A cloud coverage feature inversion method and device for a hyperspectral atmospheric environment monitoring satellite and a storage medium

Abstract

The present disclosure provides a cloud coverage feature inversion method and device for hyperspectral atmospheric environment monitoring satellites and a storage medium, relating to the technical field of earth observation and remote sensing. By introducing a radiation transfer physical model, the spectral radiation response under cloud and clear sky scenarios is simulated and analyzed, and the radiation differences between clouds and the ground under different wavelengths are clearly distinguished at the spectral level. On the basis of this physical analysis, the satellite hyperspectral observation data is preliminarily screened, and the spectral segments with significant physical response to the cloud and clear sky differences are retained. Combined with the feature importance representation and optimization mechanism driven by data, the spectrum and its derived features are adaptively screened and weighted. An artificial intelligence model is used to model and analyze the screened features, to realize the inversion of the cloud coverage features of the observation pixels, including cloud amount inversion and the derived cloud mask results, taking into account the advantages of physical mechanism and data driving, and realizing high-precision, strong stability and good generalization ability of cloud detection.

Description

Technical Field

[0001] This invention relates to the field of Earth observation and remote sensing technology, and in particular to a method, apparatus and storage medium for cloud cover feature inversion for hyperspectral atmospheric environment monitoring satellites. Background Technology

[0002] Optical remote sensing sensors mounted on satellite platforms can acquire large-scale observational data continuously and stably over long periods, playing a crucial role in fields such as climate change research, air pollution monitoring, ecological environment assessment, and resource surveys. In particular, geostationary orbit atmospheric environment monitoring satellites, which have been developed in recent years, can continuously observe fixed areas with hourly temporal resolution, providing a new technological approach for monitoring the spatiotemporal changes in atmospheric composition.

[0003] However, clouds are a ubiquitous and unavoidable source of interference in optical remote sensing observations. According to the International Satellite Cloud Climatology Project (ISCCP), approximately 66% of the Earth's surface is covered by clouds at any given time. Clouds not only obscure surface targets but also significantly alter the propagation path and energy distribution of solar radiation in the atmosphere, thus significantly impacting the remote sensing inversion results of atmospheric components such as aerosols and trace gases. Therefore, before using satellite observation data to invert atmospheric environmental parameters, accurate cloud identification and cloud masking processing of the observed pixels are essential.

[0004] Atmospheric environment monitoring satellites typically employ imaging spectrometers in the ultraviolet-visible spectral band for observation. However, their spectral coverage is relatively limited, resulting in a small number of effective bands for cloud cover feature inversion. Furthermore, the sheer number and redundancy of spectral channels pose challenges for feature selection and model construction. Existing cloud cover feature inversion methods based on fixed thresholds or single physical models struggle to fully utilize their hyperspectral information, while purely data-driven artificial intelligence methods lack physical prior constraints, leading to issues such as underestimated cloud cover and insufficient identification accuracy under complex surface conditions.

[0005] Therefore, it is urgent to address the shortcomings of existing cloud cover feature inversion and cloud detection methods under complex surface and atmospheric conditions, such as insufficient cloud identification accuracy, weak identification ability of thin and fragmented clouds, and underestimate of cloud cover, which are caused by factors such as the limited spectral range of satellite sensors, the limited effective information channels, and information redundancy between channels. These shortcomings affect the inversion accuracy and reliability of atmospheric component products such as gases and aerosols. Summary of the Invention

[0006] In view of the many shortcomings of existing technologies, this disclosure provides a method, device and storage medium for cloud cover feature inversion for hyperspectral atmospheric environment monitoring satellites. By introducing a radiative transfer physics model to simulate and analyze different atmospheric and cloud scenarios, a physical feature library that can characterize the difference between clouds and clear sky is constructed. Based on this, feature selection and optimization are performed on satellite hyperspectral observation data. Subsequently, machine learning or statistical learning models are used to model and analyze the selected features to realize cloud cover inversion and cloud mask generation for observed pixels. This approach combines the advantages of physical mechanisms and data-driven approaches, achieving high-precision, strong stability and good generalization ability in cloud scenario inversion.

[0007] The first aspect of this disclosure provides a method for cloud cover feature inversion for hyperspectral atmospheric environment monitoring satellites, comprising the following steps: physical simulation using a radiative transfer model to compare the top-level atmospheric radiative response at different wavelengths for clear-sky and cloud-containing scenarios, identifying spectral segments or combinations of spectral features highly sensitive to cloud presence; based on these highly cloud-sensitive spectral segments or combinations of spectral features, performing preliminary feature screening on the original hyperspectral observation data based on physical priors, eliminating spectral channels that contribute little to cloud identification or are irrelevant to cloud sensitivity, obtaining candidate spectral features; inputting the candidate spectral features into a data-driven feature evaluation model, and during model training for the cloud cover feature inversion task, calculating and evaluating the importance score of each candidate spectral feature in parallel, and then evaluating the score based on the importance score. Importance scores are used to perform secondary screening of candidate spectral features to obtain a set of highly important features. Based on the feature set and the training sample set, continuous cloud presence probability values ​​are output, which represent the degree of cloud coverage within a pixel and are equivalently used as pixel-level cloud scores or cloud amount. A gradient boosting decision tree model is trained to learn the statistical differences between clear-sky and cloud-containing scenes in the multispectral feature space, resulting in a trained cloud coverage feature inversion model. The training sample set is obtained by spatiotemporally matching the original hyperspectral observation data with pre-acquired cloud reference products from mature third-party sensors. Using the cloud coverage feature inversion model, cloud amount values ​​are inverted pixel by pixel from the hyperspectral observation data of the target satellite to obtain cloud amount results that represent the degree of cloud coverage within a pixel in the form of continuous pixel-level values.

[0008] Unless otherwise stated, the terms “probability of cloud presence,” “cloud score,” and “cloud volume” in this article have the same technical meaning.

[0009] Furthermore, in some embodiments, the method further includes: performing spatial domain processing or morphological optimization on the cloud cover results to enhance the spatial continuity and boundary stability of the cloud region.

[0010] In some embodiments, the method further includes: converting the cloud cover value into a binary cloud mask based on the cloud cover result and in accordance with a preset discrimination rule, so as to determine whether the pixel is a cloud pixel or a clear sky pixel.

