A method for constructing an optical remote sensing image target detection data set under atmospheric interference

Abstract

This invention relates to a method for constructing a target detection dataset for optical remote sensing images under atmospheric interference, comprising: acquiring target optical remote sensing images and cloud optical remote sensing images; preprocessing the target optical remote sensing images to obtain target surface reflectance images; performing semi-automatic annotation on target samples to obtain preliminary annotation results for the samples to be used in the target detection dataset; constructing a cloud sample library; simulating target images under aerosol and water vapor interference to obtain target optical remote sensing images under aerosol and water vapor interference; and simulating target optical remote sensing images under cloud interference to form a target detection dataset for optical remote sensing images under atmospheric interference. This invention conforms to the principles of remote sensing imaging, is closer to actual conditions, and can construct larger and more diverse datasets, effectively solving the problem of a severe shortage of target detection datasets for optical remote sensing images under atmospheric interference. It is of great significance for enhancing the robustness of remote sensing image target detection and recognition algorithms.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image simulation and target detection and recognition technology, and in particular to a method for constructing a target detection dataset for optical remote sensing images under atmospheric interference. Background Technology

[0002] Currently, optical remote sensing technology is developing rapidly, generating massive amounts of optical remote sensing image data daily. Supported by artificial intelligence algorithms, this data can be used to detect, classify, and identify ground targets. Currently available high spatial resolution optical remote sensing image datasets for target identification include: UCAS-AOD, DOTA, DIOR, TGRS-HRRSD, RSOD, and RarePlanes Dataset. These datasets provide high-resolution multispectral remote sensing images of specific targets (such as aircraft, vehicles, ships, buildings, oil tanks, etc.), where the target textures and details are clear and the image quality is good.

[0003] However, optical remote sensing satellites are inevitably affected by atmospheric interference during Earth observation, making it difficult to obtain a large number of high-quality, clear images. Using models trained on high-quality, clear datasets for target detection and recognition in Earth observation images can lead to false positives and false negatives, severely degrading model performance. A reasonable solution is to add some interference data, leveraging the powerful feature mining capabilities and excellent nonlinear fitting abilities of deep neural networks to improve the robustness of the neural network to interference. However, this places demands on the quantity and quality of the dataset. On the one hand, it requires collecting as many types of interference as possible in a substantial amount of data; on the other hand, the data must be as close as possible to real-world complex scenes.

[0004] Currently, collecting large datasets with cloud and fog interference is difficult, and manual target annotation in low-quality images is extremely costly. Therefore, there is currently no large-scale interference-based remote sensing target detection dataset. Constructing remote sensing image interference datasets based on generation methods is a feasible approach. Traditional aerosol and water vapor interference generation methods simulate fog using atmospheric degradation patterns, while cloud interference synthesis simulates noise by randomly generating it or based on texture characteristics. While these methods take into account the complexity of interference in remote sensing images to some extent, they still fall short in terms of realism, lacking physical support and failing to consider the proximity effect in high-resolution optical remote sensing imaging and the inability to add interference to thin cloud regions. Therefore, how to construct a dataset with various types of interference that most closely resembles reality, based on the imaging principles of remote sensing images, is a pressing problem that needs to be solved in the field of artificial intelligence-based remote sensing image target detection and recognition. Summary of the Invention

[0005] To address the issue of false positives and false negatives in target detection of optical remote sensing images using deep learning methods due to the lack of a large dataset of interference targets, which leads to the use of high-quality, clear datasets for training models in real-world interference images, this invention aims to provide a method for constructing a target detection dataset for optical remote sensing images under atmospheric interference. This method should be suitable for diverse interference types, conform to the principles of remote sensing imaging, closely resemble real-world conditions, and be able to construct a larger dataset with more diverse types. This method effectively solves the problem of the severe shortage of target detection datasets for optical remote sensing images under atmospheric interference.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a target detection dataset for optical remote sensing images under atmospheric interference, the method comprising the following sequential steps:

[0007] (1) Acquire target optical remote sensing images and cloud optical remote sensing images: For a remote sensor, select clear sky multispectral remote sensing images containing the target as target optical remote sensing images for the target detection dataset to be constructed, select clear sky multispectral remote sensing images containing clouds over the sea surface as cloud optical remote sensing images, and use cloud optical remote sensing images as source data for cloud sample library.

