Weak target fast remote sensing detection method and system

Abstract

The application provides a weak target fast remote sensing detection method and system, which comprises the following steps: determining whether each image element belongs to a weak target pure image element, a strong target pure image element or a mixed target non-pure image element according to the abundance coefficients of weak targets and strong targets of each image element in a remote sensing image of a region to be detected; equally dividing the remote sensing image of the region to be detected to form a plurality of weak target pure image element sub-grids, strong target pure image element sub-grids and mixed target non-pure image element sub-grids; extracting weak target pure image element samples, strong target pure image element samples and mixed target non-pure image element samples to construct a training set; training a detection model by using the training set; inputting the remote sensing image of the region to be detected into the trained detection model to obtain weak target distribution in the region to be detected. The application effectively improves the weak target remote sensing detection performance and result accuracy.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning and remote sensing image technology, specifically relating to a method and system for rapid remote sensing detection of weak targets. Background Technology

[0002] In practical applications, such as geological disaster monitoring, resource exploration, and transportation, there is often a need for rapid and accurate detection and identification of weak targets over 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 the ambiguity of target features.

[0003] In recent years, the rapid development of remote sensing and computer technologies, especially the rise of deep learning, has brought new opportunities for large-scale weak target extraction. However, deep learning target recognition and classification rely on a large amount of target sample data. Traditional sample database construction often relies on human interpretation to select suitable samples, which generally has high accuracy but also suffers from problems such as low efficiency, subjectivity, and data imbalance. In terms of feature extraction, convolutional neural networks have shown excellent performance, capable of learning features of different dimensions from large-scale data, greatly improving the accuracy of target detection. In previous studies, most focused on using single features from remote sensing images to identify remote sensing targets, such as spectral, geometric, and texture features. However, since weak targets in remote sensing images are often affected by shadows, noise, and complex backgrounds, using single features and single network structures often leads to insufficient 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 rapid and accurate detection of weak targets remains an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the aforementioned background technology and provide a rapid remote sensing detection method and system for weak targets. This method generates target sample points uniformly based on hybrid pixel decomposition and classification, automatically identifies boundary range points, and generates a deep learning sample library with balanced and representative data types. Furthermore, it constructs a cascaded structure model of a multi-dimensional convolutional neural network to achieve spectral and spatial multi-dimensional feature fusion, thereby improving the performance and accuracy of remote sensing detection of weak targets.

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

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

[0007] Based on the abundance coefficients of weak and strong targets in the remote sensing image of the area to be detected, each pixel is determined to be a pure pixel of weak targets, a pure pixel of strong targets, or a non-pure pixel of mixed targets.

[0008] The remote sensing image is divided into areas to be detected at equal intervals to form several clean pixel sub-grids for weak targets, clean pixel sub-grids for strong targets, and non-clean pixel sub-grids for mixed targets;

[0009] In each weak target clean cell subgrid, select a single cell belonging to the weak target clean cell as a weak target clean cell sample;

[0010] In each strong target clean cell subgrid, select a single cell belonging to the strong target clean cell as a strong target clean cell sample;

[0011] In each subgrid of mixed target impure pixels, select a single pixel belonging to the mixed target impure pixel and its neighborhood as a mixed target impure pixel sample;

[0012] Constructing the training set: Use clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs, and label them as weak targets or strong targets respectively; use non-clean pixel samples of mixed targets as model inputs, and further label each pixel in the non-clean pixel samples of mixed targets as a weak target or a strong target.

[0013] The detection model is trained using the training set;

[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 solution, the process of calculating the abundance coefficients of weak and strong targets in each pixel of the remote sensing image of the area to be detected includes: extracting the spectral measurement values ​​of weak and strong targets from the remote sensing image of the area to be detected, respectively, based on the known coordinate positions of weak and strong targets in the area to be detected; and calculating the abundance coefficients of weak and strong targets in each pixel of 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 weak and strong targets.

