Satellite-borne hyperspectral cloud layer identification method based on dual-band fusion technology

Abstract

The invention relates to the technical field of meteorological cloud detection, and discloses a satellite-borne hyperspectral cloud layer identification method based on a dual-band fusion technology, and the method comprises the steps: firstly obtaining a hyperspectral image, carrying out the preprocessing of the hyperspectral image, and obtaining a spectrum calibration image; respectively calculating the visible light wave band transmittance and the near-infrared wave band transmittance of each pixel in the spectrum calibration image; identifying a semitransparent weather phenomenon area and a non-semitransparent weather phenomenon area according to the visible light wave band transmittance and the near-infrared wave band transmittance of each pixel; and finally, constructing a real cloud layer spectral feature model, performing real cloud layer identification on the non-translucent weather phenomenon area by using the model, and fusing the identified real cloud layer area and the translucent weather phenomenon area to generate a final cloud detection result graph. According to the method, spectral information of visible light and near-infrared wavebands is fused, a real cloud layer spectral feature model is constructed for cloud detection, and the cloud detection precision under complex atmospheric conditions is improved.

Description

Technical Field

[0001] The present invention relates to the field of meteorological cloud detection technology, and more specifically, to a satellite-borne hyperspectral cloud layer recognition method based on dual-band fusion technology. Background Art

[0002] In the field of meteorological cloud detection technology, cloud detection using satellite-borne hyperspectral images has always been an important research topic. Traditional cloud detection methods mainly use spectral information in the visible light band for cloud detection, but under complex atmospheric conditions, it is difficult to accurately identify clouds using only visible light information.

[0003] The patent with authorization announcement number CN103604421B discloses a satellite-borne hyperspectral image cloud detection method based on spectral reflectance characteristics. This method obtains image data of a specific wavelength and the solar altitude angle, calculates the reflectance of the image of a specific wavelength, divides the entire area into a high cloud pending area and a medium and low cloud pending area, and obtains the final cloud detection result by excluding non-cloud areas. This method uses spectral reflectance information for cloud detection, which improves the accuracy of cloud detection to a certain extent. However, this method only considers the spectral characteristics of the visible light band, ignores the cloud information contained in the near-infrared band, and is difficult to accurately identify semi-transparent clouds. In addition, this method requires excluding multiple types of land objects one by one, which is relatively complicated to implement.

[0004] The patent application with publication number CN116026775A discloses a cloud detection method and device based on multispectral thresholds. The method first obtains multispectral remote sensing data from near-ultraviolet bands to short-wave infrared bands, and then uses different threshold combinations for multispectral data of different surface types to perform cloud detection. This method solves the limitations of single threshold cloud detection to a certain extent, but still does not consider the fusion of spectral information of visible light bands and near-infrared bands, and requires separate processing of different surfaces, which is relatively cumbersome to implement.

[0005] In summary, the existing satellite-borne hyperspectral image cloud detection technology only uses the spectral information of the visible light band or a single band, ignoring the important role of the near-infrared band in cloud identification. It lacks an effective method to integrate the spectral information of the visible light band and the near-infrared band, making it difficult to accurately identify complex weather phenomena such as translucent clouds. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology. The method obtains a hyperspectral image containing visible light and near-infrared bands, calculates the transmittance of each pixel in the two bands, and identifies the semi-transparent weather phenomenon area based on the transmittance difference, and then constructs a real cloud spectral feature model to further distinguish the non-semi-transparent area, and finally realizes the refined cloud detection of visible light and near-infrared band fusion. The method overcomes the shortcomings of the prior art and can provide more accurate and reliable cloud detection results under complex atmospheric conditions.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] Acquire a hyperspectral image, preprocess the hyperspectral image, and obtain a spectral calibration image; the hyperspectral image includes a visible light band image and a near infrared band image; respectively calculate the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image; and record the solar altitude angle when the hyperspectral image is acquired;

[0009] According to the visible light band transmittance and near infrared band transmittance of each pixel in the spectral calibration image, the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area are identified;

[0010] A real cloud spectral feature model is constructed, and the real cloud spectral feature model is used to identify the real cloud in the non-translucent weather phenomenon area to obtain the real cloud area. The real cloud area is then merged with the translucent weather phenomenon area to generate the final cloud detection result map.

[0011] Furthermore, the step of respectively calculating the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image comprises:

[0012] Select n reference object areas on the spectral calibration image, each reference object area contains N1 pixels, and different reference object areas are distributed at different positions of the spectral calibration image; extract the radiance values ​​of all pixels in each reference object area in the visible light band, and calculate the measured spectral reflectance of each pixel in each reference object area in the visible light band according to the radiance value of each pixel in the visible light band and the solar altitude angle; extract the radiance values ​​of all pixels in the near-infrared band in each reference object area, calculate the solar zenith angle from the solar altitude angle, and calculate the measured spectral reflectance of each pixel in the near-infrared band in each reference object area according to the radiance value and the solar zenith angle of each pixel in the near-infrared band; n is a positive integer;

[0013] Calculate the theoretical spectral reflectance of each pixel in the visible light band of each reference object area under cloud-free atmospheric conditions; Calculate the theoretical spectral reflectance of each pixel in the near-infrared band of each reference object area under cloud-free atmospheric conditions;

[0014] The ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in the visible light band of each reference object area is used as the first regionalized variable, and an interpolation model of the transmittance of the visible light band is established to obtain the transmittance of the visible light band at any pixel position in the spectral calibration image;

[0015] The ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in the near-infrared band of each reference object area is taken as the second regionalized variable, and an interpolation model of the near-infrared band transmittance is established to obtain the transmittance of each pixel in the spectral calibration image in the near-infrared band.

[0016] Further, the identifying of the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area includes:

[0017] Calculate the transmittance difference between the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image, and determine the difference threshold T th ;like >T th , then the pixel Mark as a candidate area for semi-transparent weather phenomena; otherwise, Marked as non-translucent weather phenomenon area; Pixels in the spectral calibration image The transmittance difference between the visible light band transmittance and the near infrared band transmittance, is the index of the pixel in the spectral calibration image;

[0018] Taking pixels as units, spatial clustering is performed according to the spatial neighborhood relationship of the candidate areas of semi-transparent weather phenomena to obtain semi-transparent weather phenomenon areas and non-translucent weather phenomenon areas.

[0019] Furthermore, the determination of the difference threshold T th The methods include:

[0020] Obtain historical cloud images, and extract the visible light band transmittance and near infrared band transmittance in the semi-transparent weather phenomenon area and the real cloud layer area of ​​each historical cloud image in units of pixels;

[0021] The average difference between the visible light band transmittance and the near-infrared band transmittance of all pixels in the semi-transparent weather phenomenon area of ​​the historical cloud image is calculated, which is recorded as the first difference; the average difference between the visible light band transmittance and the near-infrared band transmittance of all pixels in the real cloud layer area of ​​the historical cloud image is calculated, which is recorded as the second difference; according to the first difference and the second difference, the difference threshold T is determined. th .

[0022] Furthermore, performing spatial clustering according to the spatial neighborhood relationship of the semi-transparent weather phenomenon candidate areas includes:

[0023] Taking pixels as units, the spatial coordinates of the candidate areas of semi-transparent weather phenomena are extracted, and an undirected graph is constructed. Each node in the undirected graph corresponds to a pixel in the candidate area, and there are edge connections between adjacent pixels in the candidate area.

[0024] Calculate the weight of the edge in the undirected graph, where the weight of the edge is the spectral feature distance between adjacent candidate area pixels;

[0025] The undirected graph is divided to obtain the set of cut edges with the minimum weight and the smallest sum. According to the set of cut edges, the pixels in the candidate area of ​​semi-transparent weather phenomenon are divided into N2 connected subgraphs. The pixels in the connected subgraphs are marked as semi-transparent weather phenomenon areas, and the remaining pixels are marked as non-semi-transparent weather phenomenon areas.

[0026] Furthermore, the constructing of a real cloud layer spectral feature model includes:

[0027] Extracting cloud regions from historical cloud images to obtain cloud regions of the historical cloud images;

[0028] Extract the spectral features of the cloud area in the historical cloud image and construct the spectral feature vector of the real cloud layer; the spectral features include spectral reflectance features, spectral absorption features and spectral scattering features in the visible light band and near infrared band;

[0029] The spectral feature vectors of real clouds are used as samples to train and generate a real cloud spectral feature model.

[0030] Furthermore, the method of using the real cloud spectral feature model to identify the real cloud layer in the non-translucent weather phenomenon area includes:

[0031] The spectral features of the spectral calibration image of the non-translucent weather phenomenon area are extracted by sliding window, the spectral features of the area to be evaluated in the semi-transparent weather phenomenon area are obtained, and the spectral feature vector of the area to be evaluated is constructed; the spectral feature vector of the area to be evaluated is input into the real cloud layer spectral feature model to obtain the probability that the area to be evaluated is a real cloud layer area; if the probability that the area to be evaluated is a real cloud layer area is greater than a preset probability threshold, the area to be evaluated is determined to be a real cloud layer area, otherwise it is determined to be a non-real cloud layer area.

[0032] A spaceborne hyperspectral cloud recognition system based on dual-band fusion technology, which is used to implement the above-mentioned spaceborne hyperspectral cloud recognition method based on dual-band fusion technology, the system comprises:

[0033] Spectral image calibration module: used to obtain a hyperspectral image, pre-process the hyperspectral image, and obtain a spectral calibration image; the hyperspectral image includes a visible light band image and a near-infrared band image; respectively calculate the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image; record the solar altitude angle when the hyperspectral image is obtained;

[0034] Semi-transparent weather phenomenon area determination module: used to identify the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area according to the visible light band transmittance and near infrared band transmittance of each pixel in the spectral calibration image;

[0035] Real cloud detection module: used to build a real cloud spectral feature model, use the real cloud spectral feature model to identify the real cloud in the non-translucent weather phenomenon area, obtain the real cloud area, merge the real cloud area with the translucent weather phenomenon area, and generate the final cloud detection result map.

[0036] An electronic device comprises a memory, a central processing unit and a computer program stored in the memory and executable on the central processing unit. When the central processing unit executes the computer program, the above-mentioned satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology is implemented.

