extract the deep features of different types of samples, and effectively combine the deep features of the spectral dimension and the spatial dimension through the weighted sum feature fusion module, and utilize the information complementarity of the two features to improve the recognition ability of the model. The fully connected layer performs a linear transformation on the fused features, and combined with the normalized exponential function module (such as Softmax), it can output an accurate probability distribution, thereby accurately predicting the target category of the pixel. The model design also supports the classification of pure pixels of weak targets, pure pixels of strong targets, and impure pixels of mixed targets, providing an effective solution for multi-target detection.

[0049] Furthermore, the present invention significantly reduces the systematic errors and random noise in remote sensing images through preprocessing steps such as radiation calibration, atmospheric correction, and geometric correction. For example, cloud and snow removal can remove invalid information that interferes with target detection, fusion and mosaic operations can enhance the continuity and integrity of the target area, and cropping operations can focus the detection area. High-quality preprocessed images lay a data foundation for the realization of rapid detection of weak targets, which helps to improve the detection accuracy and stability of the entire method.

[0050] Furthermore, the present invention generates high-quality pixel-by-pixel training labels through manual annotation, which significantly improves the accuracy of samples; refines complex mixed target areas to pixel-by-pixel level, which helps the model learn target distribution characteristics more accurately; provides accurate pixel-by-pixel training labels, so that the model can learn finer-grained features, especially in the spatial dimension; is suitable for scenes with complex spectra and significant spatial structures, and improves the robustness and generalization ability of the model in complex scenes; high-quality labels reduce error propagation in model training and improve classification effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0052] Figure 2 It is a schematic diagram of the pixel purity determination process of the present invention;

[0053] Figure 3 A schematic diagram of sample construction for the present invention;

[0054] Figure 4 It is a schematic diagram of the training and use of the detection model of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but they do not constitute a limitation on the present invention.

[0056] Example 1

[0057] like Figure 1 As shown, the present invention provides a method for rapid remote sensing detection of weak targets, comprising the following steps:

[0058] Determine the ground object categories corresponding to weak targets and strong targets in the area to be detected;

[0059] According to the abundance coefficients of weak targets and strong targets of each pixel in the remote sensing image of the area to be detected, it is determined whether each pixel is a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel;

[0060] The remote sensing image of the area to be detected is divided into equal intervals to form a number of weak target pure pixel subgrids, strong target pure pixel subgrids and mixed target impure pixel subgrids;

[0061] In each weak target pure pixel subgrid, a single pixel belonging to a weak target pure pixel is selected as a weak target pure pixel sample;

[0062] Select a single pixel belonging to a strong target pure pixel in each strong target pure pixel subgrid as a strong target pure pixel sample;

[0063] In each mixed target impure pixel subgrid, a single pixel and its neighborhood belonging to the mixed target impure pixel are selected as mixed target impure pixel samples;

[0064] Constructing a training set: taking weak target pure pixel samples and strong target pure pixel samples as model inputs, and labeling them as weak targets or strong targets accordingly; taking mixed target impure pixel samples as model inputs, and further labeling each pixel in the mixed target impure pixel samples as weak targets or strong targets;

[0065] Use the training set to train the detection model;

[0066] The remote sensing image of the area to be detected is input into the trained detection model to obtain the distribution of weak targets in the area to be detected.

[0067] The principle of the present invention is further explained below with reference to specific embodiments.

[0068] This embodiment proposes a method and system for rapid remote sensing detection of weak targets that integrates hierarchical models and deep learning, so as to automatically generate high-quality and balanced deep learning samples, extract spectral and spatial dimension information, and realize recognition, detection and classification of weak targets, which specifically includes the following steps:

[0069] The first step is remote sensing data acquisition and preprocessing

[0070] Obtain high-resolution remote sensing images that fully cover the study area. The data can come from remote sensing platforms such as satellites and drones. In order to improve the quality of remote sensing images and make them clearer and more accurate, the following high-resolution remote sensing image preprocessing contents are used:

[0071] Radiation calibration: Eliminate radiation errors caused by sensors, atmosphere and other factors, so that the grayscale value of the image can truly reflect the reflectivity of the ground.