[0008] (2) Preprocessing the target optical remote sensing image: Perform radiometric correction, atmospheric correction and geometric correction preprocessing on the target optical remote sensing image to obtain the target surface reflectance image;

[0009] (3) Perform semi-automatic annotation on target samples: Select a small number of samples from the target optical remote sensing image obtained in step (1) for manual annotation. Combine the existing target detection dataset and the existing target detection model, and achieve semi-automatic annotation of the samples through multiple iterations to obtain the preliminary annotation results of the target detection dataset to be constructed.

[0010] (4) Constructing a cloud sample library: Combining cloud detection methods and image matting methods, a cloud sample library representing cloud shape and cloud energy is constructed based on the source data of the cloud sample library for cloud interference simulation;

[0011] (5) Simulate target image under aerosol and water vapor interference: Based on atmospheric parameters and target surface reflectance image, simulate atmospheric scattering and absorption through radiative transfer and reverse Monte Carlo simulation to obtain target optical remote sensing image under aerosol and water vapor interference.

[0012] (6) Simulate the target optical remote sensing image under cloud interference: randomly select cloud samples from the cloud sample library, overlay them onto the target optical remote sensing image under aerosol and water vapor interference, and correct the preliminary annotation results of the target detection dataset samples to be constructed, so as to form the target detection dataset of optical remote sensing image under atmospheric interference.

[0013] Step (2) specifically includes the following steps:

[0014] (2a) Perform absolute radiometric correction on the multispectral target optical remote sensing image to obtain the target apparent radiance image. If the absolute radiometric calibration establishes the relationship between gray values ​​and radiance, then perform absolute radiometric correction according to the following formula:

[0015]

[0016] Among them, DN i,j,k L represents the gray value corresponding to pixel (i,j) in band k; i,j,k K represents the apparent radiance corresponding to pixel (i,j) in band k; k The intercept representing the absolute radiometric calibration coefficient of band k is, for single-point calibration, its dark current; B k The slope represents the absolute radiometric calibration coefficient of band k;

[0017] (2b) Atmospheric correction is performed on the apparent radiance image of the target to obtain the surface reflectance image of the target. Atmospheric correction first calculates the apparent reflectance according to the following formula:

[0018]

[0019] Where, ρ i,j,k Es represents the apparent reflectance corresponding to pixel (i,j) in band k; k θ represents the equivalent solar irradiance outside the atmosphere at band k; s d represents the solar zenith angle; d represents the Earth-Sun distance.

[0020] Then, calculate the surface reflectance using the following formula:

[0021]

[0022] in, ρ represents the surface reflectance corresponding to pixel (i,j) in band k; a Indicates atmospheric path reflectance; T g Indicates atmospheric absorption transmittance; θ v Indicates the observed zenith angle; Indicates the solar azimuth angle; T(θ) represents the observed azimuth angle. s T(θ) represents the total downward scattering transmittance of the atmosphere; v ρ represents the total upward scattering transmittance of the atmosphere. a T g 、T(θ s ), T(θ) v All of these were calculated using the radiative transfer model.

[0023] (2c) Use the RPC / RPB file of the target optical remote sensing image to perform geometric orthorectification on the target surface reflectance image.

[0024] Step (3) specifically includes the following steps:

[0025] (3a) Using the target annotation tool labelme, manually annotate a small number of target optical remote sensing images in the target detection dataset to be constructed to obtain the initial shape of the target detection dataset to be constructed, denoted as M1;

[0026] (3b) Input M1 and the existing target detection dataset into the existing target detection models, namely Faster-RCNN, Yolov3 and SSD, to train them and obtain three coarse target detection models.

[0027] (3c) Three coarse target detection models were used to predict the remaining target optical remote sensing images in the target detection dataset to be constructed, and the model with a confidence level greater than 85% was selected as the prediction result.

[0028] (3d) Based on the prediction results, the detection boxes are fused using the Non-Maximum Suppression (NMS) algorithm;

[0029] (3e) Use the fused model prediction results as the manually labeled results to form the sub-shape of the target detection dataset to be constructed, denoted as M2, and repeat step (3b) to iterate training and detection;

[0030] (3f) After multiple iterations, the detection model obtained can detect and label the vast majority of aircraft targets, and obtain the preliminary labeling results of the target detection dataset to be constructed.

