A self-supervised cloud top height retrieval method based on double-star asymmetric spectral residual

Abstract

This invention discloses a self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals. It does not rely solely on geometric parallax or single-star observations, nor on pixel-level parallax calculations. Instead, it utilizes binary star collaborative fusion, leveraging the advantages of FY-4B in overall cloud altitude data quality and FY-4A in cloud classification reliability. Based on binary asymmetric spectral residuals and self-supervised physical constraints, it corrects errors in misclassified cloud types, effectively mitigating systematic biases caused by cloud classification errors. This improves the accuracy and robustness of overall cloud top altitude inversion, solving the challenges of fusion of heterogeneous binary star data and altitude bias issues under specific cloud classifications.

Description

Technical Field

[0001] This invention relates to the field of cloud top height estimation technology in meteorology, and in particular to a self-supervised cloud top height inversion method based on binary star asymmetric spectral residuals. Background Technology

[0002] Cloud top height (CTH) is a key physical parameter in atmospheric science and meteorological monitoring. It is not only directly related to the evolution of weather systems (such as the development height of severe convective storms), but also plays an irreplaceable role in aviation safety, Earth's radiation balance calculation, and numerical weather prediction data assimilation.

[0003] Traditional cloud top height inversion mainly relies on physical radiative transfer models, with representative methods including the infrared window method and... The slicing method. The infrared window method assumes the cloud body is a blackbody and determines altitude by matching the brightness temperature of the infrared window region with the atmospheric temperature profile. This method works well for opaque, thick clouds, but for semi-transparent cirrus clouds, the transmission of radiation from the lower warm atmosphere leads to a significantly lower inversion result. Slicing method Multiple channels along the edge of the absorption band detect cloud top pressure and exhibit good sensitivity to mid-to-high-level thin cirrus clouds. However, physical methods heavily rely on high-precision temperature and humidity profile data and accurate simulations of radiative transfer modes, resulting in high computational complexity. Furthermore, when dealing with multi-layered cloud scenarios, they often produce significant errors due to the inability to decouple the radiative contributions of upper-layer thin clouds from those of lower-layer water clouds.

[0004] While both Fengyun-4A and Fengyun-4B satellites possess high spatiotemporal resolution advantages in cloud top altitude retrieval, significant differences exist. The Advanced Radiometric Imager (ARI) on FY-4B outperforms FY-4A in spectral channel configuration and calibration accuracy, resulting in superior overall retrieval performance. However, FY-4B suffers from a systematic bias of misclassifying cirrus clouds as multi-layered clouds, leading to an underestimation of cirrus cloud top altitudes. FY-4A, on the other hand, demonstrates greater stability in cirrus cloud identification. Therefore, it is difficult for a single satellite to simultaneously achieve both overall accuracy and precise altitude recognition for specific cloud types.

[0005] With the successful networking of Fengyun-4A and Fengyun-4B satellites, the "dual-satellite parallax method" based on stereo vision principles has become a research hotspot. This method utilizes the different observation angles of the two satellites on the same cloud cluster, calculating the parallax of the cloud cluster's projected displacement on the Earth's surface to geometrically determine cloud height. However, the parallax method heavily relies on image texture feature matching. For large areas of smooth stratiform clouds or thin cirrus clouds with poor texture, stereo matching algorithms are prone to failure. Furthermore, this method is purely based on geometric relationships, completely ignoring the upgrades in spectral channels (new features) made by FY-4B compared to FY-4A. Channel Ch15) failed to take advantage of spectral physics.

[0006] With the development of artificial intelligence, single-satellite inversion methods based on deep learning have gradually emerged. These methods use multi-channel spectral data from a single satellite as input and invert cloud height by learning the nonlinear mapping between channel data and cloud height labels. However, the single-satellite perspective is limited by the "same temperature, different height" problem, meaning clouds at different altitudes may exhibit the same brightness temperature. Furthermore, deep learning models are typically "black boxes," making it difficult to use physical constraints to correct misjudgments of specific cloud types.

