A remote sensing image anomaly detection system and method integrating multi-source noise characteristics

Abstract

The present invention discloses a remote sensing image anomaly detection system and method that integrates multi-source noise feature anomalies in the field of remote sensing image detection technology, including: obtaining optical satellite remote sensing image data to be processed; parsing the optical satellite remote sensing image data to obtain remote sensing image detailed information; annotating the remote sensing image detailed information with anomaly types to construct an anomaly feature data set; based on the anomaly feature data set, performing multi-source noise feature detection using a pre-designed anomaly detection fusion framework; and outputting the anomaly area and anomaly detection results obtained by the multi-source noise feature detection. Through fusion detection, the present invention promptly detects color anomalies, cloud shadow obscurations, pixel information anomalies, and other problems in remote sensing images, thereby improving the versatility of remote sensing image anomaly detection algorithms and further increasing the efficiency of remote sensing image anomaly detection.

Description

Technical Field

[0001] The present invention relates to a remote sensing image anomaly detection system and method that fuses abnormal multi-source noise characteristics, and belongs to the technical field of remote sensing image detection. Background Art

[0002] In the field of optical satellite remote sensing imagery, detection algorithms for color anomalies, cloud shadows, and pixel information anomalies continue to evolve. In the area of ​​color anomaly detection, methods have evolved from early reliance on manually designed features and rules (such as those based on edge and texture analysis) to current deep learning-based methods. Multilayer perceptron neural network models, for example, extract feature inputs to effectively and uniformly detect all types of color anomalies, significantly improving detection efficiency and automation. For cloud shadow detection, cloud detection is achieved through deep learning, with relatively mature algorithms such as Faster R-CNN and YOLO5. For pixel information anomaly detection, statistical methods use the mean and variance to describe the overall characteristics of the image to identify anomalous areas. However, these methods are susceptible to noise and illumination variations, and their ability to identify anomalies in local areas is limited.

[0003] In the existing technology, although various algorithms have made certain progress, the current optical satellite remote sensing image anomaly detection algorithm still mainly operates through separate detection, and lacks a systematic algorithm model framework for image anomaly detection. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a remote sensing image anomaly detection system and method that integrates multi-source noise feature anomaly. In view of the uncertainty, diversity and randomness of remote sensing image anomalies, the system aims to timely discover color anomalies, cloud shadow obscurations, pixel information anomalies and other problems in remote sensing images through fusion detection, thereby improving the versatility of remote sensing image anomaly detection algorithms and further improving the efficiency of remote sensing image anomaly detection.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] In a first aspect, the present invention provides a method for detecting anomalies in remote sensing images by fusing abnormal multi-source noise features, comprising:

[0007] Obtain optical satellite remote sensing image data to be processed;

[0008] Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images;

[0009] Annotate the abnormal types of detailed information of remote sensing images and build an abnormal feature dataset;

[0010] Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework;

[0011] Output the abnormal area and abnormal detection results obtained by multi-source noise feature detection.

[0012] Furthermore, the detailed information of the remote sensing image is annotated with abnormal types, including: annotating the remote sensing image color abnormality, cloud shadow cover and pixel information abnormal sample data according to the abnormal type, wherein the annotation method includes single remote sensing image abnormal sample annotation and mixed remote sensing image abnormal sample annotation. The single remote sensing image abnormal sample annotation includes based on predefined abnormal types, using remote sensing processing software to observe the image at multiple scales, assisting in identifying abnormal areas through contrast enhancement or index analysis, using vector tools to accurately outline the boundaries of the abnormal area, and combining spatial context to exclude pseudo-anomalies, and finally ensuring the accuracy and consistency of the annotation results through cross-validation.

[0013] Furthermore, the anomaly detection fusion framework adopts a multi-branch optical satellite remote sensing image anomaly detection backbone network model that combines a residual structure with a feature rotation operation. The feature rotation operation expression is:

[0014] in, represents the input feature map, Represents the output feature map; is the convolution operation; n represents the number of feature maps, , Represents clockwise rotation , Represents counterclockwise rotation , represents the fusion operation, Represents the rotation angle;

[0015] The loss function mathematical model of the anomaly detection fusion framework is:

[0016] in, is the calculated loss value, is the true score value of the input sample structure misalignment, is the output score of the network, Key parameters set by humans;

[0017] For a single-branch network, the loss function is:

[0018] Among them, Loss represents the single-branch network loss function, represents the number of samples, Representative The true value of the sample, Representative The predicted value of samples;

[0019] For the parameter training of the fully connected layer after the fusion of the three branches, the loss function is:

[0020] in, represents the fusion branch network loss function, is the weight of the corresponding item i, Representative Branch to The true value of the sample, Representative Branch to The predicted value of the sample.