[0011] In some embodiments, a radiative transfer model is used for physical simulation to compare the top-level atmospheric radiative response of clear-sky and cloud-covered scenarios at different wavelengths, identifying spectral segments or combinations of spectral features that are highly sensitive to cloud presence. This includes: calculating the top-level atmospheric radiance or reflectivity of clear-sky and cloud-covered scenarios using multi-parameter combinations by setting the input parameters of the radiative transfer model, obtaining radiative response sequences of clear-sky and cloud-covered scenarios at different wavelengths, wherein the input parameters include at least: observation geometric parameters, atmospheric state parameters, surface parameters, cloud parameters, and spectral parameters; based on the radiative response sequences of clear-sky and cloud-covered scenarios at different wavelengths, obtaining characteristic curves of spectral radiation changing with wavelength in clear-sky and cloud-covered scenarios; and by analyzing the characteristic curves of spectral radiation changing with wavelength in clear-sky and cloud-covered scenarios, comparing the radiative differences between clear-sky and cloud-covered scenarios, identifying spectral segments or combinations of spectral features that cause significant enhancement, attenuation, or morphological changes in radiation due to cloud presence.

[0012] In some embodiments, preliminary feature screening based on physical priors includes: retaining observation channels in spectral segments identified as highly sensitive to cloud presence; eliminating channels with a signal-to-noise ratio below a preset threshold; calculating the correlation of radiative responses of adjacent channels; and merging consecutive channels with a correlation higher than a preset threshold.

[0013] In some embodiments, candidate spectral features are input into a data-driven feature evaluation model. During the model training process for the cloud coverage feature inversion task, the importance scores of each candidate spectral feature are calculated and evaluated in parallel. The candidate spectral features are then further screened based on their importance scores to obtain a set of highly important features. This includes: inputting candidate spectral features into a machine learning model for training; during the training process, calculating the contribution of each candidate spectral feature to the cloud classification or cloud amount inversion task and obtaining the importance score of each candidate spectral feature; selecting candidate spectral features with importance scores greater than a preset importance threshold to form a set of highly important features; or, sorting candidate spectral features from high to low according to their importance scores and selecting a preset proportion or number of candidate spectral features to form a set of highly important features.

[0014] In some embodiments, the gradient boosting decision tree model adopts the LightGBM framework. The training process of the gradient boosting decision tree model includes: parallelizing the data processing of the input features and compressing the features using a mutually exclusive feature bundling strategy; constructing a local feature histogram based on the histogram algorithm and the compressed input features; and merging the local feature histograms into a global feature histogram to find the globally optimal split point.

[0015] In some embodiments, the method further includes: optimizing key parameters of the gradient boosting decision tree model, wherein the key parameters of the supervised learning model include at least: the number of tree models, the maximum depth of a single tree, the learning rate, the minimum number of samples per leaf node, and the feature sampling ratio and sample sampling ratio; the optimization process includes: determining the optimal parameter combination suitable for the target satellite by performing grid search, random search, or cross-validation optimization on the key parameters.

[0016] The second aspect of this disclosure provides a cloud cover feature inversion device based on a geostationary orbit hyperspectral atmospheric environment monitoring satellite, comprising: a cloud-sensitive spectral analysis module, used to perform physical simulation using a radiative transfer model, comparing the top-level atmospheric radiative response at different wavelengths for clear-sky and cloud-containing scenarios, and identifying spectral segments or combinations of spectral features highly sensitive to cloud presence; a feature screening module, used to perform preliminary feature screening based on physical priors on the original hyperspectral observation data according to the spectral segments or combinations of spectral features highly sensitive to cloud presence, eliminating spectral channels that contribute little to cloud identification or are irrelevant to cloud sensitivity, and obtaining candidate spectral features; and a feature optimization module, used to input the candidate spectral features into a data-driven feature evaluation model, and during the model training process for the cloud cover feature inversion task, to calculate and evaluate the importance score of each candidate spectral feature in parallel, and to perform a secondary screening of the candidate spectral features based on the importance score, obtaining... The system comprises: a high-importance feature set; a cloud cover feature inversion model training module, which, based on the feature set and training sample set, outputs continuous cloud presence probability values, representing the degree of cloud coverage within a pixel, equivalently used as pixel-level cloud scores or cloud quantity; a gradient boosting decision tree model, which learns the statistical differences between clear-sky and cloud-containing scenes in the multispectral feature space, resulting in a trained cloud cover feature inversion model; a cloud quantity inference module, which uses the cloud cover feature inversion model to invert quantity values ​​pixel by pixel from the hyperspectral observation data of the target satellite, obtaining cloud quantity results representing the degree of cloud coverage within a pixel in the form of continuous pixel-level values; and an optimization module, which performs spatial domain processing or morphological optimization on the cloud quantity results to enhance the spatial continuity and boundary stability of the cloud region.

[0017] A third aspect of this disclosure provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the aforementioned first aspect.

[0018] A fourth aspect of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the aforementioned first aspect.

[0019] The fifth aspect of this disclosure provides a computer program product including a computer program that, when executed by a processor, implements the first aspect described above.

[0020] Due to the adoption of the above technical solution, the beneficial effects of this disclosure include: 1. This disclosure introduces a radiative transfer model in the feature construction stage to perform physically consistent simulation analysis of the spectral radiative response under cloud and clear sky scenarios, clearly distinguishing the radiative differences between clouds and the ground surface under different wavelength conditions at the spectral level. Through physical mechanism constraints, the selected spectral features theoretically possess high sensitivity to the presence of clouds, avoiding the problem of relying on empirical thresholds or statistical correlations for band selection in existing technologies.

[0021] 2. Based on the introduction of physical mechanism constraints, this disclosure performs preliminary screening and dimensionality compression on hyperspectral channels, retaining only spectral segments that have significant physical responses to the differences between clouds and clear skies. This can significantly reduce the dimensionality of the model input, reduce redundancy and noise features, directly reduce the complexity and overfitting risk of the machine learning model, thereby improving the stability and generalization ability of model training, and avoiding the problem of model overfitting or regional failure caused by "high-dimensional and high-noise input" in the prior art.

