Cloud detection method for spaceborne optical remote sensing images driven by imaging simulation model data
Through the imaging simulation model and structural similarity (SSIM) method, the problem that cloud detection in satellite-borne optical remote sensing images is difficult to maintain high spatial resolution, and accurate cloud detection under different conditions is achieved, suitable for single spectrum and multiple spectrum Spectral segment images improve the applicability and accuracy of detection.
Patent Information
- Application Number
- CN202410143130.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-01
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-02-01
AI Technical Summary
The prior art is difficult to maintain high spatial resolution when detecting clouds in satellite-borne optical remote sensing images, and the complex scattered radiation characteristics of clouds lead to lack of scalability and accuracy in detection results, especially in multi-spectral joint detection and intelligent identification. in network method.
Using the data-driven method of imaging simulation model, cloudless images are simulated through the Earth's background hyperspectral imaging characteristic model, compared with the measured image, structure similarity (SSIM) is calculated to generate SSIM images, and cloud area is determined based on the SSIM threshold, so as to realize cloud detection.
This method can accurately detect cloud areas while maintaining the spatial resolution of the original image, and is suitable for single-spectral and multi-spectral images, and is suitable for different tracks and background conditions, improving the applicability and accuracy of detection.
Smart Images

Figure CN117994580B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of optical remote sensing image difference detection and target detection and recognition, and in particular to a satellite-borne optical remote sensing image cloud layer detection method driven by imaging simulation model data. Background Art
[0002] The onboard optical sensors include visible light cameras, full-spectrum radiometers, multispectral and hyperspectral imaging sensors, etc. The inversion of cloud distribution and its time-varying characteristics from various optical remote sensing images provides important information support for applications such as forecasting local weather, warning of wide-area natural disasters such as typhoons, floods, droughts, and monitoring of global atmospheric conditions and water vapor cycles.
[0003] The difficulty in detecting cloud-covered areas lies mainly in the diverse shapes of clouds, large differences in height distribution, non-fixed spatial positions, and complex processes of evolution or dissipation over time. As a result, their scattering and radiation characteristics fluctuate dramatically in the time and space domains. In remote sensing images, this is manifested as a large dynamic range of contrast between clouds and the surface background, that is, the signal-to-noise ratio between cloud targets and surface clutter fluctuates greatly, making it difficult to set a reasonable detection threshold based on empirical signal-to-noise ratios. In addition, the scattering and radiation characteristics of the ground object background also vary with factors such as geographic location, observation time, and imaging geometry. The complex and changeable scattering and radiation characteristics of targets (clouds) and background (ground objects) make it extremely challenging to accurately detect cloud-covered areas from wide-area satellite-borne optical sensor images.
[0004] The scattered radiation characteristics of clouds and the ground surface vary significantly with the detection spectrum. The total radiation received in the spectrum includes the scattered solar radiation and its own thermal radiation. Under normal circumstances (with sunlight), in the visible to short-wave infrared spectrum, the scattering of solar radiation by clouds and the ground surface is significantly stronger than its own thermal radiation. Since the reflectivity of clouds is greater than that of most surface covers (except for ice, snow and other surface covers), they usually appear brighter in the detection images of these spectrums. Therefore, the bright areas can be distinguished by setting an appropriate radiance threshold to achieve cloud detection. In the medium-wave to long-wave infrared spectrum, the self-thermal radiation of clouds and the ground surface is stronger than their scattering of solar radiation. Since the temperature of clouds is usually lower than that of surface covers (except for ice, snow and other surface covers), they usually appear darker in the detection images of these spectrums. Therefore, the radiance threshold can be set to distinguish the dark areas to achieve cloud detection.
[0005] The following detection methods are available in the prior art:
[0006] 1. Joint detection method of multiple spectral color ratio image thresholds. Multiple detection spectrum bands are combined in pairs to generate multiple color ratio images. An empirical threshold is set for each color ratio image to detect suspected cloud areas. Then, the detection results of each color ratio image are combined to eliminate the interference of factors such as surface ice and snow, and the final cloud detection result is output (Reference 1: S. Ackerman, R. Frey, K. Strabala, et al. Discriminating Clear-Sky From Cloud With MODIS Algorithm Theoretical Basis Document (MOD35) [R / OL]. https: / / eospso.gsfc.nasa.gov / sites / default / files / atbd / atbd_mod06.pdf, ATBD-MOD-06, 2010. and Reference 2: A. Heidinger and W. C. Straka. ATBD GOES-R Cloud Mask v3.0 [R / OL]. https: / / www.star.nesdis.noaa.gov / goesr / docs / ATBD / Cloud_Mask.pdf, NOAA NESDIS STAR, 2012.).
