A mass remote sensing data cleaning and recovery method

Abstract

The application provides a mass remote sensing data cleaning and recovery method, which can realize efficient cleaning of mass observation data of a remote sensing satellite, and then recovers the change increment and the space-time compression. The application is based on the space-time evolution characteristics of ground objects, and extracts high-value key frame images and compression algorithms from continuous observation images, specifically: firstly, the long-term observation data is preprocessed for space spectrum consistency by using geometric registration and relative radiation correction; then, the space-time dynamic key frame is extracted by using a remote sensing image change detection method, so that the data redundancy problem of repeated recording of time series observation is solved; finally, a change increment and space-time compression recovery method is designed, so that efficient and rapid data cleaning and recovery of long-term observation data of the remote sensing satellite are realized, the valuable long-term storage space is saved, and the effective data retrieval efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image information processing, and particularly relates to a mass remote sensing data cleaning and recovery method. BACKGROUND

[0002] With the rapid development of remote sensing satellite technology, more and more optical remote sensing satellites are in orbit, and remote sensing satellites continuously acquire observation data and return to the ground. After long-term observation, a large amount of remote sensing image data is obtained. On the one hand, the mass remote sensing image data brings great pressure to the storage system, increases the cost of data storage and regular maintenance, and will continue to rise with the increase of storage time; on the other hand, the change of ground feature scene information is not continuous in time sequence, and the record of ground feature change information is high-value information, and the repeated record data without ground feature change is redundant information.

[0003] In order to solve the problem of mass remote sensing data storage and call, the existing method mainly performs lossless compression processing on all frames of observation data, but does not solve the data redundancy problem of time sequence observation repeated record. SUMMARY

[0004] Therefore, the present application provides a mass remote sensing data cleaning and recovery method, which can realize efficient cleaning of mass observation data of remote sensing satellites, and then recover according to change increment and space-time compression.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows:

[0006] A mass remote sensing data cleaning and recovery method, comprising the following steps:

[0007] The long-term observation data is preprocessed for spectral consistency by using geometric registration and relative radiation correction; the time-space dynamic key frame data is extracted from the preprocessed image by using a remote sensing image change detection method, and the time-space key frame data is losslessly compressed; the data after lossless compression is recovered by using a change increment and space-time compression recovery method, so as to realize cleaning and recovery of remote sensing data.

[0008] The geometric registration specifically includes the following operation steps:

[0009] The multi-temporal remote sensing images of the same ground scene are globally coarsely registered based on a disparity image, and the multi-resolution characteristic of the wavelet transform is used to generate a gray-based pyramid by wavelet transforming the images; the approximate components of each layer from the top layer to the bottom layer of the pyramid are coarsely registered to finely registered, and the registration result of the previous layer is used as the registration initial value of the next layer, so that the registration initial value is obtained; the initial value obtained in the second step is used to respectively perform correlation registration on the approximate image and the vertical detail image of the next-to-bottom layer, so that the corresponding registration correlation curves are obtained; the correlation registration curve of the vertical detail image of the next-to-bottom layer is used as the weight to perform weighted processing on the approximate image registration curve, and then the extreme value of the weighted correlation curve is found to obtain the stable optimal registration, complete the geometric registration, and obtain the multi-temporal remote sensing images after the geometric registration.

[0010] The specific method of the relative radiation correction is:

[0011] For the multi-temporal remote sensing images after the geometric registration, the best super parameter combination is obtained by using the genetic algorithm to automatically search the super parameters on the basis of data statistics; the histogram matching method is used, that is, the statistical parameters of the radiation response gray distribution of a plurality of detectors are selected as reference benchmarks, the statistical parameters of the gray distribution of other detectors are matched to the reference benchmarks, the radiation response relationship between the detectors is constructed, the correction coefficients are directly extracted from the relationship, and a relative radiation correction model Y(i)=X(i)M(i) is generated, wherein X(i) represents the original image data, M(i) represents the radiation correction model, and Y(i) represents the corrected image data.