[0037] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, the above-mentioned satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology is implemented.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention combines the spectral information of visible light and near-infrared bands for cloud detection, and comprehensively utilizes the different spectral characteristics of the two bands, thereby improving the accuracy of cloud detection. The visible light and near-infrared bands have different penetration capabilities and reflection characteristics for atmospheric phenomena such as clouds, fog, and haze. The combination of the two bands can more accurately identify different types of clouds. A spectral feature model of real cloud layers is constructed for cloud detection. The model takes into account the reflectivity, absorption and scattering characteristics of clouds in visible light and near-infrared bands, and can more comprehensively characterize the spectral characteristics of real cloud layers, thereby improving the recognition ability of real cloud layers. The model is used to further identify the non-translucent weather phenomenon area, and the real cloud layer under the translucent cloud can be detected. A method for identifying the translucent weather phenomenon area based on the difference in transmittance between the visible light and near-infrared bands is proposed. According to the difference in transmittance between the two bands, it can be preliminarily judged whether there are translucent weather phenomena such as thin clouds and fog. On this basis, the extraction results of the translucent phenomenon area are further optimized by spatial clustering. When calculating the transmittance, by selecting reference objects at multiple different locations and comparing the spectral reflectance of the reference objects under cloudy and cloudless conditions to obtain the transmittance, the atmospheric transmittance of different areas of the whole map is adaptively estimated, reducing the influence of atmospheric effects on the transmittance calculation. Using a sliding window to extract local spectral features of the image can cope with the situation of uneven cloud distribution under complex weather conditions, improve the accuracy of local cloud detection, and make the cloud detection results more refined. In general, the present invention forms a set of cloud detection methods that can cope with complex weather conditions by integrating visible light and near-infrared band information, constructing a real cloud spectrum model, and refining the estimation of atmospheric transmittance. It improves the accuracy of cloud detection while reducing the misjudgment rate and has good application prospects. In the fields of meteorological monitoring, satellite remote sensing, etc., the present invention can provide a new technical means for obtaining high-quality cloud detection products. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 It is a principle flow chart of a satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology in the present invention;

[0042] Figure 2 It is a flow chart of a method for performing radiometric calibration and geometric correction on a hyperspectral image in a satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology of the present invention;

[0043] Figure 3A flow chart of a method for respectively calculating the visible light band transmittance and the near infrared band transmittance of each pixel in a spectral calibration image in a spaceborne hyperspectral cloud recognition method based on dual-band fusion technology of the present invention;

[0044] Figure 4 A flow chart of a method for determining a difference threshold in a spaceborne hyperspectral cloud recognition method based on dual-band fusion technology of the present invention;

[0045] Figure 5 A flow chart of a method for performing spatial clustering according to the spatial neighborhood relationship of candidate areas of semi-transparent weather phenomena in a satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology of the present invention;

[0046] Figure 6 A flow chart of a method for establishing a real cloud spectral feature model in a satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology of the present invention;

[0047] Figure 7 This is a functional module diagram of a satellite-borne hyperspectral cloud recognition system based on dual-band fusion technology in the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Example 1

[0050] See also Figure 1 As shown, this embodiment provides a spaceborne hyperspectral cloud recognition method based on dual-band fusion technology, including:

[0051] Step S1000, acquiring a hyperspectral image, preprocessing the hyperspectral image, and obtaining a spectral calibration image; the hyperspectral image includes a visible light band image and a near infrared band image; respectively calculating the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image; and recording the solar altitude angle when the hyperspectral image is acquired;

[0052] Furthermore, step S1000 includes:

[0053] Step S1100, acquiring a hyperspectral image, recording the solar altitude angle when the hyperspectral image is acquired, and performing radiation calibration and geometric correction on the hyperspectral image;

[0054] Furthermore, if Figure 2As shown, step S1100 includes:

[0055] Step S1110, obtaining a satellite-borne hyperspectral image of the area to be detected, wherein the hyperspectral image includes a visible light band image and a near infrared band image;

[0056] Step S1120, performing radiometric calibration on the visible light band image and the near infrared band image in units of pixels, and converting the DN value into a radiance value;

[0057] Step S1130 , geometrically correcting the visible light band image and the near infrared band image after radiation calibration in units of pixels.

[0058] Specifically, step S1100 is the data basis for cloud detection. First, the satellite-borne hyperspectral imaging data of the area to be detected is obtained, and the visible light and near-infrared dual-band sensors are used for synchronous imaging to obtain visible light and near-infrared images of the same phase and spatial resolution. The visible light band usually selects 0.4-0.9μm, which provides rich spectral information of ground objects; the near-infrared band selects 0.9-1.4μm, which has strong reflection on clouds, snow, etc. and can be used for cloud detection. The original image is recorded in the form of DN value (Digital Number), and radiation calibration is required to convert the DN value into a radiance value with a clear physical meaning. The calibration process requires the use of calibration coefficients to consider the influence of factors such as sensor gain, bias, dark current, and bad pixels. Geometric distortion will be introduced during the image acquisition process due to factors such as satellite attitude changes and earth curvature, and geometric precision correction is required. Using ground control points, the mapping relationship between the image coordinate system and the geographic coordinate system is established, and the orthographic projection image is generated by resampling to provide a registered multi-band image for subsequent processing. As an important parameter to characterize the solar radiation angle, the solar altitude angle will significantly affect the radiation brightness value of the target. Therefore, the solar altitude angle information needs to be recorded when acquiring data for subsequent reflectivity calculation.

[0059] Step S1100 acquires visible light and near-infrared band data through dual-band synchronous imaging, expands the data source, and provides rich spectral information for cloud detection. Radiometric calibration eliminates inherent sensor errors, and geometric precision correction realizes image registration, ensuring the feasibility of subsequent fusion processing. Obtaining high-quality radiance images after calibration and registration is a prerequisite for cloud detection. These preprocessing tasks lay a data foundation for subsequent spectral analysis and band fusion, and are crucial to improving cloud detection accuracy. At the same time, mature radiometric calibration and geometric correction algorithms are adopted, and with the help of high-precision calibration fields and control points, processing efficiency is taken into account while ensuring accuracy, shortening the data processing cycle and improving the timeliness of cloud detection. High-quality preprocessed data is the starting point of cloud detection, providing reliable input for subsequent spectral transmittance calculations, band fusion and other processing links, and ensuring the reliability of cloud detection results from the source.

[0060] Step S1200, performing atmospheric correction on the geometrically corrected visible light band image and near infrared band image in units of pixels to obtain a spectrally calibrated image;

[0061] Specifically, step S1200 uses the 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) model for atmospheric correction. 6S is a classic atmospheric radiation transfer model that can accurately simulate the scattering and absorption process of solar radiation through the atmosphere. By inputting the sun-sensor geometric parameters, atmospheric model parameters (such as aerosol optical thickness, ozone content, etc.), surface reflection characteristics, etc., the 6S model can output atmospheric parameters such as atmospheric transmittance, path radiance, and surface coupling terms. Using these parameters, a radiation transfer equation between scene radiance and surface reflectivity is established to solve the surface reflectivity, thereby achieving atmospheric correction, eliminating atmospheric effects, and obtaining the true spectral reflection characteristics of the surface. Atmospheric correction can significantly improve the quality of spectral data and provide more reliable spectral criteria for cloud detection. At the same time, the 6S model uses a fast lookup table to improve the calculation efficiency while ensuring accuracy, so that even large-scale hyperspectral data can achieve fast and accurate atmospheric correction processing.

[0062] Step S1200 obtains the real spectral reflectance data of the surface by considering the atmospheric effect. Atmospheric scattering and absorption will change the radiation signal of the surface and introduce errors. By establishing an atmospheric radiation transmission model, simulating the interaction between solar radiation and atmospheric components, separating the atmospheric contribution and the surface contribution, realizing the inversion of surface reflectance, restoring the essential spectral characteristics of the ground objects, and improving the reliability of spectral analysis. Atmospheric correction of spectral signals is the basis of quantitative application of remote sensing, and is of great significance to improving the spectral discrimination ability of cloud detection. The atmospheric correction method based on physical processes overcomes the limitations of empirical methods, has better adaptability and robustness, and provides support for cloud detection under complex atmospheric conditions. At the same time, the differences in the temporal and spatial distribution of atmospheric parameters are fully considered, and a pixel-by-pixel atmospheric correction strategy is adopted to improve the refinement of cloud detection. High-precision atmospheric correction is the premise of spectral analysis. Reliable surface reflectance data is obtained through advanced methods such as the 6S model, thereby providing data support for cloud detection based on spectral characteristics.

[0063] Step S1300 , respectively calculating the transmittance of each pixel in the spectral calibration image in the visible light band and the transmittance of the near infrared band.

[0064] Furthermore, if Figure 3 As shown, step S1300 includes:

[0065] Step S1310, selecting n reference object regions on the spectral calibration image, each reference object region including N1 pixels, and different reference object regions are distributed at different positions of the spectral calibration image;

[0066] Step S1320, extracting the radiance values ​​of all pixels in each reference object region in the visible light band, and calculating the measured spectral reflectance of each pixel in the visible light band in each reference object region according to the radiance value of each pixel in the visible light band and the solar altitude angle;

[0067] Step S1330, extracting the radiance values ​​of all pixels in each reference object region in the near-infrared band, calculating the solar zenith angle from the solar altitude angle, and calculating the measured spectral reflectance of each pixel in the near-infrared band in each reference object region according to the radiance value and solar zenith angle of each pixel in the near-infrared band;

[0068] The method for calculating the measured spectral reflectance of each pixel in each reference object area in the visible light band includes:

[0069]

[0070] in:

[0071] : The measured spectral reflectance of pixel i in the reference feature region j in the visible light band; j is the index of the reference feature region; i is the index of the pixel in the reference feature region j; 1≤j≤n, 1≤i≤N1;

[0072] : The radiance value of the visible light band received by pixel i in the reference object area j can be obtained by radiation calibration; The larger it is, the higher the spectral reflectance is;

[0073] : The atmospheric path radiation of pixel i in reference area j in the visible light band can be simulated by the MODTRAN model; It refers to the additional radiation formed when solar radiation is scattered by atmospheric components when passing through the atmosphere. The introduction of can deduct the influence of atmospheric effects and restore the true reflectivity; The larger the value, the stronger the atmospheric scattering effect is, and the greater the degree of overestimation of reflectivity is. The smaller the reflectivity obtained;

[0074] : Earth-space distance, that is, the distance between the sun and the earth, which can be obtained by checking the astronomical almanac according to the imaging time; It reflects the difference in solar irradiance as the distance between the sun and the earth changes; when the earth revolves around the sun, the distance between the sun and the earth will change, which in turn affects the solar irradiance reaching the earth's surface. 2 Inversely proportional to the irradiance, the larger d is, the smaller the irradiance is, and vice versa; introducing d 2 The effect of irradiance changes on reflectance calculations can be corrected;

[0075] : Solar irradiance in the visible light band; Characterizes the radiation intensity of the sun in the visible light band and is a normalization factor when calculating reflectivity; The larger it is, the stronger the incident irradiance is and the lower the reflectivity is.