[0007] However, it has the following disadvantages: the spatial resolution of each spectrum band of the satellite-borne multispectral sensor is different. For example, the ABI (Advanced Baseline Imager) sensor carried by the GOES (Geostationary Operational Environmental Satellite)-16 satellite has an angular resolution of about 28μm in the short-wave spectrum band of visible light and about 56μm in the long-wave infrared band. When multiple spectrum bands are used to detect cloud cover areas, the spatial resolution (number of pixel units) of the detection image depends on the spectrum band with the lowest angular resolution. And if a dynamic detection threshold is used to determine whether it is a cloud layer by the color ratio probability density statistics of the local pixel unit of the image, the spatial resolution of the cloud detection image will be further reduced.
[0008] 2. Similar to the first technical solution, an intelligent recognition network is used instead of the empirical threshold to detect clouds. Multiple spectral image data are still used, and the color ratio images are generated by combining them in pairs. Then, intelligent recognition algorithms such as convolutional neural network (CNN) are used to classify the image pixel units. The classification map contains cloud categories (Reference 3: Z. Shao, Y. Pan, C. Diao, and J. Cai. Cloud Detection in Remote Sensing Images Based on Multiscale Features-Convolutional Neural Network [J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57 (6): 4062-4076. and Reference: 4: W. Li, F. Zhang, H. Lin, et al. Cloud Detection and Classification Algorithms for Himawari-8 Imager Measurements Based on Deep Learning [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60 (4107117): 1-17).
[0009] However, it has the following disadvantages: Similar to the first technical solution, the spatial resolution of the classified image depends on the minimum angular resolution of the adopted spectrum. In addition, the completeness of the data used in training the intelligent network is the main factor affecting the accuracy of the classification results. However, the complex and changeable scattered radiation characteristics of clouds make it difficult to verify whether the training data is complete, thus affecting the scalability of this technical solution.
[0010] 3. Based on the output image of the threshold of technical solution 1 or the intelligent network detection of technical solution 2, the cloud coverage area is detected in combination with the morphological characteristics of the cloud. (Reference 5: P.Bo, S.Fenzhen, and M.Yunshan. A Cloud and Cloud Shadow Detection Method Based on Fuzzy c-Means Algorithm[J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13:1714-1727).
[0011] However, it has the following disadvantages: the morphological characteristics of clouds, such as height, shape, area, and surface roughness, are diverse and have significant temporal uncertainty. Therefore, the cloud detection results of remote sensing images are difficult to solidify into morphological feature knowledge, and the detection method is not extensible. Summary of the invention
[0012] In order to solve the above technical problems, the present invention proposes a cloud detection method for satellite optical remote sensing images driven by imaging simulation model data, aiming at the characteristic that the scattered radiation characteristics of the ground object background in the satellite optical remote sensing detection scene change relatively slowly over time. According to the parameter information such as the geographical location, orbital altitude, imaging field of view, and detection spectrum of the satellite optical sensor when acquiring the detection image, the hyperspectral imaging characteristic model of the earth background is used to simulate the cloud-free image of the same observation scene, and the brightness value of the simulated cloud-free image is calibrated by comparing the solar radiation illumination difference obtained by the model simulation and the actual measurement. Then, the structural similarity (Structure SIMilarity, SSIM) of the simulated and measured images is calculated pixel by pixel to generate an SSIM image. The pixel unit value of the SSIM image is between 0 and 1: the closer the pixel unit value is to 0, the greater the probability that the pixel unit is a cloud layer; the closer the pixel unit value is to 1, the smaller the probability that the pixel unit is a cloud layer. By setting the SSIM threshold, the pixel units below the threshold are judged as clouds, and the cloud coverage area detection result image is obtained.