[0072] Atmospheric correction: Convert the original digital value DN into a physical quantity (such as reflectivity) to establish a quantitative relationship between the image and the reflectivity of the ground object.

[0073] Geometric correction: By obtaining DEM data in the study area, the geometric distortion of the image can be eliminated so that the positions of the objects on the image correspond to the actual objects.

[0074] Declouding: Suppress random noise (clouds, snow, etc.) in images and improve the signal-to-noise ratio of images.

[0075] Fusion, mosaicking and cropping: images of different spatial resolutions or different spectral bands are fused together to generate images with high spatial resolution and rich information. Considering that images often cross scenes in the study area, the fused images need to be further mosaicked and cropped according to the study area vector.

[0076] The second step is to identify the remote sensing pixel type.

[0077] Considering that the distinction between weak targets and other targets in remote sensing images is not obvious, there are often multiple target types in one pixel. Based on the acquired high-quality, high-resolution remote sensing data, in order to build a deep learning sample library, it is necessary to distinguish the pixel types of remote sensing images, that is, the determination of pure pixels and impure pixels. Weak targets and strong targets both refer to the classification of land object categories. In actual use, multiple land object categories can be defined as weak targets, or multiple land object categories can be defined as strong targets. In this embodiment, both weak targets and strong targets are single land object categories. Determine the land object categories corresponding to weak targets and strong targets according to the actual needs of the site.

[0078] Taking weak target A and strong target B, two remote sensing target objects, as an example, the target objects can be set according to the actual situation.

[0079] First, it is necessary to collect the coordinate information of weak target A and strong target B and the spectral information of the corresponding remote sensing image position on the spot, extract the end member spectrum, and build a spectral library on this basis. That is, by knowing the coordinate information and remote sensing image of weak target A and strong target B, the spectral measurement values ​​of weak target A and strong target B in each band can be calculated.

[0080] Furthermore, the average of the spectral measurement values ​​of several known weak targets A in a remote sensing image in a certain band can be used as the spectral measurement value of the weak target A in the band. Similarly, the spectral measurement value of the strong target B in the band can be calculated.

[0081] Then, the fully constrained least squares multi-endmember mixed pixel decomposition model is used to obtain the area ratio of each endmember in the pixel, also known as the abundance α, and the threshold T is used to divide the pixel type. In order to make the model have practical physical meaning, it is assumed that the sum of the area ratios of all endmembers is 1, and the ratios of each endmember are non-negative numbers. The calculation formula of the fully constrained least squares multi-endmember mixed pixel decomposition model is as follows:

[0082]

[0083] Among them, y i represents the spectral parameter of any pixel in the i-th band; p ij Refers to the spectral measurement value of the jth end member in the i-th band; j∈{1,2,…,p}; α j Refers to the abundance coefficient of the jth end member of the pixel; β i Refers to the error term of the pixel in the ith band; p represents the total number of endmembers; among all endmembers, at least one is a weak target and at least one is a strong target. Each unit corresponds to a different ground object category.

[0084] The mixed pixel decomposition model is used to obtain the abundance coefficients of weak target A and strong target B, and further obtain the area proportion of targets A and B in each pixel α A and α B , where α A and α B Belong to α j .

[0085] Considering that weak targets are easily affected by noise from other targets in remote sensing images, the threshold T is set to increase the number of pure pixels of weak targets and reduce the interference of non-weak target types while ensuring the quality of samples. No constraints are set for strong targets. The specific formula for distinguishing pixel types is as follows:

[0086]

[0087] α B <1,α A +α B =1)

[0088] Among them, P represents the pixel type of a single pixel in the remote sensing image; P a Refers to weak target pure pixel; P b Refers to mixed target non-pure pixels; P c Refers to strong target pure pixels; α A refers to the abundance coefficient of weak targets, α B Refers to the abundance coefficient of strong targets; T refers to the set threshold, which is generally set to no less than 0.9.

[0089] The spectral measurement values ​​of weak target A and strong target B in each band and the spectral parameters of each pixel in the remote sensing image are fitted and solved by formula 1, and the weak target abundance coefficient and strong target abundance coefficient of each pixel can be obtained. Then, the type of each pixel can be determined by formula 2.