[0021] Furthermore, the multi-source noise feature detection includes: conducting satellite remote sensing image anomaly feature detection training based on existing sample data, training a single branch of the network for anomalies appearing in the image data and establishing anomaly features, the anomalies including remote sensing image color anomalies, cloud shadow masking and pixel anomalies, the anomaly features including strip anomaly features based on frequency domain analysis, cloud anomaly features based on band thresholds and texture / anomaly point features based on local variance detection.

[0022] Furthermore, the color anomaly detection is as follows: through HSV color space conversion and dual-stream attention mechanism, hue / saturation offset features are explicitly extracted to detect color anomalies, including HSV color space conversion, hue / saturation offset feature extraction, dual-stream attention mechanism and anomaly detection and output, wherein:

[0023] The HSV color space conversion includes converting the RGB image to the HSV space and separating the three independent components of hue (H), saturation (S), and value (V):

[0024] In the formula, R represents the red component, G represents the green component, and B represents the blue component;

[0025] The hue / saturation shift feature extraction formula is:

[0026] in, is the hue shift, is the current hue, For reference normal hue, is the saturation offset, is the current saturation, is the reference saturation;

[0027] The dual-stream attention mechanism includes calculating spatial attention within the hue shift and saturation shift respectively, focusing on local anomalies, bidirectionally interacting hue shift and saturation shift, and capturing hue-saturation correlation abnormality patterns;

[0028] The anomaly detection and output includes the anomaly score being divided by a threshold segmentation or classification network Figure 2 The image is then digitized and an RGB image with abnormal area marks is output.

[0029] Furthermore, the cloud shadow detection is performed by combining multi-scale dilated convolution and the CBAM attention module to capture the contextual features of large-scale occlusions and detect cloud shadows, including multi-scale dilated convolution feature extraction, CBAM attention module feature optimization, and feature fusion and decoding, where:

[0030] The expression of the multi-scale dilated convolution feature extraction is:

[0031] Where, Represents the output feature map at position The value of , m represents the row coordinate of the output feature map, n represents the column coordinate of the output feature map, Represents the height of the dilated convolution kernel, Represents the width of the hole convolution kernel, a represents the row coordinate of the input feature map, b represents the column coordinate of the input feature map, represents the input feature map, Represents the dilated convolution kernel, r represents the dilation rate, and when r=1, it degenerates into standard convolution. When r increases, the actual receptive field satisfies:

[0032] Where K represents the actual receptive field, and k represents the size of the standard convolution kernel;

[0033] The CBAM attention module optimizes features to detect cloud shadows with variable scales and blurred boundaries:

[0034] The feature fusion and decoding include: introducing temporal difference features to enhance the cloud shadow dynamic detection capability, fusing multi-scale features, fusing the multi-scale features optimized by CBAM with shallow features through jump connections, supplementing detailed information, using transposed convolution or upsampling layers to gradually restore the resolution, and outputting a cloud shadow probability map.

[0035] Furthermore, the pixel anomaly detection is performed by combining a high-resolution preservation structure with densely connected blocks to maintain the original detail texture to detect local pixel anomalies, including high-resolution preservation structure, densely connected block feature extraction and local pixel anomaly detection, wherein:

[0036] The high-resolution preservation structure retains the detailed texture of the original image through the loss function design, avoiding the masking of minor anomalies by the structural similarity loss. The total loss function is:

[0037] in, is the total loss function, is the optimal weight, is the pixel-level loss, is the structural similarity loss;

[0038] The densely connected block feature extraction includes, in the feature extraction stage, a high-resolution image is initially downsampled by convolution, and multiple stacked densely connected blocks are input, the densely connected blocks retain multi-scale textures through feature reuse, and output enhanced features;

[0039] The local pixel anomaly detection is achieved by calculating the local gradient autocorrelation, and the formula is:

[0040] in, Represents the amplitude of the local gradient, and local pixel anomaly detection is achieved by calculating the autocorrelation of the local gradient. I represents the grayscale value of the input image. It is a Sobel operator, a convolution kernel in the x and y directions, used to calculate the gradient of the image in the horizontal and vertical directions.

[0041] In a second aspect, the present invention provides a remote sensing image anomaly detection system integrating multi-source noise feature anomaly, comprising:

[0042] Data acquisition module: acquires optical satellite remote sensing image data to be processed;

[0043] Data analysis module: Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images;

[0044] Data annotation module: annotate the abnormal types of detailed information of remote sensing images and build abnormal feature datasets;

[0045] Data detection module: Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework;

[0046] Data output module: outputs abnormal areas and abnormal detection results obtained by multi-source noise feature detection.