[0013] In order to achieve the above object, the technical solution adopted by the present invention is:
[0014] The cloud detection method of remote sensing images of spaceborne optical sensors driven by imaging simulation model data includes the following steps:
[0015] Step 1: extracting detection parameters of the satellite-borne optical sensor to obtain remote sensing images, wherein the detection parameters include time, position, field of view, and spectrum;
[0016] Step 2: Using the Earth background hyperspectral imaging simulation model, set the model parameters to the detection parameters extracted in step 1 to generate a cloud-free radiance image;
[0017] Step 3: Use the atmospheric radiation transfer code to calculate the solar spectrum irradiance used in the simulation image, and compare it with the solar spectrum irradiance in the spectrum band measured by the remote sensing satellite. Calibrate the value of the average solar spectrum irradiance in the spectrum band of the simulation image to be consistent with the measured value;
[0018] Step 4: Calculate the structural similarity of each pixel unit of the simulated and measured images to generate an SSIM image, where SSIM represents structural similarity;
[0019] Step 5: Set the structural similarity threshold to determine whether each pixel unit is the ground surface or the cloud layer;
[0020] Step 6: Output the cloud detection result map.
[0021] Compared with the existing cloud detection method of remote sensing images using spaceborne optical sensors, the present invention has the following beneficial effects:
[0022] (1) The method of the present invention has good applicability. The present invention can be used for cloud detection in single-spectrum remote sensing images, and is also applicable to cloud detection in color ratio images; it is applicable to cloud detection in geostationary orbit satellites, and is also applicable to cloud detection in sun-synchronous orbit satellites; it is applicable to cloud detection under land background conditions, and is also applicable to cloud detection under ocean background conditions.
[0023] (2) The detection image obtained by the present invention retains the spatial resolution of the original remote sensing image. Compared with the existing solution of using multiple spectral bands to jointly detect the cloud layer, the present invention can use a single spectral band remote sensing image to complete cloud layer detection, so the cloud layer detection result image retains the spatial resolution of the original remote sensing image. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 A flowchart of a cloud layer detection method for satellite-borne optical remote sensing images driven by imaging simulation model data of the present invention;
[0025] Figure 2(a) and Figure 2(b) are measured images of the 2.125μm wavelength spectrum band of the spaceborne optical sensor; Figure 2(a) is an image of the Earth disk in the 2.10-2.35μm wavelength spectrum band obtained by the AGRI sensor of the Fengyun-4B satellite, with a size of 5496×5496 pixels; Figure 2(b) is an image of the eastern coastal area of my country in the 2.105-2.155μm wavelength spectrum band obtained by the MODIS sensor of the Aqua satellite, with a size of 4060×2708 pixels;
[0026] Figure 3(a) and Figure 3(b) are simulated cloud-free images of the 2.125 μm wavelength spectrum of the spaceborne optical sensor; Figure 3(a) is a simulated image of the same observation scene as Figure 2(a); Figure 3(b) is a simulated image of the same observation scene as Figure 2(b);
[0027] Figure 4 The solar spectrum irradiance curve with wavelength of 2.10-2.35μm used in simulating cloud-free images;
[0028] Figure 5(a) and Figure 5(b) are structural similarity (SSIM) images of simulated and measured images; Figure 5(a) is the structural similarity of the scene image observed by the AGRI sensor of the Fengyun-4B satellite; Figure 5(b) is the structural similarity of the scene image observed by the MODIS sensor of the Aqua satellite;
[0029] Figure 6(a) and Figure 6(b) are cloud cover area detection result diagrams; Figure 6(a) is a cloud cover area detection result diagram of the AGRI sensor image of the Fengyun-4B satellite; Figure 6(b) is a cloud cover area detection result diagram of the MODIS sensor image of the Aqua satellite;
[0030] Figure 7 This is an example of using the cloud cover area detection result map to simulate an image of a cloud scene in the 2.125μm wavelength spectrum of the Aqua satellite MODIS sensor. DETAILED DESCRIPTION
[0031] The present invention is further described below with reference to the accompanying drawings and specific embodiments.
[0032] like Figure 1 As shown, the imaging simulation model data-driven satellite optical remote sensing image cloud detection method of the present invention comprises the following steps:
[0033] Step 1: Extract the detection parameters such as time, position, field of view, and spectrum of the remote sensing images acquired by the spaceborne optical sensor based on the Level 1 data product of the spaceborne optical sensor.
[0034] The Level 1 data products of the onboard optical sensors of optical remote sensing satellites are usually calibrated radiance images, reflectance images or equivalent temperature images. The file name of the data product usually contains the start time and end time of data acquisition. The spectral band number is described in the product attachment. The central wavelength and spectral width of the spectral band can be calculated based on the cutoff wavelength (3dB wavelength point) of the spectral response function in the spectral band. The orbital altitude information of the onboard platform (i.e., sensor platform) can be obtained from public data, such as the geostationary orbit altitude of about 35786km, the MODIS sensor orbit altitude of about 708km, etc.