[0076] : solar zenith angle, , is the solar altitude angle; the solar zenith angle represents the angle between the sun's incident direction and the surface normal, which affects the solar irradiance reaching the surface; The larger it is, the more oblique the sunlight is and the weaker the irradiance is, and vice versa; It represents the attenuation trend of irradiance with the incident angle. The bigger, The smaller it is, the more obvious the irradiance attenuation is; It can correct the irradiance difference caused by the change of solar angle and improve the angle normalization effect of reflectivity calculation.

[0077] Reflectance and radiance values Positive correlation, The larger it is, the greater the reflectivity;

[0078] Reflectivity and Path Radiance Negative correlation, The larger it is, the smaller the reflectivity;

[0079] The reflectivity is negatively correlated with the distance d between the sun and the earth. The larger d is, the smaller the reflectivity is.

[0080] Reflectivity and Solar Irradiance Negative correlation, The larger it is, the smaller the reflectivity;

[0081] Reflectivity and solar zenith angle Positive correlation, The bigger, The smaller it is, the greater the reflectivity.

[0082] The whole formula is passed represents the contribution of surface radiation brightness after deducting the atmospheric path radiation; then combined with the solar irradiance ( ), distance between the sun and the earth (d), solar zenith angle (θ S) and other parameters to finally obtain the true reflectivity. The introduction of distance and angle correction factors can eliminate the influence of imaging conditions and obtain comparable reflectivity results, which is an important basis for quantitative remote sensing analysis. On the one hand, the atmospheric path radiation is deducted to restore the intrinsic reflectivity characteristics of the ground objects; on the other hand, the introduction of multiple correction factors improves the radiation, distance, and angle normalization capabilities of the reflectivity calculation.

[0083] The measured spectral reflectance of each pixel in each reference feature area in the near-infrared band is obtained using the same calculation method as the visible light band; by selecting multiple reference areas distributed in different positions of the whole map and covering different types of features, the spectral differences of the surface of the area to be detected can be comprehensively considered. When calculating the measured reflectance, it is necessary to obtain accurate radiance, solar irradiance and solar zenith angle data to eliminate the influence of solar radiation and atmospheric effects. The difference between the measured reflectance and the true reflectance of the surface reflects the degree of attenuation of the spectral signal by the atmosphere and is the basis for estimating the spectral transmittance. Using multiple reference feature areas, the spectral characteristics of different features are integrated to improve the spatial representativeness of the estimation results. The selected reference areas have uniform spectra, rich types, and no cloud interference, which ensures the reliability of the measured reflectance. The acquisition of measured reflectance is an important link in the estimation of spectral transmittance and lays a data foundation for subsequent analysis.

[0084] Step S1340, calculating the theoretical spectral reflectance of each pixel in the visible light band of each reference object area under cloudless atmospheric conditions;

[0085] Step S1350, calculating the theoretical spectral reflectance of each pixel in the near-infrared band of each reference object area under cloudless atmospheric conditions;

[0086] Specifically, the radiative transfer model is used to calculate the theoretical spectral reflectance of each pixel in the visible light band and the theoretical spectral reflectance of each pixel in the near-infrared band in each reference area under cloudless atmospheric conditions. MODTRAN (Moderate Resolution Atmospheric Transmission) is selected as the radiative transfer model, which can accurately simulate the interaction between solar radiation and atmospheric components to obtain atmospheric parameters and surface reflectance. The input parameters of MODTRAN include atmospheric profile (air pressure, temperature, humidity, etc.), aerosol type and content, sun-sensor position geometry, surface elevation and reflectivity, etc., and the output parameters include atmospheric transmittance, path radiance, hemispheric albedo, etc. Using these parameters, the radiative transfer equation can be established. The MODTRAN model solves the radiative transfer equation by line-by-line integration, which requires detailed atmospheric profile data. Profile data can be obtained through synchronous sounding, or external data sources such as reanalysis data can be used. For aerosol parameters, you can choose the aerosol type that comes with the model, such as continental, oceanic, urban, etc., or you can input the measured aerosol optical thickness (AOD). The surface reflectance can be selected based on prior knowledge, such as 0.5 for vegetation and 0.2 for bare soil. The MODTRAN model has a high temporal and spatial resolution, and can simulate atmospheric conditions at different times and locations to obtain highly targeted theoretical radiance. After converting the radiance into theoretical reflectance, it can be directly compared with the measured reflectance.

[0087] The use of mature radiation transfer models to obtain theoretical reflectivity avoids the difficulty of synchronous field observations and is highly practical. The MODTRAN model is a standard model in the field of atmospheric correction. It has been verified for a long time, has reliable accuracy, and is suitable for different atmospheric conditions. By constructing a ground-atmosphere system and simulating the transmission process of solar radiation in the atmosphere, the atmospheric effect can be stripped off and the inherent spectral characteristics of the surface can be obtained. Converting radiance into reflectivity eliminates the influence of solar irradiance and facilitates comparison under different time and space conditions. The theoretical reflectivity obtained through model simulation provides a reference benchmark for the estimation of spectral transmittance.

[0088] Step S1360, taking the ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in each reference object region in the visible light band as the first regionalized variable, establishing an interpolation model of the transmittance of the visible light band, and obtaining the transmittance of the visible light band at any pixel position in the spectral calibration image;

[0089] Step S1370, taking the ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in each reference object area in the near-infrared band as the second regionalized variable, establishing an interpolation model of the near-infrared band transmittance, and obtaining the transmittance of each pixel in the spectral calibration image in the near-infrared band.

[0090] Specifically, step S1360 uses the measured reflectivity and theoretical reflectivity of each reference object in the visible light band to estimate the atmospheric transmittance in the visible light band at any position in the study area (area to be detected). By comparing the measured reflectivity of the reference object with the theoretical reflectivity, the degree of influence of the atmospheric effect on the spectral signal can be quantitatively characterized. The measured reflectivity is obtained under real atmospheric conditions and is affected by atmospheric absorption, scattering and other effects, while the theoretical reflectivity is simulated under ideal conditions assuming that the atmosphere is transparent and has no influence, representing the inherent spectral characteristics of the surface. The ratio of the two directly reflects the degree of attenuation of visible light radiation by the atmosphere. The smaller the ratio, the stronger the atmospheric effect and the lower the transmittance; conversely, the closer the ratio is to 1, the smaller the atmospheric effect and the higher the transmittance.

[0091] The measured reflectivity and theoretical reflectivity ratio of each reference object are used as known samples to construct a regional interpolation model, so as to obtain the transmittance estimation value of any location in the study area. Regionalized variables refer to variables that show a certain distribution pattern in space and are related to the location. By interpolating the known points, the attribute values ​​of the unknown points can be obtained. Common regionalized interpolation methods include Kriging, Inverse Distance Weighting (IDW), Spline, etc. Universal Kriging is selected for interpolation here. This method takes into account the spatial correlation and spatial trend of regionalized variables, is suitable for non-stationary data with global trends, and has high estimation accuracy. By establishing a spatial autocorrelation model of the measured reflectivity and theoretical reflectivity ratio, the correlation degree of regionalized variables at different distances is obtained, and then the unknown points are spatially weighted averaged to obtain the interpolation results. The construction of the interpolation model needs to comprehensively consider the spatial distribution characteristics of the variables, sampling density, scale effect and other factors, select appropriate theoretical models (such as Gaussian, exponential, etc.) and parameters (such as sill value, nugget value, etc.) for fitting, and perform cross-validation to ensure the reliability of the interpolation results.

[0092] The final visible light band transmittance interpolation result is the transmittance estimate at each pixel in the study area. The spatial resolution of the interpolation result is consistent with the original radiance image, reflecting the spatial distribution of transmittance. Local changes are closely related to factors such as terrain and underlying surface. The transmittance interpolation result can be used in subsequent cloud detection algorithms to provide a quantitative basis for spectral discrimination of clouds and ground objects.

[0093] The accuracy of transmittance estimation is affected by many factors, such as the selection of reference objects, the accuracy of radiation transfer models, and the spatiotemporal synchronization of atmospheric parameters. In order to improve the estimation accuracy, it is necessary to match the imaging conditions in time, space, and spectrum, and to set the necessary model parameters. Nevertheless, the transmittance interpolation results obtained based on the reference object ratio method can better reflect the spatiotemporal variation of transmittance compared with empirical models and statistical models, providing reliable prior knowledge for subsequent cloud detection algorithms and helping to improve the accuracy and reliability of cloud detection. At the same time, this method requires relatively few parameters and has high computational efficiency. It can achieve rapid estimation of transmittance and provides a feasible solution for operational cloud detection at regional or global scales.

[0094] The processing method of step S1370 is similar to that of step S1360, except that the visible light band is replaced by the near infrared band. The near infrared band has a strong reflection on targets such as clouds and snow, but a weak reflection on ground objects such as vegetation and soil. Therefore, compared with the visible light band, it can better highlight the spectral difference between clouds and ground objects, and is an important discriminant band for cloud detection.

[0095] For the near-infrared band, the ratio of the measured reflectivity to the theoretical reflectivity also reflects the level of atmospheric transmittance. The smaller the ratio, the stronger the attenuation of the atmosphere on the near-infrared radiation and the lower the transmittance; conversely, the closer the ratio is to 1, the less the near-infrared band is affected by the atmosphere and the higher the transmittance. The ratio of the measured reflectivity to the theoretical reflectivity of each reference object in the near-infrared band is used as a regionalized variable, and the same Kriging interpolation model as step S1360 is used to obtain the transmittance of each pixel position in the study area in the near-infrared band.

[0096] Transmittance estimation is an important input to the cloud detection algorithm, and its accuracy directly affects the final effect of cloud detection. Through steps S1360 and S1370, the spatial distribution information of transmittance is obtained in the two key bands of visible light and near-infrared, which not only takes into account the spatiotemporal variability of atmospheric effects, but also takes into account the differences in spectral responses, providing important physical parameters for subsequent cloud detection algorithms. This transmittance estimation method based on regional interpolation utilizes limited reference ground object information and combines the principles of spatial statistics to achieve the extrapolation and spatialization of transmittance, which not only ensures the accuracy of the estimation results, but also improves the computational efficiency and scope of application. It is of great significance for regional cloud detection under complex terrain and complex weather conditions, and is one of the key technologies to improve the level of remote sensing cloud detection.

