Satellite remote sensing image in-orbit geometric calibration method based on multi-scale matching

Abstract

The invention discloses a satellite remote sensing image in-orbit geometric calibration method based on multi-scale matching, and relates to the technical field of remote sensing data processing.The method comprises the steps that a strict imaging model is constructed for positive calculation according to the parameters of the age difference, nutation, polar movement and rotation of IAU2000A and in combination with the position of a GNSS, the attitude of a fixed star sensor, the installation angle of a camera, the focal length and the like; obtaining initial geographic coordinates of the coordinates on the map, and forming a spatial buffer area by using four corners of the map sheet and buffer distances; searching in global high-resolution panchromatic reference image tiles L10-L14 stored on a satellite, gradually matching key points from low to high, extracting an elevation by using a 90m resolution DEM, and accumulating and de-weighting to form a control point pair until a threshold value is reached; performing RFM back calculation on the control point pairs by adopting ridge regression, resolving an RPC coefficient, and performing RFM positive calculation according to the RPC coefficient to complete in-orbit geometric calibration; according to the system architecture, data decoding and time marking are achieved through an FPGA in front-end processing, satellite-borne measurement information is synchronously analyzed, and a CPU dispatches an NPU / DCU / GPU to conduct tensor calculation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing data processing, and particularly relates to a satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching. BACKGROUND

[0002] With the rapid development of commercial aerospace and earth observation technology, satellites carrying imaging payloads such as optical cameras can obtain remote sensing images with continuously improved resolution. In the traditional process, L0-level raw images usually need to be returned to the ground, and ground computing power is used to complete radiation calibration, geometric calibration and other preprocessing to generate L1-L3-level products for application. This mode is subject to link bandwidth, ground queuing and processing delay, and is difficult to meet the high timeliness and high throughput on-orbit application requirements. In recent years, COTS devices and on-board high-performance computing technology have matured, and satellites can carry on-board intelligent computing payloads composed of CPUs, FPGAs and GPUs or NPUs, DCUs and other tensor computing-oriented devices. Under the premise of controlled power consumption, this type of payload significantly improves on-orbit computing power, making it feasible to preprocess and analyze massive remote sensing data on orbit, and providing a hardware basis for on-orbit geometric calibration, on-orbit identification and rapid downlink results.

[0003] Remote sensing images have geographic coding properties, and geometric calibration is a key step to convert on-image coordinates to geographic coordinates (longitude, latitude, and elevation), and is a prerequisite for target positioning, change detection and quantitative inversion of products. Existing geometric calibration mainly has two types: one is the rational function model (RFM), which solves the rational polynomial coefficients (RPC) by uniformly laying out and measuring control points (forming on-image coordinates-geodetic coordinates pairs) when complete orbit and sensor parameters are lacking, and then performs forward calculation to obtain the geographic coordinates of the entire image. This method is relatively simple to implement, but has poor universality: the RPC coefficients of each sheet need to be solved independently, rely on a large number of control points and manual intervention, and is time-consuming and labor-intensive, making it difficult to automate batch processing; the other is a strict imaging model, which constructs the collinearity equation between the projection center, image point and object point, and realizes high-precision conversion combined with the rotation relationship between the coordinate systems. This method has high precision, but the algorithm is complex and sensitive to parameters: it needs to process the effects of precession, nutation, polar motion and earth rotation according to the IAU2000A framework to complete the conversion from the inertial coordinate system to the geocentric and fixed coordinate system; it also needs to obtain the satellite position from GNSS broadcast, calculate the attitude rotation matrix from the attitude quaternion of the star sensor, and participate in the solution together with the camera installation angle, focal length and other parameters. Influenced by the earth rotation, orbit drift and platform disturbance, the related rotation matrix and external parameters will deviate, and need to be updated regularly, otherwise the geometric correction accuracy will decrease; at the same time, the calculation amount is large, which is not conducive to on-orbit real-time and long-term autonomous operation.