[0031] Step (4) specifically includes the following steps:

[0032] (4a) Calculate the apparent reflectance of the cloud multispectral image over the clear sky sea surface;

[0033] (4b) Based on the low reflectivity of the sea surface and the high reflectivity of the cloud layer, the cloud outline is coarsely extracted by the threshold method to obtain cloud pixels, non-cloud pixels and unknown areas.

[0034] When the ratio of the apparent reflectance of a pixel in the near-infrared band to that in the red band is less than T1, the pixel is determined to be a cloudless pixel.

[0035]

[0036] Where, ρ i,j,NIR ρ represents the apparent reflectance corresponding to pixel (i,j) in the near-infrared band; i,j,R T1 represents the apparent reflectance corresponding to the red band pixel (i,j); T1 is the threshold value, which is 0.7.

[0037] When the difference in apparent reflectance between a pixel in the near-infrared band and a cloudless pixel is greater than T2, the pixel is determined to be a cloud pixel.

[0038]

[0039] in, T1 represents the average apparent reflectance of the near-infrared red band without pixels; T2 is the threshold value, which is 0.15.

[0040] When a pixel is neither identified as a cloud pixel nor as a non-cloud pixel, it is considered an unknown region.

[0041] (4c) Cloud samples are obtained by finely extracting the cloud outline using the Image Matting method, as detailed below:

[0042] Cloud pixels in cloud images are used as the foreground region F. i,j Non-cloud pixels serve as background region B i,j At this time, cloud image I i,j Represented as:

[0043] I i,j =a i,j F i,j +[1-a i,j B i,j (6)

[0044] Among them, a i,j Representing cloud energy, for a cloud pixel, a i,j For a cloudless pixel, a is 1. i,j =0;

[0045] At this point, the Image Matting method from computer graphics is used to solve for the unknown region 'a'. i,j That is, solving the Poisson equation:

[0046]

[0047] in, div is a divergence algorithm. F is the Laplace operator; for a pixel p in the unknown region, F p B represents the value of the nearest foreground pixel to p. p Let FB = F, representing the pixel value of the nearest background region to p. p -B p a is obtained by solving the Poisson equation. p If a p >0.95 Update p to the foreground area or a p<0.05 Update p to the background region, and obtain the cloud energy value corresponding to each pixel in the unknown region through multiple iterations.

[0048] Step (5) specifically includes the following steps:

[0049] (5a) Based on the observation geometry and solar illumination geometry corresponding to the multispectral image of the target, simulate atmospheric conditions and calculate the apparent reflectance of the target under the simulated conditions based on the surface reflectance image of the target.

[0050]

[0051] in, ρ represents the surface reflectance corresponding to pixel (i,j) in band k; a Indicates atmospheric path reflectance; T g Indicates atmospheric absorption transmittance; θ s θ represents the solar zenith angle. v Indicates the observed zenith angle; Indicates the solar azimuth angle; T(θ) represents the observed azimuth angle. s T(θ) represents the total downward scattering transmittance of the atmosphere; v ) represents the total upward scattering transmittance of the atmosphere, s is the atmospheric hemispherical albedo, and ρ is the total upward scattering transmittance. a T g 、T(θ s ), T(θ) v Both ) and s were calculated using the radiative transfer model;

[0052] (5b) Calculate visibility based on the 550 nm aerosol optical thickness using the following formula:

[0053]

[0054] Where V0 is visibility; τ 550 The aerosol optical thickness is 550 nm.

[0055] (5c) The visibility is substituted into the inverse Monte Carlo model to simulate photon propagation, and the atmospheric point spread function (PSF) is obtained.

[0056] (5d) Perform Fast Fourier Transform on the simulated target apparent reflectance image and point spread function, multiply the two in the frequency domain, and then perform Fast Inverse Fourier Transform to simulate the effect of proximity effect and obtain the target optical remote sensing image under aerosol and water vapor interference.

[0057] Step (6) specifically includes the following steps:

[0058] (6a) Randomly select cloud samples from the cloud sample library, based on the cloud layer energy a of the cloud samples. i,jIt is then superimposed onto the target optical remote sensing image under aerosol and water vapor interference;

[0059] (6b) The initial annotation results of the target detection dataset to be constructed are corrected, and the annotations of samples that are 80% obscured by clouds are removed to form an optical remote sensing image target detection dataset under atmospheric interference.