[0097] Step S2000, identifying a semi-transparent weather phenomenon area and a non-semi-transparent weather phenomenon area according to the visible light band transmittance and near infrared band transmittance of each pixel in the current spectral calibration image;

[0098] Furthermore, step S2000 includes:

[0099] Step S2100, calculate the transmittance difference between the visible light band transmittance and the near infrared band transmittance of each pixel in the current spectral calibration image, and determine the difference threshold T th , the transmittance difference and the determined difference threshold T th By comparison, candidate areas of semi-transparent weather phenomena are obtained;

[0100] Furthermore, if Figure 4 As shown, step S2100 includes:

[0101] Step S2110, obtaining historical cloud images, and extracting the visible light band transmittance and near infrared band transmittance in the semi-transparent weather phenomenon area and the real cloud layer area of ​​each historical cloud image in units of pixels;

[0102] Step S2120, calculating the average difference between the visible light band transmittance and the near infrared band transmittance of all pixels in the semi-transparent weather phenomenon area of ​​the historical cloud image, recorded as the first difference; calculating the average difference between the visible light band transmittance and the near infrared band transmittance of all pixels in the real cloud layer area of ​​the historical cloud image, recorded as the second difference;

[0103] Step S2130: Determine a difference threshold T according to the first difference and the second difference. th ;

[0104] The difference threshold T is determined th The methods include:

[0105]

[0106] in:

[0107] : Difference threshold, used to distinguish between semi-transparent weather phenomenon areas and real cloud areas;

[0108] : The total number of semi-transparent weather phenomenon areas in historical cloud images;

[0109] N: the total number of real cloud areas in historical cloud images;

[0110] : No. The average value of the visible light transmittance of all pixels in the semi-transparent weather phenomenon area; 1≤ ≤ ;

[0111] : No. The average value of the near-infrared transmittance of all pixels in the semi-transparent weather phenomenon area; is the first difference;

[0112] : No. The weight factor of a semi-transparent weather phenomenon area is proportional to the number of pixels in the area;

[0113] : No. The correction coefficient of a semi-transparent weather phenomenon area is inversely proportional to the discrete degree of the transmittance difference in this area;

[0114] : The average transmittance of all pixels in the kth real cloud layer in the visible light band; 1≤k≤ ;

[0115] : The average transmittance of all pixels in the kth real cloud area in the near infrared band; is the second difference;

[0116] : The weight factor of the kth real cloud area, which is proportional to the number of pixels in the area;

[0117] : The correction coefficient of the kth real cloud area, which is inversely proportional to the discrete degree of the transmittance difference in this area;

[0118] : Weight balance factor between semi-transparent weather phenomena and real cloud areas, the value range is (0,1).

[0119] Average transmittance , , , The transmittance of pixels in the corresponding area of ​​the historical cloud image is extracted and the average value is calculated, which reflects the spectral transmittance characteristics of each area in different bands. , The weight is obtained by calculating the ratio of the number of pixels in each area to the total number of pixels, which reflects the importance of each area in the entire cloud detection process. The more pixels a region has, the greater its weight. , The weight balancing factor δ is obtained by calculating the inverse of the standard deviation of the difference in pixel transmittance in each area, and is used to correct the internal non-uniformity of each area. The greater the degree of dispersion of the difference, the smaller the correction coefficient. The weight balancing factor δ is set based on the experience of the difficulty of distinguishing between semi-transparent weather phenomena and real clouds, and is used to balance the contribution of the two types of areas to the threshold. The larger δ is, the greater the contribution of the real cloud layer. , , , , It is determined by those skilled in the art through a large number of experiments.

[0120] The greater the difference in transmittance between the visible light band and the near-infrared band in the semi-transparent weather phenomenon area, the smaller the difference in transmittance between the two in the real cloud area, and the greater the difference in transmittance between the two in the real cloud area, the greater the difference in transmittance between the visible light band and the near-infrared band. th The larger the value, the more helpful it is to distinguish the two types of areas. The more pixels there are in the semi-transparent weather phenomenon area, the greater the weight, and T th The closer to the average difference of the semi-transparent weather phenomenon area. The smaller the dispersion of the difference in each area, the larger the correction coefficient, and the greater the impact of regional non-uniformity on T th The larger the weight balancing factor δ, the smaller the influence of th The closer it is to the average difference of the real cloud area, the lower the difference. δ can be adjusted dynamically according to the difficulty of distinguishing the two types of areas.

[0121] The formula comprehensively considers the transmittance difference characteristics of the semi-transparent weather phenomenon area and the real cloud area, and enhances the discrimination power of the threshold. The introduction of regional weights makes the area with a large number of pixels play a leading role in the determination of the threshold, which improves the representativeness of the threshold. The introduction of correction coefficients weakens the interference of internal non-uniformity in each region and improves the stability of the threshold. The introduction of weight balance factors can flexibly adjust the influence of the two types of regions on the threshold according to different imaging conditions, thereby improving the adaptability of the threshold. On the basis of fully considering historical prior knowledge, the formula can obtain a comprehensive, stable and adaptable semi-transparent weather phenomenon discrimination threshold through weighted averaging, correction, balance control and other measures, which provides important support for the high-precision realization of cloud detection. The parameters such as weights, correction coefficients, and balance factors in the formula provide a wealth of control freedom for the threshold, which can be optimized according to different data sources and different imaging conditions, thereby improving the versatility of the algorithm. The formula is cleverly designed to achieve the effective integration of regional discrimination characteristics and overall optimization goals, enrich the traditional threshold selection method, and has certain theoretical innovation.

[0122] Historical cloud images are multi-scene cloud image data, which refers to multiple historical remote sensing image data containing clouds. Using multi-scene cloud image data can provide rich cloud transmittance samples. By analyzing the visible and near-infrared band transmittance of the semi-transparent weather phenomenon area and the real cloud layer area in the multi-scene cloud image, a large amount of transmittance sample data can be obtained. The more samples there are, the more accurate and robust the cloud detection discriminant function constructed later will be. Covering the cloud distribution of different times, seasons, and regions. Multi-scene cloud images usually come from different times, seasons, and geographical locations, reflecting the differences in the temporal and spatial distribution of clouds. Using this difference, a cloud detection model suitable for different imaging conditions can be constructed to improve the adaptability and generalization ability of the model. Historical cloud images contain various types of clouds. Cloud images of different scenes contain various cloud types such as cirrus, stratus, cumulus, and clouds with different transmittances and different shapes. Using a variety of cloud samples, the spectral transmittance characteristics of clouds in the visible and near-infrared bands can be fully characterized, and the recognition ability of cloud detection algorithms for different cloud types can be enhanced.

[0123] Transparent weather phenomena refer to atmospheric phenomena that have a certain degree of transparency to solar radiation and surface reflected light, mainly including:

[0124] Thin clouds such as cirrus and cumulus clouds; these clouds are very thin and have strong light transmittance, and the features of the ground objects under the clouds can be received by satellite sensors through the clouds.

[0125] Ice crystal cloud; composed of ice crystals, its permeability is stronger than that of ordinary water clouds, and its transmittance varies greatly across different wavebands.

[0126] Mist; composed of liquid water or ice, with high overall transmittance and strong absorption in the near-infrared band.

[0127] Aerosols such as sand and dust; Aerosols such as sand and dust will attenuate solar radiation to a certain extent through scattering and absorption, and at the same time affect the intensity of diffuse radiation.

[0128] Although these translucent phenomena have a certain degree of light transmittance, they will also weaken and affect the surface radiation signal. In the translucent weather phenomenon area, the signal received by the satellite is a mixture of the ground object signal after passing through the atmosphere and the signal of the translucent material itself, which has a unique spectrum.

[0129] Step S2140, calculating the transmittance difference between the visible light band transmittance and the near infrared band transmittance of each pixel in the current spectral calibration image;

[0130] Step S2150, set the pixel in the current spectral calibration image The difference between the transmittance of the visible light band and the transmittance of the near infrared band is ,like >T th , then the pixel Marked as semi-transparent weather phenomenon candidate area, is the index of the pixel in the spectral calibration image; otherwise, it is marked as a non-translucent weather phenomenon area.

[0131] Specifically, step S2100 aims to obtain the prior knowledge of the difference in transmittance between semi-transparent weather phenomena and real clouds in the visible light band and near-infrared band by analyzing the spectral transmittance characteristics of historical cloud images, so as to provide a basis for determining the preset difference threshold. Historical cloud image data usually comes from long-term satellite observations and contains rich cloud spectral information. Through manual annotation or automatic extraction, the location distribution of semi-transparent weather phenomenon areas (such as cirrus, fog, etc.) and real cloud areas (such as cumulus, stratus, etc.) can be obtained. On this basis, the visible light band transmittance and near-infrared band transmittance data of the pixels in each area are extracted in units of pixels, and the average difference between the visible light band transmittance and the near-infrared band transmittance of the semi-transparent weather phenomenon area and the real cloud area is calculated, which is recorded as the first difference and the second difference. The first difference and the second difference respectively measure the statistical characteristics of the difference in transmittance between the two types of targets, semi-transparent weather phenomena and real clouds, in the visible light and near-infrared bands. Since the scattering and absorption of visible light by semi-transparent weather phenomena are relatively weak, the first difference is usually smaller than the second difference. This scheme makes full use of historical data and prior knowledge. By analyzing the statistical laws of the difference in transmittance between semi-transparent weather phenomena and real clouds in the visible and near-infrared bands, it adaptively determines the preset difference threshold so that the threshold can dynamically adapt to the changes in the spectral characteristics of different clouds, thereby improving the adaptability and robustness of the method. At the same time, taking pixels as processing units, the differences in the spatial distribution of cloud spectra are taken into account, which improves the precision of recognition. If the transmittance difference is less than or equal to the threshold T th , then the pixel may not belong to a semi-transparent weather phenomenon, but more likely to be clear sky or opaque clouds. Therefore, when the transmittance difference is less than or equal to T th , the pixel is marked as a non-translucent weather phenomenon area.

[0132] Step S2200, taking pixels as units, spatial clustering is performed according to the spatial neighborhood relationship of the semi-transparent weather phenomenon candidate areas to obtain semi-transparent weather phenomenon areas and non-semi-transparent weather phenomenon areas.