[0004] Under the background of continuous on-orbit and explosive growth of data volume of high-resolution sensors, it is difficult to meet the requirements of timeliness and stability by relying on ground calibration only; and strict imaging model is faced with the contradiction between parameter time-varying, frequent uploading and limited computing power in long-term operation in orbit. On the other hand, in order for the star-borne image analysis module (such as classification, target recognition and change detection) to directly output the available geographic positioning results, it must be based on reliable on-orbit geometric calibration, otherwise it will lead to target positioning deviation, false alarm or downstream business decision lag. Therefore, an on-orbit geometric calibration method capable of being automatically completed on a star-borne intelligent computing load, with controllable calculation amount and good compatibility for orbit error and sensor internal parameter change is urgently needed. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, which realizes automatic high-precision geometric calibration of on-orbit remote sensing images, and gets rid of ground processing and uploading requirements.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The present application provides a satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, which realizes automatic high-precision geometric calibration of on-orbit remote sensing images, and gets rid of ground processing and uploading requirements.

[0009] Step S1, according to the precession, nutation, polar motion and rotation parameters in the IAU2000A framework, the rotation relationship between the inertial coordinate system and the geocentric and fixed coordinate system is established, and the satellite position is obtained in combination with GNSS broadcast;

[0010] Step S2, the attitude quaternion output by the star sensor is obtained, and the rotation matrix between the satellite body coordinate system and the satellite orbit coordinate system is obtained according to the rigid body rotation principle;

[0011] Step S3, according to the known parameters of camera installation angle and focal length, the rotation matrix between the camera coordinate system and the satellite body coordinate system is obtained;

[0012] Step S4, the rotation matrix obtained from steps S1 to S3 is used to construct the photogrammetry collineation equation, and the geographic coordinate three elements, longitude, latitude and height, of the on-orbit sensing image are obtained by forward calculation of the image coordinates (x, y) on the image;

[0013] Step S5, taking the longitude and latitude of the target sheet four-corner point as the vertex, a spatial buffer polygon is generated according to the buffer distance d of angle unit;

[0014] Step S6, the multi-level reference image tiles intersecting with the spatial buffer zone are retrieved in the storage unit of the star-borne intelligent computing load, and are used for matching from low level to high level step by step;

[0015] Step S7, performing multi-scale key point matching between the in-orbit sensing image and the reference image tile: initializing from the lowest level and gradually increasing the resolution; for the successfully matched key points, combining the DEM image to extract the corresponding elevation, forming and accumulating the control point pair list GCPs; comparing the new control points with the existing control points to remove the duplicates; and terminating the matching when the number of control points reaches the threshold GCP_nums;

[0016] Step S8, performing rational function model RFM back calculation on the GCPs based on ridge regression, and iteratively solving the rational polynomial coefficients RPC;

[0017] Step S9, performing RFM forward calculation using the RPC coefficients to complete the geometric calibration of the in-orbit sensing image and output the standardized image data with geographic positioning.

[0018] As a preferred scheme of the satellite remote sensing image in-orbit geometric calibration method based on multi-scale matching, in the step S6, the reference image tile is a global high-resolution panchromatic 8-bit integer image, which is stored in L10-L14 levels, the spatial resolution of L10 is 152m, and the spatial resolution of L14 is 10m; the tile is indexed and retrieved in a level_row_number_column_number paging manner; when a single tile is retrieved, the tile is cropped based on the buffer area, and when multiple tiles are retrieved, the tiles are respectively cropped and spliced using a mosaic algorithm.

[0019] As a preferred scheme of the satellite remote sensing image in-orbit geometric calibration method based on multi-scale matching, in the step S5, the spatial buffer area of (x1', y1'), (x2', y2'), (x3', y3'), and (x4', y4') is generated based on the buffer distance d and the quadrilateral points (x1, y1), (x2, y2), (x3, y3), and (x4, y4) as the reference for tile retrieval and cropping.

[0020] As a preferred scheme of the satellite remote sensing image in-orbit geometric calibration method based on multi-scale matching, in the step S7, the multi-scale key point matching adopts a scale-invariant feature transform SIFT algorithm or a feature matching network based on deep learning, and is iteratively executed according to the strategy that when k key points are matched, the threshold count GCP_nums is added by k, and the on-image coordinates and geographic coordinates in the GCPs are synchronously updated until len(GCPs) is greater than or equal to GCP_nums.

[0021] As a preferred scheme of the satellite remote sensing image in-orbit geometric calibration method based on multi-scale matching, in the step S7, the newly obtained control points are compared with the recorded control points, and if the comparison is repeated, the control points are not added to the GCPs list.

[0022] As a preferred scheme of the satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, in the step S7, the DEM image spatial resolution is 90m, and the grid elevation corresponding to the longitude and latitude of the control point is extracted as the control point elevation value.