[0007] In the prior art, CN 117805940 A discloses a method for measuring cloud top height based on a dual-satellite imager. This method utilizes two geostationary satellites (specifically Fengyun-4A and Fengyun-4B) positioned at different longitudes as observation platforms to acquire satellite images recorded simultaneously by the imagers of the two satellites. The method first determines the target area in the first satellite image acquired by the first satellite, recording the first abscissa and corresponding latitude and longitude coordinates of this area in the image. Then, based on the latitude and longitude coordinates, it quickly determines the pixel coordinates of the target area in the second satellite image from the image acquired by the second satellite, and records the second abscissa. Next, it calculates the visual difference based on the difference between the first and second abscissas, and calculates the vertical distance between the cloud top of the target area and the satellite plane using a binocular vision-based ranging method (i.e., constructing a triangulation formula using the imager focal length, the optical center distance between the two satellites, and the coordinate difference). Finally, it calculates the final cloud top height by subtracting the height of the satellite plane from the calculated distance from the cloud top to the satellite plane.

[0008] Although the aforementioned existing technologies have solved the cloud height measurement problem to some extent, they have the following significant technical shortcomings in handling complex cloud phases and achieving high-precision inversion:

[0009] 1) Neglecting the utilization of spectral physical features, resulting in weak detection capability for thin cirrus clouds: Existing technologies ignore the rich spectral information contained in the Advanced Geosynchronous Radiation Imager (AGRI) carried by the FY-4B satellite. For high-altitude thin cirrus clouds, their optical thickness is extremely low, and their texture features are sparse and smooth. Traditional methods are prone to failure in such regions, leading to errors in parallax calculation.

[0010] 2) Lack of cloud phase semantic differentiation and error correction mechanisms: Existing technologies cannot solve the systemic misclassification problem in the FY-4B cloud classification product that misclassifies cirrus clouds as multi-layer clouds. When encountering a "multi-layer cloud" structure where semi-transparent cirrus clouds cover lower-level water clouds, existing technologies often lock the lower-level cloud with clearer textures, resulting in a severely underestimated height. Summary of the Invention

[0011] This invention addresses the problems of existing technologies neglecting spectral physical features, leading to failure in detecting thin cirrus clouds with poor texture, and lacking semantic error correction mechanisms, which cannot solve the systematic bias caused by misclassification of cirrus clouds with specific cloud phases. It proposes a self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals. This method no longer relies solely on geometric parallax or single-star observations, nor on pixel-level parallax calculations. Instead, it achieves error correction for misclassified cloud types through binary-star collaborative fusion, leveraging the advantages of FY-4B in overall cloud altitude data quality and FY-4A in cloud classification reliability. Based on binary asymmetric spectral residuals and self-supervised physical constraints, it effectively mitigates the systematic bias caused by cloud classification errors, thereby improving the accuracy and robustness of overall cloud top altitude inversion. This solves the challenges of fusion of heterogeneous binary-star data and the altitude bias problem under specific cloud classifications.

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] Firstly, a self-supervised cloud top height inversion method based on binary star asymmetric spectral residuals includes the following steps:

[0014] S1. Acquire primary radiometer observation data and secondary cloud classification products from FY-4A and FY-4B satellites at the same time, and perform spatiotemporal matching to obtain standardized channel data, cloud detection, and cloud phase classification labels with pixel alignment within the shared field of view;

[0015] S2. Based on asymmetric difference and A gated spectral residual fusion network processes the standardized channel data to generate the initial cloud top height;

[0016] The asymmetric difference and Gated spectral residual fusion networks include:

[0017] The first N channels of FY-4A data are used as the input base stream to extract baseline features and output the baseline cloud top height;

[0018] The correction stream is input using the differential data of corresponding channels of FY-4B and FY-4A, and the Mth channel data of FY-4B, to extract differential features and The channel's high-level semantic features are analyzed, and the cloud height correction and dynamic gating coefficient are output.