[0133] Furthermore, if Figure 5 As shown, step S2200 includes:

[0134] Step S2210, extracting the spatial coordinates of the semi-transparent weather phenomenon candidate area in units of pixels, and constructing an undirected graph, wherein each node in the undirected graph corresponds to a semi-transparent weather phenomenon candidate area pixel, and there are edge connections between adjacent candidate area pixels;

[0135] Step S2220, calculating the weight of the edge in the undirected graph in units of pixels, where the weight of the edge is the spectral feature distance between adjacent candidate area pixels;

[0136] Step S2230, divide the undirected graph to obtain a set of cut edges with the smallest weight and the smallest value, and divide the pixels in the candidate area of ​​semi-transparent weather phenomena into N2 connected subgraphs according to the cut edge set. The pixels in the connected subgraphs are marked as semi-transparent weather phenomenon areas, and the remaining pixels are marked as non-semi-transparent weather phenomenon areas; N2 is a positive integer.

[0137] Specifically, step S2200 uses graph cut theory to perform spatial clustering on the candidate area of ​​semi-transparent weather phenomenon, aiming to make full use of the spatial neighborhood information of the candidate area pixels, overcome the defects of pixel classification being susceptible to noise interference and fragmented recognition results, and obtain a semi-transparent weather phenomenon area with continuous space and consistent spectrum. First, the spatial coordinates of the candidate area of ​​semi-transparent weather phenomenon are extracted in pixels, and an undirected graph is constructed. Each node in the undirected graph corresponds to a pixel in the candidate area, and there are edge connections between adjacent candidate area pixels. The purpose of this step is to establish the spatial topological relationship of pixels in the candidate area, and provide a data structure basis for subsequent clustering and segmentation. Then, the weight of each edge in the undirected graph is calculated in pixels. The weight of the edge is used to measure the spectral similarity of adjacent candidate area pixels, and commonly used distance metrics such as Euclidean distance and Mahalanobis distance can be selected as the weight calculation method. The smaller the edge weight, the closer the spectral characteristics of adjacent pixels are, and the greater the possibility of belonging to the same class. Finally, the undirected graph is divided using the graph cut algorithm. Graph cut refers to finding a set of cut edges in a graph so that the sum of the edge weights inside the two subgraphs that are cut is the largest and the sum of the edge weights between the subgraphs is the smallest. The set of cut edges can be solved by optimization methods such as the maximum flow minimum cut algorithm. According to the obtained set of cut edges, the undirected graph can be divided into several connected subgraphs. The pixels inside the connected subgraphs are spectrally similar and spatially adjacent, and can be merged into a whole. The pixels in the connected subgraphs are marked as semi-transparent weather phenomenon areas, and the remaining pixels are marked as non-semi-transparent weather phenomenon areas, thereby realizing the spatial clustering of semi-transparent weather phenomenon areas. Non-semi-transparent weather phenomenon areas refer to areas that do not belong to semi-transparent weather phenomena, mainly including clear sky areas, ground object areas, and opaque cloud areas. The spectral characteristics of these areas are obviously different from semi-transparent weather phenomena, such as the clear sky area has low reflectivity, the ground object area has reflectivity related to the ground object type, and the opaque cloud area has high reflectivity.

[0138] This scheme cleverly uses graph theory models to describe the spatial topological relationship and spectral similarity between pixels, transforms the problem of extracting semi-transparent weather phenomenon areas into a graph partitioning problem, and realizes spatial clustering of candidate areas by optimizing the set of cut edges, thus obtaining continuous and complete semi-transparent weather phenomenon areas. Compared with the traditional pixel-by-pixel classification method, this scheme overcomes noise interference and the "salt and pepper" phenomenon through spatial constraints between pixels, making the recognition results smoother and more continuous. At the same time, spectral distance calculation and spatial relationship construction based on pixels can fully explore the subtle differences in the spectrum inside the cloud layer and realize the recognition of semi-transparent weather phenomenon areas at the sub-pixel level. The graph cut clustering method is a research hotspot in the current field of image segmentation. It has a solid theoretical foundation and rich improvement schemes, which can provide a reference for further optimization of the method.

[0139] Step S3000, construct a real cloud spectral feature model, use the real cloud spectral feature model to identify the real cloud in the non-translucent weather phenomenon area, obtain the real cloud area, merge the real cloud area with the translucent weather phenomenon area, and generate the final cloud detection result map.

[0140] Furthermore, step S3000 includes:

[0141] Step S3100, extracting the spectral features of historical cloud images and establishing a real cloud spectral feature model;

[0142] Furthermore, if Figure 6 As shown, step S3100 includes:

[0143] Step S3110, extracting the cloud region of the historical cloud image to obtain the cloud region of the historical cloud image;

[0144] Specifically, historical cloud image data refers to remote sensing image data containing clouds acquired by hyperspectral imagers over a period of time in the past. These historical data cover different time periods (such as different seasons, different years), different imaging conditions (such as different solar altitudes, different atmospheric conditions) and different geographical areas to ensure the diversity and representativeness of the samples. Hyperspectral imaging is an advanced remote sensing technology that can continuously image ground objects over a wide wavelength range (such as visible light, near infrared, short-wave infrared, etc.) with a very high spectral resolution (such as a few nanometers to tens of nanometers), and obtain image data in hundreds of bands, each of which corresponds to a narrow wavelength range. Hyperspectral images contain rich spectral information, which can accurately depict the spectral characteristics of ground objects and are ideal data sources for cloud detection.

[0145] After acquiring historical cloud image data, it is necessary to extract the cloud area of ​​interest from the image to provide data support for subsequent spectral feature analysis. Cloud area extraction refers to the use of image processing and pattern recognition technology to segment the cloud-covered area from the entire image to obtain the spatial distribution information of the cloud. Common cloud area extraction methods include threshold segmentation, texture analysis, cluster segmentation, etc. The threshold segmentation method sets a threshold based on the high reflectivity of the cloud in a specific band (such as 1.38μm) to binarize the image into cloud and non-cloud areas; the texture analysis method uses the differences in texture features such as brightness and contrast in the cloud area to achieve segmentation; the cluster segmentation method maps the pixel spectral features to the feature space through a clustering algorithm, and separates the cloud and background according to the inter-class distance. The selection of a suitable segmentation algorithm requires consideration of factors such as the spatiotemporal resolution and spectral characteristics of the cloud image. In order to improve the accuracy of cloud area extraction, it is usually necessary to manually check and correct the results of automatic extraction to obtain reliable cloud area reference data for subsequent spectral feature analysis and modeling.

[0146] The cloud area extraction results are visualized as binary images, with cloud areas as 1 and non-cloud areas as 0. The corresponding cloud areas must be extracted from each historical cloud image data to obtain a cloud area binary image dataset. These cloud area data depict the distribution patterns of real clouds under different temporal and spatial conditions, and are the basis for analyzing cloud spectral characteristics and establishing a discrimination model.

[0147] By extracting cloud regions from massive historical data, high-quality samples required for cloud spectral analysis were obtained. The extraction process fully utilized the advantages of hyperspectral data and selected the key bands that are most conducive to cloud detection from hundreds of bands. The combination of multiple segmentation algorithms overcomes the limitations of a single method and significantly improves the accuracy and reliability of cloud region extraction. Through human-computer interactive processing, the integration of computer intelligence and manual experience further improves the accuracy of cloud region data. High-quality cloud region samples are the prerequisite for cloud spectral feature analysis and lay a solid foundation for the construction of data-driven spectral discrimination models. At the same time, cloud region extraction also provides an objective reference for algorithm performance evaluation and provides data guarantee for subsequent model optimization and iteration.

[0148] Step S3120, extracting spectral features of cloud areas in historical cloud images, and constructing spectral feature vectors of real cloud layers; the spectral features include spectral reflectance features, spectral absorption features, and spectral scattering features in visible light bands and near infrared bands;

[0149] Specifically, after obtaining the cloud areas of historical cloud images, it is necessary to extract features that can reflect the essential properties of cloud spectra from these cloud areas to form spectral feature vectors to characterize the spectral radiation characteristics of real clouds. Common cloud spectral features include reflectivity features, absorption features, and scattering features, which are contained in visible light band and near-infrared band data.

[0150] The spectral reflectance characteristics describe the reflectivity of clouds at different wavelengths and characterize the scattering intensity of clouds to solar radiation. Clouds have a high and uniform reflectivity in the visible light band (0.4-0.75μm), and the reflectivity decreases slightly in the near-infrared band (0.75-1.3μm) but still maintains a high level. Several narrow bands can be selected to calculate the average reflectivity of the pixels in the cloud area in these bands to obtain the reflectivity feature vector. Some typical spectral indices (such as the normalized difference cloud index NDCI) can also be used to integrate the reflectivity information of two or more bands to highlight the differences between clouds and the surface.

[0151] The spectral absorption characteristics describe the degree of cloud absorption of electromagnetic wave energy within a specific band, and are closely related to the cloud's microphysical properties such as particle phase and effective radius. Clouds have several absorption bands in the near-infrared band, which are mainly related to the absorption of ice crystals in the clouds. Typical absorption bands include 1.6μm, 2.1μm, 3.7μm, etc. Several narrow bands near these absorption bands can be selected to calculate indicators such as absorption depth and absorption area to form an absorption characteristic vector. The selection of absorption bands needs to consider the spectral response characteristics of the sensor and the characteristics of the atmospheric window area.

[0152] The spectral scattering characteristics reflect the scattering and polarization characteristics of cloud particles to the incident light, including parameters such as scattering phase function and polarization degree. The scattering of clouds in the visible to near-infrared band is mainly Mie scattering, showing strong forward scattering. The scattering phase function describes the trend of the intensity of scattered radiation changing with the scattering angle. Clouds with different particle phases, particle sizes and morphologies have different scattering phase function curves. The degree of polarization reflects the polarization characteristics of the scattered radiation and is closely related to the particle morphology. According to the Mie scattering theory and the radiation transfer model, the scattering phase function coefficients and polarization degree parameters of the cloud area pixels can be inverted to form a scattering feature vector.

[0153] After extracting spectral features of multiple categories such as reflectivity, absorption, and scattering, a high-dimensional spectral feature vector can be obtained. This feature vector describes the spectral radiation characteristics of clouds from multiple angles such as reflection, absorption, and scattering, and contains rich cloud microphysical information. However, too high a feature dimension may introduce information redundancy and reduce the efficiency of subsequent training. Therefore, it is necessary to reduce the dimensionality of the high-dimensional feature vector to remove the correlation between features. Commonly used dimensionality reduction methods include principal component analysis (PCA), independent component analysis (ICA), linear discriminant analysis (LDA), etc. Dimensionality reduction can significantly improve the discriminative ability and computational efficiency of features while retaining the main information. The low-dimensional feature vector after dimensionality reduction will be used as a discriminant feature for identifying real clouds.