[0023] As a preferred scheme of the satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, in the step S8, the ridge regression is aimed at minimizing the control point residual error, and iteratively solves the robust RPC coefficient.

[0024] As a preferred scheme of the satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, in the step S8, the camera internal parameter and the sensor external parameter are updated according to the residual error in the iteration process, so as to improve the geometric calibration accuracy of the RFM forward calculation in the step I, wherein the camera internal parameter includes the focal length, and the sensor external parameter includes the installation angle.

[0025] As a preferred scheme of the satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, after the geometric calibration, the image is classified, target is recognized and change is detected by an image analysis module, the target recognition output includes the target category, the center image coordinate of the block image, the target height and width, and the probability, and then the RPC coefficient obtained in the step S9 is called to perform the RFM forward calculation, so as to obtain the geographic coordinates of the target, wherein the input of the target recognition includes the block image center coordinate, the height, the width and the target probability.

[0026] As a preferred scheme of the satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, the on-orbit analysis result with the geographic coordinates is transmitted to the ground station in real time or quasi-real time through a satellite-ground communication module, so as to assist decision-making.

[0027] The present application has the beneficial effects that: the present application designs a satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, the method carries out forward calculation according to a strict imaging model, obtains geographical coordinates corresponding to image coordinates, obtains a polygon geographical area of the imaging area, creates a spatial buffer area, matches key points from low to high based on multi-level reference image tiles, the key points meeting the threshold value enter the control point pair sequence after corresponding elevation data is extracted by a digital elevation model (Digital Elevation Model, DEM), and the matching is stopped when the control point pairs meet the quantity requirement. The RFM model is constructed by the ridge regression method, the RPC coefficient is obtained by solving, the on-orbit geometric correction of the remote sensing image is realized, the target recognition algorithm can be carried on the load, the geographical coordinates of the recognition result can be obtained through the RPC coefficient and the image coordinates corresponding to the recognition result, and the recognition result can be transmitted to the ground in real time through the space-ground communication to provide auxiliary decision information. The method realizes automatic and high-precision on-orbit remote sensing image geometric calibration, the method optimizes the precision of the on-orbit geometric calibration while the calculation amount is expanded limitedly, effectively compatible with the correction deviation caused by the changes of the orbit error, the sensor internal parameter and the like, and new parameters are not needed to be uploaded.

[0028] The present application realizes automatic and high-precision on-orbit remote sensing image geometric calibration, optimizes the precision of the on-orbit geometric calibration while the calculation amount is expanded limitedly by using the multi-scale matching method, effectively compatible with the correction deviation caused by the changes of the orbit error, the sensor internal parameter and the like, and new parameters are not needed to be uploaded. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating laborious work.

[0030] Figure 1 It is a flowchart of the satellite remote sensing image on-orbit geometric calibration method of the present application.

[0031] Figure 2 It is a functional module schematic diagram of the on-orbit geometric calibration system of the present application.

[0032] Figure 3 It is a flowchart of the RFM inverse calculation / forward calculation and recognition result geographical positioning processing of the present application. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0034] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be appreciated that the present application maybe practiced in a variety of ways beyond the specifics set forth herein, which can be practiced in any number of manners, and that the present application should be understood to include any such variations as fall within the scope of the teachings and / or patentable aspects. Accordingly, the present application should not be limited to the specific examples described below.

[0035] Second, the term "one embodiment" or "an embodiment" as used herein means that a particular implementation can include, but does not require, each and every feature or combination of features described herein. Thus, the term "one embodiment" or "an embodiment" as used herein is distinguished from phrases such as "in one embodiment" or "in at least one embodiment" that can appear in various places throughout this specification. The phrase "in at least one embodiment" does not necessarily refer to the same embodiment, although it may.

[0036] With the development of commercial space technology, through the carrying of imaging payloads such as optical cameras, earth observation satellites can obtain remote sensing images with higher and higher resolution. However, L0-level remote sensing images need to be pre-processed through radiation calibration and geometric calibration before they can be converted into data that can be used for practical applications. Due to the high complexity of the pre-processing algorithm and the limitation of computing resources, traditional satellite remote sensing applications usually use the original data to return to the ground, and the ground computing power is used to complete the pre-processing of L0-level data to L1-L3-level data and product inversion.

