Primary optical remote sensing image target detection simulation data set construction method and system

Abstract

The application discloses a first-level optical remote sensing image target detection simulation data set construction method and system, and the method comprises the following steps: S1, according to a preset application requirement, screening images and labels from an existing optical remote sensing target detection data set to construct a new optical remote sensing image target detection data set; S2, performing radiation degradation on the new optical remote sensing image target detection data set to obtain a radiation-degraded image data set; S3, performing geometric degradation on the radiation-degraded image data set to add geometric distortion of first-level remote sensing images to the data set to obtain a geometric-degraded image data set; S4, performing additive degradation on the geometric-degraded image data set to superimpose interference signals to obtain a degraded data set; and S5, re-labeling the degraded data set to obtain the first-level optical remote sensing image target detection simulation data set.

Description

Method and System for Constructing Simulation Datasets for Target Detection in Level 1 Optical Remote Sensing Images Technical Field

[0001] This invention relates to a method and system for constructing a simulation dataset for target detection in first-level optical remote sensing images, belonging to the field of target detection in remote sensing images. Background Technology

[0002] With the rapid development of sensor technology, the acquisition rate and data volume of remote sensing sensors have experienced explosive growth. Satellites, as edge devices, have limited computing power, storage space, payload power consumption, and transmission bandwidth, easily leading to the predicament of "unable to store, transmit, or process" high-throughput imagery data in orbit, severely restricting the effective utilization of remote sensing data. Faced with these challenges, on-orbit satellite computing has become an effective solution to the explosive growth of remote sensing data and satellite resource bottlenecks. By processing acquired data in real time on-board, the communication burden between the satellite and ground stations can be significantly reduced, greatly improving the satellite's response speed and data processing efficiency, and enhancing the satellite's intelligence level.

[0003] Deep learning-based on-orbit remote sensing target detection technology is a crucial component of satellite on-orbit computing and a vital tool for interpreting remote sensing images. Its primary task is to automatically identify and locate specific target objects, such as buildings, vehicles, ships, and vegetation, from large-scale remote sensing imagery. On-orbit target detection allows for preliminary screening of detection results on-board, selectively transmitting data to ground stations based on target priority and importance. This significantly reduces the communication burden between the satellite and ground stations, improving data transmission efficiency. Furthermore, on-orbit target detection enables real-time processing of satellite-acquired remote sensing images and automatic target detection without waiting for ground intervention, greatly enhancing satellite response speed and data processing efficiency, and providing decision support for ground command, rescue, and monitoring missions.

[0004] Most current mainstream on-orbit object detection methods are based on deep learning technology and are data-driven, making the construction of effective, accurate, and scenario-appropriate datasets crucial. Current optical remote sensing object detection datasets are often produced based on advanced remote sensing products or aerial imagery, such as DOTA, HRSC2016, NWPU VHR-10, UCAS-AOD, RSOD, and DIOR. DOTA (Dataset for Object Detection in Aerial Images) is one of the most representative and influential large-scale datasets in the field of remote sensing object detection. It contains 15 categories, is specifically designed for multi-class object detection tasks in high-resolution remote sensing images, and supports rotated bounding box detection and horizontal bounding box detection. It is widely used for algorithm evaluation and model training.

[0005] The production of the dataset first requires the acquisition of remote sensing imagery. Data sources can include high-resolution satellite imagery, drone-captured images, or public remote sensing data platforms. After acquisition, the images need to be screened and cleaned to remove low-quality, invalid, or redundant data, ensuring the accuracy and efficiency of subsequent annotation work. Next, professional annotation tools are typically used to manually annotate targets in the remote sensing images, such as aircraft, ships, vehicles, and buildings. The annotation format can be selected according to the target detection algorithm used, such as YOLO's TXT format, VOC's XML format, or COCO's JSON format, ensuring that each target includes category information and accurate location coordinates. After annotation is completed, data augmentation can be performed according to the characteristics of the remote sensing images, such as geometric transformations (rotation, scaling, cropping), spectral transformations (brightness and contrast adjustment), and noise addition, to improve the model's generalization ability. Finally, the data is divided into training, validation, and test sets, with the proportions set according to task requirements, and the distribution of targets across categories should be kept as balanced as possible. The annotation results should also be sampled and visually checked to ensure data quality meets requirements, thus providing a solid data foundation for the subsequent training and evaluation of the remote sensing target detection model.