[0154] By extracting spectral features from multiple dimensions and angles, the advantages of hyperspectral data are fully utilized to comprehensively characterize the spectral radiation characteristics of clouds from the aspects of reflection, absorption, scattering, etc. On the one hand, the information of multiple bands such as visible light and near infrared is comprehensively utilized to expand the conventional visible light band cloud detection method, greatly improving the ability to characterize cloud spectral characteristics; on the other hand, the introduction of parameters such as scattering and polarization, on the basis of macroscopic radiation, further deepens the cloud microphysical level, and enhances the description of cloud vertical structure and particle characteristics. Spectral feature extraction is the core of cloud detection algorithm and directly determines the discrimination performance of the algorithm. The multidimensional fusion strategy expands the feature space, so that reliable and stable cloud discrimination capabilities can be obtained even under complex backgrounds. Through feature dimensionality reduction optimization, the essential characteristics of cloud spectrum are further condensed, the data dimension is compressed, and the computational efficiency is improved while ensuring the discrimination accuracy. The optimized spectral feature set provides a solid data foundation for the subsequent construction of cloud discrimination model and is the key to realizing remote sensing cloud detection.

[0155] Step S3130, using the spectral feature vectors of real clouds as samples, training and generating a real cloud spectral feature model.

[0156] After constructing the spectral feature vector of the real cloud layer, it is necessary to design a suitable machine learning algorithm to construct a cloud spectral discrimination model with the spectral feature vector as input. Each spectral feature vector of the real cloud layer corresponds to a pixel and carries a category label of "cloud" or "non-cloud". Based on the training samples, the model is trained and the performance is evaluated by methods such as K-fold cross validation. The sample set is randomly divided into K mutually exclusive subsets, and K-1 subsets are selected as training sets each time, and the remaining 1 is used as a validation set. After training K models, the average performance is taken. Cross-validation can make full use of limited samples, objectively evaluate the generalization ability of the model, and optimize the model parameters. Evaluation indicators generally use confusion matrix, overall accuracy, Kappa coefficient, etc. The confusion matrix counts the degree of conformity between the model prediction results and the reference true value, from which indicators such as accuracy and recall can be calculated. The Kappa coefficient corrects the influence of random consistency on the basis of overall accuracy, and is more robust to the situation of unbalanced sample distribution.

[0157] For cloud detection, it is necessary to find a decision boundary with the maximum interval between cloud and non-cloud spectral samples. Support vector machine (SVM) is used. SVM transforms nonlinear problems into linear problems in high-dimensional space through kernel techniques, which can handle the nonlinear relationship of spectral features well. At the same time, SVM is based on the structural risk minimization criterion, strikes a balance between empirical risk and model complexity, and has strong generalization ability.

[0158] The training process of the real cloud spectral feature model is as follows:

[0159] Data preparation: Spectral features are extracted from historical cloud images, and each pixel is labeled as "cloud" or "non-cloud" to form a training sample set. To ensure the representativeness and adequacy of the samples, a sufficient number of cloud spectral samples need to be collected under different temporal and spatial conditions.

[0160] Feature selection: Screen high-dimensional spectral features, remove redundant and noise features, and retain only the most discriminative key band combinations. Commonly used feature selection methods include Relief and mRMR, which select the optimal feature subset by calculating the correlation between features and categories and the redundancy between features.

[0161] Parameter optimization: The SVM model contains hyperparameters such as kernel functions and penalty coefficients, which directly affect the model performance. Using methods such as grid search, we can automatically search and optimize in the parameter space to find the parameter combination that optimizes the performance of the validation set. Common kernel functions include linear kernel, Gaussian kernel RBF, etc.

[0162] Model training: Based on the preferred features and parameters, the SVM model is trained using training set samples. SVM maps samples to a high-dimensional feature space and finds the maximum margin hyperplane between different categories. The model training process is achieved by solving a convex quadratic programming problem.

[0163] Performance evaluation: Use cross-validation and other methods to objectively evaluate the performance of the trained SVM model. Divide the sample set into K folds, take K-1 folds for training in turn, and the remaining 1 fold for validation, and take the average performance. Use indicators such as confusion matrix, overall accuracy, and Kappa coefficient to measure the cloud detection accuracy of the model.

[0164] Model prediction: The trained and optimized SVM model is applied to new remote sensing images to identify the “cloud” or “non-cloud” category to which the spectral characteristics of each pixel belong, thereby achieving cloud detection for the entire scene.

[0165] The SVM cloud detection method can well model the nonlinear and complex relationship of spectral features. It does not require too much prior knowledge. Feature selection and parameter optimization can improve the discriminative performance and robustness of the model. It has fast training and prediction speeds, relatively low computational complexity, and is easy to implement in engineering. In short, SVM can effectively overcome the nonlinear problem of spectral features in cloud detection by learning the optimal discriminant hyperplane from high-dimensional spectral features. The data-driven SVM method provides a new idea for the intelligent interpretation of complex cloud scenes, and is expected to further improve the accuracy and efficiency of remote sensing cloud detection. In the next step, we can consider combining SVM with deep learning and other methods to fully explore spectral-spatial information and realize further intelligent cloud detection.

[0166] Machine learning shifts cloud detection from experience-driven to data-driven. By learning from massive historical samples, it automatically learns new ways to improve detection performance through human-computer interaction. Machine learning methods based on spectral features overcome the limitations of physical modeling and can extract universal cloud discrimination rules from complex and changeable remote sensing big data. On the one hand, machine learning models can adaptively weigh multiple spectral features and discover key discriminant information implicit in band combinations; on the other hand, the data-driven framework enables the algorithm to continuously learn and iterate from new samples and autonomously evolve intelligence to deal with unknown and complex scenes. Compared with traditional threshold methods, machine learning cloud detection methods have stronger adaptability, robustness, and scalability. At the same time, the explicit modeling of cloud spectral knowledge also provides new ideas for the study of physical mechanisms. The integration of machine learning and spectral analysis will surely promote the intelligent development of cloud remote sensing detection.

[0167] By integrating machine learning and spectral analysis technology, an intelligent method for automatically learning cloud discrimination rules from massive historical data has been realized, creating a new paradigm for remote sensing cloud detection. A hyperspectral cloud image dataset with a large temporal and spatial span and strong representativeness has been obtained, and a variety of segmentation algorithms have been used to extract the true value of the cloud area, and a high-quality cloud spectrum knowledge sample library has been constructed. Multiple bands such as visible light and near-infrared have been fully considered, and rich cloud spectrum features have been extracted from multiple angles such as reflection, absorption, and scattering, forming a high-dimensional cloud spectrum feature space. The key features have been condensed by using dimensionality reduction optimization technology, which improves the discriminability and computational efficiency while ensuring the amount of information. The innovative introduction of machine learning into remote sensing cloud detection has overcome the limitations of traditional physical modeling, greatly improved the adaptive ability and intelligent level of cloud detection, and provided new ideas and methods for reliable cloud detection under complex conditions. The cloud spectrum discrimination rules learned from massive historical data have formed a set of cloud spectrum basic datasets under complex conditions, which has high application value. This knowledge can not only guide cloud detection by optical remote sensing, but can also be extended to other remote sensing methods such as microwaves and lasers. The cloud spectrum knowledge base brings together multi-source and multi-scale cloud information, providing an important reference for in-depth research on cloud radiation, microphysics and other mechanisms. From a single visible light band to the near-infrared and even the entire solar reflection band, from simple reflectivity characteristics to multi-dimensional characteristics such as absorption, scattering, and polarization, the role of spectral analysis in cloud detection has been unprecedentedly exerted. Studies have shown that many band combinations and spectral parameters have unique contributions to cloud discrimination, expanding the conventional band combination method based on albedo thresholds. Multidimensional spectral analysis provides new ideas for finely characterizing the radiation characteristics and microphysical properties of clouds.

[0168] Step S3200, using a sliding window method to extract spectral features from the spectral calibration image of the non-translucent weather phenomenon area, obtain the spectral features of the area to be evaluated in the translucent weather phenomenon area, and construct a spectral feature vector of the area to be evaluated; input the spectral feature vector of the area to be evaluated into a real cloud layer spectral feature model to obtain the probability that the area to be evaluated is a real cloud layer area; compare the probability that the area to be evaluated is a real cloud layer area with a preset probability threshold, and determine whether the area to be evaluated is a real cloud layer area based on the comparison result.

[0169] If the probability that the area to be evaluated is a real cloud area is greater than a preset probability threshold, the area to be evaluated is determined to be a real cloud area, otherwise it is determined to be a non-real cloud area.

[0170] Specifically, step S3200 uses the real cloud spectral feature model established in step S3100 to identify the real cloud area from the non-translucent weather phenomenon area. The non-translucent weather phenomenon area refers to the area that cannot be clearly determined as a translucent weather phenomenon after preliminary screening based on translucent optical thickness. There may be real clouds in these areas, which need to be further identified using spectral features. In order to gradually scan each local area of ​​the non-translucent weather phenomenon area, a sliding window method can be used to slide in the area with a certain step length to extract the spectral features of the window coverage area. The size of the sliding window needs to be reasonably set according to the scale distribution characteristics of the real cloud layer. The horizontal scale of the cloud layer ranges from tens of meters to hundreds of kilometers. The probability distribution of the cloud layer scale can be obtained by statistical analysis of historical cloud images. According to the peak interval of the distribution, sliding windows of different scales are set to meet the recognition requirements of clouds of different scales. For the extracted spectral features of the area to be evaluated, a spectral feature vector is constructed according to the method of step S3120, and then input into the real cloud layer spectral feature model for classification and discrimination. The output of the model is the probability that the area to be evaluated belongs to the real cloud layer. In order to convert the probability value into a certain category label, a probability threshold needs to be set. When the probability value is greater than the threshold, the area to be evaluated is considered to be a real cloud area, otherwise it is considered to be a non-real cloud area. The selection of the probability threshold requires a trade-off between detection accuracy and missed detection rate. The larger the threshold, the stricter the judgment standard for the real cloud layer, the higher the detection accuracy, but the missed detection rate will also increase; the smaller the threshold, the lower the detection accuracy and the missed detection rate. Usually, the probability threshold can be optimized based on historical cloud image data using a cross-validation method. A certain proportion of data is used as a validation set to calculate the accuracy and missed detection rate under different thresholds, and a threshold that is relatively balanced in both indicators is selected as the optimal threshold. Through the step-by-step evaluation of the sliding window and the comparison of the probability threshold, the real cloud area can eventually be identified in the non-translucent weather phenomenon area to supplement and improve the detection results of the translucent weather phenomenon area.