[0012] The remote sensing image change detection method comprises the following specific steps: a homogeneous image difference representation learning network model integrating difference estimation, difference representation learning and unsupervised clustering is established, wherein the homogeneous image difference representation learning network maps two-time images to a suitable feature space to extract key change information and suppress noise; the two-time data are compared and analyzed in the feature space based on the denoising automatic coding to highlight the changes; the K-means clustering is performed on the difference representation friendly to clustering; the spatial motion vector information of the current pixel block is recorded, and the prediction residual of the current block and the matching block is obtained, so as to remove the time redundancy of adjacent images; the image is transformed from the spatial domain to the frequency domain by the transformation method, the high-energy area is integrated and distributed, the visual redundancy information of the image is removed, the ring filter SAO is used to reduce the ringing effect, and the spatio-temporal dynamic key frame data extraction is completed.

[0013] The spatio-temporal key frame data are losslessly compressed by using the context-based adaptive binary arithmetic coding to realize entropy coding, and the compressed data are obtained.

[0014] The spatial domain decompression processing is specifically: according to the high-value information high-efficiency compression processing algorithm design, reverse decompression processing is carried out, the compressed data is input, and the decompressed key frame image and the change increment data are output; in the change increment recovery, the change increment is based on the space-time key frame image, and a plurality of frames of change increments corresponding to each frame of space-time key frame image are sequentially recovered according to the time sequence relationship, and the specific recovery method is key frame local increment replacement; the time domain continuity recovery is a recovery module designed for the same or similar image pre-removal module, and the same or similar frames removed during the removal process are recovered according to the semantic level parameters, and the semantic level parameters are specifically represented as the time sequence reserved frame serial number and the removal frame serial number.

[0015] Beneficial effects

[0016] 1. The application is based on the space-time evolution characteristics of ground objects, and high-value key frame images and compression algorithms in continuous observation images are extracted, specifically: firstly, the long-term observation data is pre-processed for spatial spectrum consistency by using geometric registration and relative radiation correction; then, the space-time dynamic key frame is extracted by using a remote sensing image change detection method, solving the problem of data redundancy caused by repeated recording in time sequence observation; finally, the change increment and space-time domain compression recovery method is designed, realizing efficient and rapid data cleaning and recovery of long-term observation data of remote sensing satellites, saving valuable long-term storage space and improving the effective data retrieval efficiency.

[0017] 2. In the application, the geometric registration method combined with wavelet transform solves the problem of fast and robust registration of parallax images in the geometric registration process of high-resolution remote sensing images in the traditional method; a relative radiation correction method combined with data statistical strategy, histogram matching method and genetic algorithm is used, thereby greatly improving the accuracy of relative radiation correction.

[0018] 3. The remote sensing image change detection method based on deep learning integrates the traditional feature extraction and feature analysis into one process, directly learns the change features from the multi-temporal remote sensing images, and finally obtains the change map by segmenting the images through the change features, and has good robustness. Compared with the traditional change detection method, the deep learning method can eliminate the influence of the dependence on the difference map of the detection result, can process remote sensing data obtained by multiple sensors, and has strong applicability. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is a schematic diagram of the geometric registration method of multi-temporal remote sensing data of the application.

[0020] Figure 2 It is a schematic diagram of the relative radiation correction method of multi-temporal remote sensing data of the application.

[0021] Figure 3 It is a schematic diagram of the lossless compression method of remote sensing images of the application.

[0022] Figure 4 The figure is a variation increment recovery method of the present application.

[0023] Figure 5 The figure is a time domain continuity recovery method of the present application. DETAILED DESCRIPTION

[0024] The present application is described in detail below with reference to the accompanying drawings and examples.

[0025] For long-term observation image data of remote sensing satellites of the same scene on the ground, there are a large number of same / similar redundant images, so data cleaning is needed to save valuable long-term storage space and improve effective data retrieval efficiency. The present application provides a mass remote sensing data cleaning and recovery method to solve the problem of low compression efficiency and repeated recording of data redundancy in time sequence observation of traditional remote sensing data cleaning algorithm, including the following steps:

[0026] Step 1: Preprocessing continuous multi-temporal remote sensing images by using geometric registration and relative radiation correction to realize the spectral consistency of remote sensing data, and obtaining preprocessed images.

