A remote-sensing-satellite-oriented on-board image recognition processing method and system

Abstract

The application discloses a kind of satellite-borne image recognition processing method and system for remote sensing satellite.The method collects original image, imaging parameter and on-board resource state, cross-band and cross-scan line image is mapped to unified ground unit, and ground evidence chain is constructed;Shielding deduction is implemented to single evidence dimension and associated evidence dimension combination, evidence fracture spectrum is calculated, fragile ground object unit and fracture evidence dimension are identified;From original image cache, associated image data is extracted, and local compensation view is generated;Candidate identification result is generated through basic identification branch and local review branch;Identification closure gain is calculated, target identification result and on-board computing resource quota are determined;Identification evidence information is output, and identification configuration is updated according to ground verification result.The application can reduce misidentification caused by evidence loss, reduce invalid review calculation, improve the accuracy, stability and resource utilization of satellite-borne image recognition.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and spaceborne intelligent computing technology, specifically to a spaceborne image recognition and processing method and system for remote sensing satellites. Background Technology

[0002] Remote sensing satellites can acquire images of the Earth's surface using optical or multispectral imaging equipment, and perform image preprocessing, target detection, and ground feature classification using onboard processing equipment. Due to limitations in satellite-to-ground transmission bandwidth, onboard computing power, cache capacity, and energy conditions, on-orbit identification typically requires registration, slicing, downsampling, or compression of remote sensing images.

[0003] Existing spaceborne image recognition methods typically input preprocessed images into lightweight recognition models and select target regions based on recognition confidence levels. However, cloud cover, attitude perturbations, cross-band imaging time differences, and reduced resolution can cause loss of spectral, texture, or positional evidence, making it difficult for recognition models to distinguish between the absence of a target and the lack of recognition evidence. Furthermore, performing high-precision verification on the entire image increases the onboard computational load, while selecting local regions solely based on confidence levels makes it difficult to accurately pinpoint the dimensions of evidence requiring compensation.

[0004] Therefore, how to identify the combined evidence of breaks between different bands, scan lines and image scales, and generate local compensation views for the broken evidence, in order to improve the accuracy and stability of remote sensing image recognition under limited on-board resources, is a technical problem that needs to be solved. Summary of the Invention

[0005] This invention provides a satellite image recognition and processing method for remote sensing satellites, comprising: S10. Collect raw images, imaging parameters and on-board resource status from remote sensing satellites, map cross-band and cross-scan line images to a unified ground unit, and construct a chain of ground feature evidence. S20. Perform shielding deduction on the single evidence dimension and the combination of related evidence dimensions in the evidence chain of ground features, calculate the evidence fracture spectrum based on the deviation of the sum of the combined shielding loss and the single shielding loss, and identify vulnerable ground feature units and their fracture evidence dimensions. S30. Extract associated image data of vulnerable land cover units from the original image cache based on the evidence fracture spectrum, and generate a local compensated view corresponding to the fracture evidence dimension. S40. Input the ground feature units and local compensation views into the basic recognition branch and local verification branch of the spaceborne image recognition model to generate corresponding candidate recognition results; S50. Calculate the identification closure gain based on the consistency of categories, position stability and evidence recovery degree before and after local compensation, and determine the target identification result and on-board computing resource quota. S60: Output the target identification result and the corresponding identification evidence information, and update the evidence breakage threshold and local verification branch configuration based on the ground verification result.

[0006] The above-described method for satellite-borne image recognition and processing for remote sensing satellites includes: acquiring raw images, imaging parameters, and onboard resource status from the remote sensing satellite; mapping cross-band and cross-scan line images to a unified ground unit; and constructing a chain of ground feature evidence, including: S101. Based on the imaging time, satellite attitude, and band position, register cross-band and cross-scan line images to a unified ground unit; S102. Establish evidence mapping between ground units and bands, scan lines, image scales, and original buffer locations to generate ground feature evidence chains.

[0007] The above-described satellite-based image recognition and processing method for remote sensing satellites includes the following steps: a shielding deduction is performed on single evidence dimensions and combinations of associated evidence dimensions in the ground feature evidence chain. Based on the deviation between the sum of combined shielding loss and single shielding loss, the evidence fracture spectrum is calculated to identify vulnerable ground feature units and their fracture evidence dimensions. S201. Separately mask single evidence dimensions and combinations of evidence dimensions with related relationships, and extract the changes in identification features under each masking state; S202. Calculate the evidence fracture spectrum based on the deviation between the sum of the combined shielding loss and the corresponding single shielding loss, and determine the land cover units and evidence dimensions that reach the fracture threshold as vulnerable land cover units and fracture evidence dimensions.

[0008] The above-described satellite-based image recognition and processing method for remote sensing satellites includes extracting associated image data of vulnerable ground features from the original image cache based on the evidence fracture spectrum, and generating a local compensated view corresponding to the fracture evidence dimension, comprising: S301. Based on the vulnerable ground features and fracture evidence dimensions, locate the corresponding bands, scan lines, image scales, and original cache locations; S302. Extract the corresponding image data from the original image cache, perform spectral compensation, geometric compensation or scale compensation according to the fracture evidence dimension, and generate a local compensated view.