[0060] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, based on the coupling process of solar radiation with target objects, background, atmosphere, and clouds, and using target surface reflectance images, the present invention constructs a target detection dataset for interfering optical remote sensing images through numerical simulation and computer simulation. It also considers the influence of proximity effects, resulting in clear physical meaning, conforming to the principles of remote sensing imaging, and being closer to the actual situation. Second, the present invention can achieve interference simulation under different visibility and water vapor conditions. Based on cloud images over the sea surface, it constructs a cloud sample library that can characterize cloud energy, enabling simulation of various cloud interference morphologies. Compared with traditional methods, it can construct a larger and more diverse dataset. Third, the present invention applies semi-automatic manual annotation technology to dataset annotation. It trains the model using a small number of manually annotated samples, and through multiple iterations, achieves the annotation of interference dataset samples. After adding cloud interference, the annotation results are corrected, resulting in high annotation efficiency and more accurate results. Fourth, the present invention can effectively solve the problem of a severe shortage of target detection datasets for optical remote sensing images under atmospheric interference, which is of great significance for enhancing the robustness of remote sensing image target detection and recognition algorithms. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method of the present invention;

[0062] Figure 2 Image of clouds over the sea surface;

[0063] Figure 3 These are the cloudless areas, cloud areas, and unknown areas generated after cloud detection.

[0064] Figure 4 The cloud energy map obtained by solving the Poisson equation;

[0065] Figure 5 This is a photon distribution diagram from the reverse Monte Carlo simulation of this invention;

[0066] Figure 6 This is a flowchart of the semi-automatic annotation process for samples in this invention;

[0067] Figure 7 These are the samples labeled after the first iteration;

[0068] Figure 8 These are the samples labeled after the fourth iteration;

[0069] Figure 9 , 10 11 are target images under aerosol and water vapor interference according to the present invention;

[0070] Figure 12 , 13 Image 14 is a target image under cloud interference according to the present invention;

[0071] Figure 15 This is a simulation result of the target image authenticity evaluation under atmospheric interference, based on the present invention. Detailed Implementation

[0072] like Figure 1 As shown, a method for constructing a target detection dataset from optical remote sensing images under atmospheric interference is described. The method includes the following sequential steps:

[0073] (1) Acquire target optical remote sensing images and cloud optical remote sensing images: For a remote sensor, select clear sky multispectral remote sensing images containing the target as target optical remote sensing images for the target detection dataset to be constructed, select clear sky multispectral remote sensing images containing clouds over the sea surface as cloud optical remote sensing images, and use cloud optical remote sensing images as source data for cloud sample library.

[0074] (2) Preprocessing the target optical remote sensing image: Perform radiometric correction, atmospheric correction and geometric correction preprocessing on the target optical remote sensing image to obtain the target surface reflectance image;

[0075] (3) Perform semi-automatic annotation on target samples: Select a small number of samples from the target optical remote sensing image obtained in step (1) for manual annotation. Combine the existing target detection dataset and the existing target detection model, and achieve semi-automatic annotation of the samples through multiple iterations to obtain the preliminary annotation results of the target detection dataset to be constructed.

[0076] (4) Constructing a cloud sample library: Combining cloud detection methods and image matting methods, a cloud sample library representing cloud shape and cloud energy is constructed based on the source data of the cloud sample library for cloud interference simulation;

[0077] (5) Simulate target image under aerosol and water vapor interference: Based on atmospheric parameters and target surface reflectance image, simulate atmospheric scattering and absorption through radiative transfer and reverse Monte Carlo simulation to obtain target optical remote sensing image under aerosol and water vapor interference.

[0078] (6) Simulate the target optical remote sensing image under cloud interference: randomly select cloud samples from the cloud sample library, overlay them onto the target optical remote sensing image under aerosol and water vapor interference, and correct the preliminary annotation results of the target detection dataset samples to be constructed, so as to form the target detection dataset of optical remote sensing image under atmospheric interference.

[0079] Step (2) specifically includes the following steps:

[0080] (2a) Perform absolute radiometric correction on the multispectral target optical remote sensing image to obtain the target apparent radiance image. If the absolute radiometric calibration establishes the relationship between gray values ​​and radiance, then perform absolute radiometric correction according to the following formula:

[0081]

[0082] Among them, DN i,j,k L represents the gray value corresponding to pixel (i,j) in band k; i,j,k K represents the apparent radiance corresponding to pixel (i,j) in band k; k The intercept representing the absolute radiometric calibration coefficient of band k is, for single-point calibration, its dark current; B k The slope represents the absolute radiometric calibration coefficient of band k;

[0083] (2b) Atmospheric correction is performed on the apparent radiance image of the target to obtain the surface reflectance image of the target. Atmospheric correction first calculates the apparent reflectance according to the following formula:

[0084]

[0085] Where, ρ i,j,k Es represents the apparent reflectance corresponding to pixel (i,j) in band k; k θ represents the equivalent solar irradiance outside the atmosphere at band k; s d represents the solar zenith angle; d represents the Earth-Sun distance.