[0047] In a third aspect, the present invention provides a remote sensing image anomaly detection device that integrates abnormal multi-source noise characteristics, including a processor and a storage medium;

[0048] The storage medium is used to store instructions;

[0049] The processor is configured to operate according to the instructions to execute the steps of any of the above methods.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. This solution addresses the need for anomaly detection in optical satellite remote sensing images. By integrating multi-source noise features into an optical satellite remote sensing image anomaly detection method, it can detect visual phenomena such as color anomalies, pixel distribution anomalies, detail information destruction, and pixel information anomalies caused by various factors, including atmospheric disturbances, space radiation, electromagnetic interference, satellite platform disturbances, and optical imaging system anomalies. This approach further improves image preprocessing efficiency and supports the production of value-added remote sensing image products and intelligent interpretation and interpretation.

[0053] 2. This solution can quickly and timely detect color anomalies, cloud shadow obscurations, and pixel information anomalies in remote sensing images, solving the problems of the single anomaly detection algorithm for conventional remote sensing images and the low efficiency of image preprocessing caused by anomaly detection. Through the research of this method, it can effectively solve business problems such as value-added product production, quantitative remote sensing inversion applications, and intelligent image interpretation and interpretation caused by optical satellite remote sensing image anomalies in the actual business production process. At the same time, this technology can also be used in business systems such as large-scale production processing, cleaning and management, and application services of massive optical satellite remote sensing images to support subsequent remote sensing big data applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings, which constitute part of the present invention, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0055] Figure 1 A schematic diagram of anomaly types in remote sensing satellite images according to a method for detecting anomalies in remote sensing images by integrating abnormalities in multi-source noise characteristics provided in the first embodiment of the present invention;

[0056] Figure 2 A flowchart of a method for detecting anomalies in remote sensing images by integrating abnormal multi-source noise features provided in the first embodiment of the present invention;

[0057] Figure 3 A schematic diagram of anomaly samples of hybrid remote sensing images provided by a method for detecting anomalies in remote sensing images integrating abnormalities with multi-source noise features according to the first embodiment of the present invention;

[0058] Figure 4A schematic diagram of a residual network combining a residual structure and a feature rotation operation for a remote sensing image anomaly detection method integrating multi-source noise feature anomaly provided in the first embodiment of the present invention;

[0059] Figure 5 A schematic diagram of feature rotation operation of a method for detecting anomalies in remote sensing images by integrating multi-source noise feature anomalies provided in the first embodiment of the present invention;

[0060] Figure 6 A schematic diagram of color anomaly detection using HSV color space conversion and a dual-stream attention mechanism, provided in accordance with the first embodiment of the present invention, in a remote sensing image anomaly detection method integrating multi-source noise features;

[0061] Figure 7 A schematic diagram of a dual-stream attention mechanism for detecting abnormalities in remote sensing images by integrating multi-source noise features, provided in the first embodiment of the present invention;

[0062] Figure 8 A schematic diagram of cloud shadow mask detection combining multi-scale dilated convolution and CBAM attention in a remote sensing image anomaly detection method integrating multi-source noise characteristics provided in the first embodiment of the present invention;

[0063] Figure 9 A schematic diagram of pixel anomaly detection using a high-resolution preserving structure and densely connected blocks in a remote sensing image anomaly detection method that integrates multi-source noise feature anomaly detection provided by the first embodiment of the present invention;

[0064] Figure 10 A schematic diagram of low-resolution image to high-resolution image reconstruction for a remote sensing image anomaly detection method integrating multi-source noise characteristics provided in the first embodiment of the present invention;

[0065] Figure 11 A schematic diagram of abnormal area positioning and result output of a remote sensing image anomaly detection method that integrates multi-source noise characteristics provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0066] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other.

[0067] The following detailed description is an exemplary description and is intended to provide further detailed description of the present invention. Unless otherwise indicated, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present invention belongs. The terms used in the present invention are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0068] Example 1:

[0069] This embodiment addresses the need for unified intelligent detection of remote sensing image noise. A remote sensing image anomaly detection method that integrates multi-source noise feature anomalies is designed. This method integrates computer vision, deep learning, and remote sensing domain knowledge to construct a Rotational Residual Multi-branch Anomaly Detection Network (RRMAD-Net) that combines "physical model guidance, multi-task learning, and anomaly detection." Different types of remote sensing image anomalies are trained separately to obtain network parameters suitable for different anomaly detection methods, thereby improving the accuracy of remote sensing anomaly detection. To improve detection efficiency, the network parameters for different anomaly detection methods are then regressed in a fully connected partial branch anomaly detection method to obtain the optimal model parameters. This allows for the timely and rapid detection of remote sensing image color anomalies, cloud shadow obscuration, and pixel information anomalies, supporting the production of subsequent basic products (L2-level products) and value-added products (DOM, DSM / DEM). The overall technical approach is as follows:

[0070] (1) First, the optical satellite remote sensing image data to be processed is analyzed to obtain detailed information of the remote sensing image.