[0006] Most optical remote sensing satellites use cameras that do not yet support on-orbit color imaging, and current onboard computing power cannot support the large-scale on-orbit production of advanced products. Therefore, there is a lack of advanced data available in orbit, limiting on-orbit target detection tasks to low-level remote sensing images that lack processing capabilities. Furthermore, low-level remote sensing datasets have not yet been developed, leading to a weak data foundation for the development of on-orbit target detection algorithms: existing target detection datasets are all based on high-level remote sensing or aerial imagery, which are highly processed and lack the characteristics of raw remote sensing data. If models are developed and trained using existing datasets, they will perform poorly when detecting low-level remote sensing images on satellite. Therefore, the development of on-orbit target detection models needs to be combined with real-world scenarios and trained and evaluated using datasets based on low-level remote sensing data to reflect the model's true performance and achieve good inference performance on satellites.

[0007] Existing remote sensing target detection datasets are often built based on advanced remote sensing products or aerial imagery. However, in on-orbit target detection scenarios, due to limitations in satellite resources, advanced remote sensing products cannot be generated on-orbit, and only low-level remote sensing products with lower processing capabilities can be used for detection. If ground-based detection work is to be migrated to the satellite, datasets based on low-level remote sensing products are essential and form the data foundation for developing on-orbit target detection models. Summary of the Invention

[0008] The main objective of this invention is to propose a method for constructing a simulation dataset for target detection in first-level optical remote sensing images. This method addresses the urgent need for training data in satellite on-orbit remote sensing target detection by constructing a simulation dataset for target detection in first-level optical remote sensing images, thereby providing reliable data support and knowledge transfer pathways for the development of data-driven on-orbit detection models.

[0009] To achieve the above objectives, the present invention proposes the following technical solution:

[0010] A method for constructing a simulation dataset for target detection in first-level optical remote sensing images includes the following steps: S1, selecting images and labels from existing optical remote sensing target detection datasets according to preset application requirements to construct a new optical remote sensing image target detection dataset; S2, performing radiometric degradation on the new optical remote sensing image target detection dataset to obtain a radiometrically degraded image dataset; S3, performing geometric degradation on the radiometrically degraded image dataset to obtain a geometrically degraded image dataset; S4, performing additive degradation on the geometrically degraded image dataset to superimpose interference signals to obtain a degraded dataset; S5, re-labeling the degraded dataset to obtain the simulation dataset for target detection in first-level optical remote sensing images.

[0011] Furthermore, the radiation degradation mentioned in step S2 specifically includes: in the simulation, the first radiance received by the sensor at the top of the atmosphere is regarded as the image pixel value after radiation degradation, and the surface reflectance is regarded as the image pixel value to be radiation-degraded. Based on the conversion relationship between the first radiance and the surface reflectance, a radiation-degraded image is generated.

[0012] Furthermore, the conversion relationship between the first radiance and the surface reflectance is as follows: ;

[0013] Among them, L m L0 represents the first radiance, L0 represents the second radiance of atmospheric path radiation entering the sensor, ρ represents the surface reflectivity, and F represents the second radiance. d T represents the downward radiation transmittance from the sun to the ground, T represents the upward radiation transmittance from the ground to the sensor, and s represents the albedo of the balloon surface.

[0014] Further, in step S2, a 6S radiative transfer model is used for radiative degradation. The 6S radiative transfer model receives preset aerosol optical thickness, aerosol type, ground elevation, solar zenith angle, sensor height, and correction wavelength as input parameters, calculates atmospheric path radiation, spherical albedo, upward radiative transmittance, and downward radiative transmittance, and organizes the calculation results and their corresponding input parameters into a lookup table. When performing radiative degradation, the corresponding calculation results are obtained using the lookup table according to the specific values ​​of the input parameters, and then substituted into the transformation relationship to calculate the image pixel values ​​after radiative degradation.

[0015] Furthermore, the geometric degradation mentioned in step S3 includes at least one of atmospheric refraction degradation based on 2D turbulent field, geometric degradation based on quadratic polynomial, and motion blur degradation; wherein, the atmospheric refraction degradation based on 2D turbulent field is used to restore turbulent disturbances in the imaging process, the geometric degradation based on quadratic polynomial is used to restore image distortion, and the motion blur degradation is used to restore image blur caused by motion in the imaging process.