[0086] Then, calculate the surface reflectance using the following formula:

[0087]

[0088] in, ρ represents the surface reflectance corresponding to pixel (i,j) in band k; a Indicates atmospheric path reflectance; T g Indicates atmospheric absorption transmittance; θ v Indicates the observed zenith angle; Indicates the solar azimuth angle; T(θ) represents the observed azimuth angle. s T(θ) represents the total downward scattering transmittance of the atmosphere; v ρ represents the total upward scattering transmittance of the atmosphere. a T g 、T(θ s ), T(θ) v All of these were calculated using the radiative transfer model.

[0089] (2c) Use the RPC / RPB file of the target optical remote sensing image to perform geometric orthorectification on the target surface reflectance image.

[0090] like Figure 6 As shown, step (3) specifically includes the following steps:

[0091] (3a) Using the target annotation tool labelme, manually annotate a small number of target optical remote sensing images in the target detection dataset to be constructed to obtain the initial shape of the target detection dataset to be constructed, denoted as M1;

[0092] (3b) Input M1 and the existing target detection dataset into the existing target detection models, namely Faster-RCNN, Yolov3 and SSD, to train them and obtain three coarse target detection models.

[0093] (3c) Three coarse target detection models were used to predict the remaining target optical remote sensing images in the target detection dataset to be constructed, and the model with a confidence level greater than 85% was selected as the prediction result.

[0094] (3d) Based on the prediction results, the detection boxes are fused using the Non-Maximum Suppression (NMS) algorithm;

[0095] (3e) Use the fused model prediction results as the manually labeled results to form the sub-shape of the target detection dataset to be constructed, denoted as M2, and repeat step (3b) to iterate training and detection;

[0096] (3f) After multiple iterations, the detection model obtained can detect and label the vast majority of aircraft targets, and obtain the preliminary labeling results of the target detection dataset to be constructed.

[0097] Figure 7 The image shows the labeled samples after the first round of iterations. Figure 8 The image shows the samples labeled after the fourth iteration. As you can see, 46 aircraft targets were identified after the first iteration, and 128 aircraft targets were identified after the fourth iteration. Most aircraft targets have been successfully labeled, and the positions of the labeled boxes are very close to the real boxes.

[0098] Step (4) specifically includes the following steps:

[0099] (4a) such as Figure 2 As shown, the apparent reflectance of the cloud multispectral image over the clear sky above the sea surface is calculated.

[0100] (4b) Based on the low reflectivity of the sea surface and the high reflectivity of clouds, a thresholding method is used to coarsely extract the cloud outline, obtaining cloud pixels, non-cloud pixels, and unknown areas, such as... Figure 3 As shown;

[0101] When the ratio of the apparent reflectance of a pixel in the near-infrared band to that in the red band is less than T1, the pixel is determined to be a cloudless pixel.

[0102]

[0103] Where, ρ i,j,NIR ρ represents the apparent reflectance corresponding to pixel (i,j) in the near-infrared band; i,j,R T1 represents the apparent reflectance corresponding to the red band pixel (i,j); T1 is the threshold value, which is 0.7.

[0104] When the difference in apparent reflectance between a pixel in the near-infrared band and a cloudless pixel is greater than T2, the pixel is determined to be a cloud pixel.

[0105]

[0106] in, T1 represents the average apparent reflectance of the near-infrared red band without pixels; T2 is the threshold value, which is 0.15.

[0107] When a pixel is neither identified as a cloud pixel nor as a non-cloud pixel, it is considered an unknown region.