[0009] The above-described satellite image recognition processing method for remote sensing satellites includes inputting ground feature units and local compensation views into the basic recognition branch and local verification branch of the satellite image recognition model to generate corresponding candidate recognition results, including: S401. Input the ground feature unit into the basic identification branch to generate basic category, basic location and basic evidence response; S402. Input the vulnerable feature unit and local compensation view into the local verification branch that matches the fracture evidence dimension, and generate the verification category, verification location and compensation evidence response.

[0010] The above-described satellite-based image recognition processing method for remote sensing satellites includes calculating the recognition closure gain based on the consistency of categories, positional stability, and evidence recovery degree before and after local compensation, and determining the target recognition result and on-board computing resource quota, including: S501. Compare the category changes, positional shifts, and fracture evidence response changes between the basic identification results and each candidate identification result to form an identification closure state; S502. Calculate the identification closure gain based on the identification closure state and the resource occupation of the local verification branch, select the corresponding candidate identification result as the target identification result, and determine the on-board computing resource quota.

[0011] The above-described satellite-based image recognition processing method for remote sensing satellites includes outputting target recognition results and corresponding recognition evidence information, and updating the evidence breakage threshold and local verification branch configuration based on ground verification results, comprising: S601. Extract the band segments, local responses and imaging parameters that support the target recognition results, and generate recognition evidence information associated with the target recognition results; S602. Correct the evidence breakage threshold based on the ground verification results, and adjust the matching relationship between the broken evidence dimension and the local review branch.

[0012] This invention also provides a spaceborne image recognition and processing system for remote sensing satellites, comprising: The ground feature evidence chain construction module is used to collect raw images, imaging parameters, and on-board resource status to construct the ground feature evidence chain; The evidence break spectrum calculation module is used to perform masking and deduction on single evidence dimensions and combinations of related evidence dimensions, and calculate the evidence break spectrum. The local compensation view generation module is used to extract associated image data based on the evidence fracture spectrum and generate a local compensation view. The candidate recognition result generation module is used to generate candidate recognition results through the basic recognition branch and the local verification branch; The closed-loop gain calculation module is used to calculate the closed-loop gain and determine the target identification result and the on-board computing resource quota. The identification result feedback and update module is used to output the target identification result and identification evidence information, and update the evidence breakage threshold and local verification branch configuration based on the ground verification result.

[0013] The beneficial effects achieved by this invention are as follows: By performing masking inference on single evidence dimensions and combinations of related evidence dimensions, this invention identifies vulnerable ground features and fracture evidence dimensions in remote sensing images. It also generates local compensated views from the original image cache for fracture evidence and determines the target identification results and on-board computing resource quotas by combining the identification of closure gain. This can reduce invalid verification calculations, reduce misidentification caused by the loss of spectral, texture, or location evidence, and improve the accuracy, stability, and resource utilization of remote sensing satellite on-board image recognition. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a satellite image recognition and processing method for remote sensing satellites provided in Embodiment 1 of this application.

[0016] Figure 2 This is a schematic diagram of the structure of a spaceborne image recognition and processing system for remote sensing satellites, provided in Embodiment 2 of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 like Figure 1 As shown, Embodiment 1 of this application provides a satellite image recognition and processing method for remote sensing satellites, including the following steps: S10. Collect raw images, imaging parameters and on-board resource status from remote sensing satellites, map cross-band and cross-scan line images to a unified ground unit, and construct a chain of ground feature evidence. Specifically, the process involves acquiring multi-band raw images, imaging times for each scan line, band positions, satellite attitude, orbital positions, and onboard computing and buffering status. Ground position alignment is performed on cross-band and cross-scan line images, and a mapping relationship is established between ground units and various identification evidence items. This includes the following sub-steps: S101. Based on the imaging time, satellite attitude, and band position, register cross-band and cross-scan line images to a unified ground unit; Specifically, the original scan lines for each band are read according to the imaging time, and the initial spatial offset between bands is determined based on the positional relationship of the detection units of each band in the remote sensing imaging equipment. Combining the satellite orbital position, attitude angle, and imaging line of sight at the corresponding imaging time, each original scan line is projected onto the ground coordinate system.

[0019] The continuous ground region is divided according to a preset ground resolution, generating ground cells with unique ground location identifiers. Based on the ground projection position of each pixel, pixels covering the same area in different bands and different scan lines are assigned to the corresponding ground cells.

[0020] When cross-band images are locally misaligned due to imaging time difference or attitude changes, the ground projection position is corrected based on the stable ground feature edges in adjacent scan rows, and the corrected band image is remapped to the corresponding ground unit. For image data that cannot meet the preset registration accuracy, the corresponding band, scan row, and position deviation are recorded, and they are not directly used in the basic evidence fusion.

[0021] Each ground unit corresponds to one or more band image segments and adjacent scan line image segments, and retains the imaging time, ground coverage area and registration deviation of each image segment.

[0022] S102. Establish evidence mapping between ground units and bands, scan lines, image scales, and original buffer locations to generate ground feature evidence chains. Specifically, for each ground unit, the spectral response of each band image segment, the spatial continuity information of adjacent scan line image segments, and the texture and edge information at different image scales are extracted, and the corresponding band identifier, scan line identifier, and image scale identifier are recorded according to the source of evidence.