[0019] The fusion module is used to calculate and output the initial cloud top height based on the base cloud top height, cloud height correction amount and dynamic gating coefficient through weighted residual calculation;

[0020] S3. Process the standardized channel data, cloud classification labels, and the initial cloud top height based on a spectral-semantic dual-stream adaptive network to generate the final corrected cloud top height;

[0021] The spectral-semantic dual-stream adaptive network includes:

[0022] Semantic streams are used to encode and embed high-dimensional semantics into the cloud classification labels of FY-4A and FY-4B, generating semantic difference features.

[0023] The physical stream is used to extract spectral features from at least two specified channels of FY-4B data and output a cirrus confidence feature map.

[0024] The arbitration fusion module is used to generate a conflict region mask based on the semantic difference features, fuse the semantic difference features and the cirrus confidence feature map within the conflict region to generate arbitration weights, and use the arbitration weights to perform weighted compensation on the initial cloud top height difference between FY-4A and FY-4B, and output the final corrected cloud top height.

[0025] Further, step S1 includes:

[0026] Acquire primary multichannel radiometer observation data and secondary cloud classification products at the same time of FY-4A and FY-4B; the primary multichannel radiometer observation data includes visible light channels, near-infrared channels, and infrared channels; the secondary cloud classification products include cloud detection and cloud phase classification labels.

[0027] The observation data from the two satellites are reprojected onto the same standard latitude and longitude grid, and the shared viewing area is cropped to ensure that the pixels correspond one-to-one.

[0028] The observation data from the first-level multichannel radiometer were Z-score standardized to have a mean of 0 and a variance of 1.

[0029] Further, in step S2, the correction stream includes:

[0030] The differential encoder is used to process the differential data of the corresponding channels of FY-4B and FY-4A and extract differential features that characterize the systematic deviation between the two satellites.

[0031] The channel semantic encoder is used to process the Mth channel data of FY-4B and extract high-level semantic features.

[0032] The feature fusion layer is used to concatenate the differential features with high-level semantic features;

[0033] A correction predictor is used to predict cloud height correction based on the output of the feature fusion layer.

[0034] A gating controller is used to generate dynamic gating coefficients with values ​​ranging from 0 to 1 based on high-level semantic features.

[0035] Furthermore, in step S2, the formula for the fusion module to calculate the initial cloud top height is:

[0036]

[0037] in, Baseline predictions from FY-4A, Cloud height correction derived from channel differential features and Ch15. High cloud confidence weights from Ch15.

[0038] Further, in step S3, the semantic flow includes:

[0039] One-Hot encoding units are used to convert the cloud classification label maps of FY-4A and FY-4B into multi-channel sparse tensors, respectively;

[0040] The semantic embedding unit is used to map multi-channel sparse tensors to dense semantic feature vectors through learnable convolutional layers and to calculate the difference between two semantic feature vectors to generate semantic difference features.

[0041] Furthermore, in step S3, the physical flow is a lightweight convolutional neural network, whose input is the 4th and 15th channel data of FY-4B. It extracts spectral physical features through several convolutional layers and outputs a cirrus confidence feature map with a value range of 0 to 1 through the Sigmoid activation function.

[0042] Further, in step S3, the arbitration fusion module includes:

[0043] The conflict mask generation unit is used to identify pixel regions with inconsistent cloud classification labels between FY-4A and FY-4B based on semantic difference features, and generate conflict region masks.

[0044] The arbitration weight calculation unit is used to stitch together semantic difference features and cirrus confidence feature maps within the conflict region mask, and calculate the arbitration weight coefficients through a process that includes weighting, biasing and sigmoid activation functions.

[0045] The cloud height correction unit is used to calculate the final corrected cloud top height based on the arbitration weight coefficient and a relative correction strategy.

[0046] Furthermore, the formula for calculating the final corrected cloud top height by the cloud height correction unit is as follows:

[0047]

[0048]

[0049]

[0050] in, It is a semantic difference feature conflict mask. It is a cirrus cloud feature map. This is the arbitration weighting coefficient corrected by YunGao. and It is a Fengyun Level 2 Cloud High-end product. This is the adaptively corrected cloud height label.

[0051] In a second aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of the method described in any of the preceding claims.