[0171] Step S3200 makes up for the shortcomings of a single physical model and further improves the accuracy and completeness of cloud detection. Although translucent optical thickness can identify most translucent weather phenomena, there may be a risk of missing detection for real clouds with smaller optical thickness. The real cloud spectral feature model is automatically learned from massive historical data, and can discover real cloud patterns that cannot be identified by translucent optical thickness. Combining the two methods, the physical model and the machine learning model complement each other and can maximize the detection of cloud coverage areas in remote sensing images. At the same time, the post-processing strategy combining sliding windows and probability thresholds can adapt to the scale diversity of real clouds and the uncertainty of model discrimination, while improving detection accuracy, controlling the risk of missed detection, and making cloud detection results more reliable and complete. The introduction of multi-scale sliding windows can meet the detection needs of clouds of different sizes and improve the adaptability and flexibility of the algorithm.

[0172] Step S3300, merging the semi-transparent weather phenomenon area and the real cloud area to generate a cloud detection result map.

[0173] Specifically, step S3300 merges the recognition results of the semi-transparent weather phenomenon area and the real cloud area to form a final cloud detection result map. The semi-transparent weather phenomenon area is obtained based on the semi-transparent optical thickness threshold segmentation, corresponding to the semi-transparent clouds, fog and other weather phenomena with large optical thickness; the real cloud area is obtained based on the spectral feature model recognition, corresponding to the cloud coverage area with small optical thickness but with real cloud spectral characteristics. By merging the two recognition results, a complete cloud detection result can be obtained, including two types of cloud coverage information of semi-transparent weather phenomena and real clouds. In order to intuitively express the spatial distribution pattern of different categories of cloud coverage, different colors or grayscale values ​​can be used for identification in the cloud detection result map. Common identification schemes include pseudo-color synthesis and grayscale mapping. Pseudo-color synthesis is to map different category information to three color channels of red, green and blue to generate a color identification image. You can select a blue series for the semi-transparent weather phenomenon area, a red series for the real cloud area, and black or other background colors for the non-cloud area to intuitively express the type differences and spatial distribution patterns of cloud coverage. Grayscale mapping is to map different category information to the grayscale value range of 0-255 to generate a grayscale identification image. Similarly, a brighter grayscale value (such as 200) can be selected for the semi-transparent weather phenomenon area, a second-brighter grayscale value (such as 150) can be selected for the real cloud area, and a darker grayscale value (such as 50) can be selected for the non-cloud area. The change of light and dark intuitively expresses the difference between different cloud cover types. Using a visual identification strategy, the quantitative results of cloud detection can be converted into intuitive image expressions, which are easy for users to understand and interpret.

[0174] Step S3300 realizes the seamless integration of the detection results of the semi-transparent weather phenomenon area and the real cloud area, forming a complete and unified cloud detection result. By integrating the recognition results of different methods, the complementary advantages of the physical model and the machine learning model are fully utilized to make up for the shortcomings of a single method, thereby obtaining a comprehensive and reliable cloud detection result. The cloud detection result map adopts an intuitive visualization expression strategy, which clearly shows the spatial distribution pattern of different types of cloud coverage through color or grayscale encoding, making it easy for users to quickly understand and apply the detection results. The visualization result map not only intuitively reflects the distribution location and degree of cloud pollution, but also clearly identifies the type differences of cloud pollution, providing important prior information for the cloud removal and cloud correction processing of quantitative remote sensing products. Based on the cloud detection result map, an adaptive mask of the cloud area can be realized to shield or replace the radiation value of the cloud-covered area, eliminate the influence of clouds on the inversion of the spectral reflectance of the ground objects, and improve the accuracy of quantitative remote sensing products.

[0175] Example 2

[0176] This embodiment provides a satellite-borne hyperspectral cloud recognition system based on dual-band fusion technology on the basis of embodiment 1. Figure 7 As shown, including:

[0177] Spectral image calibration module: used to obtain a hyperspectral image, pre-process the hyperspectral image, and obtain a spectral calibration image; the hyperspectral image includes a visible light band image and a near-infrared band image; respectively calculate the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image; record the solar altitude angle when the hyperspectral image is obtained;

[0178] Semi-transparent weather phenomenon area determination module: used to identify the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area according to the visible light band transmittance and near infrared band transmittance of each pixel in the spectral calibration image;

[0179] Real cloud detection module: used to build a real cloud spectral feature model, use the real cloud spectral feature model to identify the real cloud in the non-translucent weather phenomenon area, obtain the real cloud area, merge the real cloud area with the translucent weather phenomenon area, and generate the final cloud detection result map.

[0180] The spectral image calibration module includes: a geometric correction unit, an atmospheric correction unit and a transmittance calculation unit;

[0181] The geometric correction unit is used to obtain a hyperspectral image, record the solar altitude angle when the hyperspectral image is obtained, and perform radiation calibration and geometric correction on the hyperspectral image;

[0182] The obtaining of the hyperspectral image, recording the solar altitude angle when the hyperspectral image is obtained, and performing radiation calibration and geometric correction on the hyperspectral image include:

[0183] Step S1110, obtaining a satellite-borne hyperspectral image of the area to be detected, wherein the hyperspectral image includes a visible light band image and a near infrared band image;

[0184] Step S1120, performing radiometric calibration on the visible light band image and the near infrared band image in units of pixels, and converting the DN value into a radiance value;

[0185] Step S1130 , geometrically correcting the visible light band image and the near infrared band image after radiation calibration in units of pixels.

[0186] The atmospheric correction unit is used to perform atmospheric correction on the geometrically corrected visible light band image and near infrared band image in units of pixels to obtain a spectral calibration image;

[0187] The transmittance calculation unit is used to respectively calculate the transmittance of each pixel in the spectral calibration image in the visible light band and the transmittance of the near infrared band.

[0188] The step of respectively calculating the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image comprises:

[0189] Step S1310, selecting n reference object regions on the spectral calibration image, each reference object region includes N1 pixels, and different reference object regions are distributed at different positions of the spectral calibration image; N1 is a positive integer;

[0190] Step S1320, extracting the radiance values ​​of all pixels in each reference object region in the visible light band, and calculating the measured spectral reflectance of each pixel in the visible light band in each reference object region according to the radiance value of each pixel in the visible light band and the solar altitude angle;

[0191] Step S1330, extracting the radiance values ​​of all pixels in each reference object region in the near-infrared band, calculating the solar zenith angle from the solar altitude angle, and calculating the measured spectral reflectance of each pixel in the near-infrared band in each reference object region according to the radiance value and solar zenith angle of each pixel in the near-infrared band;

[0192] Step S1340, calculating the theoretical spectral reflectance of each pixel in the visible light band of each reference object area under cloudless atmospheric conditions;

[0193] Step S1350, calculating the theoretical spectral reflectance of each pixel in the near-infrared band of each reference object area under cloudless atmospheric conditions;

[0194] Step S1360, taking the ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in each reference object region in the visible light band as the first regionalized variable, establishing an interpolation model of the transmittance of the visible light band, and obtaining the transmittance of the visible light band at any pixel position in the spectral calibration image;

[0195] Step S1370, taking the ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in each reference object area in the near-infrared band as the second regionalized variable, establishing an interpolation model of the near-infrared band transmittance, and obtaining the transmittance of each pixel in the spectral calibration image in the near-infrared band.

[0196] The semi-transparent weather phenomenon area determination module includes: a candidate area determination unit and a spatial clustering unit;

[0197] The candidate area determination unit is used to calculate the transmittance difference between the visible light band transmittance and the near infrared band transmittance of each pixel in the current spectral calibration image, and determine the difference threshold T th , the transmittance difference and the determined difference threshold T th By comparison, candidate areas of semi-transparent weather phenomena are obtained;

[0198] The spatial clustering unit is used to perform spatial clustering based on the spatial neighborhood relationship of the semi-transparent weather phenomenon candidate areas with pixels as units, so as to obtain the semi-transparent weather phenomenon areas and the non-translucent weather phenomenon areas.

[0199] The obtaining of the semi-transparent weather phenomenon candidate area comprises:

[0200] Step S2110, obtaining historical cloud images, and extracting the visible light band transmittance and near infrared band transmittance in the semi-transparent weather phenomenon area and the real cloud layer area of ​​each historical cloud image in units of pixels;

[0201] Step S2120, calculating the average difference between the visible light band transmittance and the near infrared band transmittance of all pixels in the semi-transparent weather phenomenon area of ​​the historical cloud image, recorded as the first difference; calculating the average difference between the visible light band transmittance and the near infrared band transmittance of all pixels in the real cloud layer area of ​​the historical cloud image, recorded as the second difference;

[0202] Step S2130: Determine a difference threshold T according to the first difference and the second difference. th ;

[0203] Step S2140, calculating the transmittance difference between the visible light band transmittance and the near infrared band transmittance of each pixel in the current spectral calibration image;

[0204] Step S2150, set the pixel in the current spectral calibration image The difference between the transmittance of the visible light band and the transmittance of the near infrared band is ,like >T th , then the pixel Marked as semi-transparent weather phenomenon candidate area, The index of the pixel in the spectral calibration image.

[0205] The spatial clustering according to the spatial neighborhood relationship of the semi-transparent weather phenomenon candidate area includes:

[0206] Step S2210, extracting the spatial coordinates of the semi-transparent weather phenomenon candidate area in units of pixels, and constructing an undirected graph, wherein each node in the undirected graph corresponds to a semi-transparent weather phenomenon candidate area pixel, and there are edge connections between adjacent candidate area pixels;

[0207] Step S2220, calculating the weight of the edge in the undirected graph in units of pixels, where the weight of the edge is the spectral feature distance between adjacent candidate area pixels;

[0208] Step S2230, divide the undirected graph to obtain a set of cut edges with the minimum weight and the smallest sum. According to the set of cut edges, the pixels in the candidate area of ​​semi-transparent weather phenomena are divided into N2 connected sub-graphs. The pixels in the connected sub-graphs are marked as semi-transparent weather phenomenon areas, and the remaining pixels are marked as non-semi-transparent weather phenomenon areas.

[0209] In the real cloud layer detection module, the construction of the real cloud layer spectral feature model includes:

[0210] Step S3110, extracting the cloud region of the historical cloud image to obtain the cloud region of the historical cloud image;

[0211] Step S3120, extracting spectral features of cloud areas in historical cloud images, and constructing spectral feature vectors of real cloud layers; the spectral features include spectral reflectance features, spectral absorption features, and spectral scattering features in visible light bands and near infrared bands;

[0212] Step S3130, using the spectral feature vectors of real clouds as samples, training and generating a real cloud spectral feature model.