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

[0017]

[0018] Among them, y i p represents the spectral parameters of any pixel in the i-th band; ij The spectral measurement value of the j-th endmember in the i-th band; j∈{1,2,…,p}; α jThe abundance coefficient of the j-th endmember of that pixel; β i This refers to the participation 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 determining whether each pixel belongs to a clean pixel of a weak target, a clean pixel of a strong target, or a non-clean pixel of a mixed target are as follows:

[0020]

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

[0022] In the above technical solution, the process of dividing the remote sensing image into equally spaced detection areas includes: obtaining the endpoint positions of the specified target category pixels in the detection area, 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 to be divided in the remote sensing image of the detection area for the target category.

[0023] Based on the preset sub-grid size and row and column spacing, the area to be detected is divided into several sub-target category grids at equal intervals until the entire total pixel range to be divided is covered; the preset sub-grid size and row and column spacing ensure that the divided target category sub-grids cover all pixels belonging to that target category;

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

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

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

[0027] The 1D-CNN model is used to extract the depth features of clean pixels of weak targets and clean pixels of strong targets in the spectral dimension during the training of the detection model, and to extract the depth 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 spatial depth features of mixed target impure pixel samples during detection model training, and to extract the spectral depth features of the input remote sensing image during detection model application.

[0029] The weighted summation feature fusion module is used to fuse deep features in the spectral and spatial dimensions;

[0030] The fully connected layer is used to flatten the fused feature vectors 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 the clean pixels of weak targets and the clean pixels of strong targets.

[0031] In the above technical solution, after performing radiometric calibration, atmospheric correction, geometric correction, cloud and snow removal, fusion, mosaicking and cropping preprocessing operations on the remote sensing image of the area to be detected, the steps of the rapid remote sensing detection method for weak targets are then executed.

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

[0033] This invention also provides a rapid remote sensing detection system for weak targets, which is used to implement the rapid remote sensing detection method for weak targets described in the above technical solution; including:

[0034] The feature category identification module is used to determine the feature categories corresponding to weak and strong targets in the area to be detected;

[0035] The 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 non-pure pixel based on the abundance coefficients of weak and strong targets in the remote sensing image of the area to be detected.

[0036] The sub-grid division module is used to divide the remote sensing image of the area to be detected into several clean pixel sub-grids for weak targets, clean pixel sub-grids for strong targets, and non-clean pixel sub-grids for mixed targets at equal intervals.

[0037] The 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; to 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; and to select a single pixel belonging to a mixed target non-pure pixel and its neighborhood in each mixed target non-pure pixel subgrid as mixed target non-pure pixel samples.

[0038] The training set construction module is used to take clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs and label them as weak targets or strong targets respectively; it takes impure pixel samples of mixed targets as model inputs and further labels each pixel in the impure pixel samples of mixed targets as a weak target or a strong target.

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

[0040] The weak target identification module in the detection area 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 having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the rapid remote sensing detection method for weak targets described in the above technical solution.

[0042] The beneficial effects of this invention are as follows: By combining spectral library establishment and multi-endpoint hybrid model construction, the purity of pixels can be accurately determined, improving the representativeness of samples, reducing the false detection rate, and providing a reliable basis and data foundation for the selection of samples for subsequent deep learning; by generating object-level samples for hybrid targets, the spatial correlation of samples is improved, which helps the model to better capture the spatial distribution characteristics of weak targets. By combining hybrid pixel decomposition abundance maps and statistical rules, a multi-level "point-area" sample set is constructed, ensuring the balance and diversity of samples and improving the model's ability to identify targets of different scales and complexities. By extracting spectral features and spatial geometric features, weak targets in remote sensing images are identified through feature fusion, enhancing the model's ability to represent weak targets and making full use of the multi-dimensional information of high-resolution remote sensing images.

[0043] Furthermore, this invention proposes a method for calculating the abundance coefficients of weak and strong targets in each pixel. This method can enhance the accuracy of target recognition. By extracting spectral measurements of weak and strong targets and using the spectral information to calculate the abundance coefficients, the mixing ratio of different target categories in each pixel can be quantitatively reflected, thereby improving the accuracy of target classification. This method also supports the accuracy of subsequent classification. The calculated abundance coefficients assign specific target feature weights to each pixel, making subsequent classification of weak, strong, and mixed targets more reliable.