[0037] With the continuous improvement of satellite payload technology and the continuous development of COTS devices, satellites can carry high-performance computing payloads. Such computing payloads are usually composed of CPUs, FPGAs, and GPUs or NPUs suitable for tensor computing. They can be designed to respond to power consumption according to demand. Compared with traditional on-board computers, they effectively improve the computing performance and are suitable for the application scenario of on-orbit pre-processing of massive remote sensing data.

[0038] Different from other types of data and information, remote sensing image has geographic coding, so geometric calibration is one of the key steps in remote sensing image preprocessing. The process of realizing the conversion of coordinates on the remote sensing image map and geographic coordinates is called geometric calibration. There are two methods for current geometric calibration. The first method is rational function model (RFM). In the case of missing satellite orbit parameters and sensor parameters, the user sets control points on the remote sensing image uniformly, and carries out geodetic survey on the control points to form a control point pair of map coordinates-geodetic coordinates. When the number of control point pairs exceeds a certain number, the RFM can be constructed, and the optimal solution of the polynomial is the rational polynomial coefficient (RPC). The geographic coordinates of the whole remote sensing image can be obtained by the RPC coefficient direct calculation. The advantage of RFM method is simple, but the universality of this method is poor. The RPC coefficient of each map needs to be calculated separately, which requires a lot of manual processing and is time-consuming and laborious. The second method is strict imaging model. This method realizes the conversion between map coordinates and geodetic coordinates by constructing the collinear equation among the projection center, the image point and the object point. This method has high precision but high algorithm complexity, and involves many rotation matrix coefficients. For example, the user needs to realize the conversion between inertial coordinate system and geocentric coordinate system. The rotation matrix will change with the errors of earth rotation and orbit offset, so the new rotation matrix parameters need to be uploaded regularly to ensure the accuracy of geometric correction.

[0039] Embodiment 1, refer to Figure 2 As the first embodiment of the present application, the embodiment provides a satellite remote sensing image on-orbit geometric calibration system based on multi-scale matching, which comprises an image acquisition module, a front-end processing module, an image preprocessing module, an image analysis module and a satellite-ground communication module, wherein:

[0040] The image acquisition module realizes the control and imaging of the optical sensor, converts the perceived light signal into a digital image, and provides a data source for subsequent processing and analysis;

[0041] The front-end processing module realizes the conversion processing of image data from a special protocol to a system internal format, and includes error checking, decoding and time marking functions. In addition, the module also supports synchronous analysis of global navigation satellite system broadcast data, star sensor observation and gyroscope attitude information, which is an important input for image geometric calibration;

[0042] The image preprocessing module realizes the data preprocessing functions of radiation calibration and geometric calibration, and outputs standardized image data that is physically consistent and spatially accurate, for subsequent recognition, classification and other algorithm calls;

[0043] The image analysis module implements analysis tasks such as classification, target recognition and change detection, and extracts information of regions of interests (ROIs) according to a task scheduling algorithm; and the satellite-ground communication module is responsible for transmitting the ROIs with geometric positioning to the ground.

[0044] The front-end processing module is implemented by an FPGA of the on-board intelligent computing payload;

[0045] The image preprocessing module and the image analysis module are implemented by scheduling NPU, DCU or GPU and the like acceleration computing units of the on-board intelligent computing CPU, and complete tensor calculation.

[0046] Embodiment 2, with reference to Figure 1 and Figure 3 The present application provides a satellite remote sensing image on-orbit geometric calibration method based on multi-scale matching, which realizes automatic and high-precision on-orbit remote sensing image geometric calibration, and optimizes the precision of on-orbit geometric calibration while expanding the calculation amount within a limited range, effectively compatible with correction deviation caused by changes in orbit error and sensor parameters, without the need to upload new parameters, including:

[0047] Step S1, according to the precession, nutation, polar motion and rotation parameters in the IAU2000A framework, the rotation matrix between the inertial coordinate system and the geocentric and fixed coordinate system is solved, and the satellite position information obtained from GNSS broadcast is acquired;

[0048] Step S2, the satellite attitude quaternion obtained by the star sensor is acquired, and the rotation matrix between the satellite body coordinate system and the satellite orbit coordinate system is solved by using the rigid body rotation principle;

[0049] Step S3, the rotation matrix between the camera coordinate system and the satellite body coordinate system is solved by using the known parameters such as camera installation angle and focal length;