[0213] In the real cloud detection module, the real cloud recognition in the non-translucent weather phenomenon area using the real cloud spectral feature model includes:

[0214] The spectral feature extraction of the spectral calibration image of the non-translucent weather phenomenon area is performed by using a sliding window method, the spectral features of the area to be evaluated in the semi-transparent weather phenomenon area are obtained, and the spectral feature vector of the area to be evaluated is constructed; the spectral feature vector of the area to be evaluated is input into the real cloud layer spectral feature model to obtain the probability that the area to be evaluated is a real cloud layer area; the probability that the area to be evaluated is a real cloud layer area is compared with a preset probability threshold, and whether the area to be evaluated is a real cloud layer area is determined according to the comparison result.

[0215] Example 3

[0216] This embodiment discloses an electronic device, which may include one or more processors and one or more memories. The memories store computer-readable codes, which, when executed by the one or more processors, may execute the above-mentioned spaceborne hyperspectral cloud recognition method based on dual-band fusion technology.

[0217] The method or system according to the embodiment of the present application can also be implemented with the help of the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, an input / output component, a hard disk, etc. A storage device in the electronic device, such as a ROM or a hard disk, can store a spaceborne hyperspectral cloud recognition method based on dual-band fusion technology provided in the present application. A satellite-borne hyperspectral cloud recognition method based on dual-band fusion technology may, for example, include: acquiring a hyperspectral image, preprocessing the hyperspectral image, and obtaining a spectral calibration image; the hyperspectral image includes a visible light band image and a near-infrared band image; respectively calculating the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image; recording the solar altitude angle when the hyperspectral image is acquired; identifying the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area according to the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image; constructing a real cloud spectral feature model, using the real cloud spectral feature model to perform real cloud recognition on the non-semi-transparent weather phenomenon area to obtain a real cloud area, and fusing the real cloud area with the semi-transparent weather phenomenon area to generate a final cloud detection result map.

[0218] Furthermore, the electronic device may also include a user interface. Of course, the architecture disclosed in the present invention is only exemplary, and when implementing different devices, one or more components in the electronic device disclosed in the present invention may be omitted according to actual needs.

[0219] Example 4

[0220] This embodiment discloses a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, a method for identifying satellite-borne hyperspectral cloud layers based on dual-band fusion technology according to an embodiment of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache). Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0221] In addition, according to the implementation mode of the present application, the process described in the above reference flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, the non-transitory machine-readable storage medium stores machine-readable instructions, and the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application, for example: acquiring a hyperspectral image, preprocessing the hyperspectral image, and obtaining a spectral calibration image; the hyperspectral image includes a visible light band image and a near-infrared band image; respectively calculating the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image; recording the solar altitude angle when the hyperspectral image is acquired; according to the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image, identifying the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area; constructing a real cloud layer spectral feature model, using the real cloud layer spectral feature model to identify the real cloud layer in the non-semi-transparent weather phenomenon area, obtaining the real cloud layer area, merging the real cloud layer area with the semi-transparent weather phenomenon area, and generating the final cloud detection result map. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are performed.

[0222] The methods, systems, and devices of the present application may be implemented in many ways. For example, the methods, systems, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0223] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0224] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A spaceborne hyperspectral cloud recognition method based on dual-band fusion technology, characterized in that: The method comprises: Acquire a hyperspectral image, preprocess the hyperspectral image, and obtain a spectral calibration image; the hyperspectral image includes a visible light band image and a near infrared band image; respectively calculate the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image; and record the solar altitude angle when the hyperspectral image is acquired; According to the visible light band transmittance and near infrared band transmittance of each pixel in the spectral calibration image, the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area are identified; A real cloud spectral feature model is constructed, and the real cloud spectral feature model is used to identify the real cloud in the non-translucent weather phenomenon area to obtain the real cloud area. The real cloud area is then merged with the translucent weather phenomenon area to generate the final cloud detection result map.

2. The method for identifying cloud layers by using a spaceborne hyperspectral method based on dual-band fusion technology according to claim 1, characterized in that: The step of respectively calculating the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image comprises: Select n reference object areas on the spectral calibration image, each reference object area contains N1 pixels, and different reference object areas are distributed at different positions of the spectral calibration image; extract the radiance values ​​of all pixels in each reference object area in the visible light band, and calculate the measured spectral reflectance of each pixel in each reference object area in the visible light band according to the radiance value of each pixel in the visible light band and the solar altitude angle; extract the radiance values ​​of all pixels in the near-infrared band in each reference object area, calculate the solar zenith angle from the solar altitude angle, and calculate the measured spectral reflectance of each pixel in the near-infrared band in each reference object area according to the radiance value and the solar zenith angle of each pixel in the near-infrared band; Calculate the theoretical spectral reflectance of each pixel in the visible light band of each reference object area under cloud-free atmospheric conditions; calculate the theoretical spectral reflectance of each pixel in the near-infrared band of each reference object area under cloud-free atmospheric conditions; The ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in the visible light band of each reference object area is used as the first regionalized variable, and an interpolation model of the transmittance of the visible light band is established to obtain the transmittance of the visible light band at any pixel position in the spectral calibration image; The ratio of the measured spectral reflectance and the theoretical spectral reflectance of each pixel in the near-infrared band of each reference object area is taken as the second regionalized variable, and an interpolation model of the near-infrared band transmittance is established to obtain the transmittance of each pixel in the spectral calibration image in the near-infrared band.

3. The method for identifying cloud layers by using a satellite-borne hyperspectral method based on dual-band fusion technology according to claim 1, characterized in that: The identification of the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area includes: Calculate the transmittance difference between the visible light band transmittance and the near infrared band transmittance of each pixel in the spectral calibration image, and determine the difference threshold T th ;like >T th , then the pixel Mark as a candidate area for semi-transparent weather phenomena; otherwise, Marked as non-translucent weather phenomenon area; Pixels in the spectral calibration image The transmittance difference between the visible light band transmittance and the near infrared band transmittance, is the index of the pixel in the spectral calibration image; Taking pixels as units, spatial clustering is performed according to the spatial neighborhood relationship of the candidate areas of semi-transparent weather phenomena to obtain semi-transparent weather phenomenon areas and non-translucent weather phenomenon areas.

4. The method for identifying cloud layers by using a spaceborne hyperspectral method based on dual-band fusion technology according to claim 3, characterized in that: The difference threshold T is determined th The methods include: Obtain historical cloud images, and extract the visible light band transmittance and near infrared band transmittance in the semi-transparent weather phenomenon area and the real cloud layer area of ​​each historical cloud image in units of pixels; The average difference between the visible light band transmittance and the near-infrared band transmittance of all pixels in the semi-transparent weather phenomenon area of ​​the historical cloud image is calculated, which is recorded as the first difference; the average difference between the visible light band transmittance and the near-infrared band transmittance of all pixels in the real cloud layer area of ​​the historical cloud image is calculated, which is recorded as the second difference; according to the first difference and the second difference, the difference threshold T is determined. th .

5. The method for identifying cloud layers by using a spaceborne hyperspectral method based on dual-band fusion technology according to claim 3, characterized in that: The spatial clustering according to the spatial neighborhood relationship of the semi-transparent weather phenomenon candidate area includes: Taking pixels as units, the spatial coordinates of the candidate areas of semi-transparent weather phenomena are extracted, and an undirected graph is constructed. Each node in the undirected graph corresponds to a pixel in the candidate area, and there are edge connections between adjacent pixels in the candidate area. Calculate the weight of the edge in the undirected graph, where the weight of the edge is the spectral feature distance between adjacent candidate area pixels; The undirected graph is divided to obtain the set of cut edges with the minimum weight and the smallest sum. According to the set of cut edges, the pixels in the candidate area of ​​semi-transparent weather phenomenon are divided into N2 connected subgraphs. The pixels in the connected subgraphs are marked as semi-transparent weather phenomenon areas, and the remaining pixels are marked as non-semi-transparent weather phenomenon areas.

6. The method for identifying cloud layers by using a spaceborne hyperspectral method based on dual-band fusion technology according to claim 4, characterized in that: The constructing of a real cloud layer spectral characteristic model comprises: Extracting cloud regions from historical cloud images to obtain cloud regions of the historical cloud images; Extract the spectral features of the cloud area in the historical cloud image and construct the spectral feature vector of the real cloud layer; the spectral features include spectral reflectance features, spectral absorption features and spectral scattering features in the visible light band and near infrared band; The spectral feature vectors of real clouds are used as samples to train and generate a real cloud spectral feature model.

7. The method for identifying cloud layers by using a spaceborne hyperspectral method based on dual-band fusion technology according to claim 6, characterized in that: The method of using the real cloud spectral feature model to identify the real cloud layer in the non-translucent weather phenomenon area includes: The spectral features of the spectral calibration image of the non-translucent weather phenomenon area are extracted by sliding window, the spectral features of the area to be evaluated in the semi-transparent weather phenomenon area are obtained, and the spectral feature vector of the area to be evaluated is constructed; the spectral feature vector of the area to be evaluated is input into the real cloud layer spectral feature model to obtain the probability that the area to be evaluated is a real cloud layer area; if the probability that the area to be evaluated is a real cloud layer area is greater than a preset probability threshold, the area to be evaluated is determined to be a real cloud layer area, otherwise it is determined to be a non-real cloud layer area.

8. A spaceborne hyperspectral cloud recognition system based on dual-band fusion technology, which is used to implement a spaceborne hyperspectral cloud recognition method based on dual-band fusion technology as described in any one of claims 1 to 7, characterized in that: The system comprises: Spectral image calibration module: used to obtain a hyperspectral image, pre-process the hyperspectral image, and obtain a spectral calibration image; the hyperspectral image includes a visible light band image and a near-infrared band image; respectively calculate the visible light band transmittance and the near-infrared band transmittance of each pixel in the spectral calibration image; record the solar altitude angle when the hyperspectral image is obtained; Semi-transparent weather phenomenon area determination module: used to identify the semi-transparent weather phenomenon area and the non-semi-transparent weather phenomenon area according to the visible light band transmittance and near infrared band transmittance of each pixel in the spectral calibration image; Real cloud detection module: used to build a real cloud spectral feature model, use the real cloud spectral feature model to identify the real cloud in the non-translucent weather phenomenon area, obtain the real cloud area, merge the real cloud area with the translucent weather phenomenon area, and generate the final cloud detection result map.

Images