[0016] Furthermore, the atmospheric refraction degradation based on the 2D turbulent field includes: calculating the standard deviation of pixel displacement of the image; generating a Gaussian white noise field, and performing Gaussian blurring on the Gaussian white noise field to restore the continuity of the turbulent field; then adjusting the blurred noise field to the calculated standard deviation of pixel displacement to obtain the adjusted white noise field, i.e., the 2D turbulent field; and then superimposing the 2D turbulent field onto the image to achieve the atmospheric refraction degradation based on the 2D turbulent field.

[0017] Furthermore, the geometric degradation based on the quadratic polynomial includes: adding geometric distortion to the image using a quadratic polynomial with preset parameters, so that the position of the main subject of the image remains unchanged while the image is slightly distorted globally.

[0018] Furthermore, the motion blur degradation includes: convolving the image with a motion blur kernel, which is a filter used for convolution that spatially simulates the blur effect produced in the image when an object or camera moves along a certain direction; the parameters of the motion blur kernel include length and angle, where the length represents the intensity of the blur and the angle is the angle of the target relative to the horizontal line, that is, the direction of motion.

[0019] Further, the additive degradation in step S4 includes: strip degradation and / or noise degradation; the strip degradation includes: randomly selecting several pixel columns of the image and adjusting the pixel values ​​of the several pixel columns to simulate the difference in pixel response relative to before calibration; the noise degradation includes: adding different types of noise to the image.

[0020] In another aspect, this invention proposes a system for constructing a simulation dataset for target detection in first-level optical remote sensing images, comprising: a dataset acquisition module, used to select images and labels from existing optical remote sensing target detection datasets according to preset application requirements, and construct a new optical remote sensing image target detection dataset; a radiometric degradation module, used to perform radiometric degradation on the new optical remote sensing image target detection dataset to obtain a radiometrically degraded image dataset; a geometric degradation module, used to perform geometric degradation on the radiometrically degraded image dataset to add geometric distortions present in first-level remote sensing images to the dataset, and obtain a geometrically degraded image dataset; an additive degradation module, used to perform additive degradation on the geometrically degraded image dataset to superimpose interference signals, and obtain a degraded dataset; and a dataset processing module, used to re-label the degraded dataset to obtain the simulation dataset for target detection in first-level optical remote sensing images.

[0021] The beneficial effects of this invention are as follows: Based on the forward processing of remote sensing images, this invention simulates the characteristics of real low-level images through multi-level degradation of existing high-level remote sensing product datasets, performing reverse degradation, scientific simulation, and reasonable approximation. This technology can produce first-level remote sensing image simulation data for the development of on-orbit target detection models without the need for actual data collection, cleaning, and annotation, significantly reducing the data production cycle and cost, and providing reliable data support and knowledge transfer pathways for data-driven on-orbit detection model development. The method and system proposed in this invention are original, addressing the lack of data foundation for on-orbit target detection model development by considering the specific characteristics of on-orbit scenarios. This invention constructs datasets based on simulation principles and processes existing data, offering advantages such as low cost, controllable variables, good scalability, and large-scale production capability. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the image degradation process for a target detection dataset based on advanced remote sensing images in the simulation dataset construction method of this embodiment of the invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The embodiments provided are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0024] The production of advanced remote sensing imagery involves data processing steps such as noise suppression, radiometric correction, atmospheric correction, and geometric correction. This invention decomposes the forward process, analyzes the impact of each step on visual perception, selects processes that may affect model detection performance, and designs a reverse degradation algorithm to restore the visual effect of Level 1 remote sensing imagery. This algorithm is then applied to existing optical remote sensing target detection datasets to create a Level 1 optical remote sensing imagery target detection simulation dataset. Referring to Figure 1, based on remote sensing image processing theory, this invention decomposes the advanced remote sensing image degradation process into three steps: geometric degradation, radiometric degradation, and additive degradation.