[0044] Furthermore, this invention calculates abundance coefficients through formula fitting, which improves the accuracy of abundance coefficient calculation. By combining multi-band spectral information with endmember spectral measurements using a formula-based fitting method, the abundance coefficients of weak and strong targets can be separated more accurately, reducing the impact of noise and model errors. The inclusion of noise and error terms in the formula makes this method applicable to complex spectral environments (such as mixed pixels) that actually exist in remote sensing images.

[0045] Furthermore, this invention proposes a standard for determining pixel type. This invention provides a clear classification standard, based on the abundance coefficients of weak and strong targets, and a threshold setting, making the pixel classification method uniform and easy to implement. Utilizing clear threshold judgment rules, this invention can quickly distinguish between clean pixels with weak targets, clean pixels with strong targets, and impure pixels with mixed targets, providing efficient preprocessing for subsequent operations (such as subgrid partitioning and sample selection).

[0046] Furthermore, the remote sensing image of the detection area divided at equal intervals in this invention can ensure a uniform distribution of samples. By dividing the image into grids at equal intervals, the sample distribution of the target category within each sub-grid is uniform, avoiding sample selection bias. This invention obtains the endpoint positions of the target categories and divides the image based on the sub-grid size, ensuring that all pixels of all target categories are covered, thus guaranteeing sample integrity. The adjustability of the sub-grid size and interval allows the division process to adapt to different target distribution densities, improving the flexibility of the method.

[0047] Furthermore, by merging pixels with strong associations to impure pixels of a single mixed target into a single sample, this invention can more comprehensively capture the spatial distribution characteristics of mixed targets. 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 modular design of the detection model employed in this invention, using both 1D-CNN and 2D-CNN models, effectively extracts deep features from different types of samples. Through a weighted summation feature fusion module, it effectively combines deep features from both spectral and spatial dimensions, leveraging the complementary nature of these two types of features to improve the model's recognition capability. The fully connected layer performs a linear transformation on the fused features, combined with a normalized exponential function module (such as Softmax), enabling the output of an accurate probability distribution, thereby accurately predicting the target category of pixels. The model design simultaneously supports the classification of clean pixels representing weak targets, clean pixels representing strong targets, and non-clean pixels representing mixed targets, providing an effective solution for multi-target detection.

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

[0050] Furthermore, this invention significantly improves sample accuracy by manually generating high-quality pixel-by-pixel training labels; it refines complex mixed target regions to the pixel level, helping the model to learn target distribution features more accurately; it provides accurate pixel-by-pixel training labels, enabling the model to learn finer-grained features, especially in spatial dimensions; it is suitable for scenes with complex spectra and significant spatial structures, improving the model's robustness and generalization ability in complex scenes; and the high-quality labels reduce error propagation during model training, improving classification performance. Attached Figure Description

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

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

[0053] Figure 3 This is a schematic diagram of sample construction for the present invention;

[0054] Figure 4 This is a schematic diagram illustrating the training and use of the detection model of the present invention. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but these descriptions 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 land cover categories corresponding to weak and strong targets in the area to be detected;

[0059] Based on the abundance coefficients of weak and strong targets in the remote sensing image of the area to be detected, each pixel is determined to be a pure pixel of weak targets, a pure pixel of strong targets, or a non-pure pixel of mixed targets.

[0060] The remote sensing image is divided into areas to be detected at equal intervals to form several clean pixel sub-grids for weak targets, clean pixel sub-grids for strong targets, and non-clean pixel sub-grids for mixed targets;

[0061] In each weak target clean cell subgrid, select a single cell belonging to the weak target clean cell as a weak target clean cell sample;

[0062] In each strong target clean cell subgrid, select a single cell belonging to the strong target clean cell as a strong target clean cell sample;

[0063] In each subgrid of mixed target impure pixels, select a single pixel belonging to the mixed target impure pixel and its neighborhood as a mixed target impure pixel sample;

[0064] Constructing the training set: Use clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs, and label them as weak targets or strong targets respectively; use non-clean pixel samples of mixed targets as model inputs, and further label each pixel in the non-clean pixel samples of mixed targets as a weak target or a strong target.