[0090] Step 3: Deep learning sample set construction

[0091] In order to construct the hierarchical sample set, the endpoint positions of the pure pixels and impure pixels in the remote sensing image are first located, which is conducive to speeding up the sample point generation speed, that is, obtaining the endpoint position of the weak target on the remote sensing image: [(X min1 ,Y min1 ),(X max1 ,Y max1 )], get the endpoint position of the strong target on the remote sensing image: [(X min2 ,Y min2 ),(X max2 ,Y max2 )]The endpoint position of the mixed target on the remote sensing image: [(X min3 ,Y min3 ),(X max3 ,Y max3 )].

[0092] According to the endpoint pixel coordinates of the three targets, the pure pixel samples of weak targets and strong targets, as well as the mixed pixel of the mixed target object level sample benchmark are screened according to the equally spaced grid division method. Figure 3 Specifically, considering that the generation principles of pure pixel samples of weak targets and strong targets are the same, the following content takes the generation of pure pixel samples of weak targets as an example:

[0093] ①Query the grid size containing all weak target pure pixels in the entire remote sensing image:

[0094] M=Y max1 -Y min1 , N = X max1 -X min1 (3)

[0095] Among them, M refers to the number of vertical pixels, that is, the number of rows; N refers to the number of horizontal pixels, that is, the number of columns.

[0096] ② Assuming that the subgrid size is a rows and b columns, and the subgrid interval is n pixels, find the number of subgrids W in the entire grid:

[0097]

[0098] W=MM*NN(5)

[0099] Among them, the subgrid size and interval size are determined according to the distribution of pure pixels of weak targets in the entire image, ensuring that the divided subgrids cover all pixels belonging to weak targets; MM refers to the number of effective subgrids that can be placed in the number of grid rows, and NN refers to the number of effective subgrids that can be placed in the number of grid columns.

[0100] ③ Select each weak target pure pixel in each sub-grid as a weak target pure pixel sample.

[0101] The process of generating strong target pure pixel samples is the same as above.

[0102] Different from the above, since mixed targets include weak targets and strong targets, the effect of using a single pixel as a sample is often not ideal. Therefore, after using ① to ③ to screen out the benchmark mixed pixels, another step is required, as follows:

[0103] ④ Generate object-level samples according to the “first law of geography” that “everything is related to other things, but similar things are more closely related”.

[0104] C(u,v)=f(D(u,v))(6)

[0105] Among them, C(u,v) represents the association strength between pixels u and v, which can be utilized; D(u,v) represents the distance between pixels u and v; f is a monotonically decreasing function, indicating that the association strength gradually weakens with the increase of distance.

[0106] Assume that two pixels u(u 1 ,u 2 ,…,u n ) and v(v 1 ,v 2 ,…,v n ), the association strength between pixels u and v can be measured using the Euclidean distance, and the specific calculation formula is as follows:

[0107]

[0108] Among them, u i is the value of pixel u in the i-th band; v i is the value of pixel v in the i-th band; n is the number of bands.

[0109] Furthermore, by locating the reference pixel by the reference mixed pixel, and combining the idea of ​​"the first law of geography" and the Euclidean distance to determine the object-level neighborhood size S, the reference pixel can be diverged into object-level samples of a certain neighborhood size; that is, the pixels whose correlation strength with a single mixed target impure pixel is greater than the set value are combined with the mixed target impure pixel to form a mixed target impure pixel sample.

[0110] The fourth step is to build a weak target detection model and achieve final classification.

[0111] According to the first to fourth steps, high-quality remote sensing images and a certain number of sample libraries (including pure weak and strong target sample points and mixed weak and strong target object-level samples) can be obtained.

[0112] The training set construction process of the detection model includes: taking weak target pure pixel samples and strong target pure pixel samples as model input, and labeling weak targets or strong targets accordingly; taking mixed target non-pure pixel samples as model input, and further labeling each pixel in the mixed target non-pure pixel samples as weak targets or strong targets.

[0113] Specifically, on the basis of high-quality remote sensing images, based on pure weak and strong target pixel-level samples and automatically generate corresponding target labels, the label is the target category (strong target or weak target) corresponding to the sample.