[0071] (2) Then, the abnormal feature dataset is constructed. According to the abnormal type, the sample data of remote sensing image color abnormality, cloud shadow cover, and pixel information abnormality are labeled to construct the satellite remote sensing image abnormal feature dataset.

[0072] (3) Secondly, the design of the Rotation Residual Multi-branch Anomaly Detection Network (RRMAD-Net). The residual structure is combined with the feature rotation operation to design a multi-branch anomaly detection backbone network combined with the feature rotation operation.

[0073] (4) Secondly, multi-source noise feature detection. Based on the existing sample data, satellite remote sensing image anomaly feature detection training is carried out. A single branch of the network is trained for anomalies such as remote sensing image color anomalies, cloud shadow masking, and pixel information anomalies that appear in the image data. Features such as strip anomaly features based on frequency domain analysis, cloud anomaly features based on band thresholds, and texture / anomaly point features based on local variance detection are established.

[0074] (5) Finally, the abnormal area is located and the results are output. Fusion is performed at the end of the network to locate abnormal areas such as color abnormalities, cloud shadows, and pixel information abnormalities in remote sensing images, and the remote sensing image anomaly detection result map is output.

[0075] At present, large-scale satellite remote sensing image anomaly detection mainly manifests itself in optical satellite remote sensing image color anomalies, cloud shadow occlusion, pixel information anomalies and other problems. This scheme adopts an optical satellite remote sensing image anomaly detection method that integrates multi-source noise features. By designing a rotation residual multi-branch anomaly detection network (RRMAD-Net), it can detect remote sensing image color anomalies, cloud shadow occlusion, pixel information anomalies and other contents. It can effectively detect abnormal noise in optical satellite remote sensing images, preserve the original ground object details of the image, and quickly and accurately locate the anomaly position of optical satellite remote sensing images.

[0076] Based on the above analysis, if Figure 2 As shown in the figure, the present invention designs a method for detecting anomalies in optical satellite remote sensing images by integrating multi-source noise features. The specific steps are as follows:

[0077] S1: Analyze the optical satellite remote sensing image data to be processed and obtain detailed information of the remote sensing image.

[0078] S2: Construction of satellite remote sensing image anomaly feature dataset. According to the anomaly type, the remote sensing image color anomaly, cloud shadow mask, pixel information anomaly sample data (such as Figure 1 As shown in Figure 3, there are two main methods to construct satellite remote sensing image abnormal feature datasets: manual annotation and automatic annotation.

[0079] S2-1: Annotate a single remote sensing image for anomaly samples. Based on predefined anomaly types (such as color anomalies, cloud shadows, and pixel information anomalies), use remote sensing processing software (such as QGIS and ENVI) to perform multi-scale observations of the image. Use contrast enhancement or index analysis to assist in identifying anomalous areas. Use vector tools to precisely delineate the boundaries of the anomalous areas, and incorporate spatial context to eliminate spurious anomalies. Finally, cross-validate to ensure the accuracy and consistency of the annotation results, providing high-quality samples for model training.

[0080] S2-2: Anomaly sample annotation of mixed remote sensing images. S2-1 mainly describes remote sensing images with anomaly problems. In fact, optical satellite remote sensing images are very large. The image width of wide-band satellites can even reach 150km. There will be anomalies such as color anomalies, cloud shadows, and pixel information anomalies. Typical samples are as follows: Figure 3 As shown, there are multiple sample anomalies in one image.

[0081] S3: Construction of anomaly detection fusion framework. Since optical satellite remote sensing images are overhead images, ground objects exhibit rotation invariance during the imaging process. When using a deep learning network, both the target pattern characteristics and the imaging angle of the image must be considered, resulting in multiple kernels representing a pattern. Therefore, a convolutional neural network is required. Since the number of convolution kernels is fixed, the convolution kernel's ability to express feature diversity is limited. A multi-branch optical satellite remote sensing image anomaly detection backbone network model is adopted, combining the residual structure with the feature rotation operation. A multi-branch anomaly detection backbone network combined with the feature rotation operation is designed to enrich the convolution kernel's ability to express feature diversity, thereby improving the feature extraction ability of the neural network and improving the performance of the designed convolution network.