[0022] 3. Building upon the initial physical screening, this disclosure further introduces feature importance analysis based on the model training process to quantitatively evaluate the actual contribution of each spectral feature to the cloud recognition task. While ensuring physical rationality, the feature set is further optimized through a data-driven approach, ensuring that the final input features possess both physical interpretability and statistical discriminative power. This overcomes the limitations of existing technologies where "pure physical models are difficult to adapt to complex observation conditions" or "pure data-driven methods lack physical constraints." Unlike traditional physical methods that rely on iterative optimization based on relative entropy and mutual information, requiring recalculation of the probability distribution in each iteration, this application uses machine learning methods and radiative transfer simulation for automated screening, employing a combination of "initial physical screening + data-driven secondary screening."

[0023] 4. Unlike fixed threshold algorithms that are based on cloud cover profile inversion and then set a threshold for judgment, which can only report whether clouds exist, this application uses a machine learning algorithm based on LightGBM for cloud detection in hyperspectral images from geostationary satellites and can output cloud probability information.

[0024] 5. Unlike the binary cloud / clear weather determination method commonly used in existing technologies, this disclosure directly inverts the pixel-level cloud existence probability in the model output stage and interprets this probability as cloud score or cloud coverage probability. This can effectively characterize the cloud-clear weather transition area, reduce the impact of boundary pixel misjudgment on subsequent applications, and provide an adjustable discrimination basis for different application scenarios.

[0025] 6. This disclosure does not rely on a fixed satellite platform or a specific algorithm model, but provides a general technical framework that can flexibly select input features and learning models according to the characteristics of different satellite sensors. While ensuring computational efficiency, it significantly improves the accuracy and stability of cloud detection, has good cross-satellite applicability and promotional value, and can effectively improve the reliability of tropospheric atmospheric composition monitoring results. Attached Figure Description

[0026] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which: Figure 1 A flowchart illustrating a cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites according to an embodiment of the present disclosure is shown. Figure 2 An example diagram illustrating the comparison between a radiative transfer model simulation of a cloud-covered scenario and a clear-sky scenario according to an embodiment of the present disclosure is shown. Figure 3 An example diagram illustrating a machine learning-based feature importance evaluation result according to an embodiment of the present disclosure is shown schematically. Figure 4 The diagram schematically illustrates a flow example of a feature selection stage according to an embodiment of the present disclosure; Figure 5 This schematically illustrates a flowchart example of a model building phase according to an embodiment of the present disclosure; Figure 6 An example diagram illustrating a cloud recognition result according to an embodiment of the present disclosure is shown schematically, wherein, Figure 6 (a) is the RGB image of the observed scene. Figure 6 (b) is a diagram showing the cloud cover inversion results for the observed scene. Figure 6 (c) is a cloud mask image of the observed scene; Figure 7 This illustration schematically shows an example diagram for comparing and verifying cloud identification results according to an embodiment of the present disclosure, wherein... Figure 7 (a) is an example diagram of GOES-R cloud mask. Figure 7 (b) is an example image of an official TEMPO product. Figure 7 (c) is an example image of AERONET observations; Figure 8 This schematic diagram illustrates the architecture of a cloud cover feature inversion device for a hyperspectral atmospheric environment monitoring satellite according to an embodiment of the present disclosure; Figure 9 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0027] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0028] First, a detailed introduction to the technical background involved in this disclosure will be given.

[0029] Currently, in the field of satellite remote sensing cloud detection and cloud amount inversion technology, the common technical routes mainly include the following categories: (1) Cloud detection method based on multispectral threshold discrimination. This type of method usually sets a fixed or adaptive threshold according to the difference in radiation characteristics of clouds in different spectral bands, distinguishes cloud pixels from clear sky pixels, and thus generates a cloud mask. This type of method is simple to implement and has high computational efficiency, and has been widely used in multispectral satellites and their related business products. However, this method is mostly used in multispectral satellites and is highly dependent on empirical thresholds. It is difficult to efficiently transfer when the spectral range of hyperspectral sensors is limited and the bands are redundant, and its adaptability is poor. (2) Cloud parameter inversion method based on physical model. This type of method uses the radiative transfer model to perform forward or inverse calculations of satellite-observed radiation, and realizes cloud identification by estimating physical parameters such as cloud optical thickness, cloud top height, and effective cloud amount. This type of method has strong physical interpretability, but has high computational complexity and is sensitive to prior assumptions and input parameters. It has insufficient efficiency in real-time processing and large-scale applications. (3) Cloud detection method based on machine learning or deep learning. With the development of artificial intelligence technology, more and more research is applying supervised learning models to cloud detection tasks. By constructing feature inputs and training the model using existing cloud products or manually labeled samples, automatic classification of cloud pixels can be achieved. Such methods can achieve high accuracy under certain conditions, but they often rely on a large number of training samples, have limited model generalization ability, and lack physical constraints, making them prone to performance degradation when applied across regions or satellites.

[0030] Overall, existing cloud detection technologies are relatively mature on traditional meteorological satellites or polar-orbiting satellite platforms, and have led to various operational products. However, for geostationary orbit atmospheric environment monitoring satellites that have emerged in recent years, the above methods still have significant shortcomings. These satellites typically use imaging spectrometers in the ultraviolet-visible spectral band for observation, resulting in a relatively limited spectral coverage and a small number of effective bands for cloud identification. Furthermore, the large number and high redundancy of spectral channels pose challenges for feature selection and model construction. Existing methods based on fixed thresholds or single physical models struggle to fully utilize their hyperspectral information, while purely data-driven machine learning methods lack physical prior constraints, easily leading to problems such as underestimated cloud cover and insufficient identification accuracy under complex surface conditions.

[0031] Therefore, how to construct a cloud identification method that combines physical rationality and data-driven advantages, based on fully utilizing the observation characteristics of hyperspectral atmospheric environment monitoring satellites, and achieve high-precision, highly generalizable cloud detection, has become a key technical problem that urgently needs to be solved in this field.

[0032] To address cloud pollution issues in hyperspectral atmospheric monitoring satellite data, the industry typically employs a physical inversion method based on the absorption characteristics of the dual oxygen absorption channel (O2-O2) to obtain cloud cover. A cloud mask is then created by setting quality control thresholds. This is the main technical approach used in operational cloud products such as TEMPO (Tropospheric Emissions: Monitoring of Pollution), GEMS (Geostationary Environment Monitoring Spectrometer), and OMI (Ozone Monitoring Instrument). Specifically, TEMPO is the first geostationary hyperspectral observation instrument used to monitor tropospheric composition in North America, providing hourly daytime observation data. GEMS is another geostationary hyperspectral atmospheric monitoring instrument used for atmospheric pollution observation in East Asia. OMI is an ultraviolet-visible imaging spectrometer aboard NASA's Aura satellite, used to monitor global atmospheric composition. It is primarily used to invert various trace gases such as ozone (O3), nitrogen dioxide (NO2), and sulfur dioxide (SO2), as well as cloud and aerosol parameters.