[0065] The detection model is trained using the training set;

[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 principles of the present invention will be further explained below with reference to specific embodiments.

[0068] This embodiment proposes a rapid remote sensing detection method and system for weak targets that integrates hierarchical models and deep learning. It automatically generates high-quality, balanced deep learning samples, extracts spectral and spatial dimensional information, and achieves the identification, detection, and classification of weak targets. Specifically, it includes the following steps:

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

[0070] Acquire high-resolution remote sensing imagery that fully covers the study area. This data can come from remote sensing platforms such as satellites and drones. To improve the quality of the remote sensing imagery and make it clearer and more accurate, the following high-resolution remote sensing image preprocessing methods are employed, including but not limited to:

[0071] Radiometric calibration: Eliminates radiometric errors caused by factors such as sensors and the atmosphere, so that the gray values ​​of the image can truly reflect the reflectivity of the ground objects.

[0072] Atmospheric correction: Converting the original digital (DN) values ​​into physical quantities (such as reflectance) to establish a quantitative relationship between the reflectance of images and ground features.

[0073] Geometric correction: By acquiring DEM data within the study area, geometric distortion of the image is eliminated, so that the locations of ground features in the image correspond to the actual ground features.

[0074] Cloud removal: Suppresses random noise (clouds, snow, etc.) in images and improves the signal-to-noise ratio of images.

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

[0076] The second step is remote sensing pixel type identification.

[0077] Considering the lack of clear distinction between weak targets and other targets in remote sensing imagery, a single pixel often contains multiple target types. Based on the acquired high-quality, high-resolution remote sensing data, to construct a deep learning sample library, it is necessary to perform remote sensing image pixel type discrimination, i.e., determining whether a pixel is clean or impure. "Weak target" and "strong target" both refer to the classification of land cover categories. In practical applications, multiple land cover categories can be defined as weak targets, or multiple land cover categories can be defined as strong targets. In this embodiment, both weak and strong targets are single land cover categories. The land cover categories corresponding to weak and strong targets are determined according to the actual needs of the field.

[0078] This paper takes two types of remote sensing targets, weak target A and strong target B, as examples. The target 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 in the field, as well as the spectral information of their corresponding remote sensing image locations, and extract the endmember spectra. Based on this, a spectral library is established. That is, using the known coordinate information and remote sensing images of weak target A and strong target B, the spectral measurements of weak target A and strong target B in each band can be calculated.

[0080] Furthermore, the average of the spectral measurements of several known weak targets A in a specific band in a remotely sensed image can be used as the spectral measurement of weak target A in that band. Similarly, the spectral measurement of strong target B in that band can be calculated.

[0081] Then, a fully constrained least squares multi-endmember mixed pixel decomposition model is used to obtain the area proportion of each endmember in the pixel, also known as abundance α, and a threshold T is set based on this to classify the pixel type. To give the model practical physical meaning, it is assumed that the sum of the area proportions of all endmembers is 1, and that each endmember proportion is non-negative. The calculation formula for the fully constrained least squares multi-endmember mixed pixel decomposition model is as follows:

[0082]

[0083] Among them, y i p represents the spectral parameters of any pixel in the i-th band; ij The spectral measurement value of the j-th endmember in the i-th band; j∈{1,2,…,p}; α j The abundance coefficient of the j-th endmember of that pixel; β i This refers to the participation 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. Each cell corresponds to a different land cover category.

[0084] The abundance coefficients of weak target A and strong target B are obtained using a mixed pixel decomposition model, and the area proportion α of targets A and B in each pixel is further obtained. A and α B , where α A and α B Belongs to α j .

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

[0086]

[0087] α B <1,α A +α B =1)

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

[0089] By fitting the spectral measurements of weak target A and strong target B in each band and the spectral parameters of each pixel in the remote sensing image using Formula 1, the abundance coefficients of weak targets and strong targets for each pixel can be obtained. Then, Formula 2 can be used to determine the type of each pixel.