[0082] S3-1: Design of a backbone network model for anomaly detection in multi-branch optical satellite remote sensing images. Residual neural networks (ResNet) solve the training problems of deep networks, especially the network degradation problem, by introducing residual blocks. This simplifies the network learning objectives, greatly reduces the difficulty of network training, and enables the number of network layers to reach hundreds. By combining the residual structure with feature rotation operations (such as Figure 4 As shown in Figure 3), we design a multi-branch anomaly detection backbone network combined with feature rotation operations.

[0083] S3-2: Combining residual structure with feature rotation. This solves the problem of deep network training and retains low-level detail information. The feature rotation operation definition is:

[0084] in, represents the input feature map, Represents the output feature map. It is a convolution operation used for feature extraction. It can be a simple convolution unit or a set of independent convolution structures. n represents the number of feature maps. , Represents clockwise rotation , Represents counterclockwise rotation , represents the fusion operation, Represents the rotation angle, and we plan to use the mean fusion method to fuse the feature maps. Figure 5 A schematic diagram of a rotation operation with four angles is given.

[0085] In this operation, the rotation and reverse rotation of the feature map keep the spatial relationship unchanged. The feature extraction part is the convolution operation unit, such as Figure 5As shown in the figure, each block m represents a convolutional structure. Multiple feature maps are fed into this feature extraction structure, and parameters are trained using a weight-sharing approach. The output is a fusion of the feature maps. By rotationally expanding and fusing the feature maps, the convolution kernel's sensitivity to the rotation angle of the pattern features is reduced, allowing the convolution kernel to focus more on extracting the pattern features.

[0086] S2-3: Loss function construction. The Huber loss function is selected. Based on its good derivative continuity, outlier robustness, and the consistency between the linear gradient decrease and network accuracy, the model performance is evaluated to further improve the accuracy of the multi-branch optical satellite remote sensing image anomaly detection model. Its mathematical model is as follows:

[0087] in, is the calculated loss value, is the true score value of the input sample structure misalignment, is the output score of the network, These are key parameters set artificially.

[0088] For a single-branch network, the loss function is:

[0089] Among them, Loss represents the single-branch network loss function, represents the number of samples, Representative The true value of the sample, Representative The predicted value of the sample.

[0090] For the parameter training of the fully connected layer after the fusion of the three branches, the loss function uses the weighted Huber loss function:

[0091] in, represents the fusion branch network loss function, is the weight of the corresponding item i, Representative Branch to The true value of the sample, Representative Branch to The predicted value of the sample.

[0092] S4: Multi-source noise feature detection. Based on existing sample data, we conduct training for satellite remote sensing image anomaly feature detection. We train a single branch of the network to detect anomalies such as color anomalies, cloud shadows, and pixel information anomalies in the image data. We establish features such as strip anomaly features based on frequency domain analysis, cloud anomaly features based on band thresholds, and texture / outlier features based on local variance detection.

[0093] S4-1: Color anomaly detection branch. Through HSV color space conversion and dual-stream attention mechanism (specific technical route is as follows Figure 6 ), explicitly extracting hue / saturation shift features to detect color anomalies.

[0094] S4-1-1: HSV color space conversion, converting the RGB image to HSV space, separating the three independent components of hue (H), saturation (S), and lightness (V):

[0095] Where R represents the red component, G represents the green component, and B represents the blue component.

[0096] S4-1-1-2: Hue / saturation offset feature extraction. Hue offset ΔH: the difference between the current hue H value and the reference normal hue The difference.

[0097] Saturation offset , the current saturation S value and the reference saturation The difference.

[0098] S4-1-3: Dual-stream attention mechanism (see the specific technical process Figure 7 ), respectively in and Calculate spatial attention within the stream to focus on local anomalies; bidirectional interaction and flow,capturing unusual patterns of hue-saturation correlations.

[0099] S4-1-4: Anomaly detection and output. The anomaly score is calculated by threshold segmentation or classification network (such as 1×1 convolution). Figure 2 The image is then digitized and an RGB image with abnormal area marks is output.

[0100] S4-2: Cloud shadow detection branch. Combining multi-scale dilated convolution and CBAM attention module (specific technical route is as follows Figure 8 As shown in Figure 3), it effectively captures the contextual features of large-scale occlusions and detects cloud shadows.

[0101] S4-2-1: Multi-scale dilated convolution feature extraction to detect cloud shadows with variable scales and blurred boundaries. , dilated convolution kernel (void rate r), the output Y is:

[0102] Where, Represents the output feature map at position The value of , m represents the row coordinate of the output feature map, n represents the column coordinate of the output feature map, Represents the height of the dilated convolution kernel, Represents the width of the hole convolution kernel, a represents the row coordinate of the input feature map, and b represents the column coordinate of the input feature map.