[0025] The image degradation function designed in this invention is shown in the following formula: ;

[0026] Where: F(x,y) is the degraded image, f(x,y) is the clear image in the existing target detection dataset, G represents geometric degradation, R represents radiometric degradation, and n represents additive degradation. The dataset degradation process is regarded as the superposition of three types of degradation, as shown in Figure 1. The input is a target detection dataset based on advanced remote sensing products (highly processed satellite or aerial images, such as Level 3 remote sensing images, images from Google Maps, drone aerial images, etc.). After three degradation processes, the output is a Level 1 remote sensing image target detection simulation dataset.

[0027] The following section provides a detailed explanation of the method for constructing the simulation dataset based on the degradation process shown in Figure 1.

[0028] This invention proposes a method for constructing a simulation dataset for target detection in first-level optical remote sensing images, comprising the following steps:

[0029] Step 1: Acquire image data:

[0030] Based on specific application requirements (such as target category, resolution requirements, target scene, etc.), images and labels are selected from existing optical remote sensing target detection datasets to construct a new optical remote sensing image target detection dataset, denoted as I. During the selection and construction process, special attention must be paid to sample balance to ensure that targets of each category are reasonably distributed in the dataset, avoiding data skew that could lead to model training biased towards high-frequency categories. Specifically, balancing can be achieved through methods such as category proportion visualization, category frequency statistics, and manual screening of scarce categories. Category proportion visualization refers to visually presenting the number of samples for each target category using bar charts, pie charts, etc., to facilitate the assessment of any imbalances. Category frequency statistics involve real-time counting of the number of targets for each category during the initial data collection and labeling process. Manual screening of scarce categories involves expert manual review or rule-based selection of images of specific areas for categories with low numbers (such as oil tankers, nuclear power plants, etc.). Furthermore, for target categories with low frequency of occurrence, more samples should be generated through targeted screening, supplementary collection, or data augmentation (such as rotation, translation, affine transformation, Mosaic, Copy-Paste, etc.) to improve their representativeness in the dataset. Furthermore, by combining information such as target scale and target location distribution in the image, multi-dimensional sample distribution analysis can be performed to improve the diversity and balance of the dataset from the source, laying a solid foundation for subsequent model training.

[0031] Step 2: Perform radiometric degradation on the image:

[0032] In a specific embodiment of the present invention, the 6S radiative transfer model is used to perform radiative degradation on dataset I obtained in step one, in order to restore the characteristics of illumination differences, atmospheric absorption, atmospheric scattering, and color distortion in the first-level remote sensing image. The 6S radiative transfer model is a classic physical model for atmospheric correction. It simulates the scattering and absorption effects of solar radiation as it passes through the atmosphere and can be used for bidirectional conversion between the top atmospheric radiation received by the satellite and the surface reflectivity.

[0033] First, the color images in dataset I are fused into a single-channel grayscale image using a weighted method. The pixel values ​​of the grayscale image reflect the brightness variations of the ground and can be approximated as the surface reflectance.

[0034] Next, the 6S radiative transfer model is used to receive calculation parameters such as aerosol optical thickness, ground elevation, solar zenith angle, sensor height, aerosol type, and correction wavelength. These parameters are used as input parameters to calculate atmospheric path radiation, spherical albedo, upward radiative transmittance, and downward radiative transmittance. The calculation results and their corresponding input parameters are then organized into a lookup table to facilitate quick retrieval of calculation results in subsequent use without having to re-call the 6S radiative transfer model for calculation. The aforementioned calculation parameters can be manually set, as long as they conform to scientific principles. For example, aerosol optical thickness (dimensionless) can be set to 0.01~1.0, ground elevation (km) can be set to less than 0 (a negative value, the absolute value represents the target height), solar zenith angle to 0°~90° (90° represents the horizon), and sensor height can be set as follows: -1000 for satellite observation, 0 for ground observation, and -100~0 for aircraft observation (absolute value is aircraft altitude). Aerosol type can be set as follows: 0—no aerosols, 1—land model, 2—ocean model, 3—city model, 4—custom model, 5—desert model, 6—biosphere model, 7—stratospheric model. The correction wavelength is the spectral parameter file of the specific sensor input by the user. For parameters with continuous values, the calculation can take a number at intervals of 5 to calculate the corresponding result, such as taking several discrete values ​​for the solar zenith angle, such as 0, 5, 10, 15, etc., and so on. Other calculation parameters can be calculated in the same way. For calculation parameters with discrete values, each value can be calculated once. In this way, there are multiple sets of discrete calculation parameter combinations, and the 6S radiation transfer module can output multiple sets of discrete calculation results.