[0033] The typical technical process includes: acquiring atmospheric top-layer radiation observation data using hyperspectral instruments in the ultraviolet to visible light band, especially narrow bands containing O2-O2 absorption characteristics; extracting the O2-O2 absorption band and adjacent continuous spectrum bands located around 477nm from the hyperspectral observation data, using the radiation contrast of these bands as the main physical basis for cloud identification. This step usually only uses a small number of fixed bands and does not fully utilize other large spectral channels; comparing the actual observed radiation with the calculation results of the radiative transfer model, and using least squares fitting or lookup table methods to invert and obtain the effective cloud fraction and related quality control parameters for each pixel; setting quality control parameters (Quality Assurance, QA) for the inverted cloud amount and setting a fixed threshold. When the cloud amount is greater than the threshold, the pixel is identified as a cloud; when the cloud amount is less than the threshold, the pixel is identified as clear sky; based on the threshold discrimination results, generating the final cloud detection product, mainly including: effective cloud coverage product; binarized cloud mask product obtained based on threshold conversion; and corresponding quality control identification information.

[0034] While this method can provide cloud cover and cloud mask information to some extent, it still has the following technical limitations in hyperspectral geostationary satellite applications: 1. High dependence on radiometric calibration accuracy leads to instability in cloud cover retrieval. Existing retrieval algorithms based on dual-oxygen absorption channels require matching observed radiometric values ​​with those calculated by the radiative transfer model. However, hyperspectral satellites have inherent radiometric calibration uncertainties. When calibration errors are introduced into the model, they cause systematic biases in the retrieved cloud cover. Therefore, even if the cloud physical state remains unchanged, the cloud cover retrieval results may still fluctuate with calibration errors, exhibiting poor stability.

[0035] 2. Fixed threshold strategies are prone to underestimating cloud coverage. Most existing hyperspectral satellite cloud masks are based on fixed thresholds (e.g., 0.5) for cloud cover. Thin clouds, edge clouds, and fragmented clouds are observed weakly, and their retrieved cloud cover is often lower than this threshold, thus being misclassified as clear skies. Therefore, there is a systematic underestimation of cloud coverage.

[0036] 3. Lack of effective utilization of multidimensional features. Because hyperspectral satellites have a large number of narrow bands, while existing physical models only utilize a few feature bands, a large amount of spectral information is not effectively used, which limits the algorithm's ability to represent complex cloud fields.

[0037] 4. Insufficient cross-satellite generalization capability. Due to differences in spectral range, sensor noise levels, and radiometric calibration accuracy among different satellites, the performance of existing algorithms varies significantly across different platforms, resulting in weak generalization ability.

[0038] To this end, this disclosure provides a cloud identification method, device, and storage medium applicable to geostationary orbit hyperspectral atmospheric environment monitoring satellites. It introduces physical mechanism constraints through a radiative transfer model for screening hyperspectral cloud-sensitive features; combines physical constraint feature screening with machine learning optimization based on feature importance to form a physical-data fusion feature construction mechanism; uses cloud existence probability rather than binary determination as the core output form of the model; introduces spatial consistency optimization and mask generation strategies based on cloud quantity results; and uses mature cloud products as supervisory labels to achieve cross-sensor cloud information migration and inversion, thereby improving the accuracy and stability of cloud quantity inversion and cloud mask generation, while enhancing applicability across satellite platforms.

[0039] The following description, with reference to the accompanying drawings, describes a method, apparatus, and electronic device for cloud cover feature inversion for hyperspectral atmospheric environment monitoring satellites, according to embodiments of the present disclosure.

[0040] Figure 1 A flowchart illustrating a cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites according to an embodiment of the present disclosure is shown.

[0041] like Figure 1 As shown, the method includes the following steps 101-104.

[0042] Step 101: Use the radiative transfer model to perform physical simulation, compare the top-level atmospheric radiative response at different wavelengths for clear sky and cloud-covered scenarios, and identify spectral segments or combinations of spectral features that are highly sensitive to the presence of clouds.

[0043] In some embodiments of this disclosure, a physical simulation model is constructed based on radiative transfer simulation, encompassing two typical scenarios: clouds and clear skies. The radiative response of the top atmosphere under different wavelength conditions is then calculated. Radiative transfer simulation refers to a method that simulates the propagation, absorption, and scattering processes of solar radiation in the atmosphere and on the Earth's surface based on an atmospheric physical model to obtain theoretical observation signals.

[0044] In some embodiments, step 101 includes: calculating the radiative brightness or reflectance of the top atmospheric layer in clear-sky and cloud-covered scenarios by setting input parameters of the radiative transfer model, respectively, to obtain radiative response sequences of clear-sky and cloud-covered scenarios at different wavelengths, wherein the input parameters include at least: observation geometric parameters, atmospheric state parameters, surface parameters, cloud parameters, and spectral parameters; based on the radiative response sequences of clear-sky and cloud-covered scenarios at different wavelengths, obtaining characteristic curves of spectral radiation changing with wavelength in clear-sky and cloud-covered scenarios; by analyzing the characteristic curves of spectral radiation changing with wavelength in clear-sky and cloud-covered scenarios, comparing the radiation differences between clear-sky and cloud-covered scenarios, and identifying spectral segments or combinations of spectral features that cause significant enhancement, attenuation, or morphological changes in radiation due to the presence of clouds.

[0045] Furthermore, the observed geometric parameters include the solar zenith angle, satellite-observed zenith angle, and azimuth angle; atmospheric state parameters include the standard atmospheric profile, aerosol type, and optical properties; surface parameters include surface albedo or bidirectional reflectance; cloud parameters include cloud optical thickness, cloud top height, cloud type (such as water cloud or ice cloud), and effective particle radius; and spectral parameters include the hyperspectral band range and spectral resolution to be analyzed.

[0046] In one model implementation, mature radiative transfer calculation tools such as SCIATRAN, MODTRAN, 6S, or other equivalent models are used to simulate clear-sky and cloud-covered scenarios under defined observation geometry conditions. By performing multi-parameter combination calculations on the two scenarios, radiative response sequences at different wavelengths are obtained.