[0023] The storage address, data length, and effective retention time of each image fragment in the onboard image cache are associated with the corresponding evidence source, enabling each piece of identification evidence to be located in a specific data area within the original image cache. When cache space is insufficient, original image fragments covering multiple evidence dimensions or with registration deviations exceeding preset values ​​are preferentially retained.

[0024] Evidence mapping is established based on ground units, evidence sources, and original cache locations. Spectral evidence, scanning time-series evidence, scale evidence, and spatial structure evidence corresponding to the same ground unit are connected according to imaging time sequence to generate a ground feature evidence chain.

[0025] The ground feature evidence chain records the location identifier, evidence dimension, evidence source, registration deviation, image quality, original cache location and cache validity time of the ground unit, and outputs the evidence chain of the corresponding ground unit to the subsequent shielding simulation process.

[0026] S20. Perform shielding deduction on the single evidence dimension and the combination of related evidence dimensions in the evidence chain of ground features, calculate the evidence fracture spectrum based on the deviation of the sum of the combined shielding loss and the single shielding loss, and identify vulnerable ground feature units and their fracture evidence dimensions. Specifically, taking the land cover evidence chain generated by S10 as input, single evidence dimensions and combinations of related evidence dimensions are masked respectively. The features, categories, and location changes under each masking state are extracted, and the evidence break spectrum is calculated based on the higher-order interaction effects generated by the combined masking. This includes the following sub-steps: S201. Separately mask single evidence dimensions and combinations of evidence dimensions with related relationships, and extract the changes in identification features under each masking state; Specifically, the band evidence, scan timing evidence, image scale evidence, and spatial structure evidence corresponding to each ground unit are each identified as an independent evidence dimension. Two or more evidence dimensions that are correlated in spectral response, have adjacent scan timings, or jointly cover the same identification area are identified as a combination of related evidence dimensions. The number of evidence dimensions in a combination of related evidence dimensions does not exceed a preset combination limit. When the number of candidate evidence dimensions exceeds the preset combination limit, evidence dimensions are selected from largest to smallest based on the changes in identification features under a single shielding state to form a combination of related evidence dimensions.

[0027] By inputting the complete ground feature evidence chain into the shared feature extraction layer and basic recognition branch of the spaceborne image recognition model, the baseline features, baseline category boundaries, and baseline target locations in the unmasked state are obtained. The baseline category boundary is the difference between the target category response and the second-highest category response.

[0028] When the basic identification branch does not output the target category and target location, the category with the highest category response is determined as the baseline response category, the center of the local response area corresponding to the category is determined as the baseline response location, and the baseline response category and baseline response location are used for subsequent category boundary and location change analysis under the shielding state.

[0029] For a single evidence dimension, its corresponding image data is replaced with non-discriminatory reference data within the same ground unit, while keeping other evidence dimensions unchanged, thus forming a single shielded state. The non-discriminatory reference data is generated based on the local background mean of the corresponding band or scan line to avoid introducing abnormal grayscale or artificial edges during the shielding process.

[0030] For each combination of related evidence dimensions, simultaneous masking is applied to the combination and its constituent evidence subsets, forming a combined masking state and a subset masking state. Each masking state is input into the same shared feature extraction layer and basic recognition branch to extract feature direction offset, category boundary descent, and target position offset relative to the unmasked state, and to form the corresponding masking response vector.

[0031] S202. Calculate the evidence fracture spectrum based on the deviation between the sum of the combined shielding loss and the corresponding single shielding loss, and determine the land cover units and evidence dimensions that reach the fracture threshold as vulnerable land cover units and fracture evidence dimensions.

[0032] Specifically, for ground units The combination of related evidence dimensions G is used to determine the alternation coefficients based on the difference in the number of dimensions between the combination of related evidence dimensions and each subset of evidence. An inclusion-exclusion superposition is then performed on the combined masking response and the masking responses of each subset to eliminate response changes that can be explained by a single evidence dimension and a low-order subset of evidence, thus obtaining the high-order interaction residual corresponding to combination G. .

[0033] Alternative evidence dimensions are selected from the ground feature evidence chain that are not included in combination G, cover the same ground area as combination G, and can provide similar spectral, temporal, or spatial structure information. An alternative evidence response space is established based on the shielding response of the alternative evidence dimensions. The portion of the higher-order interaction residuals that can be explained by alternative evidence is determined through regularized projection. The residuals retained after projection are identified as uncompensable break vectors. .

[0034] A baseline response covariance matrix is ​​established based on the normal identification response changes of adjacent ground units in the unshielded state. The scarcity of alternative evidence is determined based on the number of available alternative evidence dimensions, remaining cache time, and registration bias. .

[0035] Ground Unit The evidence breakage value under the combination of related evidence dimensions G is expressed as: In the formula, Indicates the value of evidence breakage; This represents the uncompensable fracture vector; I represents the reference response covariance matrix; I represents the identity matrix. Represents higher-order interactive residuals; Represents the masked response vector for a single evidence dimension d; Indicates the scarcity of alternative evidence; and These represent the feature, category, and location components in the uncompensable fracture vector, respectively. and The adjustment coefficient is greater than zero; To prevent the matrix from being non-invertible or from having a positive number with a zero denominator.