[0090] The third step is to construct a deep learning sample set.

[0091] To construct a hierarchical sample set, the endpoints of clean and non-clean pixels in the remote sensing image are first located. This helps to accelerate the generation of sample points, i.e., to obtain the endpoint positions of weak targets in the remote sensing image: [(X min1 ,Y min1 ),(X max1 ,Y max1 [(X)], to obtain the endpoint positions of strong targets on remote sensing images: [(X)] min2 ,Y min2 ),(X max2 ,Y max2 The endpoint locations of the blended targets on the remote sensing image: [(X min3 ,Y min3 ),(X max3 ,Y max3 )).

[0092] Based on the endpoint pixel coordinates of the three types of targets, pure pixel samples of weak and strong targets are selected using an equidistant grid division method, along with the baseline mixed pixels for the mixed target object level samples. The overall result is as follows: Figure 3 As shown. Specifically, considering that the generation principles for clean pixel samples of weak and strong targets are the same, the following content takes the generation of clean pixel samples of weak targets as an example:

[0093] ① Query the raster size containing all clean pixels of weak targets within the entire remote sensing image:

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

[0095] Where M refers to the number of vertical pixels, i.e., the number of rows; and N refers to the number of horizontal pixels, i.e., the number of columns.

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

[0097]

[0098] W = MM * NN(5)

[0099] The size and spacing of the subgrid are determined based on the distribution of clean pixels of weak targets in the entire image, ensuring that the subgrids cover all pixels belonging to weak targets; MM refers to the number of effective subgrids that can be placed in the number of raster rows, and NN refers to the number of effective subgrids that can be placed in the number of raster columns.

[0100] ③ Select each clean pixel of weak target in each sub-grid and use it as a sample of clean pixel of weak target.

[0101] The process for generating clean pixel samples with strong targets is the same as above.

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

[0103] ④ Generate object-level samples according to the idea of ​​"everything is related to other things, but things that are similar are more closely related".

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

[0105] Where 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 as the distance increases.

[0106] Suppose there are two pixels u(u1, u2, ..., u...). n ) and v(v1,v2,…,v n If the correlation strength between pixels u and v is such that the Euclidean distance can be used as a measure, the specific calculation formula is as follows:

[0107]

[0108] Among them, u i It 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 mixed pixel, and combining the idea of ​​the "first law of geography" and 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, pixels with a correlation strength greater than a set value with a single mixed target impure pixel together with the mixed target impure pixel form mixed target impure pixel samples.

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

[0111] By following the steps from the first to the fourth, 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 clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs, and labeling them as weak targets or strong targets respectively; taking non-clean pixel samples of mixed targets as model inputs, and further labeling each pixel in the non-clean pixel samples of mixed targets as a weak target or a strong target.

[0113] Specifically, based on high-quality remote sensing imagery, and using clean, weak and strong target pixel-level samples and automatically generated corresponding target labels, the labels represent the target category (strong or weak target) corresponding to the sample.

[0114] This method is based on mixed weak and strong target object-level samples and manual interpretation to delineate corresponding target labels. Specifically, for impure pixels of mixed targets, each pixel within the mixed target impure pixel sample is labeled one by one through manual annotation, clarifying its corresponding category label as "weak target" or "strong target". The manually annotated label matrix is ​​used to generate pixel-by-pixel training labels for mixed target samples.

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

[0116] The 1D-CNN model is used to extract the depth features of clean pixels of weak targets and clean pixels of strong targets in the spectral dimension during the training of the detection model, and to extract the depth 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 spatial depth features of mixed target impure pixel samples during detection model training, and to extract the spectral depth features of the input remote sensing image during detection model application.

[0118] The weighted summation feature fusion module is used to fuse deep features in the spectral and spatial dimensions;

[0119] The fully connected layer is used to flatten the fused feature vectors 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 the clean pixels of weak targets, the clean pixels of strong targets, and the non-clean pixels of mixed targets.