[0047] The multi-parameter combination calculation involves changing cloud parameters, such as cloud optical thickness and cloud top height, by setting input parameters to form multiple parameter combinations. Radiative transfer simulations are then performed on each parameter combination to obtain a physical simulation dataset covering various cloud-containing scenarios, such as typical cases from thin to thick clouds and from low to high clouds. This contrasts sharply with existing technologies that typically use only a single fixed cloud model for table lookup or fitting, avoiding the problem of systematic underestimation of cloud coverage.

[0048] Furthermore, in terms of model verification, the results of the clear sky simulation can be compared with the actual satellite clear sky observations, the consistency analysis of the simulated radiation characteristics of cloud-containing scenarios with existing cloud products or measured data can be performed, or the physical rationality of the simulated radiation changing with cloud parameters can be verified by changing parameters such as cloud optical thickness and cloud top height.

[0049] In some embodiments, characteristic curves of spectral radiation as a function of wavelength are analyzed to identify spectral segments or features that are highly sensitive to the presence of clouds. This analysis focuses on the modulation effect of cloud presence on the radiation field, identifying significant radiative enhancement, attenuation, or morphological changes caused by clouds in different spectral segments, thereby determining the band range or combination of spectral features that are highly sensitive to the presence of clouds.

[0050] For example, within the ultraviolet to visible spectral range, the following types of cloud-sensitive features can be identified: in the ultraviolet bands such as 340-360 nm and 380-400 nm, regions where the enhancement effect of clouds on radiation is significant can be considered as sensitive segments for cloud identification; in the visible light bands of 440-480 nm and 500-550 nm, the contrast in reflectance between clouds and the ground surface is high, making them suitable for distinguishing clouds from the ground background; near 650-680 nm, the high reflectance characteristics of clouds are significantly different from those of vegetation and water bodies, and can be used as auxiliary discrimination features; at the same time, the difference between clouds and clear skies can be characterized by combining ratios or differences between different bands, i.e., spectral feature combinations, such as reflectance of band A / reflectance of band B.

[0051] For example, Figure 2 This paper describes a comparative analysis of the spectral radiance variations with wavelength for clear-sky and cloud-covered scenes. The horizontal axis represents wavelength (nm), and the vertical axis represents radiance at the top of the atmosphere (W / m²). 2 As shown in the figure, the red curve represents the spectral radiation of a cloud scene as a function of wavelength, and the blue curve represents the spectral radiation of a clear sky scene as a function of wavelength. The green and purple areas can be identified as regions that are highly sensitive to the presence of clouds.

[0052] Understandably, the above steps reveal the essential differences in spectral response between clouds and the Earth's surface from a physical mechanism perspective, providing a theoretical basis for subsequent feature selection and effectively reducing reliance on empirical thresholds or purely data-driven methods.

[0053] It should be noted that the radiative transfer model used for cloud-sensitive spectral analysis can be replaced by other atmospheric radiation models with multiple scattering calculation capabilities, as long as they can characterize the physical differences between clouds and clear skies in spectral response. This disclosure does not impose any restrictions on such models.

[0054] Step 102: Based on the spectral segments or combinations of spectral features that are highly sensitive to cloud presence, perform preliminary feature screening on the original hyperspectral observation data based on physical priors, and eliminate spectral channels that contribute little to cloud identification or are unrelated to cloud sensitivity to obtain candidate spectral features.

[0055] In the disclosed embodiments, preliminary feature screening achieves initial dimensionality reduction of hyperspectral features by eliminating spectral channels that contribute little to cloud identification, have no significant difference between clouds and clear skies, or are highly redundant. This step significantly reduces the dimensionality of input features while ensuring the integrity of physical information, alleviating the "curse of dimensionality" problem of hyperspectral data, improving the training efficiency and generalization ability of subsequent machine learning models, and reducing the potential interference of noise bands on cloud identification results.

[0056] In the disclosed embodiments, hyperspectral observation refers to a satellite observation method that acquires spectral information within a continuous and narrow band range, providing data from hundreds to thousands of spectral channels.

[0057] In some embodiments, preliminary feature screening includes directly retaining observation channels in the original hyperspectral observation data that are identified as cloud-sensitive spectral regions, while eliminating channels that contribute little to cloud identification. For example, if step 1 analysis determines that 340-360nm, 440-480nm, 500-550nm, and 650-680nm are sensitive regions, then all observation channels within these wavelength ranges are retained. The signal-to-noise ratio (SNR) of each observation channel is obtained or calculated, and all channels below a set threshold are eliminated to exclude channels with excessive noise or SNR below the practical level. Within the retained channels, the radiative response correlation of adjacent channels is calculated, and consecutive observation channels with high correlation are merged to eliminate redundant channels, further compressing the dimensionality with almost no loss of information.

[0058] Step 103: Input the candidate spectral features into the data-driven feature evaluation model. During the model training process for the cloud cover feature inversion task, calculate and evaluate the importance score of each candidate spectral feature in parallel, and perform secondary screening of the candidate spectral features based on the importance score to obtain a set of highly important features.

[0059] In the disclosed embodiments, machine learning refers to a type of computational method that trains an algorithm model using sample data, enabling it to automatically learn feature patterns and perform classification or prediction.

[0060] In some embodiments, step 103 includes: inputting candidate spectral features into a machine learning model for training; during the training process, calculating the contribution of each candidate spectral feature to the cloud classification or cloud amount inversion task, and obtaining the importance score of each candidate spectral feature; selecting candidate spectral features with an importance score greater than a preset importance threshold to form a high-importance feature set; or, sorting the candidate spectral features from high to low according to their importance scores, and selecting the top-ranked candidate spectral features by a preset proportion or number to form a high-importance feature set.

[0061] Specifically, this disclosure introduces a machine learning-based feature importance assessment method to quantitatively analyze the relative contribution of each candidate spectral feature to the cloud identification task. This includes constructing a feature importance ranking by statistically analyzing the contribution of features to the improvement of classification or regression performance during model training. Based on the feature importance assessment results, features that significantly contribute to cloud classification or cloud quantity inversion are further selected according to preset screening rules. For example, feature importance scores can be used as a screening criterion, setting an importance threshold (e.g., retaining features with importance values ​​greater than the set threshold), or selecting a certain proportion or number of features ranked high according to importance (e.g., the top N most important features). This eliminates bands that contribute little to cloud identification or have high redundancy, forming the final set of model input features. This process achieves an organic combination of physical priors and data-driven methods, improving the objectivity and relevance of feature selection while ensuring physical interpretability.