[0036] An evidence fracture spectrum is established based on ground units, combinations of associated evidence dimensions, higher-order interactive residuals, uncompensated fracture vectors, scarcity of alternative evidence, and evidence fracture values. Ground units whose evidence fracture values ​​reach a preset fracture threshold are identified as vulnerable ground units. Each evidence dimension in the combination of associated evidence dimensions is restored sequentially, and the uncompensated fracture vector is redefined. The evidence dimension with the largest decrease in evidence fracture value after restoration is identified as the fracture evidence dimension. When the decrease values ​​corresponding to multiple evidence dimensions are within a preset range, the corresponding evidence dimensions are collectively identified as the fracture evidence dimension.

[0037] S30. Extract associated image data of vulnerable land cover units from the original image cache based on the evidence fracture spectrum, and generate a local compensated view corresponding to the fracture evidence dimension. Specifically, based on the vulnerable land cover units and their fracture evidence dimensions output by S20, the corresponding image source and original cache location are located in the land cover evidence chain, and a local compensated view is generated for the spectral, location, or scale evidence of the fracture, including the following sub-steps: S301. Based on the vulnerable ground features and fracture evidence dimensions, locate the corresponding bands, scan lines, image scales, and original cache locations; Specifically, the location identifiers of vulnerable ground features, combinations of associated evidence dimensions, uncompensated fracture vectors, and fracture evidence dimensions are read from the evidence fracture spectrum. Based on the ground feature evidence chain, the corresponding band identifiers, scan line identifiers, image scale identifiers, and original cache locations are queried for the fracture evidence dimensions.

[0038] When the fracture evidence dimension is band evidence, locate the corresponding band and the original image fragment with adjacent spectral responses to that band; when the fracture evidence dimension is scan time sequence evidence, locate the scan line where the vulnerable ground feature unit is located and its adjacent scan lines before and after it; when the fracture evidence dimension is image scale or spatial structure evidence, locate the original resolution image of the corresponding ground area and its surrounding continuous area.

[0039] Based on the registration deviation and ground coverage recorded in S10, the range of the located image segments is expanded to cover vulnerable ground features and their adjacent stable ground features. Image segments whose original cache locations are still within the effective retention time and whose image quality meets preset requirements are identified as associated image data.

[0040] When the same vulnerable land cover unit corresponds to multiple fracture evidence dimensions, the correspondence between each fracture evidence dimension and the associated image data is preserved, and the extraction order of the associated image data is determined according to the evidence fracture value.

[0041] S302. Extract the corresponding image data from the original image cache, perform spectral compensation, geometric compensation or scale compensation according to the fracture evidence dimension, and generate a local compensated view.

[0042] Specifically, associated image data is read according to the original cache location determined in S301, and the corresponding compensation method is selected based on the dimension of the fault evidence. The compensation process only processes vulnerable ground features and their extended ranges, without reprocessing the entire remote sensing image.

[0043] For band evidence of breakage, the local response of the broken band is constrained and reconstructed based on the spectral response relationship of adjacent bands in the stable ground area, and the original unbroken band is retained as the spectral reference for the compensation view to generate a spectral compensation view.

[0044] For breaks in scanning timing or location evidence, the local positional deviations of vulnerable ground features between different bands and scanning lines are corrected based on the stable ground feature edges in adjacent scan lines, satellite attitude changes, and imaging time differences, generating a geometrically compensated view.

[0045] For images with broken scale or spatial structure evidence, local regions containing vulnerable land cover units and their surrounding textures are extracted from the original resolution image. The local sampling scale is adjusted according to the target size and edge continuity direction to generate a scale-compensated view.

[0046] When a vulnerable ground feature unit simultaneously possesses multiple fracture evidence dimensions, the associated image data is processed in the order of geometric compensation, spectral compensation, and scale compensation to avoid positional bias affecting subsequent spectral and texture compensation. For each local compensated view, the corresponding ground unit, compensation type, original image source, fracture evidence dimension, and compensation data range are recorded.

[0047] S40. Input the ground feature units and local compensation views into the basic recognition branch and local verification branch of the spaceborne image recognition model to generate corresponding candidate recognition results; Specifically, the ground feature units are input into the basic identification branch to complete the routine identification, and the corresponding local verification branch is selected according to the fault evidence dimension to process the vulnerable ground feature units and their local compensation views, generating candidate identification results with a unified output format, including the following sub-steps: S401. Input the ground feature unit into the basic identification branch to generate basic category, basic location and basic evidence response; Specifically, the various surface units generated by S10 are fed into the shared feature extraction layer of the spaceborne image recognition model according to a preset input scale to extract the corresponding spectral features, texture features and spatial structure features, and the extraction results are input into the basic recognition branch.

[0048] The basic identification branch performs category determination and location regression on shared features, generating the basic category, basic category response, and basic location for each ground unit. For ground units where no target is identified, the peak value of the category response and its location are recorded.