[0114] The corresponding target labels are outlined based on the mixed weak and strong target object-level samples and manual interpretation. That is, for the mixed target impure pixels, each pixel inside the mixed target impure pixel sample is labeled one by one through the manual labeling method, and its corresponding category label is clearly "weak target" or "strong target"; the manually labeled label matrix is ​​used to generate pixel-by-pixel training labels for the mixed target samples.

[0115] The detection model includes a 1D-CNN model, a 2D-CNN model, a weighted sum feature fusion module, a fully connected layer and a normalized exponential function module;

[0116] The 1D-CNN model is used to extract the deep features of weak target pure pixel and strong target pure pixel samples in the spectral dimension when the detection model is trained, and to extract the deep features of the input remote sensing image in the spectral dimension when the detection model is applied;

[0117] The 2D-CNN model is used to extract the depth features of the mixed target impure pixel samples in the spatial dimension when the detection model is trained, and to extract the depth features of the input remote sensing image in the spectral dimension when the detection model is applied;

[0118] Extracting the weighted sum feature fusion module to fuse the depth features in the spectral dimension and the spatial dimension;

[0119] The fully connected layer is used to flatten the fused feature vector and then perform a linear transformation; the normalized exponential function module is used to convert the output of the fully connected layer into the probability distribution of the model input belonging to weak target pure pixels, strong target pure pixels and mixed target impure pixels.

[0120] Specifically, in 1D-CNN and 2D-CNN, each layer uses multiple filters to convolve the input remote sensing image, aiming to extract local shallow features, and gradually increase the number of layers of the convolutional neural network to achieve deep extraction of spectral and spatial features. The overall convolution operation calculation idea is as follows:

[0121]

[0122] In the formula, Z l and Z l+1 It represents the feature map, which is also the input and output of the l+1th convolution operation; Z(i,j) is the pixel corresponding to the feature map; s is the convolution layer parameter; k is the number of channels of the feature map; b is the bias term.

[0123] After each convolution operation obtains the feature map, the pooling layer is used to reduce the size of the feature map after convolution to increase the calculation speed and enhance the robustness of the feature. Furthermore, in order to enhance the learning ability of the network's features, a nonlinear activation function is usually added after the convolution layer to enhance the nonlinearity of the feature information. Common activation functions include Sigmoid function, Tanh function, and ReLU function. Compared with Sigmoid and Tanh, the ReLU function is simpler, has a faster convergence speed, and avoids the gradient explosion and gradient disappearance problems to a certain extent. The specific calculation is as follows:

[0124] f(x)=max{0,x}(9)

[0125] In the above formula, x represents the input feature vector.

[0126] Based on the spectral features and spatial features extracted by 1D-CNN and 2D-CNN, a weighted sum feature fusion module (WSM) is constructed to fuse the output feature maps of 1D-CNN and 2D-CNN, combining the spectral features and spatial features. The calculation formula of the weighted sum feature fusion module is as follows:

[0127] W(g)=α*U(g)+(1-α)*V(G)(10)

[0128] Where W(g) refers to the fused features, U(g) represents the spectral features, V(G) represents the spatial features, and α is a weight parameter between 0 and 1, which can be adjusted experimentally to obtain the best effect.

[0129] Then, the fused feature map is linearly transformed on the flattened feature vector in the fully connected layer (FC), and finally the output of the FC layer is converted into the probability distribution of each category in the softmax layer to achieve the final classification. The overall process diagram is shown in the figure. Figure 4 shown.