[0027] The geometric registration has the following specific operation steps:

[0028] The multi-temporal remote sensing images of the same ground scene are globally coarsely registered based on the disparity image of wavelet transform. The multi-resolution characteristics of wavelet transform are used to generate a gray-based pyramid for the image. At the top layer of the pyramid, the image size is the smallest and the resolution is the lowest, while at the bottom layer of the original image, the resolution is the highest. On this basis, the correlation-based technology is used to coarsely to finely register the approximate components of each layer from the top layer to the bottom layer of the pyramid. The registration result of the previous layer can be used as the initial value of the registration of the next layer, so that a better initial value of registration can be obtained. Further, in order to improve the accuracy and reliability of the registration, the initial value obtained in the second step is used to respectively perform correlation registration on the approximate image and the vertical detail image at the second bottom layer of the pyramid, and the corresponding registration correlation curve is obtained. The correlation registration curve of the vertical detail image at the second bottom layer is used as a weight to perform weighted processing on the approximate image registration curve, and then the extreme value of the weighted correlation curve is obtained to obtain a stable optimal registration, complete the geometric registration, and obtain the multi-temporal remote sensing images after geometric registration. The multi-temporal remote sensing data geometric registration method of the present application is shown in Figure 1 .

[0029] The specific method of the relative radiation correction is as follows:

[0030] For a large number of geometrically registered multi-temporal remote sensing images, on the basis of good data statistics, the genetic algorithm is used to automatically search for the best super parameter combination; the histogram matching method is used, that is, the radiation response gray distribution statistical parameters of a plurality of probes are selected as reference benchmarks, and the gray distribution statistical parameters of other probes are matched to the reference benchmarks, so that the radiation response relationship between the probes can be constructed, the correction coefficient can be directly extracted therefrom, and a relative radiometric correction model Y(i) = X(i)M(i) is generated, wherein X(i) represents original image data, M(i) represents a radiometric correction model, and Y(i) represents corrected image data; the correction model is used for relative radiometric correction of two temporal images, and preprocessing is completed, and the multi-temporal remote sensing relative radiometric correction method of the application is as shown in Figure 2 .

[0031] Step 2, using a remote sensing image change detection method to extract the spatio-temporal dynamic key frame data from the preprocessed images, and performing lossless compression processing on the spatio-temporal key frame data, as shown in Figure 3 .

[0032] The specific steps of the remote sensing image change detection method include: establishing a homogeneous image difference representation learning network model integrating difference estimation, difference representation learning (DRL) and unsupervised clustering, wherein the homogeneous image difference representation learning network maps two temporal images to a suitable feature space to extract key change information and suppress noise; based on the denoising automatic encoding (DAE), the two temporal data are compared and analyzed in the feature space to highlight the changes; the K-means clustering is performed on the difference representation friendly to clustering; the spatial motion vector information of the current pixel block is recorded, the prediction residual of the current block and the matching block is obtained, so as to remove the temporal redundancy of adjacent images; by transforming the image from the spatial domain to the frequency domain, the high-energy region is integrated and distributed, the visual redundancy information of the image is removed, and the loop filter SAO is used to reduce the ringing effect, and the spatio-temporal dynamic key frame data extraction is completed.

[0033] Specifically, in the embodiment, the context-based adaptive binary arithmetic coding is used for entropy coding to realize lossless compression processing of the spatio-temporal key frame data, and compressed data is obtained.

[0034] Step 3, using spatial domain decompression, change increment recovery and time domain continuity recovery to perform recovery processing for massive remote sensing data cleaning.

[0035] Specifically, the spatial domain decompression processing is: according to the reverse decompression processing of the high-value information efficient compression processing algorithm design, the compressed data is input, and the decompressed key frame image and change increment data are output.