[0103] When r=1, it degenerates into standard convolution;

[0104] When the void rate r increases, the actual receptive field K satisfies:

[0105] Where K represents the actual receptive field and k represents the size of the standard convolution kernel.

[0106] S4-2-2: The CBAM attention module optimizes features to detect cloud shadows with variable scales and blurred boundaries.

[0107] S4-2-3: Feature fusion and decoding: Introducing temporal difference features to enhance cloud shadow dynamic detection capabilities. Multi-scale features are integrated, using skip connections to fuse the CBAM-optimized multi-scale features with shallow features to supplement detailed information. Transposed convolution or upsampling layers are used to gradually restore resolution, outputting a cloud shadow probability map.

[0108] S4-3: Pixel anomaly detection branch. It uses a high-resolution preservation structure and dense connection block combination (specific technical route is as follows Figure 9 ), maintaining the original detail texture to detect local pixel anomalies.

[0109] S4-3-1: High-resolution structure preservation. The loss function is designed to preserve the detailed texture of the original image and avoid the structural similarity loss (SSIM) from masking minor anomalies. The total loss function is:

[0110] in, is the total loss function, is the optimal weight, This is a pixel-level loss that forces the network to restore local details (such as tiny scratches and noise) and is sensitive to anomalies. For structural similarity loss, over-optimization will blur local differences.

[0111] S4-3-2: Densely connected block feature extraction. In the feature extraction stage, the high-resolution image is initially downsampled by convolution and input into multiple stacked densely connected blocks (specifically, Figure 10 As shown in Figure 3, the densely connected blocks retain multi-scale textures through feature reuse, output enhanced features, reduce gradient disappearance, and improve detail recovery efficiency.

[0112] S4-3-3: Local pixel anomaly detection. This is achieved by calculating the local gradient autocorrelation (LGA):

[0113] in, Represents the amplitude of the local gradient, and local pixel anomaly detection is achieved by calculating the autocorrelation of the local gradient. I represents the grayscale value of the input image. It is a Sobel operator with convolution kernels in the x and y directions, which is used to calculate the gradient of the image in the horizontal and vertical directions and detect abnormally smooth edges.

[0114] S5: Abnormal area positioning and result output. Fusion is performed at the end of the network to locate abnormal areas such as remote sensing image color anomaly, cloud shadow cover, pixel information anomaly, and output the remote sensing image anomaly detection result map (specifically, Figure 11 shown).

[0115] Example 2:

[0116] A remote sensing image anomaly detection system integrating multi-source noise feature anomaly can implement the remote sensing image anomaly detection method integrating multi-source noise feature anomaly described in Example 1, comprising:

[0117] Data acquisition module: acquires optical satellite remote sensing image data to be processed;

[0118] Data analysis module: Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images;

[0119] Data annotation module: annotate the abnormal types of detailed information of remote sensing images and build abnormal feature datasets;

[0120] Data detection module: Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework;

[0121] Data output module: outputs abnormal areas and abnormal detection results obtained by multi-source noise feature detection.

[0122] Example 3:

[0123] The embodiment of the present invention further provides a device for detecting anomalies in remote sensing images by fusing abnormal multi-source noise characteristics, which can implement the method for detecting anomalies in remote sensing images by fusing abnormal multi-source noise characteristics described in the first embodiment, including a processor and a storage medium;

[0124] The storage medium is used to store instructions;

[0125] The processor is configured to operate according to the instructions to execute the steps of the following method:

[0126] Obtain optical satellite remote sensing image data to be processed;

[0127] Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images;

[0128] Annotate the abnormal types of detailed information of remote sensing images and build an abnormal feature dataset;

[0129] Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework;

[0130] Output the abnormal area and abnormal detection results obtained by multi-source noise feature detection.

[0131] Example 4:

[0132] The embodiment of the present invention further provides a computer-readable storage medium that can implement the method for detecting anomalies in remote sensing images by integrating abnormal multi-source noise characteristics as described in the first embodiment. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the following steps:

[0133] Obtain optical satellite remote sensing image data to be processed;

[0134] Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images;

[0135] Annotate the abnormal types of detailed information of remote sensing images and build an abnormal feature dataset;

[0136] Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework;

[0137] Output the abnormal area and abnormal detection results obtained by multi-source noise feature detection.