[0108] (4c) Cloud samples are obtained by finely extracting the cloud outline using the Image Matting method, as detailed below:

[0109] Cloud pixels in cloud images are used as the foreground region F. i,j Non-cloud pixels serve as background region B i,j At this time, cloud image I i,j Represented as:

[0110] I i,j =a i,j F i,j +[1-a i,j B i,j (6)

[0111] Among them, a i,j Representing cloud energy, for a cloud pixel, a i,j For a cloudless pixel, a is 1. i,j =0;

[0112] At this point, the Image Matting method from computer graphics is used to solve for the unknown region 'a'. i,j That is, solving the Poisson equation, such as Figure 4 As shown:

[0113]

[0114] in, div is a divergence algorithm. F is the Laplace operator; for a pixel p in the unknown region, F p B represents the value of the nearest foreground pixel to p. p Let FB = F, representing the pixel value of the nearest background region to p. p -B p a is obtained by solving the Poisson equation. p If a p >0.95 Update p to the foreground area or a p <0.05 Update p to the background region, and obtain the cloud energy value corresponding to each pixel in the unknown region through multiple iterations.

[0115] Step (5) specifically includes the following steps:

[0116] (5a) Based on the observation geometry and solar illumination geometry corresponding to the multispectral image of the target, simulate atmospheric conditions and calculate the apparent reflectance of the target under the simulated conditions based on the surface reflectance image of the target.

[0117]

[0118] in, ρ represents the surface reflectance corresponding to pixel (i,j) in band k; a Indicates atmospheric path reflectance; T g Indicates atmospheric absorption transmittance; θ s θ represents the solar zenith angle. v Indicates the observed zenith angle; Indicates the solar azimuth angle; T(θ) represents the observed azimuth angle. s T(θ) represents the total downward scattering transmittance of the atmosphere; v ) represents the total upward scattering transmittance of the atmosphere, s is the atmospheric hemispherical albedo, and ρ is the total upward scattering transmittance. a T g 、T(θ s ), T(θ) v Both ) and s were calculated using the radiative transfer model;

[0119] (5b) Calculate visibility based on the 550 nm aerosol optical thickness using the following formula:

[0120]

[0121] Where V0 is visibility; τ 550 The aerosol optical thickness is 550 nm.

[0122] (5c) such as Figure 5As shown, visibility is substituted into the inverse Monte Carlo model to simulate photon propagation, and the atmospheric point spread function (PSF) is obtained.

[0123] (5d) Perform Fast Fourier Transform on the simulated target apparent reflectance image and point spread function, multiply the two in the frequency domain, and then perform Fast Inverse Fourier Transform to simulate the effect of proximity effect and obtain the target optical remote sensing image under aerosol and water vapor interference.

[0124] Figure 9 It is an image of surface reflectance. Figure 10 These are simulated images of atmospheric absorption and path radiation. Figure 11 This is the image after the proximity effect simulation. As you can see, the proximity effect still has a fairly obvious impact on the image.

[0125] like Figure 12 , 13 As shown in Figure 14, step (6) specifically includes the following steps:

[0126] (6a) Randomly select cloud samples from the cloud sample library, based on the cloud layer energy a of the cloud samples. i,j It is then superimposed onto the target optical remote sensing image under aerosol and water vapor interference;

[0127] (6b) The initial annotation results of the target detection dataset to be constructed are corrected, and the annotations of samples that are 80% obscured by clouds are removed to form an optical remote sensing image target detection dataset under atmospheric interference.

[0128] Figure 12 , Figure 13 , Figure 14 These are three airports extracted from the object detection dataset.

[0129] like Figure 15 As shown, clear-sky imagery of Dubai was selected and atmospheric correction was performed to obtain a target surface reflectance image as a baseline map; an image of the region under high aerosol conditions was selected as a reference map; aerosol optical depth (AOD) and water vapor content (CWV) under high aerosol conditions in this region were obtained using MODIS products; and a simulated image of the target under atmospheric interference conditions was obtained based on the baseline map. From the baseline map, reference map, and simulated image, four types of land features—vegetation, ground, rooftops, and sand—were selected, and their reflectance was compared. The results are as follows. Figure 15 As shown in (a), (b), (c), and (d), the reflectance of features in the simulated image is closer to that in the reference image, indicating that the simulation method is more realistic.