[0049] Based on the position of the inverse response of the basic category response in the shared features, the bands, scanning areas, and scale levels that contribute to the basic identification results and meet the preset requirements are extracted to generate basic evidence responses. The basic evidence responses record the correlation between the identification results and the corresponding evidence dimensions in the ground feature evidence chain.

[0050] For ground units not identified as vulnerable ground features by S20, the basic identification results are used as the identification results to be output; for vulnerable ground features, the basic category, basic location, and basic evidence response are sent to the corresponding local verification branch.

[0051] S402. Input the vulnerable feature unit and local compensation view into the local verification branch that matches the fracture evidence dimension, and generate the verification category, verification location and compensation evidence response.

[0052] Specifically, based on the fracture evidence dimension of vulnerable ground features, a spectral verification branch, a geometric verification branch, or a scale verification branch is selected from preset local verification branches. Each local verification branch shares the basic feature extraction parameters with the basic identification branch, and only loads the local feature enhancement layer corresponding to the evidence dimension.

[0053] When the fracture evidence dimension is band evidence, vulnerable ground features and spectral compensation views are input into the spectral verification branch to enhance the difference response between different bands; when the fracture evidence dimension is scan time sequence or location evidence, vulnerable ground features and geometric compensation views are input into the geometric verification branch to correct the target response position; when the fracture evidence dimension is image scale or spatial structure evidence, vulnerable ground features and scale compensation views are input into the scale verification branch to restore the target edge and local texture response.

[0054] The local verification branch outputs the verification category, verification category response, and verification location, and generates a compensation evidence response based on the newly added valid responses in the local compensation view. The compensation evidence response records the changes in the contribution of each evidence dimension to the recognition result before and after local compensation.

[0055] When a vulnerable feature unit corresponds to multiple fracture evidence dimensions, multiple candidate identification results are generated through the corresponding local verification branches, and the verification category, verification location, compensation evidence response, verification branch and resource occupation information of each candidate identification result are recorded uniformly.

[0056] S50. Calculate the identification closure gain based on the consistency of categories, position stability and evidence recovery degree before and after local compensation, and determine the target identification result and on-board computing resource quota. Specifically, using the basic identification results and candidate identification results generated by S40 as input, the changes in category response, target position, and fracture evidence before and after local compensation are compared to form the identification closure state of each local verification branch. The identification closure gain is then calculated in conjunction with onboard resource usage, including the following sub-steps: S501. Compare the category changes, positional shifts, and fracture evidence response changes between the basic identification results and each candidate identification result to form an identification closure state; Specifically, for vulnerable land cover units, the basic category response is compared with the verification category response corresponding to each local compensation view. The basic identification results, vulnerable land cover unit identification results, and candidate identification results of the corresponding local compensation views are mapped to a unified category response space. The category closure is determined based on the consistency of the dominant categories among the candidate identification results and their convergence with the basic category response. The more consistent the dominant categories of the candidate identification results and the smaller the differences in category responses, the higher the category closure.

[0057] The base position and each verification position are mapped to a unified ground coordinate system. The position closure is determined based on the degree of overlap between the verification positions and their deviation from the geometric compensation target position. The more concentrated the verification positions are and the smaller their deviation from the geometric compensation target position, the higher the position closure.

[0058] The compensated evidence response is matched with the uncompensated fracture vector generated in S20, and the degree of evidence recovery is determined based on the attenuation ratio of the uncompensated fracture vector after the addition of the local compensation view. The more consistent the direction of the compensated evidence response with the fracture evidence dimension, and the less anomalous responses are generated in other evidence dimensions, the higher the degree of evidence recovery.

[0059] The category closure, location closure, and evidence recovery are normalized to 0 to 1, and the computation time, cache space, and energy consumed by the corresponding local verification branch are recorded to form the identification closure state.

[0060] S502. Calculate the identification closure gain based on the identification closure state and the resource occupation of the local verification branch, select the corresponding candidate identification result as the target identification result, and determine the on-board computing resource quota.

[0061] Specifically, targeting vulnerable ground features The k-th local verification branch is used to calculate and identify the closure gain based on category closure, location closure, evidence recovery, and on-board resource usage. In the formula, Represents vulnerable ground features The identification closing gain after processing by the kth local verification branch; Indicates category closure; Indicates positional closure; Indicates the degree of evidence recovery; and These represent the weights of category, location, and evidence recovery, respectively, and the sum of the three is 1; and These represent the computation time, cache space, and energy consumed by the local verification branch, respectively. and These represent the corresponding remaining available resources; This represents the closed-loop equilibrium adjustment coefficient; and This represents the resource utilization adjustment coefficient; This indicates a positive number to prevent the denominator from being zero. The weights and adjustment coefficients are preset parameters greater than zero, determined based on the recognition results of historical remote sensing images or the simulation results of preset on-orbit processing scenarios.

[0062] Arrange the local verification branches in descending order of recognition closure gain. The candidate recognition result corresponding to the local verification branch with the highest recognition closure gain that reaches the preset closure threshold is determined as the target recognition result. If the recognition closure gain of any local verification branch does not reach the preset closure threshold, the candidate recognition result with the highest recognition closure gain is retained and marked as pending verification.