[0035] The instantaneous field of view of the sensor (i.e., spatial angular resolution) can be inverted from the sensor platform orbital height and the given image spatial resolution:
[0036]
[0037] Where H is the orbital height of the satellite; Δx is the spatial resolution of the optical remote sensing image; and α is the instantaneous field of view of the sensor.
[0038] The orbit of an artificial satellite is usually pre-set and cannot be changed at will. The geographical location (latitude and longitude coordinates) of the sub-satellite point of a geostationary orbit satellite does not change over time; the geographical location of the sub-satellite point of a sun-synchronous orbit satellite can be calculated by querying the orbit information and observing the time.
[0039] Step 2: Using the Earth background hyperspectral imaging simulation model, set the model parameters to the extraction results of step 1 and simulate the cloud-free radiance image of the same sensor observation scene.
[0040] The remote sensing images acquired by the spaceborne optical sensor can usually be represented as a two-dimensional matrix. Assuming that the simulated image is X and the measured image is Y, the number of pixel units of the two images is M×N, we have:
[0041]
[0042] Where M represents the total number of rows; N represents the total number of columns; i represents the number of rows; j represents the number of columns; x (i)(j) represents the pixel unit in the i-th row and j-th column of the simulated image X; y (i)(j) Represents the pixel unit in the i-th row and j-th column in the simulated image Y.
[0043] For a pixel unit x in the simulated image X (i)(j) The corresponding sensor instantaneous field of view entrance pupil radiance Lx ij , which can be expressed as a discrete summation of spectra:
[0044]
[0045] Where λ represents the wavelength; λ1 and λ2 are the starting and ending wavelengths of the spectrum integration respectively; Δλ is the spectrum sampling interval; E p E is the radiance of the transmission path radiation spectrum; g is the surface spectral radiance, calculated using the simplified expression:
[0046] E g (λ)=[E0(λ)·τ s (λ)·ρ r (λ)+ε g (λ)·B(λ,T g )]·τ v (λ) (4)
[0047] Where E0 is the solar spectral irradiance at the top of the atmosphere; τ s With τ v are the atmospheric transmittances of the solar radiation incident path and observation path respectively; ρ r is the bidirectional spectral reflectance of the surface, which is related to the geometric relationship and terrain undulation; ε g is the surface spectral emissivity; T g is the surface temperature; B is the spectral radiance of the object with temperature T, which can be calculated using Planck's formula:
[0048]
[0049] Where h is Planck's constant, which is approximately 6.626×10 -34 J·s; c is the speed of light in a vacuum, about 2.998×10 8 m·s -1 ; K B is the Boltzmann constant, about 1.38×10 -23 J.K. -1 .
[0050] The solar spectral irradiance E0 at the top of the atmosphere and the atmospheric transmittance τ of the solar radiation incident path in formula (4) are s , atmospheric transmittance τ of the observation path v And the transmission path radiation spectrum radiance E in formula (3) p All of them can be obtained using typical atmospheric radiation transfer calculation codes, while the surface temperature, emissivity and bidirectional reflectivity of the surface cover can be obtained by querying geographic maps.
[0051] Step 3: Use the atmospheric radiation transfer code to calculate the solar spectral irradiance used in the simulated image, and compare it with the solar spectral irradiance in the spectrum band measured by the remote sensing satellite. Calibrate the solar irradiance value of the simulated image to be consistent with the actual measurement.
[0052] The solar spectrum radiance brightness given in the atmospheric radiation transfer code is an empirical value, while the solar spectrum radiance brightness in the actual observation scene varies with the distance between the sun and the earth and the solar activity. Therefore, in the visible light to short-wave infrared band, the brightness value of the simulated image X needs to be calibrated to make it as close as possible to the brightness value of the measured image Y:
[0053]
[0054] Where Lx′ ij is the simulated image after radiation brightness calibration; E′0 is the solar irradiance value in the spectrum given by the measured data product; Lx ij is the pixel unit x in the simulation image calculated using formula (3) ij radiance.
[0055] Step 4: Calculate the structural similarity (SSIM) of each pixel unit of the simulated and measured images to generate an SSIM image.