[0138] It is understood from common technical knowledge that the present invention may be implemented by other embodiments that do not depart from its spirit or essential features. Therefore, the embodiments disclosed above are, in all respects, merely illustrative and not exclusive. All modifications within the scope of the present invention or equivalent to the scope of the present invention are intended to be encompassed by the present invention.

[0139] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, 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 magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0140] 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 flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, 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 flowcharts and / or block diagrams. 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.

[0141] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work 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 The function specified in one or more boxes.

[0142] 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 the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for detecting anomalies in remote sensing images by integrating abnormal multi-source noise features, characterized in that: include: Obtain optical satellite remote sensing image data to be processed; Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images; Anomaly types are annotated on the detailed information of remote sensing images to construct anomaly feature dataset. The anomaly types include color anomaly, cloud shadow obscuration, and pixel information anomaly. Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework. The anomaly detection fusion framework adopts a multi-branch optical satellite remote sensing image anomaly detection backbone network model that combines residual structure with feature rotation operation. The feature rotation operation expression is: in, represents the input feature map, Represents the output feature map; is the convolution operation; n represents the number of feature maps, , Represents clockwise rotation , i=0,1,..,n-1, Represents counterclockwise rotation , represents the fusion operation, Represents the rotation angle; The loss function mathematical model of the anomaly detection fusion framework is: in, is the calculated loss value, is the true value of the input sample score, is the output score of the network, that is, the model is based on the input x The predicted value, It is a key parameter set manually to control the sensitivity of the loss function; For a single-branch network, the loss function is: Among them, Loss represents the single-branch network loss function, represents the number of samples, =0,1,2,……,N, Representative The true value of the sample, Representative The predicted value of samples; For the parameter training of the fully connected layer after the fusion of the three branches, the loss function is: Among them, Loss 总 represents the fusion branch network loss function, is the weight of the corresponding item i, representing the Branch to The true value of the sample, Representative Branch to The predicted value of samples; Output the abnormal area and abnormal detection results obtained by multi-source noise feature detection.

2. The method for detecting anomalies in remote sensing images by integrating abnormal multi-source noise characteristics according to claim 1 is characterized in that: The detailed information of remote sensing images is annotated with abnormal types, including: annotating remote sensing image color anomaly, cloud shadow cover and pixel information abnormal sample data according to the anomaly type, wherein the annotation methods include single remote sensing image abnormal sample annotation and mixed remote sensing image abnormal sample annotation. The single remote sensing image abnormal sample annotation includes multi-scale observation of the image based on predefined anomaly types using remote sensing processing software, auxiliary identification of abnormal areas through contrast enhancement or index analysis, and precise delineation of the boundaries of the abnormal area using vector tools. At the same time, pseudo anomalies are excluded in combination with spatial context, and finally cross-validation is performed to ensure the accuracy and consistency of the annotation results.

3. The method for detecting anomalies in remote sensing images by integrating abnormal multi-source noise features according to claim 1 is characterized in that: The multi-source noise feature detection includes: conducting satellite remote sensing image anomaly feature detection training based on existing sample data, training a single network branch for anomalies appearing in the image data and establishing anomaly features, wherein the anomalies include remote sensing image color anomalies, cloud shadow masking, and pixel anomalies, and the anomaly features include strip anomaly features based on frequency domain analysis, cloud anomaly features based on band thresholds, and texture / anomaly point features based on local variance detection.

4. The method for detecting anomalies in remote sensing images by fusing abnormal multi-source noise features according to claim 3 is characterized in that: The color anomaly detection is as follows: through HSV color space conversion and dual-stream attention mechanism, hue / saturation offset features are explicitly extracted to detect color anomalies, including HSV color space conversion, hue / saturation offset feature extraction, dual-stream attention mechanism and anomaly detection and output, wherein: The HSV color space conversion includes converting the RGB image into the HSV space and separating the three independent components of hue H, saturation S, and lightness V: In the formula, R represents the red component, G represents the green component, and B represents the blue component; The hue / saturation shift feature extraction formula is: Among them, ΔH(x,y) is the hue shift, H(x,y) is the current hue, is the reference normal hue, ΔS(x,y) is the saturation offset, S(x,y) is the current saturation, is the reference saturation; The dual-stream attention mechanism includes calculating spatial attention within the hue shift and saturation shift respectively, focusing on local anomalies, bidirectionally interacting hue shift and saturation shift, and capturing hue-saturation correlation abnormality patterns; The anomaly detection and output includes binarizing the anomaly score map through threshold segmentation or classification network, and outputting an RGB image with abnormal area marks.