[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. The number of layers in the convolutional neural network is gradually increased 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 Z(i,j) represents the feature map, which is also the input and output of the (l+1)th convolutional operation; Z(i,j) is the pixel of the corresponding feature map; s is the convolutional layer parameter; k is the number of channels of the feature map; b is the bias term.

[0123] After obtaining feature maps from each convolutional operation, pooling layers are used to reduce the size of the convolutional feature maps, thereby improving computational speed and enhancing the robustness of the features. Furthermore, to strengthen the network's feature learning ability, non-linear activation functions are typically added after the convolutional layers to enhance the non-linearity of the feature information. Common activation functions include the Sigmoid function, Tanh function, and ReLU function. Compared to Sigmoid and Tanh, the ReLU function is simpler, has a faster convergence speed, and to some extent avoids the gradient explosion and gradient vanishing problems. Its 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 and spatial features extracted by 1D-CNN and 2D-CNN, a Weighted Sum Module (WSM) is constructed to fuse the output feature maps of 1D-CNN and 2D-CNN, combining the spectral and spatial features. The calculation formula for the Weighted Sum Module is as follows:

[0127] W(g)=α*U(g)+(1-α)*V(G)(10)

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

[0129] Then, the fused feature map is linearly transformed into flattened feature vectors in a fully connected layer (FC). Finally, the output of the FC layer is transformed into the probability distribution of each class in a softmax layer to achieve the final classification. The overall process diagram is shown below. Figure 4 As shown.

[0130] Example 2

[0131] This invention also provides a rapid remote sensing detection system for weak targets, which is used to implement the rapid remote sensing detection method for weak targets described in the above technical solution; including:

[0132] The feature category identification module is used to determine the feature categories corresponding to weak and strong targets in the area to be detected;

[0133] The 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 non-pure pixel based on the abundance coefficients of weak and strong targets in the remote sensing image of the area to be detected.

[0134] The sub-grid division module is used to divide the remote sensing image of the area to be detected into several clean pixel sub-grids for weak targets, clean pixel sub-grids for strong targets, and non-clean pixel sub-grids for mixed targets at equal intervals.

[0135] The 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; to 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; and to select a single pixel belonging to a mixed target non-pure pixel and its neighborhood in each mixed target non-pure pixel subgrid as mixed target non-pure pixel samples.

[0136] The training set construction module is used to take clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs and label them as weak targets or strong targets respectively; it takes impure pixel samples of mixed targets as model inputs and further labels each pixel in the impure pixel samples of mixed targets as a weak target or a strong target.

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

[0138] The weak target identification module in the detection area 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 having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the rapid remote sensing detection method for weak targets described in the above technical solution.

[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 computer instructions to perform the rapid remote sensing detection method for weak targets described above.