[0063] For vulnerable ground features, the basic resource quota is determined based on the actual resource usage of the corresponding local verification branch according to the target identification results, and the subsequent processing resources are adjusted in conjunction with its evidence break value; the larger the evidence break value and the higher the identification closure gain, the higher the on-board computing resource quota obtained by the corresponding ground feature unit. For non-vulnerable ground feature units, processing is performed according to the preset resource quota of the basic identification branch. When the remaining resources are insufficient, local verification branches with higher identification closure gain are given priority.

[0064] S60, output the target recognition result and the corresponding recognition evidence information, and update the evidence breakage threshold and local verification branch configuration according to the ground verification result; Specifically, the target identification result determined by S50 is correlated with the local image, evidence response, and imaging parameters supporting the result to generate identification evidence information. The subsequent spaceborne image recognition configuration is then corrected based on the verification results returned from the ground station, including the following sub-steps: S601. Extract the band segments, local responses and imaging parameters that support the target recognition results, and generate recognition evidence information associated with the target recognition results; Specifically, based on the ground units and local verification branches corresponding to the target identification results, band image segments, scan line image segments, and local compensation views that effectively contribute to the target category, target location, and evidence recovery are extracted from the ground feature evidence chain.

[0065] Based on the compensation evidence response, select local response regions that meet the preset requirements, and record the corresponding evidence dimension, response intensity, and compensation type. Map the local response regions to the original image cache, and extract the minimum image range that can cover the effective response region.

[0066] The imaging time, satellite attitude, orbital position, band identifier, ground position, registration deviation, and image scale of the corresponding image segment are read and associated with the target category, target position, recognition closure gain, and local verification branch to generate recognition evidence information.

[0067] The target identification results and identification evidence information are written to the onboard identification results storage area, and the output order is set according to the target mission level, identification closure gain, and verification identifier. Only local data that can support the target identification results are retained in the identification evidence information, and the complete original remote sensing image is not saved repeatedly.

[0068] S602. Correct the evidence breakage threshold based on the ground verification results, and adjust the matching relationship between the broken evidence dimension and the local review branch.

[0069] Specifically, the ground terminal verifies the results based on the target identification results and identification evidence information, and returns the target category verification results, target position deviation, evidence validity and compensation type evaluation.

[0070] The onboard processing unit matches the ground verification results with the corresponding evidence breakage values, identification closure gain, and local verification branches. When the target identification result is verified correctly and the identification evidence is valid, the corresponding evidence breakage threshold and branch matching relationship are retained; when the target identification result is incorrect or the identification evidence is invalid, the evidence breakage threshold is corrected based on the difference between the corresponding evidence breakage value and the breakage value of historical correct samples.

[0071] For different dimensions of fracture evidence, the verification results of spectral verification branches, geometric verification branches, and scale verification branches are statistically analyzed. Local verification branches that consistently produce correct identification results are given higher matching priority, while local verification branches that consistently produce incorrect results or whose resource consumption exceeds a preset range are given lower matching priority.

[0072] Based on the updated evidence breakage threshold and branch matching priority, the correspondence between the broken evidence dimension and the local verification branch is adjusted, and the adjustment results are written into the spaceborne recognition configuration table. When subsequent remote sensing images enter the processing flow, evidence breakage determination and local verification are performed according to the updated configuration.

[0073] Example 2 like Figure 2 As shown, Embodiment 2 of this application provides a spaceborne image recognition and processing system for remote sensing satellites, including: The ground feature evidence chain construction module collects raw images, imaging parameters, and on-board resource status from remote sensing satellites, maps cross-band and cross-scan line images to a unified ground unit, and constructs a ground feature evidence chain, including a ground unit registration submodule and an evidence chain construction submodule: The ground unit registration submodule reads the original scan lines of each band according to the imaging time, determines the initial spatial offset between bands based on the positional relationship of each band detection unit, and projects each original scan line onto the ground coordinate system in combination with the satellite orbit position, attitude angle and imaging line of sight.

[0074] The ground unit registration submodule divides continuous ground regions according to a preset ground resolution, assigning pixels covering the same region in different bands and scan lines to the corresponding ground units. When local misalignment occurs in cross-band images, the ground projection position is corrected based on the stable ground feature edges in adjacent scan lines, and the imaging time, ground coverage area, and registration deviation of each image segment are recorded.

[0075] The evidence chain construction submodule extracts the spectral response, scanning time sequence, texture, edge and image scale information corresponding to each surface unit, and records the corresponding band identifier, scan line identifier and image scale identifier.

[0076] The evidence chain construction submodule associates the original cache location, data length, and effective retention time of each image fragment with the corresponding evidence dimension, and connects the spectral evidence, scanning time sequence evidence, scale evidence, and spatial structure evidence corresponding to the same ground unit according to the imaging time sequence to generate a ground feature evidence chain.