5. The method for detecting anomalies in remote sensing images by fusing abnormal multi-source noise features according to claim 3 is characterized in that: The cloud shadow detection method combines multi-scale dilated convolution and CBAM attention module to capture the contextual features of large-scale occlusion and detect cloud shadow. The method includes multi-scale dilated convolution feature extraction, CBAM attention module feature optimization, and feature fusion and decoding. The expression of the multi-scale dilated convolution feature extraction is: Where, Represents the value of the output feature map at position (m,n), m represents the row coordinate of the output feature map, n represents the column coordinate of the output feature map, k h Represents the height of the dilated convolution kernel, k w Represents the width of the hole convolution kernel, a represents the row coordinate of the input feature map, b represents the column coordinate of the input feature map, X represents the input feature map, Represents the dilated convolution kernel, r represents the dilation rate, and when r=1, it degenerates into standard convolution. When r increases, the actual receptive field satisfies: Where K represents the actual receptive field, and k represents the size of the standard convolution kernel; The CBAM attention module optimizes features to detect cloud shadows with variable scales and blurred boundaries: The feature fusion and decoding include: introducing temporal difference features to enhance the cloud shadow dynamic detection capability, fusing multi-scale features, fusing the multi-scale features optimized by CBAM with shallow features through jump connections, supplementing detailed information, using transposed convolution or upsampling layers to gradually restore the resolution, and outputting a cloud shadow probability map.

6. The method for detecting anomalies in remote sensing images by fusing abnormal multi-source noise features according to claim 3 is characterized in that: The pixel anomaly detection method is to maintain the original detail texture to detect local pixel anomalies by combining high-resolution preservation structure and densely connected blocks, including high-resolution preservation structure, densely connected block feature extraction and local pixel anomaly detection, wherein: The high-resolution preservation structure retains the detailed texture of the original image through the loss function design, avoiding the masking of minor anomalies by the structural similarity loss. The total loss function is: Among them, L total is the total loss function, which is a comprehensive indicator used to measure the difference between the predicted results and the actual results during model training; λ is the optimal weight, a parameter between 0 and 1, used to balance the weight between pixel-level loss and structural similarity loss; It is a pixel-level loss, which is used to measure the difference between the predicted image and the real image at the pixel level; is the structural similarity loss; The densely connected block feature extraction includes, in the feature extraction stage, a high-resolution image is initially downsampled by convolution, and multiple stacked densely connected blocks are input, the densely connected blocks retain multi-scale textures through feature reuse, and output enhanced features; The local pixel anomaly detection is achieved by calculating the local gradient autocorrelation, and the formula is: Among them, G mag Represents the amplitude of the local gradient, and local pixel anomaly detection is achieved by calculating the autocorrelation of the local gradient. I represents the grayscale value of the input image. 、 It is the convolution kernel of the Sobel operator in the x and y directions, which is used to calculate the gradient of the image in the horizontal and vertical directions.

7. A remote sensing image anomaly detection system integrating multi-source noise characteristics, characterized by: include: Data acquisition module: acquires optical satellite remote sensing image data to be processed; Data analysis module: Analyze optical satellite remote sensing image data to obtain detailed information of remote sensing images; Data annotation module: annotates the remote sensing image details with abnormal types and constructs an abnormal feature dataset. The abnormal types include color abnormalities, cloud shadows, and pixel information abnormalities. Data detection module: Based on the abnormal feature dataset, multi-source noise feature detection is performed through a pre-designed anomaly detection fusion framework. The anomaly detection fusion framework adopts a multi-branch optical satellite remote sensing image anomaly detection backbone network model that combines residual structure with feature rotation operation. The feature rotation operation expression is: in, represents the input feature map, Represents the output feature map; is the convolution operation; n represents the number of feature maps, , Represents clockwise rotation , i=0,1,..,n-1, Represents counterclockwise rotation , represents the fusion operation, Represents the rotation angle; The loss function mathematical model of the anomaly detection fusion framework is: in, is the calculated loss value, is the true value of the input sample score, The output score of the network, that is, the model according to the input x The predicted value, It is a key parameter set manually to control the sensitivity of the loss function; For a single-branch network, the loss function is: Among them, Loss represents the single-branch network loss function, represents the number of samples, =0,1,2,……,N, Representative The true value of the sample, Representative The predicted value of samples; For the parameter training of the fully connected layer after the fusion of the three branches, the loss function is: Among them, Loss 总 represents the fusion branch network loss function, is the weight of the corresponding item i, Representative Branch to The true value of the sample, Representative Branch to The predicted value of samples; Data output module: outputs abnormal areas and abnormal detection results obtained by multi-source noise feature detection.

8. A remote sensing image anomaly detection device integrating multi-source noise characteristics, characterized by: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images