[0052] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. The self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals proposed in this invention utilizes the unique 15th channel of Fengyun-4B satellite (FY-4B). Leveraging the spectral advantages of the absorption band and the fourth channel (1.38 μm cirrus channel), and through the construction of an asymmetric spectral residual network, spectral fusion can be achieved. The transmission characteristics of the channel accurately capture cirrus clouds, making up for the difficulty of matching in weak texture areas by the pure geometric parallax method from the perspective of radiation physics, and significantly improving the detection sensitivity of high-level thin clouds.

[0055] 2. The self-supervised cloud top height inversion method based on binary star asymmetric spectral residuals proposed in this invention proposes a self-supervised physical constraint correction mechanism that does not require the participation of external truth values. Through the joint constraint of semantic embedding and physical channel features, it automatically identifies and corrects the cloud height labeling error in the FY-4B cloud classification product that misclassifies "cirrus cloud" as "multi-layer cloud", thereby eliminating the systematic underestimation of cloud height caused by classification errors. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0057] Figure 1 The asymmetric difference and... provided for embodiments of the present invention Diagram of gated spectral residual fusion network.

[0058] Figure 2 The spectral-semantic dual-stream adaptive network graph provided in this embodiment of the invention. Detailed Implementation

[0059] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0060] The self-supervised cloud top altitude inversion method based on binary star asymmetric spectral residuals proposed in this invention has the following system architecture: Figure 1 As shown, it includes the following steps:

[0061] S1. Data preprocessing and spatiotemporal matching: Obtain the first-level (L1) radiometer observation data and second-level cloud classification products (L2) of FY-4A and FY-4B satellites at the same time, and perform spatiotemporal matching to obtain pixel-aligned standardized channel data, cloud detection and cloud phase classification labels within the common viewing area.

[0062] Because the observation perspectives of Fengyun-4A and Fengyun-4B satellites are different, strict data alignment is required first.

[0063] Specifically, step S1 includes:

[0064] 1) Data Acquisition

[0065] Acquire primary multichannel radiometer observation data and secondary cloud classification products at the same time for FY-4A and FY-4B; the primary multichannel radiometer observation data includes visible light channels, near-infrared channels and infrared channels; the secondary cloud classification products include cloud detection and cloud phase classification labels (discrete values ​​from 0 to 9).

[0066] 2) Reprojection and cropping

[0067] A standard latitude and longitude grid is constructed. Using the satellite's built-in positioning lookup table and orbital parameters, the observation data of the two satellites, FY-4A and FY-4B, are reprojected onto the same standard latitude and longitude grid, and the common viewing area is cropped to ensure that the pixels correspond one-to-one.

[0068] 3) Normalization of physical quantities

[0069] The observation data from the first-level multi-channel radiometer were Z-score standardized to have a mean of 0 and a variance of 1, in order to accelerate the convergence of the subsequent neural network.

[0070] S2. Based on asymmetric difference and A gated spectral residual fusion network processes the standardized channel data to generate the initial cloud top height.

[0071] To address the issue that the first 14 channels of FY-4A and FY-4B in binary satellite data have the same physical meaning but different calibration, this invention does not adopt equal dual-stream input, but instead constructs an asymmetric architecture of "baseline + correction".

[0072] The asymmetric difference and Gated spectral residual fusion networks include:

[0073] The first N channels of FY-4A data are used as the input base stream to extract baseline features and output the baseline cloud top height;

[0074] The correction stream is input using the differential data of corresponding channels of FY-4B and FY-4A, and the Mth channel data of FY-4B, to extract differential features and The channel's high-level semantic features are analyzed, and the cloud height correction and dynamic gating coefficient are output.

[0075] The fusion module is used to calculate and output the initial cloud top height based on the base cloud top height, cloud height correction amount, and dynamic gating coefficient through weighted residual calculation.

[0076] The correction stream includes:

[0077] The differential encoder is used to process the differential data of the corresponding channels of FY-4B and FY-4A and extract differential features that characterize the systematic deviation between the two satellites.

[0078] The channel semantic encoder is used to process the Mth channel data of FY-4B and extract high-level semantic features.