[0077] The evidence fracture spectrum calculation module performs shielding simulation on single evidence dimensions and combinations of related evidence dimensions in the ground feature evidence chain. It calculates the evidence fracture spectrum based on the deviation between the sum of combined shielding loss and single shielding loss, including a shielding state generation submodule and a fracture spectrum value calculation submodule: The shielding state generation submodule identifies band evidence, scan timing evidence, image scale evidence, and spatial structure evidence as independent evidence dimensions, and combines evidence dimensions that are correlated in spectral response, have adjacent scan timing, or jointly cover the same identification area into associated evidence dimension combinations. The number of evidence dimensions in an associated evidence dimension combination does not exceed a preset combination limit.

[0078] The masking state generation submodule inputs the complete ground feature evidence chain into the shared feature extraction layer and the basic identification branch to obtain the baseline features, baseline category boundaries, and baseline target locations. When the basic identification branch does not output a target, the highest category response and the center of its local response region are determined as the baseline response category and baseline response location, respectively.

[0079] The shielding state generation submodule uses non-discriminatory reference data within the same ground unit to replace the image data corresponding to the single evidence dimension and the combination of related evidence dimensions, respectively, to generate single shielding state, combined shielding state and evidence subset shielding state, and extracts the shielding response vector under each shielding state.

[0080] The fracture spectrum value calculation submodule determines the positive and negative alternation coefficients according to the difference in the number of dimensions between the combination of related evidence dimensions and each evidence subset, and performs inclusion-exclusion superposition on the combined shielding response and the shielding response of the evidence subset to obtain the higher-order interactive residual.

[0081] The fracture spectrum calculation submodule establishes the alternative evidence response space based on the shielding response of the alternative evidence dimension, extracts the uncompensated fracture vector in the high-order interaction residual through regularized projection, and generates an alternative evidence scarcity value of 0 to 1 based on the number of alternative evidence, the remaining cache time, and the registration deviation.

[0082] The fracture spectrum calculation submodule processes the uncompensated fracture vector, higher-order interactive residual, benchmark response covariance and alternative evidence scarcity according to the evidence fracture value calculation method described in Example 1 S202, generates the evidence fracture spectrum, and determines the vulnerable ground feature unit and fracture evidence dimension by recovering the associated evidence dimension one by one.

[0083] The local compensated view generation module extracts associated image data of vulnerable land cover units from the original image cache based on the evidence fracture spectrum, and generates a local compensated view corresponding to the fracture evidence dimension. This module includes a cache data localization submodule and a compensated view generation submodule. The cached data location submodule reads the location identifier of the vulnerable ground feature unit, the combination of associated evidence dimensions, and the fracture evidence dimensions, and locates the corresponding band, scan line, image scale, and original cache location based on the ground feature evidence chain.

[0084] The cached data localization submodule expands the local extraction area based on registration deviation and ground coverage, and identifies image segments within the effective retention time and whose image quality meets preset requirements as associated image data. When the same vulnerable ground feature unit corresponds to multiple fracture evidence dimensions, the extraction order of associated image data is determined according to the evidence fracture value.

[0085] The compensation view generation submodule selects the corresponding compensation method based on the dimension of the fracture evidence. For fracture evidence in bands, a spectral compensation view is generated based on the spectral response relationship of adjacent bands within the stable ground object region; for fracture evidence in scanning time sequence or location, a geometric compensation view is generated based on the edge of stable ground objects, satellite attitude changes, and imaging time difference; for fracture evidence in image scale or spatial structure, a scale compensation view is generated based on the original resolution image and local texture direction.

[0086] When the same vulnerable ground feature unit has multiple fracture evidence dimensions, the compensation view generation submodule performs geometric compensation, spectral compensation and scale compensation in sequence, and records the ground unit, compensation type, image source and compensation data range corresponding to the local compensation view.

[0087] The candidate recognition result generation module inputs the ground feature units and local compensation views into the basic recognition branch and local verification branch of the spaceborne image recognition model to generate corresponding candidate recognition results, including a basic result generation submodule and a verification result generation submodule: The basic results generation submodule inputs each ground unit into the shared feature extraction layer, extracts spectral features, texture features and spatial structure features, and generates basic categories, basic category responses and basic locations through the basic recognition branch.

[0088] The basic result generation submodule generates basic evidence responses based on the inverse response position of the basic category response in the shared features. For non-vulnerable land cover units, the basic identification results are directly determined as target identification results; for vulnerable land cover units, the basic identification results are sent to the corresponding local verification branch.

[0089] The verification result generation submodule selects the spectral verification branch, geometric verification branch, or scale verification branch based on the dimension of fracture evidence, and inputs the vulnerable ground feature unit and the corresponding local compensation view into the selected local verification branch.

[0090] The verification result generation submodule generates verification categories, verification locations, and compensation evidence responses through local verification branches. When a vulnerable feature unit corresponds to multiple fracture evidence dimensions, corresponding candidate identification results are generated for each, and the verification branches and resource usage information of each candidate identification result are recorded.