[0130] Example 2

[0131] The present invention also provides a weak target rapid remote sensing detection system, which is used to implement the weak target rapid remote sensing detection method described in the above technical solution; comprising:

[0132] A ground object category recognition module is used to determine the ground object categories corresponding to weak targets and strong targets in the area to be detected;

[0133] A pixel classification module is used to determine whether each pixel belongs to a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel according to the abundance coefficient of weak targets and strong targets of each pixel in the remote sensing image of the area to be detected;

[0134] A subgrid division module is used to divide the remote sensing image of the area to be detected at equal intervals to form a number of weak target pure pixel subgrids, strong target pure pixel subgrids and mixed target impure pixel subgrids;

[0135] A sample acquisition module is used to select a single pixel belonging to a weak target pure pixel in each weak target pure pixel subgrid as a weak target pure pixel sample; select a single pixel belonging to a strong target pure pixel in each strong target pure pixel subgrid as a strong target pure pixel sample; select a single pixel belonging to a mixed target non-pure pixel and its neighborhood in each mixed target non-pure pixel subgrid as a mixed target non-pure pixel sample;

[0136] The training set construction module is used to take weak target pure pixel samples and strong target pure pixel samples as model inputs, and label weak targets or strong targets accordingly; take mixed target impure pixel samples as model inputs, and further label each pixel in the mixed target impure pixel samples as weak targets or strong targets;

[0137] A model training module is used to train the detection model using a training set;

[0138] The detection area weak target recognition module is used to input the remote sensing image of the area to be detected into the trained detection model to obtain the distribution of weak targets in the area to be detected.

[0139] Example 3

[0140] The present invention also provides a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the method for rapid remote sensing detection of weak targets described in the above technical solution is implemented.

[0141] Example 4

[0142] The present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the weak target rapid remote sensing detection method described in the above technical solution by executing the computer instructions.

[0143] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0145] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0147] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

[0148] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A method for rapid remote sensing detection of weak targets, characterized in that: The following steps are involved: Determine the ground object categories corresponding to weak targets and strong targets in the area to be detected; According to the abundance coefficients of weak targets and strong targets of each pixel in the remote sensing image of the area to be detected, it is determined whether each pixel is a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel; The remote sensing image of the area to be detected is divided into equal intervals to form a number of weak target pure pixel subgrids, strong target pure pixel subgrids and mixed target impure pixel subgrids; In each weak target pure pixel subgrid, a single pixel belonging to a weak target pure pixel is selected as a weak target pure pixel sample; Select a single pixel belonging to a strong target pure pixel in each strong target pure pixel subgrid as a strong target pure pixel sample; In each mixed target impure pixel subgrid, a single pixel and its neighborhood belonging to the mixed target impure pixel are selected as mixed target impure pixel samples; Constructing a training set: taking weak target pure pixel samples and strong target pure pixel samples as model inputs, and labeling them as weak targets or strong targets accordingly; taking mixed target impure pixel samples as model inputs, and further labeling each pixel in the mixed target impure pixel samples as weak targets or strong targets; Use the training set to train the detection model; The remote sensing image of the area to be detected is input into the trained detection model to obtain the distribution of weak targets in the area to be detected.

2. A method according to claim 1, characterized in that: The process of calculating the abundance coefficients of weak targets and strong targets for each pixel in the remote sensing image of the area to be detected includes: extracting the spectral measurement values ​​of the weak targets and strong targets from the remote sensing image of the area to be detected respectively through the coordinate positions of the weak targets and strong targets in the area to be detected; and calculating the abundance coefficients of the weak targets and strong targets for each pixel in the remote sensing image of the area to be detected according to the spectral information of the remote sensing image of the area to be detected and the spectral measurement values ​​of the weak targets and strong targets.

3. A method according to claim 2, characterized in that: The abundance coefficient was calculated by fitting the following formula: Among them, y i represents the spectral parameter of any pixel in the i-th band; p ij Refers to the spectral measurement value of the jth end member in the i-th band; j∈{1,2,…,p}; α j Refers to the abundance coefficient of the jth end member of the pixel; β i Refers to the participating error term of the pixel in the i-th band; p represents the total number of endmembers; among all endmembers, at least one is a weak target and at least one is a strong target.

4. A method according to claim 1, characterized in that: The criteria for determining whether each pixel is a weak target pure pixel, a strong target pure pixel, or a mixed target impure pixel are: Among them, P represents the pixel type of a single pixel in the remote sensing image; P a Refers to weak target pure pixel; P b Refers to mixed target non-pure pixels; P c Refers to strong target pure pixels; α A Refers to the abundance coefficient of weak targets, α B refers to the abundance coefficient of strong targets; T refers to the set threshold.