[0050] Step S4, according to the rotation matrices obtained in steps S1, 2 and 3, the photogrammetry collinearity equation is constructed, and the point (longitude, latitude, height) in the geocentric and fixed coordinate system corresponding to the point (x, y) in the image coordinate system is solved;

[0051] Step S5, according to the sheet four-corner point longitude and latitude coordinates [(x1, y1), (x2, y2), (x3, y3), (x4, y4)] obtained in step S4, the parameter buffer distance d is set, the unit is degree, the buffer coordinates of the four-corner points are obtained, and the space buffer [(x1+d, y1+d), (x2+d, y2+d), (x3+d, y3+d), (x4+d, y4+d)] is formed, which is simply denoted as [(x ′ 1, y′ 1),(x ′ 2,y ′ 2),(x ′ 3,y ′ 3),(x ′ 4,y ′ 4)];

[0052] Step S6, high-resolution image tile retrieval, the storage unit is deployed on the spaceborne intelligent computing payload, the global high-resolution panchromatic 8-bit integer image is stored on the storage unit, the image data is stored in the form of tiles, which is composed of L10 level to L14 level, wherein the spatial resolution of L10 level is 152m, and the spatial resolution of L14 level is 10m, the image tile is stored in the form of its level_row_number_column_number (for example: L10_R00001_C00001) in the form of paging, which is convenient for quick retrieval;

[0053] The spatial buffer [(x ′ 1,y ′ 1),(x ′ 2,y ′ 2),(x ′ 3,y ′ 3),(x ′ 4,y ′ 4)] obtained in step S5 and the image level are input into the tile retrieval algorithm as input parameters, and the index dictionary Tiles with the image level, row number and column number is returned, for example, Tiles={“L10”:[R0000001_C00001,...],“L11”:[R0000001_C00001,...],....};

[0054] If there is only one tile in the specified level tile list, the buffer obtained in step S5 is cropped;

[0055] If there are multiple tiles in the specified level tile list, the tiles are cropped, and finally the tiles are spliced by using the mosaic algorithm;

[0056] Step S7, key point matching based on multiple scales, the output data obtained by the front-end processing module and the multi-level tile data obtained in step S6 are used as inputs, a key point matching algorithm is used to identify the key points level by level, the number of control point pairs is set as GCP_nums, and the control point pairs are a list

[0057] GCPs

[0058] = [(x0,y0,longtitude0,latitude0,height0),(

[0059] (x1, y1, longtitude1, latitude1, height1),...,

[0060] (x n ,y n ,longtitude n ,latitude n ,height n )],

[0061] The algorithm of multi-scale key point matching is initialized from the lowest level tile data. If k key points are successfully matched, the control point pair GCP_nums+=k, the x, y, longitude and latitude in the GCPs list are updated, the 90m resolution DEM image is saved on the storage unit, the longitude and latitude are input into the algorithm of extracting grid point data to obtain the corresponding height data, which is updated to the height in the GCPs list. When the number of GCPs is greater than GCP_nums, the matching is terminated. Otherwise, return to step S6 to match the tile with higher resolution. In the matching process, the new control pair needs to be compared with the known control point pair. If there is a repetition, the control point pair sequence cannot be added. Until len(GCPs)>=GCP_nums.

[0062] The image key point matching algorithm can use the traditional scale invariant theory algorithm (SIFT), or use deep learning method.

[0063] Step S8, RFM model inverse calculation:

[0064] Based on the ridge regression algorithm, the control point pair list obtained in step S7 is input into the ridge regression algorithm to iteratively solve the RPC coefficient; the camera intrinsic parameter and sensor extrinsic parameter are updated for the next step.

[0065] Step S9, RFM model forward calculation:

[0066] Based on the RPC coefficient obtained in step S8, the RFM model forward calculation is completed, thereby realizing the on-orbit geometric calibration of remote sensing image.

[0067] Step S10, the image analysis module is equipped with image classification, target recognition and other algorithms. Taking target recognition as an example, the data obtained by the front-end processing module is read by data block, the category of the target is obtained by data analysis algorithm, and the target obtained by reasoning such as the center of the block image on the image coordinate, the height and width of the block image on the image coordinate and the probability is taken as input. Return to step S9 for RFM model forward calculation to obtain the geographic position of the target.