[0062] In one implementation, a decision tree-type machine learning model is used to evaluate feature importance, as follows: Assuming the training dataset is D, during model training, for each split node of the decision tree, the algorithm traverses candidate features f and selects the optimal split feature by maximizing the information gain IG. The formula for calculating information gain is shown in Formula 1 below: in, This represents the impurity of dataset D. Represents the set of all possible values ​​for feature f. This represents the subset of samples corresponding to the value v of feature f. This indicates the number of samples in the subset. This represents the total number of samples in dataset D.

[0063] Gini index It can measure the impurity of nodes, and the specific calculation formula is shown in Formula 2: Where C represents the total number of categories, This represents the proportion of samples belonging to category c in dataset D.

[0064] Based on the above impurity measurement methods, the overall importance score of feature f is... The impurity reduction of this feature across all decision trees and all split nodes is calculated by accumulating the impurity reduction across all decision trees and all split nodes, as shown in Formula 3. Where T represents the set of all decision trees in the gradient boosting tree model, This represents the reduction in impurity (i.e., the gain) resulting from splitting at node t using feature f.

[0065] For example, Figure 3 A characteristic curve depicting the importance assessment result as a function of wavelength is presented, with the horizontal axis representing wavelength (in nm) and the vertical axis representing the importance assessment result. For example... Figure 3 As shown, the importance threshold was set to 1000, and the marked wavelengths corresponding to the peak points in the graph with importance scores higher than the importance threshold were selected, highlighting the features that the model depends on most.

[0066] It should be noted that feature importance analysis is not limited to the implementation method based on gradient boosting decision trees, but can also be completed using other interpretable machine learning methods or statistical contribution evaluation methods, and this disclosure does not impose any restrictions.

[0067] Step 104: Based on the feature set and training sample set, the continuous cloud presence probability value is output. The cloud presence probability value represents the degree of cloud coverage within a pixel and is equivalent to being used as a pixel-level cloud score or cloud quantity. The gradient boosting decision tree model is trained to learn the statistical difference features between clear sky scenes and cloud scenes in the multispectral feature space, and a well-trained cloud coverage feature inversion model is obtained.

[0068] The cloud presence probability value represents the degree of cloud coverage within a pixel, and is equivalent to being used as a pixel-level cloud score or cloud amount. The training sample set is obtained by matching raw hyperspectral observation data with cloud mask labels used for reference.

[0069] In some embodiments, the process of acquiring the training sample set includes: selecting existing mature and reliable cloud products or multi-source observation data as cloud / clear sky reference data; and accurately matching these data with hyperspectral satellite observation data through spatial registration, time synchronization, and observation geometry consistency to construct the training sample set required for supervised learning. The samples obtained through this process not only have consistency in spatial and temporal scales but also possess high reliability in cloud identification, providing a reliable reference label foundation for the stable training of the cloud coverage feature inversion model.

[0070] In some embodiments, the supervised learning model may employ a machine learning model architecture suitable for tabular high-dimensional features, such as a gradient boosting decision tree model, a random forest model, a support vector machine model, or a neural network model.

[0071] In one implementation, a gradient boosting decision tree framework is used as the main structure of the model, and nonlinear modeling of cloud and clear sky features is achieved through the cascading ensemble of multiple weak classifiers. The model parameters are iteratively optimized by minimizing a preset loss function, enabling the model to fully learn the statistical differences between clouds and clear sky in the multispectral feature space. The loss function can be a cross-entropy loss function, a log-likelihood loss function, or a mean squared error loss function, etc.

[0072] Furthermore, the key parameters of the supervised learning model include at least: the number of tree models, for example, 100 to 2000 trees; the maximum depth of a single tree, for example, 3 to 20 trees; the learning rate, for example, 0.01 to 0.2 trees; the minimum number of samples per leaf node, for example, 10 to 100 trees; and the feature sampling ratio and sample sampling ratio. In the process of training the supervised learning model, the optimal combination of parameters suitable for the target satellite is determined by optimizing the key parameters through grid search, random search, or cross-validation, thereby obtaining a stable and reliable cloud coverage feature inversion model.

[0073] It should be noted that the goal of model training in this step is not simple binary classification, but to output continuous cloud presence probabilities, thereby more precisely characterizing the cloud-clear transition region and related cloud coverage features, providing a foundation for subsequent cloud mask generation and uncertainty analysis.

[0074] The feature selection stage of this disclosure combines physically constrained feature screening with machine learning optimization based on feature importance. For example, Figure 4A flowchart illustrating the feature selection stage is provided, as shown in the figure. Explainable artificial intelligence (XAI) is employed, taking raw satellite radiation data (such as TEMPO / GEMS radiances) and reference cloud mask labels (such as GOES-R / Himawari cloud masks) as input, and outputting a feature set obtained through feature importance calculation. The process includes: initial feature screening of the raw satellite radiation data based on the comparison results of radiative transfer simulation; aligning the raw satellite radiation data with the reference cloud mask labels using spatiotemporal matching under physical constraints; constructing samples using random sampling; and optimizing the features through feature importance calculation to obtain the feature importance assessment results. Among them, XAI refers to artificial intelligence methods that can reveal the basis of model decision-making, explain the importance of features, and improve transparency; feature selection refers to the process of screening out the variables or spectral channels that are most useful for model prediction from a large amount of input data; GOES-R cloud products refer to mature cloud detection and cloud characteristic inversion products provided by the GOES-R series of geostationary meteorological satellites, which are often used as reference or validation data.

[0075] For example, Figure 5 Provides a flowchart example of the model building phase, such as Figure 5As shown, the supervised learning model adopts the gradient boosting decision tree framework LightGBM. The model training and optimization process includes: inputting core observation features selected by feature selection, including the observed radiometric values, observation geometric angles, and reference cloud mask labels of the hyperspectral sensor in the ultraviolet and visible bands; segmenting and numbering the input features as Feature1 to Feature6; performing data parallelization on the features; and allocating the features to the working nodes of multiple parallel computing tasks (including Worker1-Worker3) for parallel optimization. Parallel optimization includes using exclusive feature bundling for feature compression, selecting the most important local features (local top features) at each task node through local voting, and then selecting the most important local features as candidate features through global voting. Local feature histograms are constructed based on a histogram-based algorithm, and the local feature histograms are merged to obtain global feature histograms to find and output the global best split.