[0035] When performing radiometric degradation on an image, a lookup table can be used to obtain the corresponding calculation result by accepting a certain input parameter from the image. Specifically, if the value of the user-provided input parameter can be found in the lookup table, the calculation result can be obtained directly; if the value of the user-provided input parameter (such as a solar zenith angle of 13 degrees) is not in the lookup table, the calculation result for a solar zenith angle of 13 degrees can be approximated by interpolation, without actually calling the 6S radiative transfer model. After obtaining the required calculation result, it is substituted into the following formula for inverting the radiance of the upper atmosphere to adjust the image pixel values: ;

[0036] Among them, L m L0 represents the radiance received by the sensor at the top of the atmosphere, which is treated as the degraded image pixel value in the simulation; ρ represents the surface reflectance, which is treated as the image pixel value to be radiated in the simulation; F dLet T represent the downward radiative transmittance from the sun to the ground, T represent the upward radiative transmittance from the ground to the sensor, and s represent the albedo of the balloon surface. Finally, to ensure the consistency of the dataset format, the pixel values ​​of the degraded images need to be limited to between 0 and 255 and converted to 8 bits.

[0037] A specific embodiment of this invention also provides a radiation degradation experiment of the 6S radiative transfer model: Using aerosol optical thickness (AOD) and zenith angle as user-adjustable parameters, with the aerosol optical thickness (AOD) fixed at 0.2, the radiation degradation effect is displayed when the solar zenith angle (SZ) is changed to 10 degrees, 50 degrees, and 80 degrees. With other parameters fixed, the larger the zenith angle (closer to dusk), the darker the image appears, consistent with reality. Furthermore, with the solar zenith angle (SZ) fixed at 10 degrees, the radiation degradation effect is displayed when the aerosol optical thickness (AOD) is changed to 0.1, 0.3, and 0.5. With other parameters fixed, the larger the aerosol optical thickness, the stronger the scattering effect, resulting in a slight decrease in image brightness, consistent with reality.

[0038] Step 3: Perform geometric degradation on the radiometrically degraded image:

[0039] Geometric degradation primarily adds geometric distortion effects to the dataset, reflecting the characteristics of first-level remote sensing imagery. This invention can add at least one of three different distortion methods to the image: atmospheric refraction degradation based on a 2D turbulent field, geometric degradation based on a quadratic polynomial, and motion blur degradation. Specifically, atmospheric refraction degradation based on a 2D turbulent field is used to restore turbulent disturbances during the imaging process; geometric degradation based on a quadratic polynomial is used to restore image distortion caused by factors such as satellite attitude changes, Earth curvature, and terrain undulations; and motion blur degradation is used to restore image blur caused by rapid relative movement of the platform to the target during imaging.

[0040] (1) Atmospheric refraction degradation based on 2D turbulent field:

[0041] This invention employs a 2D turbulence field simulation method. By setting various parameters in the turbulence field, it can simulate turbulence fields of different intensities and morphologies. First, the mean square image displacement in the image is calculated using four parameters. : ;

[0042] Where f is the focal length. Where L is the structural coefficient, L is the distance between the target and the sensor, and D is the aperture size.

[0043] Then The standard deviation of the pixel displacement can be obtained by taking the square root and dividing by the pixel size. As shown in the following formula: ;

[0044] Next, Gaussian white noise fields are generated in both the X and Y directions. This can be achieved using the Python code `np.random.normal(0, 1, shape)`, which generates a two-dimensional matrix of the same size as `shape` that follows a standard normal distribution N(0,1). Here, `shape` represents the predefined shape of the matrix; for example, a 2x3 matrix with `shape` is [2,3]. This characterizes the size of the image, treating it as a matrix where the image's length and width are recorded by pixel values.

[0045] Then, Gaussian blurring is applied to the white noise field to restore the continuity of the turbulent field. After that, the blurred noise field is adjusted to the calculated standard deviation of pixel displacement, as shown in the following formula:

[0046] ;

[0047] in, Here, `std` represents the standard deviation of pixel displacement, `old_noise` represents the white noise field, and `new_noise` represents the adjusted white noise field, i.e., the 2D turbulent field. Finally, this 2D turbulent field is superimposed on the radiometrically degraded image to restore the effect of turbulence on the imaging effect and obtain the geometrically degraded image.