[0056] Image structure similarity (Structure SIMilarity, SSIM) is an indicator commonly used to evaluate the similarity between two images in digital image processing. When generating an SSIM image, a two-dimensional matrix with a pixel size of 11×11 is extracted with the pixel unit to be evaluated as the center:
[0057]
[0058] In the formula, xij With y ij They are all 11×11 two-dimensional matrices extracted from the simulated and measured images, centered on the pixel unit in the i-th row and j-th column, that is:
[0059]
[0060] In the formula, x′ (u)(v) and y′ (u)(v) They represent the values of the u-th row and v-th column matrix elements in the 11×11 two-dimensional matrix of simulated and measured size after extraction.
[0061] Then a circularly symmetric weight matrix W of size 11×11 is used 11×11 Weight the extracted two-dimensional matrix and calculate its SSIM value:
[0062]
[0063] And there are:
[0064]
[0065]
[0066]
[0067] In the formula, w (u)(v) Represents the element in the uth row and vth column of the weight coefficient matrix W; and They are two-dimensional matrices x ij With y ij The mean of and They are two-dimensional matrices x ij With y ij The variance of is a two-dimensional matrix x ij With y ij The mutual correlation coefficient of ; K1 and K2 are constants, and their empirical values are 0.01 and 0.03 respectively; L is the dynamic range of the image, which depends on the maximum and minimum values of the simulated and measured images:
[0068] L=max((max(X)-min(X)),(max(Y)-min(Y))) (13)
[0069] Where max(·) is the maximum value operation; min(·) is the minimum value operation.
[0070] After calculating the SSIM value of each pixel unit, the SSIM image can be obtained. Its image size is (M-10)×(N-10) pixel units, which is slightly smaller than the size of the measured remote sensing image. The size of the SSIM image can be made consistent with the measured remote sensing image by padding zeros around the simulated and measured remote sensing images.
[0071] Step 5: Set the threshold of structural similarity (SSIM) to determine whether each pixel unit is the ground surface or the cloud layer.
[0072] SSIM describes the difference between the simulated and measured images: the closer the SSIM is to 1, the closer they are, and the closer it is to 0, the greater the difference between them. Considering that the scattered radiation characteristics of the ground cover background change slowly compared to the cloud layer, when the value of a certain pixel unit in the SSIM image is large, the area can be considered to be the ground (no cloud); when the value of a certain pixel unit in the SSIM image is small, the area can be considered to be the cloud layer:
[0073]
[0074] Where γ is the SSIM threshold for determining whether a pixel unit is a cloud layer.
[0075] Step 6: Output the cloud detection result map.
[0076] The detection results obtained in step 5 can be output in different ways. For example, the cloud-free area in the measured image is set to zero to obtain a surface mask image, or the cloud area in the measured image is set to zero to obtain a cloud mask image, or the cloud coverage area data is output in combination with the longitude and latitude coordinates of the surface projection point corresponding to each pixel unit to provide information for subsequent identification of cloud types and forecasting of meteorological conditions in the area of interest.
[0077] Example:
[0078] The following uses a short-wave infrared wavelength 2.125μm spectrum band, and uses the method of the present invention to process an AGRI (Advanced Geostationary Radiation Imager) sensor image carried by the Fengyun-4B satellite and a MODIS (Moderate Resolution Imaging Spectroradiometer) sensor image carried by the Aqua satellite, respectively, to further illustrate the present invention. Among them, the Fengyun-4B satellite is a geostationary orbit satellite with an orbit altitude of about 35786km, and the subsatellite point is located at the equator and 133° east longitude. The spectral response function passband of the AGRI sensor in the selected spectrum band is 2.10~2.35μm, the angular resolution is about 56μrad (the spatial resolution of the subsatellite point is about 2km), and the size of a scene image is 5496×5496 pixels. The Aqua satellite is a sun-synchronous orbit satellite with an orbit altitude of about 705.3km and a revisit period of 16 days. The spectral response function of the MODIS sensor in the selected spectrum has a passband of 2.105 to 2.155 μm, an angular resolution of about 709 μrad (the spatial resolution at the sub-satellite point is about 0.5 km), and the size of a scene image is 4060×2708 pixels.
[0079] Step 1: Extract the detection parameters such as time, position, field of view, spectrum, etc. of the remote sensing images acquired by the satellite-borne optical sensor.