Claims

1. A method for spaceborne image recognition and processing for remote sensing satellites, characterized in that, Includes the following steps: S10. Collect raw images, imaging parameters and on-board resource status from remote sensing satellites, map cross-band and cross-scan line images to a unified ground unit, and construct a chain of ground feature evidence. S20. Perform shielding deduction on the single evidence dimension and the combination of related evidence dimensions in the evidence chain of ground features, calculate the evidence fracture spectrum based on the deviation of the sum of the combined shielding loss and the single shielding loss, and identify vulnerable ground feature units and their fracture evidence dimensions. S30. Extract associated image data of vulnerable land cover units from the original image cache based on the evidence fracture spectrum, and generate a local compensated view corresponding to the fracture evidence dimension. S40. Input the ground feature units and local compensation views into the basic recognition branch and local verification branch of the spaceborne image recognition model to generate corresponding candidate recognition results; S50. Calculate the identification closure gain based on the consistency of categories, position stability and evidence recovery degree before and after local compensation, and determine the target identification result and on-board computing resource quota. S60: Output the target identification result and the corresponding identification evidence information, and update the evidence breakage threshold and local verification branch configuration based on the ground verification result.

2. The satellite image recognition and processing method for remote sensing satellites according to claim 1, characterized in that, Acquire raw images, imaging parameters, and onboard resource status from remote sensing satellites, map cross-band and cross-scan line images to a unified ground unit, and construct a chain of ground feature evidence, including the following sub-steps: S101. Based on the imaging time, satellite attitude, and band position, register cross-band and cross-scan line images to a unified ground unit; S102. Establish evidence mapping between ground units and bands, scan lines, image scales, and original buffer locations to generate ground feature evidence chains.

3. The satellite image recognition and processing method for remote sensing satellites according to claim 1, characterized in that, The process involves masking and extrapolating the single and combined evidence dimensions within the evidence chain of ground features. Based on the deviation between the sum of the combined masking loss and the single masking loss, the evidence fracture spectrum is calculated to identify vulnerable ground feature units and their fracture evidence dimensions. This includes the following sub-steps: S201. Separately mask single evidence dimensions and combinations of evidence dimensions with related relationships, and extract the changes in identification features under each masking state; S202. Calculate the evidence fracture spectrum based on the deviation between the sum of the combined shielding loss and the corresponding single shielding loss, and determine the land cover units and evidence dimensions that reach the fracture threshold as vulnerable land cover units and fracture evidence dimensions.

4. The satellite image recognition and processing method for remote sensing satellites according to claim 1, characterized in that, Based on the evidence fracture spectrum, associated image data of vulnerable land cover units are extracted from the original image cache, and a local compensated view corresponding to the fracture evidence dimension is generated, including the following sub-steps: S301. Based on the vulnerable ground features and fracture evidence dimensions, locate the corresponding bands, scan lines, image scales, and original cache locations; S302. Extract the corresponding image data from the original image cache, perform spectral compensation, geometric compensation or scale compensation according to the fracture evidence dimension, and generate a local compensated view.

5. The satellite image recognition and processing method for remote sensing satellites according to claim 1, characterized in that, The ground feature units and local compensated views are input into the basic recognition branch and local verification branch of the spaceborne image recognition model to generate corresponding candidate recognition results, including the following sub-steps: S401. Input the ground feature unit into the basic identification branch to generate basic category, basic location and basic evidence response; S402. Input the vulnerable feature unit and local compensation view into the local verification branch that matches the fracture evidence dimension, and generate the verification category, verification location and compensation evidence response.

6. The satellite image recognition and processing method for remote sensing satellites according to claim 1, characterized in that, The identification closure gain is calculated based on the consistency of categories, position stability, and degree of evidence recovery before and after local compensation. The target identification results and on-board computing resource quotas are then determined, including the following sub-steps: S501. Compare the category changes, positional shifts, and fracture evidence response changes between the basic identification results and each candidate identification result to form an identification closure state; S502. Calculate the identification closure gain based on the identification closure state and the resource occupation of the local verification branch, select the corresponding candidate identification result as the target identification result, and determine the on-board computing resource quota.

7. The satellite image recognition and processing method for remote sensing satellites according to claim 1, characterized in that, Output the target recognition result and corresponding recognition evidence information, and update the evidence breakage threshold and local verification branch configuration based on the ground verification result, including the following sub-steps: S601. Extract the band segments, local responses and imaging parameters that support the target recognition results, and generate recognition evidence information associated with the target recognition results; S602. Correct the evidence breakage threshold based on the ground verification results, and adjust the matching relationship between the broken evidence dimension and the local review branch.

8. A spaceborne image recognition and processing system for remote sensing satellites, characterized in that, include: The ground feature evidence chain construction module is used to collect raw images, imaging parameters, and on-board resource status to construct the ground feature evidence chain; The evidence break spectrum calculation module is used to perform masking and deduction on single evidence dimensions and combinations of related evidence dimensions, and calculate the evidence break spectrum. The local compensation view generation module is used to extract associated image data based on the evidence fracture spectrum and generate a local compensation view. The candidate recognition result generation module is used to generate candidate recognition results through the basic recognition branch and the local verification branch; The closed-loop gain calculation module is used to calculate the closed-loop gain and determine the target identification result and the on-board computing resource quota. The identification result feedback and update module is used to output the target identification result and identification evidence information, and update the evidence breakage threshold and local verification branch configuration based on the ground verification result.

Images