[0048] (2) Geometric degeneration based on quadratic polynomials:

[0049] Geometric correction of remote sensing images based on quadratic polynomials is a commonly used image registration and correction method. It uses a mathematical model for overall fitting without delving into the underlying type and cause of distortion. This invention adopts the same approach, using quadratic polynomials to add geometric distortion to images without geometric distortion. The quadratic polynomial model is shown below:

[0050] ;

[0051] in, , To add geometric distortion to the image pixels based on a quadratic polynomial, x and y are the coordinates of the pixels in the image to be degraded; a0 and b0 are translation terms controlling image translation; a1, b1, a2, and b2 are linear terms controlling image scaling, rotation, and cropping; and a3, b3, a4, b4, a5, and b5 are quadratic terms controlling curve distortion. The desired degradation effect is a slight global distortion while maintaining the position of the main image element. Therefore, the parameters are kept at a0=0, a1=1, a2=0, b0=0, b1=0, and b2=1. This eliminates translation, rotation, and scaling, with distortion contributed by the quadratic terms. The setting of the quadratic terms needs to consider the image size; larger images may require smaller quadratic term coefficients because the larger the coordinate values, the more pronounced the effect of the quadratic terms.

[0052] This invention provides a geometric degradation method based on a quadratic polynomial, adding geometric distortion to the image and a mesh for easier observation. Furthermore, this invention can also superimpose turbulence degradation onto the image with added geometric distortion to simulate atmospheric turbulence.

[0053] (3) Motion blur degradation:

[0054] Motion blur degradation can simulate image blurring caused by the rapid movement of a camera or satellite platform relative to the target during imaging; this is a common type of degradation. Common scenarios include blurring caused by linear motion (such as uniform linear movement), which can be simulated using a linear motion blur kernel. The degradation process is typically expressed as:

[0055] ;

[0056] in, It is a motion blur kernel (PSF: Point Spread Function), whose main parameters include: length: the intensity of the blur (number of pixels); angle: the direction of motion (angle relative to the horizontal line).

[0057] A motion blur kernel is a filter used in convolution that spatially simulates the blurring effect produced in an image when an object or camera moves along a certain direction. For example, a 9x9 horizontal motion blur kernel would look like this:

[0058] ;

[0059] The length controls the size of the convolution kernel; for example, the kernel length in the above example is 9. The angle controls the direction of motion, that is, the direction of blur, usually expressed in degrees or radians; for example, the blur direction in the above example is 0, that is, the horizontal direction.

[0060] Motion blur is essentially the integral effect of an object in an image along a certain direction. During the imaging process, if the camera or object moves along a certain direction, then within the exposure time, each pixel captured by the sensor will be the average brightness of all points along that pixel's motion trajectory. Therefore, mathematically, it can be represented as a convolution operation, where the convolution kernel is the weighted distribution of the motion path.

[0061] Step 4: Perform additive degradation on the geometrically degraded image:

[0062] (1) Strip degeneration:

[0063] Striping occurs when inconsistent detector pixel response values ​​result in regular bright and dark bands in an image. These bands typically extend along the scanning direction, making the image appear as distinct "horizontal" or "vertical" stripes, affecting the visual quality and subsequent analysis. To recreate this effect, we randomly selected a certain number of pixel columns and adjusted their pixel values. Specifically, we used `np.random.randint` to generate random numbers, which were then superimposed onto the original pixel columns to simulate the difference in pixel response compared to before calibration.

[0064] (2) Noise degradation:

[0065] The noise degradation process was simulated by adding different types of noise to interfere with the image, helping to generate remote sensing images under different noise interference conditions. Poisson noise simulates photon noise. First, the image is normalized, then noise is added using a Poisson distribution, and finally the result is scaled back to the range of 0 to 255. Gaussian noise simulates electronic and thermal noise; Gaussian noise is generated by specifying the mean and standard deviation and added to the image. Salt and pepper noise simulates impulse noise. Black dots (pepper) and white dots (salt) are randomly added to the image.