[0068] In step S11, the satellite-ground communication module is equipped with satellite-ground communication hardware devices, ground stations, etc., and the in-orbit analysis result with geographic coordinates obtained in step S10 can be transmitted to the ground through the communication module to provide real-time, efficient and accurate remote sensing information for personnel.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A multi-scale matching based on-orbit geometric calibration method for satellite remote sensing images, characterized in that, Comprise: Step S1, according to the IAU2000A framework in the precession, nutation, polar motion and rotation parameters, the establishment of inertial coordinate system and the rotation relationship of the earth fixed coordinate system, and combined with GNSS broadcast satellite position; Step S2, get the attitude quaternion output by the star sensor, and obtain the rotation matrix of the satellite body coordinate system and the satellite orbit coordinate system according to the rigid body rotation principle; Step S3, according to the known parameters of camera installation angle and focal length, the rotation matrix of camera coordinate system and satellite body coordinate system is obtained; Step S4, the rotation matrix obtained from step S1 to step S3 is used to construct the photogrammetry collinear equation, and the geographical coordinate three elements, longitude, latitude and height, are obtained by forward calculation of the on-orbit sensing image graph coordinate (x, y); Step S5, taking the target sheet four corner points longitude and latitude as the vertex, a spatial buffer polygon is generated according to the buffer distance d of angle unit; Step S6, retrieve the multi-level reference image tile intersecting with the spatial buffer in the storage unit of the on-board intelligent computing load, and use it from low level to high level for matching; Step S7, perform multi-scale key point matching between the on-orbit sensing image and the reference image tile: initialize from the lowest level and improve the resolution level by level; For the successfully matched key points, combine the corresponding elevation extracted from the DEM image to form and accumulate the control point pair list GCPs; compare the new control points with the existing control points to remove the repeated ones; when the number of control points reaches the threshold GCP_nums, terminate the matching; Step S8, based on the ridge regression, the GCPs are inversely calculated by the rational function model RFM, and the rational polynomial coefficient RPC is solved iteratively; Step S9, using the RPC coefficient to perform RFM forward calculation, complete the geometric calibration of the on-orbit sensing image and output the standardized image data with geographical positioning.

2. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, The reference image tile in step S6 is global high-resolution panchromatic 8-bit integer image, which is stored according to L10~L14 levels, the spatial resolution of L10 is 152m, and the spatial resolution of L14 is 10m; the tile is indexed and retrieved in the paging mode of level_row_number_column_number; when a single tile is retrieved, it is cropped based on the buffer, and when multiple tiles are retrieved, they are spliced after being cropped respectively.

3. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, In step S5, taking the sheet four corner points (x1, y1), (x2, y2), (x3, y3), (x4, y4) as the reference, the spatial buffer of (x1', y1'), (x2', y2'), (x3', y3'), (x4', y4') is generated according to the buffer distance d for tile retrieval and cropping.

4. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, In step S7, the multi-scale key point matching adopts scale invariant feature transform SIFT algorithm or feature matching network based on deep learning, and the threshold count GCP_nums is accumulated by k when k key points are matched, and the graph coordinate and geographical coordinate in GCPs are updated synchronously, until len(GCPs) is greater than or equal to GCP_nums.

5. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, In step S7, compare the newly obtained control points with the recorded control points, and if they are repeated, do not add them to the GCPs list.

6. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, The DEM image used in step S7 has a spatial resolution of 90 m, and the grid elevation corresponding to the longitude and latitude of the control points is extracted as the control point elevation value.

7. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, The ridge regression in step S8 aims to minimize the residual error of the control points, and iteratively solves the robust RPC coefficients.

8. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, During the iteration process in step S8, the camera intrinsic parameters including focal length and the sensor extrinsic parameters including installation angle are updated according to the residual error, so as to improve the geometric calibration accuracy of the RFM forward calculation in step I.

9. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 1, characterized in that, The images after geometric calibration are classified, target recognized and change detected by the image analysis module, and the target recognition output includes target category, center image coordinate of the block image, target height and width, and probability, and then the RPC coefficients obtained in step S9 are called to perform RFM forward calculation to obtain the geographic coordinates of the target, wherein the input of the target recognition includes the center coordinate of the block image, height, width and target probability.

10. The multi-scale matching based geometric calibration method for satellite remote sensing images on-orbit according to claim 9, characterized in that, The on-orbit analysis results with geographic coordinates are transmitted to the ground station in real time or quasi-real time through the satellite-ground communication module for decision assistance.

Images