[0080] The start time of the image acquisition by the AGRI sensor of the Fengyun-4B satellite is 1:30:00 on May 30, 2023, Universal Time (Greenwich Mean Time), and the end time is 1:44:59 on May 30, 2023, Universal Time; the orbital position is the geostationary orbit above 133° East longitude (the orbital position remains unchanged during the imaging process), the instantaneous field of view is 56 μrad, the detection spectrum is 2.10-2.35 μm, and the remote sensing image is shown in Figure 2(a), the unit is W / m 2 / sr / μm. The Aqua satellite MODIS sensor started acquiring images at 05:30:00 on January 19, 2015, and ended at 05:34:59 on January 19, 2015, with the orbit altitude maintained at 705.3 km. The subsatellite position moved from the starting point of 117.945°E and 23.266°N to 113.002°E and 41.261°N (ascending orbit). The instantaneous field of view was 709 μrad, and the detection spectrum was 2.105-2.155 μm. The remote sensing image is shown in Figure 2(b), and the unit is W / m 2 / sr / μm.
[0081] Step 2: Using the Earth background hyperspectral imaging simulation model, set the model parameters to the extraction results of step 1 to generate a cloud-free radiance image.
[0082] The short-wave infrared wavelength 2.125μm band has high transmittance, and the scattering of atmospheric gas molecules and aerosol particles is very weak. Therefore, the radiance received by the sensor in this band mainly comes from the scattering of the ground surface and clouds. Figure 3(a) shows the simulated cloud-free Earth disk radiance image of the 2.10-2.35μm band of the AGRI sensor of the Fengyun-4B satellite. The unit is the same as the measured image, both W / m 2 / sr / μm. Figure 3(b) shows the simulated cloud-free earth background radiance image of the Aqua satellite MODIS sensor in the 2.105-2.155μm spectral band, in W / m 2 / sr / μm.
[0083] Step 3: Use the atmospheric radiation transfer code to calculate the solar spectral irradiance used in the simulated image, and compare it with the solar spectral irradiance in the spectrum band measured by the remote sensing satellite. Calibrate the solar irradiance value of the simulated image to be consistent with the actual measurement.
[0084] Figure 4 The figure shows the solar spectrum irradiance curve of wavelength 2.10-2.35μm used in simulating cloudless images. The dashed line is the solar spectrum irradiance at the top of the atmosphere, the dotted line is the solar spectrum irradiance at the bottom of the atmosphere, and the solid line is the irradiance formed by the atmosphere scattering solar radiation to the surface. For the wavelength 2.10-2.35μm spectral band of the AGRI sensor of the Fengyun-4B satellite, the solar spectrum irradiance at the top of the atmosphere is 25.42W / m 2 / sr / μm, the solar spectrum irradiance at the bottom of the atmosphere is 15.47W / m 2 / sr / μm, the average atmospheric transmittance in the spectral band is 0.61; for the wavelength 2.105~2.155μm spectral band of the Aqua satellite MODIS sensor, the solar spectrum irradiance at the top of the atmosphere is 29.54W / m 2 / sr / μm, the solar spectrum irradiance at the bottom of the atmosphere is 19.55W / m 2 / sr / μm, and the average atmospheric transmittance in the spectrum is 0.66. According to the solar radiation brightness value given by the measured image, the overall brightness value of the simulated image is calibrated using formula (6).
[0085] Step 4: Calculate the structural similarity (SSIM) of each pixel unit of the simulated and measured images to generate an SSIM image.
[0086] Figure 5(a) and Figure 5(b) are the observation scenes of the AGRI sensor of the Fengyun-4B satellite and the MODIS sensor of the Aqua satellite, respectively. The SSIM images obtained by comparing the structural similarity of the simulated and measured images. The bright areas in the image indicate that the simulated and measured pixel units are highly similar, and the dark areas indicate that the two are very different. Since there are no clouds in the simulated images, the dark areas in the SSIM images are cloud areas, and the bright areas are cloudless areas.
[0087] Step 5: Set the structural similarity (SSIM) threshold to determine whether each pixel unit is the ground surface or the cloud layer.
[0088] The cloud layer detection SSIM threshold γ in equation (14) is set to 0.6. Figures 6(a) and 6(b) are the cloud layer detection results of the Fengyun-4B satellite AGRI sensor image and the Aqua satellite MODIS sensor image, respectively. The bright area with a pixel unit value of 1 in the figure represents the cloud layer.
[0089] Step 6: Output the cloud detection result map.