[0079] The feature fusion layer is used to concatenate the differential features with high-level semantic features;

[0080] A correction predictor is used to predict cloud height correction based on the output of the feature fusion layer.

[0081] A gating controller is used to generate dynamic gating coefficients with values ​​ranging from 0 to 1 based on high-level semantic features.

[0082] The asymmetric input stream construction process is as follows: The baseline stream uses the first 14 channels of FY-4A data as the baseline input, extracts the basic atmospheric radiation features through the Baseline Encoder, and predicts a baseline cloud top height through a regression head. This represents the cloud height obtained solely from the inversion of FY-4A data and serves as the base anchor point for subsequent corrections.

[0083] Both FY-4A and FY-4B possess the first 14 channels, which constitute the basic feature space for cloud top altitude inversion. Since these 14 channels have the same physical meaning on both satellites, this invention uses Ch1-14 of FY-4A as the "reference stream" to extract the basic atmospheric radiation characteristics of clouds. Instead of simply merging the 28 channels from both satellites, which would undoubtedly increase network parameters and computational resources, this invention calculates the difference in the same channels between the two satellites (FY4B - FY4A) as the "spectral difference stream." By learning the differences in these 14 channels, the network can automatically eliminate systematic errors caused by differences in instrument calibration and observation angles between the two satellites.

[0084] The correction predictor mechanism is as follows: spectral difference flow is used to construct differential feature inputs for the same channel. This branch specifically learns the systematic biases between two satellites (such as calibration errors and differences in observation angles) through the Correction Encoder, rather than learning the radiance values ​​from scratch. Additionally, it utilizes FY-4B's unique 15th channel (…). The channel extracts high-level semantic features using the Ch15 Semantic Encoder. The "difference features" and "Ch15 features" are then concatenated and fused (Feature Fusion) and input into the correction predictor to predict a cloud height correction. . The fitted value represents the altitude deviation caused by differences in satellite calibration, observation angle, and FY-4B-specific channel information.

[0085] The gating controller mechanism is as follows: The FY-4B's unique 15th channel (13.5μm) is used... The absorption band is extracted as a separate high-level semantic feature. A specific gating mechanism network is used to process the Ch15 feature, outputting a dynamic gating coefficient. (Value range 0-1). The physical logic is that when the Ch15 signal is strong (extremely low brightness temperature, indicating the presence of high clouds), The increased size allows for significant correction of the differential flow characteristics of FY-4B; when the Ch15 signal is weak, the baseline results of FY-4A are maintained.

[0086] Regarding channel 4 (1.38μm, cirrus channel), this band is heavily influenced by water vapor absorption, making the ground and low clouds invisible; only the reflection from towering cirrus clouds is visible. The abundant water vapor in the lower atmosphere completely absorbs solar radiation in this band, causing the background (ground and low clouds) to appear black. Only the ice crystal cirrus clouds at higher levels (above the water vapor) reflect sunlight. Therefore, the presence of a signal in this channel almost certainly confirms the presence of high-altitude cirrus clouds. Regarding channel 15 (13.5μm, (channel), this band is located in the strong absorption band of carbon dioxide, due to the atmosphere The radiation in this band is evenly distributed and has difficulty penetrating the lower atmosphere. It acts as a natural "upper-level filter." If this channel shows a low brightness temperature (cold), it indicates the presence of high-altitude clouds (cirrus); if it shows a high brightness temperature, it indicates the presence of mid-level clouds. Therefore, these two channels are fused to correct the systematic error of FY-4B in classifying cirrus clouds as multi-layered clouds.

[0087] Based on the outputs of the three branches mentioned above, the final initial cloud height is calculated using the weighted residual formula. This formula achieves asymmetric "soft correction." Instead of simply piecing together data from two satellites, the network learns "under what conditions (by...)..." The amount of correction to be added (determined by) The decision was made to maximize the detection advantages of FY-4B for high-altitude clouds while retaining the baseline accuracy of FY-4A.

[0088] The formula for calculating the initial cloud top height by the fusion module is:

[0089]

[0090] in, Baseline predictions from FY-4A, Cloud height correction derived from channel differential features and Ch15. High cloud confidence weights from Ch15.