[0143] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0148] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for rapid remote sensing detection of weak targets, characterized in that: Includes the following steps: Determine the land cover categories corresponding to weak and strong targets in the area to be detected; Based on the abundance coefficients of weak and strong targets in the remote sensing image of the area to be detected, each pixel is determined to be a pure pixel of weak targets, a pure pixel of strong targets, or a non-pure pixel of mixed targets. The remote sensing image is divided into areas to be detected at equal intervals to form several clean pixel sub-grids for weak targets, clean pixel sub-grids for strong targets, and non-clean pixel sub-grids for mixed targets; In each weak target clean cell subgrid, select a single cell belonging to the weak target clean cell as a weak target clean cell sample; In each strong target clean cell subgrid, select a single cell belonging to the strong target clean cell as a strong target clean cell sample; In each subgrid of mixed target impure pixels, select a single pixel belonging to the mixed target impure pixel and its neighborhood as a mixed target impure pixel sample; Constructing the training set: Use clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs, and label them as weak targets or strong targets respectively; use non-clean pixel samples of mixed targets as model inputs, and further label each pixel in the non-clean pixel samples of mixed targets as a weak target or a strong target. The detection model is trained using the training set; 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. The process of calculating the abundance coefficients of weak and strong targets in each pixel of the remote sensing image of the region to be detected includes: extracting the spectral measurements of weak and strong targets from the remote sensing image of the region to be detected, respectively, based on the known coordinates of weak and strong targets in the region to be detected; and calculating the abundance coefficients of weak and strong targets in each pixel of the remote sensing image of the region to be detected based on the spectral information of the remote sensing image of the region to be detected and the spectral measurements of weak and strong targets. The abundance coefficients are calculated by fitting the solution using the following formula: ； in, This represents the spectral parameters of any pixel in the i-th band; The spectral measurement value of the j-th endmember in the i-th band; The abundance coefficient of the j-th endmember of that pixel; This refers to the participation 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; The criteria for determining whether each pixel belongs to a clean pixel of a weak target, a clean pixel of a strong target, or a non-clean pixel of a mixed target are as follows: ； Where P represents the pixel type of a single pixel in the remote sensing image; Refers to pure pixels of weak targets; Refers to mixed target non-pure pixels; This refers to a pure pixel with a strong target. The abundance coefficient of weak targets. The abundance coefficient of strong targets; T refers to the set threshold. The process of dividing a remote sensing image into equally spaced detection areas includes: obtaining the endpoint positions of pixels of a specified target category in the detection area, 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 to be divided in the remote sensing image of the detection area for that target category; Based on the preset sub-grid size and row and column spacing, the area to be detected is divided into several sub-target category grids at equal intervals until the entire total pixel range to be divided is covered; the preset sub-grid size and row and column spacing ensure that the divided target category sub-grids cover all pixels belonging to that target category; The target category is a weak target pure pixel, a strong target pure pixel, or a mixed target non-pure pixel; In the process of generating mixed target impure pixel samples, pixels with a correlation strength greater than a set value with a single mixed target impure pixel are combined with that mixed target impure pixel to form mixed target impure pixel samples; The detection model includes a 1D-CNN model, a 2D-CNN model, a weighted summation feature fusion module, a fully connected layer, and a normalized exponential function module; The 1D-CNN model is used to extract the depth features of clean pixels of weak targets and clean pixels of strong targets in the spectral dimension during the training of the detection model, and to extract the depth 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 spatial depth features of mixed target impure pixel samples during detection model training, and to extract the spectral depth features of the input remote sensing image during detection model application. The weighted summation feature fusion module is used to fuse deep features in the spectral and spatial dimensions; The fully connected layer is used to flatten the fused feature vectors 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 the clean pixels of weak targets and the clean pixels of strong targets.

2. The method according to claim 1, characterized in that: After performing radiometric calibration, atmospheric correction, geometric correction, cloud and snow removal, fusion, mosaicking, and cropping preprocessing operations on the remote sensing image of the area to be detected, the steps of the rapid remote sensing detection method for weak targets are then executed.

3. The method according to claim 1, characterized in that: The process of generating training samples for mixed target impure pixels includes: manually labeling each pixel within the mixed target impure pixel sample to determine 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.

4. A rapid remote sensing detection system for weak targets, characterized in that: This system is used to implement the rapid remote sensing detection method for weak targets as described in any one of claims 1-3; comprising: The feature category identification module is used to determine the feature categories corresponding to weak and strong targets in the area to be detected; The 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 non-pure pixel based on the abundance coefficients of weak and strong targets in the remote sensing image of the area to be detected. The sub-grid division module is used to divide the remote sensing image of the area to be detected into several clean pixel sub-grids for weak targets, clean pixel sub-grids for strong targets, and non-clean pixel sub-grids for mixed targets at equal intervals. The 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; to 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; and to select a single pixel belonging to a mixed target non-pure pixel and its neighborhood in each mixed target non-pure pixel subgrid as mixed target non-pure pixel samples. The training set construction module is used to take clean pixel samples of weak targets and clean pixel samples of strong targets as model inputs and label them as weak targets or strong targets respectively; it takes impure pixel samples of mixed targets as model inputs and further labels each pixel in the impure pixel samples of mixed targets as a weak target or a strong target. The model training module is used to train the detection model using the training set; The weak target identification module in the detection area 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