[0076] Step 105: Using the cloud cover feature inversion model, cloud cover values ​​are inverted pixel by pixel from the hyperspectral observation data of the target satellite to obtain cloud cover results that characterize the degree of cloud cover within a pixel in the form of continuous numerical values ​​at the pixel level.

[0077] Optionally, the cloud probability results can be further optimized for spatial consistency to enhance the spatial continuity and boundary stability of the cloud region; the spatial consistency optimization can be implemented by different forms of neighborhood constraints, filtering or structured post-processing methods.

[0078] In some embodiments, the method further includes: converting the cloud cover value into a binary cloud mask based on the cloud cover result and in combination with a preset or adaptive discrimination rule, so as to determine whether the pixel is a cloud pixel or a clear sky pixel.

[0079] In one implementation, a fixed threshold discrimination rule can be used to binarize cloud cover. For example, pixels with cloud cover greater than or equal to a preset threshold are identified as cloud pixels, and pixels with cloud cover less than the threshold are identified as clear sky pixels. The preset threshold can be 0.5, or it can be set to other values ​​between 0.3 and 0.7 depending on different satellite platforms, observation scenarios, or application requirements.

[0080] For example, Figure 6 An example image of cloud detection results is provided, wherein, Figure 6 (a) is an RGB image showing the actual observation scene; Figure 6 (b) is the cloud fraction result map, where each pixel is a continuous probability value; Figure 6 (c) is a cloud mask image, which is a classification image generated by converting continuous probability values ​​into binary cloud masks. Red marks the pixels containing clouds and blue marks the pixels with clear skies.

[0081] For example, Figure 7 An example image is provided for comparing and verifying cloud recognition results, using red to mark pixels containing clouds and blue to mark pixels with clear skies. Figure 7 (a) is an example diagram of GOES-R cloud mask. The GOES-R product algorithm has been rigorously verified as an industry benchmark and can be used as reference or verification data. Figure 7 (b) is an example image of the official TEMPO product. The official TEMPO algorithm uses a physical inversion algorithm based on specific absorption bands such as O2-O2, which is the existing mainstream algorithm and is used as reference data. Figure 7 (c) is an example diagram of AERONET observations, representing ground-based measured data.

[0082] To verify the accuracy, the GOES-R cloud mask was first used as the ground truth. The cloud mask generated using the method disclosed in this publication achieved a Balanced Overall Accuracy (BOA) of 84.43%, significantly better than the 73.86% BOA of the official TEMPO cloud mask. This improvement in accuracy directly demonstrates the improvement of the method disclosed in this publication compared to existing technologies. To further confirm this, independent verification was performed with AERONET observation data. The cloud mask BOA generated using the method disclosed in this publication was 78.54%, also much higher than the 54.88% BOA of the official TEMPO product, confirming this improvement.

[0083] It should be noted that if other supervised or semi-supervised learning models are used, they should also be considered to fall within the scope of protection of this invention under the following preconditions: the input features are derived from hyperspectral information screened through physical mechanism analysis; the model output is the probability of continuous cloud presence or equivalent cloud cover information; and finally, cloud mask results are generated through discrimination rules. All alternative solutions that meet these preconditions do not deviate from the core technical approach of this invention, which combines physical constraint feature construction with probabilistic cloud inversion to achieve high-precision cloud detection.

[0084] Figure 8 This schematically illustrates a cloud cover feature inversion device 200 for a hyperspectral atmospheric environment monitoring satellite according to an embodiment of the present disclosure, such as... Figure 8 As shown, the device includes: The cloud-sensitive spectral analysis module 210 is used to simulate and compare the top-level atmospheric radiation response at different wavelengths in clear sky and cloud-containing scenarios using a radiative transfer model, and to identify spectral segments or combinations of spectral features that are highly sensitive to clouds. The feature screening module 220 is used to perform preliminary feature screening on the original hyperspectral observation data based on spectral segments or combinations of spectral features that are highly sensitive to the presence of clouds, and to obtain candidate spectral features. The feature optimization module 230 is used to evaluate the importance of candidate spectral features in the cloud recognition task based on machine learning, and to perform secondary screening of candidate spectral features based on the evaluation results to obtain a set of highly important features. The cloud cover feature inversion model training module 240 is used to train a supervised learning model based on the feature set and training sample set, with the continuous cloud presence probability as the output, to learn the statistical difference features between clear sky scenes and cloud-containing scenes in the multispectral feature space, and obtain a trained cloud cover feature inversion model. The training sample set is obtained by matching the original hyperspectral observation data with cloud mask labels used for reference. The cloud cover inference module 250 is used to invert the probability value of cloud presence on a pixel-by-pixel basis from the hyperspectral observation data of the target satellite using a cloud cover feature inversion model, and obtain the cloud cover result at the pixel level.

[0085] It should be noted that the specific implementation methods of the embodiments disclosed herein are different from those of the embodiments in this paper. Figure 1 The principle of the embodiments shown is the same, and will not be repeated here.

[0086] Figure 9A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0087] like Figure 9 As shown, device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 302 or a computer program loaded from storage unit 308 into RAM (Random Access Memory) 303. RAM 303 can also store various programs and data required for the operation of device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. I / O (Input / Output) interface 305 is also connected to bus 304.

[0088] Multiple components in device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of monitors, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0089] The computing unit 301 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processors, controllers, microcontrollers, etc. The computing unit 301 performs the various methods and processes described above, such as the cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites. For example, in some embodiments, the cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by computing unit 301, one or more steps of the method described above can be performed. Alternatively, in other embodiments, computing unit 301 can be configured by any other suitable means (e.g., by means of firmware) to perform the aforementioned cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites.