[0036] In the change increment recovery, the change increment relies on the space-time key frame image, and a plurality of frames of change increments corresponding to each frame of space-time key frame image are sequentially recovered according to the time sequence relationship. The specific recovery method is key frame local increment replacement, and the change increment recovery method is as shown in Figure 4 .

[0037] The time domain continuity recovery is a recovery module designed for the same / similar image pre-de-duplication module. According to the semantic level parameters generated in the de-duplication process, the same / similar frames cleaned up are recovered. The semantic level parameters are specifically represented as time sequence reserved frame sequence numbers, de-duplicated frame sequence numbers and the like. The time domain continuity recovery method is as shown in Figure 5 .

[0038] To sum up, the above is only a preferred embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A mass remote sensing data cleaning and recovery method, characterized in that, Comprise the following steps: The spatio-temporal dynamic key frame data is extracted from the preprocessed images by using a remote sensing image change detection method, and the lossless compression processing is performed on the spatio-temporal key frame data; the data after the lossless compression processing is recovered by using a change increment and spatio-temporal field compression recovery method, so that the remote sensing data is cleaned and recovered; The specific operation steps of the geometric registration are as follows:

2. The method of claim 1, wherein, The multi-temporal remote sensing images of the same ground scene are globally roughly registered based on a wavelet transform disparity image, the multi-resolution characteristics of the wavelet transform are used to generate a gray-based pyramid by performing wavelet transform on the images; the approximate components of each layer are roughly to finely registered from the top layer to the bottom layer of the pyramid, and the registration result of the upper layer is used as the registration initial value of the lower layer, so that the registration initial value is obtained; The approximate image and the vertical detail image of the second bottom layer of the pyramid are respectively registered using the initial value obtained in the second step, and the corresponding registration correlation curves are obtained; the registration correlation curve of the vertical detail image of the second bottom layer is used as the weight to process the approximate image registration curve, and then the extreme value of the weighted correlation curve is obtained, so that the stable optimal registration is obtained, the geometric registration is completed, and the multi-temporal remote sensing images after the geometric registration are obtained; The specific method of the relative radiation correction is as follows: The specific steps of the remote sensing image change detection method include: establishing a homogeneous image difference representation learning network model integrating difference estimation, difference representation learning and unsupervised clustering, wherein the homogeneous image difference representation learning network maps two-time images to a suitable feature space to extract key change information and suppress noise; the two-time data are compared and analyzed in the feature space based on the denoising automatic coding to highlight the changes; the K-means clustering is performed on the difference representation friendly to clustering; the spatial motion vector information of the current pixel block is recorded, the prediction residual of the current block and the matching block is obtained, so that the temporal redundancy of the adjacent images is removed; by the transformation method, the image is transformed from the spatial domain to the frequency domain, the high-energy region is integrated and distributed, the visual redundancy information of the image is removed, the ring filter SAO is used to reduce the ringing effect, and the extraction of the spatio-temporal dynamic key frame data is completed. For multi-temporal remote sensing images after geometric registration, on the basis of data statistics, the genetic algorithm is used to automatically search for the best combination of hyperparameters; the histogram matching method is used, that is, the statistical parameters of the radiation response gray scale distribution of a plurality of detectors are selected as reference benchmarks, the statistical parameters of the gray scale distribution of other detectors are matched to the reference benchmarks, the radiation response relationship between the detectors is constructed, the correction coefficients are directly extracted therefrom, and a relative radiometric correction model is generated , wherein X(i) represents original image data, M(i) represents a radiometric correction model, and Y(i) represents corrected image data. The correction model is used to perform relative radiometric correction on two temporal images, and preprocessing is completed.

3. The method of claim 2, wherein, The entropy coding is implemented by using the context-based adaptive binary arithmetic coding to realize the lossless compression processing on the spatio-temporal key frame data, and the compressed data is obtained.

4. The method according to any one of claims 1 to 3, characterized in that, ​

Images