[0130] In summary, this invention, based on the coupling process of solar radiation with target features, background, atmosphere, and clouds, and using target surface reflectance images, constructs a target detection dataset for interfering optical remote sensing images through numerical simulation and computer simulation. It achieves simulation of interference under different visibility conditions, water vapor conditions, and various cloud formations, while also considering the proximity effect. The physical meaning is clear, consistent with remote sensing imaging principles, and closer to reality. Furthermore, this invention applies semi-automatic manual annotation technology to the dataset annotation. The model is trained on a small number of manually annotated samples, and through multiple iterations, the annotation of samples in the interfering dataset is achieved. After adding cloud interference, the annotation results are corrected, resulting in high annotation efficiency and more accurate results. This invention effectively solves the problem of a severe shortage of target detection datasets for optical remote sensing images under atmospheric interference, and is of great significance for enhancing the robustness of remote sensing image target detection and recognition algorithms.

Claims

1. A method for constructing a target detection dataset from optical remote sensing images under atmospheric interference, characterized in that: The method includes the following steps in sequence: (1) Acquire target optical remote sensing images and cloud optical remote sensing images: For a remote sensor, select clear sky multispectral remote sensing images containing the target as target optical remote sensing images for the target detection dataset to be constructed, select clear sky multispectral remote sensing images containing clouds over the sea surface as cloud optical remote sensing images, and use cloud optical remote sensing images as source data for cloud sample library. (2) Preprocessing the target optical remote sensing image: Perform radiometric correction, atmospheric correction and geometric correction preprocessing on the target optical remote sensing image to obtain the target surface reflectance image; (3) Perform semi-automatic annotation on target samples: Select a small number of samples from the target optical remote sensing image obtained in step (1) for manual annotation. Combine the existing target detection dataset and the existing target detection model, and achieve semi-automatic annotation of the samples through multiple iterations to obtain the preliminary annotation results of the target detection dataset to be constructed. (4) Constructing a cloud sample library: Combining cloud detection methods and image matting methods, a cloud sample library representing cloud shape and cloud energy is constructed based on the source data of the cloud sample library for cloud interference simulation; (5) Simulate target image under aerosol and water vapor interference: Based on atmospheric parameters and target surface reflectance image, simulate atmospheric scattering and absorption through radiative transfer and reverse Monte Carlo simulation to obtain target optical remote sensing image under aerosol and water vapor interference. (6) Simulate the target optical remote sensing image under cloud interference: randomly select cloud samples from the cloud sample library, overlay them onto the target optical remote sensing image under aerosol and water vapor interference, and correct the preliminary annotation results of the target detection dataset samples to be constructed to form the target detection dataset under atmospheric interference. Step (4) specifically includes the following steps: (4a) Calculate the apparent reflectance image for the multispectral image of clouds over the clear sky above the sea surface; (4b) Based on the low reflectivity of the sea surface and the high reflectivity of the cloud layer, the cloud outline is coarsely extracted by the threshold method to obtain cloud pixels, non-cloud pixels and unknown areas. When the ratio of the apparent reflectance of a pixel in the near-infrared band to the red band is less than When this occurs, the pixel is determined to be a cloudless pixel: (4) in, Indicates near-infrared band pixels ( The corresponding apparent reflectance; Indicates red band pixels ( The corresponding apparent reflectance; The threshold value is 0.

7. When the apparent reflectance difference between a pixel in the near-infrared band and a pixel without clouds is greater than When this occurs, the pixel is determined to be a cloud pixel: (5) in, This represents the average apparent reflectance of pixels in the near-infrared red band. The threshold value is 0.

15. When a pixel is neither identified as a cloud pixel nor as a non-cloud pixel, it is considered an unknown region. (4c) Cloud samples are obtained by finely extracting the cloud outline using the Image Matting method, as follows: Cloud pixels in cloud images are used as the foreground region. Non-cloud pixels as background area At this time, cloud images Represented as: (6) in, Representing cloud energy, for cloud pixels, For a cloudless pixel, the value is 1. =0; At this point, the Image Matting method from computer graphics is used to solve for the unknown region. That is, solving the Poisson equation: (7) in, , For divergence algorithm, For the Laplace operator; for a cell in an unknown region , Representative and The pixel value of the nearest foreground region. Representative and The recent background pixel value, let = The solution obtained by solving the Poisson equation is... ;like renew To the foreground area or renew In the background region, the cloud energy value corresponding to each pixel in the unknown region is obtained through multiple iterations.