[0091] S3. Based on the spectral-semantic dual-stream adaptive network, process the standardized channel data, cloud classification labels, and the initial cloud top height to generate the final corrected cloud top height.

[0092] This invention proposes a self-supervised automatic cloud top height correction method that leverages the complementary advantages of observations from the Fengyun-4 dual satellites (FY-4A / 4B). This method addresses the systematic bias in FY-4B satellite cloud classification products that misclassify cirrus clouds as multi-layered clouds by constructing a dual-stream deep learning network comprising a label semantic stream (semantic stream) and a spectral physical stream (physical stream). The label semantic stream aims to enable the neural network to understand the physical meaning behind discrete digital cloud classification labels and quantify the cognitive differences between the two satellites in cloud classification. The core logic of the method is as follows: using observational data from FY-4A and FY-4B in regions with consistent cloud classification as "anchor samples," self-supervised learning is performed in these regions to train the network to master the spectral physical characteristics of cirrus clouds through self-supervised transfer learning. The learned physical discrimination ability is then transferred to regions with inconsistent cloud classification, i.e., label conflict regions. The network generates arbitration weights based on the learned physical characteristics, thereby adaptively correcting cloud height errors caused by FY-4B cloud classification misclassification.

[0093] The spectral-semantic dual-stream adaptive network includes:

[0094] Semantic streams are used to encode and embed high-dimensional semantics into the cloud classification labels of FY-4A and FY-4B, generating semantic difference features.

[0095] The physical stream is used to extract spectral features from at least two specified channels of FY-4B data and output a cirrus confidence feature map.

[0096] The arbitration fusion module is used to generate a conflict region mask based on the semantic difference features, fuse the semantic difference features and the cirrus confidence feature map within the conflict region to generate arbitration weights, and use the arbitration weights to perform weighted compensation on the initial cloud top height difference between FY-4A and FY-4B, and output the final corrected cloud top height.

[0097] The semantic stream includes:

[0098] One-Hot encoding units are used to convert the cloud classification label maps of FY-4A and FY-4B into multi-channel sparse tensors, respectively;

[0099] The semantic embedding unit is used to map multi-channel sparse tensors to dense semantic feature vectors through learnable convolutional layers and to calculate the difference between two semantic feature vectors to generate semantic difference features.

[0100] The One-Hot encoding process is as follows:

[0101] First, obtain the FY-4A cloud classification label image after projection registration at the same time. And FY-4B cloud category tag image One-Hot encoding is performed. Since cloud classification labels are discrete integers ("6" represents cirrus clouds, "7" represents multi-layered clouds), there is no linear numerical relationship between them. To avoid the network misinterpreting category differences as numerical differences, the single-channel label image... Convert to multi-channel One-Hot sparse tensor In the new tensor, each pixel is set to 1 only on the channel corresponding to its class index, and 0 on the other channels, thus achieving class orthogonality.

[0102] The semantic embedding unit processing procedure is as follows:

[0103] Since the One-Hot vector is too sparse and inconvenient for feature computation, the system uses a learnable convolutional layer to reduce the dimensionality of the sparse tensor and reconstruct its features, mapping it into a dense 32-channel semantic feature vector. and This step enables the network to capture the nonlinear relationships between different cloud types in a high-dimensional feature space. The system calculates the difference between the semantic feature vectors of two stars. This generates semantic difference features for the entire image. These features can accurately locate "conflict regions" where two stars have inconsistent judgments about cloud phases, providing location guidance for subsequent corrections.

[0104] The physical flow is a lightweight convolutional neural network. Its input is the 4th and 15th channel data of FY-4B. It extracts spectral physical features through several convolutional layers and outputs a cirrus confidence feature map with values ​​ranging from 0 to 1 through the Sigmoid activation function.