5. A method according to claim 1, characterized in that: The process of equally-spaced division of the remote sensing image of the area to be detected includes: obtaining the endpoint positions of the pixels of the specified target category in the area to be detected, including the minimum and maximum values ​​of the horizontal range, and the minimum and maximum values ​​of the vertical range, so as to determine the total pixel range of the target category to be divided in the remote sensing image of the area to be detected; According to the preset sub-grid size and row and column spacing, the area to be detected is divided into a number of sub-target category grids in an equally spaced manner until the entire total pixel range to be divided is covered; the preset sub-grid size and row and column spacing ensure that the target category sub-grid obtained by division covers all pixels belonging to the target category; The target category is a weak target pure pixel, a strong target pure pixel, or a mixed target impure pixel.

6. A method according to claim 1, characterized in that: In the process of generating mixed target impure pixel samples, pixels whose correlation strength with a single mixed target impure pixel is greater than a set value are combined with the mixed target impure pixel to form mixed target impure pixel samples.

7. A method according to claim 1, characterized in that: The detection model includes a 1D-CNN model, a 2D-CNN model, a weighted sum feature fusion module, a fully connected layer and a normalized exponential function module; The 1D-CNN model is used to extract the deep features of weak target pure pixel and strong target pure pixel samples in the spectral dimension when the detection model is trained, and to extract the deep features of the input remote sensing image in the spectral dimension when the detection model is applied; The 2D-CNN model is used to extract the depth features of the mixed target impure pixel samples in the spatial dimension when the detection model is trained, and to extract the depth features of the input remote sensing image in the spectral dimension when the detection model is applied; Extracting the weighted sum feature fusion module to fuse the depth features in the spectral dimension and the spatial dimension; The fully connected layer is used to flatten the fused feature vector and perform linear transformation; The normalized exponential function module is used to convert the output of the fully connected layer into the probability distribution of the model input belonging to weak target pure pixels and strong target pure pixels.

8. A method according to claim 1, characterized in that: After the remote sensing images of the area to be detected are pre-processed with radiation calibration, atmospheric correction, geometric correction, cloud and snow removal, fusion, mosaicking and cropping, the steps of the weak target rapid remote sensing detection method are executed.

9. A method according to claim 1, characterized in that: The generation process of mixed target non-pure pixel training samples includes: using manual labeling methods, labeling each pixel inside the mixed target non-pure pixel sample one by one, and clarifying its corresponding category label as "weak target" or "strong target"; the manually labeled label matrix is ​​used to generate pixel-by-pixel training labels for the mixed target samples.

10. A weak target rapid remote sensing detection system, characterized in that: The system is used to implement the weak target rapid remote sensing detection method according to any one of claims 1 to 9; comprising: A ground object category recognition module is used to determine the ground object categories corresponding to weak targets and strong targets in the area to be detected; A pixel classification module is used to determine whether each pixel belongs to a weak target pure pixel, a strong target pure pixel or a mixed target impure pixel according to the abundance coefficient of weak targets and strong targets of each pixel in the remote sensing image of the area to be detected; A subgrid division module is used to divide the remote sensing image of the area to be detected at equal intervals to form a number of weak target pure pixel subgrids, strong target pure pixel subgrids and mixed target impure pixel subgrids; A sample acquisition module is used to select a single pixel belonging to a weak target pure pixel in each weak target pure pixel subgrid as a weak target pure pixel sample; select a single pixel belonging to a strong target pure pixel in each strong target pure pixel subgrid as a strong target pure pixel sample; select a single pixel belonging to a mixed target non-pure pixel and its neighborhood in each mixed target non-pure pixel subgrid as a mixed target non-pure pixel sample; The training set construction module is used to take weak target pure pixel samples and strong target pure pixel samples as model inputs, and label weak targets or strong targets accordingly; take mixed target impure pixel samples as model inputs, and further label each pixel in the mixed target impure pixel samples as weak targets or strong targets; A model training module is used to train the detection model using a training set; The detection area weak target recognition module is used to input the remote sensing image of the area to be detected into the trained detection model to obtain the distribution of weak targets in the area to be detected.

Images