[0066] Step 5: Dataset Preparation

[0067] Because of geometric degradation, the original bounding boxes no longer accurately enclose the targets (due to deformation). Therefore, the dataset needs to be re-labeled based on the target locations in the image. Afterward, the dataset is segmented and quality checked. During the training / validation / test set partitioning, consistency in the distribution of each category is maintained to avoid the complete absence of certain categories in the validation or test sets. This process ultimately forms a complete Level 1 (L1) optical remote sensing image target detection simulation dataset.

[0068] In a specific embodiment of the present invention, the DOTA dataset was degraded using the aforementioned simulation dataset construction method, and experimental verification was performed using YOLOv11. The results are shown in Table 1 below:

[0069] Table 1

[0070] mAP50Map50-95DOTA0.750.586DOTA_degrade0.6950.532 As can be seen from the table, the detection accuracy of the degraded dataset decreased by about 5%, indicating that the first-level remote sensing image target detection simulation dataset produced according to the simulation dataset construction method of the present invention has various degradation effects that increase the difficulty of model detection, and can be used for the development of on-orbit target detection models.

[0071] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program can implement the steps of the method for constructing a simulation dataset for first-level remote sensing image target detection in the foregoing embodiments. Based on this understanding, the construction method of the present invention can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) and includes several instructions to cause a computer device (which may be a personal computer, server, network device, etc.) to execute the steps of the method in various implementation scenarios of the technical solution of the present invention and achieve the purpose of the present invention.

[0072] Another embodiment of the present invention provides a system for constructing a simulation dataset for target detection in first-level remote sensing images, comprising: a dataset acquisition module, used to select images and labels from existing optical remote sensing target detection datasets according to preset application requirements, and construct a new optical remote sensing image target detection dataset; a radiometric degradation module, used to perform radiometric degradation on the new optical remote sensing image target detection dataset to obtain a radiometrically degraded image dataset; a geometric degradation module, used to perform geometric degradation on the radiometrically degraded image dataset to add geometric distortions present in first-level remote sensing images to the dataset, and obtain a geometrically degraded image dataset; an additive degradation module, used to perform additive degradation on the geometrically degraded image dataset to superimpose interference signals, and obtain a degraded dataset; and a dataset processing module, used to re-label the degraded dataset to obtain the simulation dataset for target detection in first-level optical remote sensing images. Based on this understanding, the construction method of the present invention can be embodied in the form of a virtual system, which includes multiple virtual modules corresponding to the steps of the aforementioned construction method, each module being configured to perform functions corresponding to the aforementioned steps.

[0073] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a simulation dataset for target detection in first-level optical remote sensing images, characterized in that, Includes the following steps: S1. Based on the preset application requirements, select images and labels from the existing optical remote sensing target detection dataset to construct a new optical remote sensing image target detection dataset. S2. Perform radiometric degradation on the new optical remote sensing image target detection dataset to obtain a radiometrically degraded image dataset; The radiation degradation specifically includes: in the simulation, the first radiance received by the sensor at the top of the atmosphere is regarded as the image pixel value after radiation degradation, and the surface reflectance is regarded as the image pixel value to be radiation degraded. Based on the conversion relationship between the first radiance and the surface reflectance, a radiation degradation image is generated; S3, geometric degradation is performed on the radiation degradation image dataset to obtain a geometric degradation image dataset; the geometric degradation includes at least one of atmospheric refraction degradation based on 2D turbulent field, geometric degradation based on quadratic polynomial, and motion blur degradation; wherein, the atmospheric refraction degradation based on 2D turbulent field is used to restore the turbulent disturbance in the imaging process, The geometric degradation based on quadratic polynomials is used to restore image distortion, and the motion blur degradation is used to restore image blur caused by motion during imaging; S4, additive degradation is performed on the geometrically degraded image dataset to superimpose interference signals to obtain a degraded dataset; the additive degradation includes: strip degradation and / or noise degradation; the strip degradation includes: randomly selecting several pixel columns of the image and adjusting the pixel values ​​of the several pixel columns to simulate the difference in pixel response relative to before calibration; the noise degradation includes: adding different types of noise to the image; S5, the degraded dataset is re-labeled to obtain the first-level optical remote sensing image target detection simulation dataset.