2. The method for constructing a target detection dataset for optical remote sensing images under atmospheric interference according to claim 1, characterized in that: Step (2) specifically includes the following steps: (2a) Perform absolute radiometric correction on the multispectral target optical remote sensing image to obtain the target apparent radiance image. If the absolute radiometric calibration establishes the relationship between gray value and radiance, then perform absolute radiometric correction according to the following formula: (1) in, Indicates band Pixel ( The corresponding grayscale value; Indicates band Pixel ( The corresponding apparent radiance; Indicates band The intercept of the absolute radiation calibration coefficient, for a single-point calibration, is its dark current; Indicates band The slope of the absolute radiation calibration coefficient; (2b) Atmospheric correction is performed on the apparent radiance image of the target to obtain the surface reflectance image of the target. Atmospheric correction first calculates the apparent reflectance according to the following formula: (2) in, Indicates band Pixel ( The corresponding apparent reflectance; Indicates band Equivalent solar irradiance outside the atmosphere; Indicates the solar zenith angle; Indicates the Earth-Sun distance; Then, calculate the surface reflectance using the following formula: (3) in, Indicates band Pixel ( The corresponding surface reflectance; Indicates atmospheric path reflectivity; Indicates atmospheric absorption transmittance; Indicates the observed zenith angle; Indicates the solar azimuth angle; Indicates the observed azimuth angle; This represents the total downward atmospheric scattering transmittance; This represents the total upward scattering transmittance of the atmosphere. , , , All were calculated using the radiative transfer model; (2c) Use the RPC / RPB file of the target optical remote sensing image to perform geometric orthorectification on the target surface reflectance image.

3. The method for constructing a target detection dataset for optical remote sensing images under atmospheric interference according to claim 1, characterized in that: Step (3) specifically includes the following steps: (3a) Using the target annotation tool labelme, manually annotate a small number of target optical remote sensing images in the target detection dataset to be constructed to obtain the initial shape of the target detection dataset to be constructed, denoted as M1; (3b) Input M1 and the existing object detection dataset into the existing object detection models, namely Faster-RCNN, Yolov3 and SSD, to train them and obtain three coarse object detection models. (3c) Three coarse target detection models were used to predict the remaining target optical remote sensing images in the target detection dataset to be constructed, and the model with a confidence level greater than 85% was selected as the prediction result. (3d) Based on the prediction results, the detection boxes are fused using the Non-Maximum Suppression (NMS) algorithm; (3e) Use the fused model prediction results as the manually labeled results to form the sub-shape of the target detection dataset to be constructed, denoted as M2, and repeat step (3b) to iterate training and detection; (3f) After multiple iterations, the detection model obtained can detect and label the vast majority of aircraft targets, and obtain the preliminary labeling results of the target detection dataset to be constructed.

4. The method for constructing a target detection dataset for optical remote sensing images under atmospheric interference according to claim 1, characterized in that: Step (5) specifically includes the following steps: (5a) Based on the observation geometry and solar illumination geometry corresponding to the multispectral image of the target, simulate atmospheric conditions and calculate the apparent reflectance of the target under the simulated conditions based on the surface reflectance image of the target. (8) in, Indicates band Pixel ( The corresponding surface reflectance; Indicates atmospheric path reflectivity; Indicates atmospheric absorption transmittance; Indicates the solar zenith angle; Indicates the observed zenith angle; Indicates the solar azimuth angle; Indicates the observed azimuth angle; This represents the total downward atmospheric scattering transmittance; This represents the total upward scattering transmittance of the atmosphere. The atmospheric hemispherical albedo. , , , , All were calculated using the radiative transfer model; (5b) Calculate visibility based on the 550 nm aerosol optical thickness according to the following formula: (9) in, Visibility; The aerosol optical thickness is 550 nm. (5c) The visibility is substituted into the inverse Monte Carlo model to simulate photon propagation, and the atmospheric point spread function (PSF) is obtained; (5d) Perform fast Fourier transform on the simulated target apparent reflectance image and point spread function, multiply the two in the frequency domain, and then perform fast inverse Fourier transform to simulate the effect of proximity effect and obtain the target optical remote sensing image under aerosol and water vapor interference.

5. The method for constructing a target detection dataset for optical remote sensing images under atmospheric interference according to claim 1, characterized in that: Step (6) specifically includes the following steps: (6a) Randomly select cloud samples from the cloud sample library, based on the cloud layer energy of the cloud samples. It is then superimposed onto the target optical remote sensing image under aerosol and water vapor interference; (6b) Correct the initial annotation results of the target detection dataset to be constructed, remove the annotations of samples that are 80% obscured by clouds, and form an optical remote sensing image target detection dataset under atmospheric interference.

Images