[0090] The cloud detection result map has many uses, and simulating the earth background radiation image containing clouds is an important application. Taking the observation scene of the Aqua satellite MODIS sensor as an example, the simulated 2.105-2.155μm spectral image containing clouds is as follows Figure 7 shown.
[0091] The present invention may also employ alternative schemes that can also accomplish the purpose of the invention: employing image data from other satellite-borne optical sensors or other detection spectrum bands to simulate cloud-free images of the same scene, generate SSIM images, and complete cloud detection; simulating data from two spectrum bands to generate a simulated data color ratio diagram, and completing cloud detection by generating an SSIM image of the measured and simulated data color ratio diagrams; other image similarity evaluation methods, such as correlation coefficient, probability density, etc., may be employed in the implementation scheme to generate a similarity image and complete cloud detection.
Claims
1. A cloud detection method for spaceborne optical remote sensing images driven by imaging simulation model data, characterized in that: The steps include: Step 1: extracting detection parameters of the satellite-borne optical sensor to obtain remote sensing images, wherein the detection parameters include time, position, field of view, and spectrum; Step 2: Using the Earth background hyperspectral imaging simulation model, set the model parameters to the detection parameters extracted in step 1 to generate a cloud-free radiance image, including: The remote sensing image acquired by the spaceborne optical sensor is represented as a two-dimensional matrix. Assume that the simulated image is X and the measured image is Y, and the number of pixel units of the two images is M×N. For a pixel unit x in the simulated image X, ij The corresponding sensor instantaneous field of view entrance pupil radiance Lx ij , expressed as a spectral discrete summation: (3) Where λ represents the wavelength; λ1 and λ2 are the start and end wavelengths of the spectrum integration respectively; Δλ is the spectrum sampling interval; E p E is the radiant spectral emittance of the transmission path; g is the surface spectral radiance, calculated by the simplified expression: (4) Where E0 is the solar spectral irradiance at the top of the atmosphere; τ s With τ v are the atmospheric transmittances of the solar radiation incident path and observation path respectively; ρ r is the bidirectional spectral reflectance of the surface, which is related to the geometric relationship and terrain undulation; ε g is the surface spectral emissivity; T g is the surface temperature; B is the spectral radiance of the object with temperature T; Step 3: Use the atmospheric radiation transfer code to calculate the solar spectrum irradiance used in the simulation image, and compare it with the solar spectrum irradiance in the spectrum band measured by the remote sensing satellite. Calibrate the value of the average solar spectrum irradiance in the spectrum band of the simulation image to be consistent with the actual measurement; Step 4: Calculate the structural similarity of each pixel unit of the simulated and measured images to generate an SSIM image. SSIM represents structural similarity, including: Centered on the pixel unit to be evaluated, extract a two-dimensional matrix with a pixel size of 11×11; ij With y ij They are all 11×11 two-dimensional matrices extracted from the simulated and measured images, centered on the pixel unit in the i-th row and j-th column, that is: (8) In the formula, x′ (u)(v) and y′ (u)(v) Respectively represent the values of the matrix elements in the uth row and vth column in the simulated and measured two-dimensional matrices after extraction; Then, a 11×11 circular symmetric weight matrix W is used 11×11 Weight the extracted two-dimensional sub-graph matrix and calculate its SSIM value: (9) And there are: (10) (11) (12) In the formula, w (u)(v) Represents the element in the uth row and vth column of the weight coefficient matrix; and are the two-dimensional matrices x of the subgraphs respectively. ij With y ij The mean of and are the two-dimensional matrices x of the subgraphs respectively. ij With y ij The variance of is the subgraph 2D matrix x ij With y ij The mutual correlation coefficient of ; K1 and K2 are constants, and their empirical values are 0.01 and 0.03 respectively; L is the dynamic range of the two images: (13) Wherein, max(·) is the maximum value operation; min(·) is the minimum value operation; X is the simulated image, and Y is the measured image; Step 5: Set the structural similarity threshold to determine whether each pixel unit is the surface or the cloud layer; Step 6: Output the cloud detection result map.
2. The method for cloud layer detection of spaceborne optical remote sensing images driven by imaging simulation model data according to claim 1, characterized in that: The step 5 includes: the SSIM value of each pixel unit is between 0 and 1. When the SSIM value of a pixel unit is large, the pixel unit is determined to be clear sky; when the SSIM value of a pixel unit is small, the pixel unit is determined to be cloud layer, and: (14) Where γ represents the threshold of the SSIM value for determining whether a pixel unit is a cloud layer.