[0090] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0091] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0092] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0093] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0094] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0095] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0096] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0097] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps disclosed in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0098] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for inverting cloud cover characteristics for hyperspectral atmospheric environment monitoring satellites, characterized in that, The method includes: Physical simulations were performed using a radiative transfer model to compare the top-level atmospheric radiative response at different wavelengths in clear-sky and cloud-covered scenarios, identifying spectral segments or combinations of spectral features that are highly sensitive to cloud presence. Based on the spectral segments or combinations of spectral features that are highly sensitive to cloud presence, preliminary feature screening based on physical priors is performed on the original hyperspectral observation data to remove spectral channels that contribute little to cloud identification or are irrelevant to cloud sensitivity, thereby obtaining candidate spectral features. The candidate spectral features are input into a data-driven feature evaluation model. During the model training process for the cloud coverage feature inversion task, the importance scores of each candidate spectral feature are calculated and evaluated in parallel. The candidate spectral features are then screened a second time based on the importance scores to obtain a set of highly important features. Based on the feature set and training sample set, a continuous cloud presence probability value is output, which represents the degree of cloud coverage within a pixel and is equivalently used as a pixel-level cloud score or cloud amount. A gradient boosting decision tree model is trained to learn the statistical difference features between the clear sky scene and the cloud-containing scene in the multispectral feature space, thereby obtaining a trained cloud coverage feature inversion model. The training sample set is obtained by spatiotemporal matching of the original hyperspectral observation data with pre-acquired cloud reference products from mature third-party sensors. Using the cloud cover feature inversion model, cloud cover values ​​are inverted pixel by pixel from the hyperspectral observation data of the target satellite, resulting in cloud cover results that characterize the degree of cloud cover within a pixel in the form of continuous numerical values ​​at the pixel level.

2. The cloud scene inversion method for hyperspectral atmospheric environment monitoring satellites according to claim 1, characterized in that, The physical simulation using a radiative transfer model compares the top-level atmospheric radiative response at different wavelengths for clear-sky and cloud-covered scenarios, identifying spectral segments or combinations of spectral features highly sensitive to cloud presence, including: By setting the input parameters of the radiative transfer model, the radiance or reflectance of the top atmospheric layer in clear sky and cloud-covered scenarios are calculated using a combination of multiple parameters to obtain the radiative response sequences of the clear sky and cloud-covered scenarios at different wavelengths. The input parameters include at least: observation geometric parameters, atmospheric state parameters, surface parameters, cloud parameters, and spectral parameters. Based on the radiation response sequences of the clear sky scene and the cloud-containing scene at different wavelengths, characteristic curves of spectral radiation as a function of wavelength under the clear sky scene and the cloud-containing scene are obtained; By analyzing the characteristic curves of spectral radiation as a function of wavelength in the clear sky scene and the cloud-covered scene, and comparing the radiation differences between the clear sky scene and the cloud-covered scene, spectral segments or combinations of spectral features that cause significant enhancement, attenuation or morphological changes in radiation due to the presence of clouds can be identified.

3. The cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites according to claim 1, characterized in that, The preliminary feature screening based on physical priors includes: Observation channels within spectral ranges identified as highly sensitive to cloud presence are retained; channels with a signal-to-noise ratio below a preset threshold are removed; the correlation of radiative responses of adjacent channels is calculated, and continuous channels with a correlation above a preset threshold are merged.

4. The cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites according to claim 1, characterized in that, The candidate spectral features are input into a data-driven feature evaluation model. During model training for the cloud cover feature inversion task, the importance score of each candidate spectral feature is calculated and evaluated in parallel. Based on the importance score, the candidate spectral features are further screened to obtain a set of highly important features, including: The candidate spectral features are input into the machine learning model for training. During the training process, the contribution of each candidate spectral feature to the cloud cover inversion or cloud classification task is statistically analyzed to obtain the importance score of each candidate spectral feature. Candidate spectral features with importance scores greater than a preset importance threshold are selected to form a set of highly important features; or, the candidate spectral features are sorted from high to low according to their importance scores, and a preset proportion or number of candidate spectral features are selected to form the set of highly important features.

5. The cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites according to claim 1, characterized in that, The gradient boosting decision tree model uses the LightGBM framework, and the training process of the gradient boosting decision tree model includes: The input features are processed in parallel and compressed using a mutually exclusive feature bundling strategy. Based on the histogram algorithm and the compressed input features, a local feature histogram is constructed. The local feature histograms are merged into a global feature histogram to find the globally optimal split point.

6. The cloud cover feature inversion method for hyperspectral atmospheric environment monitoring satellites according to claim 1, characterized in that, The method further includes: The key parameters of the gradient boosting decision tree model are optimized, wherein the key parameters of the gradient boosting decision tree model include at least: the number of tree models, the maximum depth of a single tree, the learning rate, the minimum number of samples in a leaf node, and the feature sampling ratio and the sample sampling ratio; the optimization includes: determining the optimal parameter combination suitable for the target satellite by performing grid search, random search, or cross-validation optimization on the key parameters.

7. A cloud cover feature inversion device for hyperspectral atmospheric environment monitoring satellites, characterized in that, The device includes: The cloud-sensitive spectral analysis module is used to perform physical simulations using a radiative transfer model, compare the top-level atmospheric radiative response at different wavelengths in clear-sky and cloud-containing scenarios, and identify spectral segments or combinations of spectral features that are highly sensitive to clouds. The feature screening module is used to perform preliminary feature screening on the original hyperspectral observation data based on physical priors according to the spectral segments or combinations of spectral features that are highly sensitive to the presence of clouds, and to remove spectral channels that contribute little to cloud identification or are unrelated to cloud sensitivity, so as to obtain candidate spectral features. The feature optimization module is used to input the candidate spectral features into the data-driven feature evaluation model. During the model training process for the cloud coverage feature inversion task, the module calculates and evaluates the importance score of each candidate spectral feature in parallel, and performs secondary screening of the candidate spectral features based on the importance score to obtain a set of highly important features. A cloud cover feature inversion model training module is used to train a gradient boosting decision tree model to learn the statistical difference features between the clear sky scene and the cloud-containing scene in the multispectral feature space, based on the feature set and the training sample set, with continuous cloud presence probability values ​​as the output. The cloud presence probability values ​​represent the degree of cloud coverage within a pixel and are equivalently used as a pixel-level cloud score or cloud amount. The training sample set is obtained by spatiotemporal matching of the original hyperspectral observation data and pre-acquired cloud reference products from mature third-party sensors. The cloud amount inference module is used to invert the cloud coverage feature inversion model to retrieve the cloud amount values ​​pixel by pixel from the hyperspectral observation data of the target satellite, so as to obtain the cloud amount result in the form of continuous numerical values ​​at the pixel level to represent the degree of cloud coverage within the pixel.

8. An electronic device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images