2. The method for constructing a simulation dataset for target detection in first-level optical remote sensing images as described in claim 1, characterized in that, The conversion relationship between the first radiance and the surface reflectance is as follows: Among them, L m L0 represents the first radiance, L0 represents the second radiance of atmospheric path radiation entering the sensor, ρ represents the surface reflectivity, and F represents the second radiance. d T represents the downward radiation transmittance from the sun to the ground, T represents the upward radiation transmittance from the ground to the sensor, and s represents the albedo of the balloon surface.

3. The method for constructing a simulation dataset for target detection in first-level optical remote sensing images as described in claim 2, characterized in that, In step S2, the 6S radiative transfer model is used for radiative degradation. The 6S radiative transfer model receives preset aerosol optical thickness, aerosol type, ground elevation, solar zenith angle, sensor height and correction wavelength as input parameters, calculates atmospheric path radiation, spherical albedo, upward radiative transmittance and downward radiative transmittance, and organizes the calculation results and their corresponding input parameters into a lookup table. When performing radiometric degradation, the corresponding calculation result is obtained using the lookup table based on the specific value of the input parameter, and then substituted into the transformation relationship to calculate the image pixel value after radiometric degradation.

4. The method for constructing a simulation dataset for target detection in first-level optical remote sensing images as described in claim 1, characterized in that, The atmospheric refraction degradation based on the 2D turbulent field includes: calculating the standard deviation of pixel displacement of the image; generating a Gaussian white noise field and performing Gaussian blurring on the Gaussian white noise field to restore the continuity of the turbulent field; then adjusting the blurred noise field to the calculated standard deviation of pixel displacement to obtain the adjusted white noise field, i.e., the 2D turbulent field; and then superimposing the 2D turbulent field onto the image to achieve the atmospheric refraction degradation based on the 2D turbulent field.

5. The method for constructing a simulation dataset for target detection in first-level optical remote sensing images as described in claim 1, characterized in that, The geometric degradation based on quadratic polynomials includes: adding geometric distortion to the image using a quadratic polynomial with preset parameters, so that the position of the main subject of the image remains unchanged while the image is slightly distorted globally.

6. The method for constructing a simulation dataset for target detection in first-level optical remote sensing images as described in claim 1, characterized in that, The motion blur degradation includes: convolving the image with a motion blur kernel, which is a filter used for convolution that spatially simulates the blur effect produced in the image when an object or camera moves along a certain direction; the parameters of the motion blur kernel include length and angle, where the length represents the intensity of the blur and the angle is the angle of the target relative to the horizontal line, that is, the direction of motion.

7. A system for constructing a simulation dataset for target detection in first-level optical remote sensing images, characterized in that, include: The dataset acquisition module is used to select images and labels from existing optical remote sensing target detection datasets according to preset application requirements, and construct new optical remote sensing image target detection datasets. The radiometric degradation module is used to perform radiometric degradation on the new optical remote sensing image target detection dataset to obtain a radiometric degradation image dataset. The radiation degradation specifically includes: in the simulation, the first radiance received by the sensor at the top of the atmosphere is regarded as the image pixel value after radiation degradation, and the surface reflectance is regarded as the image pixel value to be radiation degraded. Based on the conversion relationship between the first radiance and the surface reflectance, a radiation-degraded image is generated; a geometric degradation module is used to perform geometric degradation on the radiation-degraded image dataset to add the geometric distortion of the first-order remote sensing image to the dataset, thereby obtaining a geometrically degraded image dataset; the geometric degradation includes at least one of atmospheric refraction degradation based on 2D turbulent field, geometric degradation based on quadratic polynomial, and motion blur degradation; wherein, the atmospheric refraction degradation based on 2D turbulent field is used to restore the image during the imaging process. The turbulent disturbance is addressed by the geometric degradation based on a quadratic polynomial to restore image distortion, and the motion blur degradation to restore image blur caused by motion during imaging. An additive degradation module is used to perform additive degradation on the geometrically degraded image dataset to superimpose interference signals and obtain a degraded dataset. The additive degradation includes strip degradation and / or noise degradation. Strip degradation includes randomly selecting several pixel columns of the image and adjusting the pixel values ​​of these columns to simulate different pixel responses relative to their pre-calibration state. Noise degradation includes adding different types of noise to the image. A dataset processing module is used to re-label the degraded dataset to obtain the first-level optical remote sensing image target detection simulation dataset.

Images