[0105] Specifically, the spectral physical flow construction process is as follows:

[0106] Based on the binary star consistency assumption, regions labeled as cirrus clouds in both FY-4A and FY-4B are selected as positive sample anchors, while regions labeled as multi-layered clouds or low clouds in both FY-4A and FY-4B are selected as negative sample anchors, as these regions are indeed not cirrus clouds. A lightweight convolutional neural network (Spectral CNN) is designed, taking the 4th and 15th channels of FY-4B data as input. The network contains 2 to 3 convolutional layers to extract local texture and spectral intensity features. The network ends with a sigmoid activation function, outputting a single-channel cirrus cloud feature map. The value ranges from 0 to 1. The physical meaning of this feature map is "the probability that the current pixel belongs to a high-altitude thin cirrus cloud." When the input 15th channel shows extremely low brightness temperature (indicating high altitude) and the 4th channel shows high reflectivity (indicating the presence of thin clouds), the network will output a value close to 1; otherwise, it will output a value close to 0. This allows the network to output a cirrus cloud confidence level close to 1 in positive sample anchor regions and a value close to 0 in negative sample anchor regions, providing an objective physical basis for resolving classification conflicts.

[0107] The arbitration fusion module includes:

[0108] The conflict mask generation unit is used to identify pixel regions with inconsistent cloud classification labels between FY-4A and FY-4B based on semantic difference features, and generate conflict region masks.

[0109] The arbitration weight calculation unit is used to stitch together semantic difference features and cirrus confidence feature maps within the conflict region mask, and calculate the arbitration weight coefficients through a process that includes weighting, biasing and sigmoid activation functions.

[0110] The cloud height correction unit is used to calculate the final corrected cloud top height based on the arbitration weight coefficient and a relative correction strategy.

[0111] After training, the network has mastered the spectral fingerprint of "what cirrus clouds are". Now, when applied to "label conflict regions" (i.e., regions where one star identifies cirrus clouds and another identifies multi-layered clouds), the network no longer relies on labels but instead provides independent and objective judgments based on spectral features. Finally, using the objective judgment provided by the physical flow, the cloud height in the label conflict regions is adaptively corrected. The arbitration fusion module's correction and cloud height output process is as follows:

[0112] Generate conflict masks using semantic difference features The correction process is initiated only in regions where the two star labels are inconsistent to save computational resources and maintain the original accuracy of non-conflict regions. The "cirrus confidence" output from the physical stream is concatenated with the "semantic difference" output from the semantic stream. The final arbitration weight coefficients are then calculated using a fusion layer (containing weighting and bias) and a sigmoid activation function. , A value close to 1 indicates that the physical characteristics support the region being cirrus clouds. A "relative correction strategy" is used to calculate the final cloud height. The system calculates the difference between the original cloud heights of the two satellites and utilizes arbitration weights. Weighted compensation shall be applied.

[0113] The formula used by the cloud height correction unit to calculate the final corrected cloud top height is:

[0114]

[0115]

[0116]

[0117] in, It is a semantic difference feature conflict mask. It is a cirrus cloud feature map. This is the arbitration weighting coefficient corrected by YunGao. and It is a Fengyun Level 2 Cloud High-end product. This is the adaptively corrected cloud height label.

[0118] The formula achieves smooth soft correction: it updates to more accurate physical inversion values ​​in the cirrus cloud misjudgment area where physical evidence is conclusive, and performs weighted fusion in the ambiguous area, thus effectively solving the problem of FY-4B systematically underestimating cloud height.

[0119] In summary, this invention provides a cloud top altitude inversion method based on binary star asymmetric spectral fusion and self-supervised physical semantic correction, which has the following significant advantages compared with existing technologies:

[0120] 1) An asymmetric difference and A gated spectral residual fusion network, through an asymmetric architecture, learns the spectral residuals between two stars, effectively overcoming interference from instrument calibration differences and observation angles. This significantly improves the detection sensitivity to upper-level thin cirrus clouds and weakly textured regions.

[0121] 2) An unsupervised cloud height correction method based on a spectral-semantic dual-stream adaptive network innovatively utilizes self-supervised transfer learning in regions of consistent binary star observations. It understands the nonlinear relationship of cloud classification through a label semantic stream and directly transforms the objective physical information provided by the spectral physical stream (Ch4 / Ch15) into specific height correction values. This solves the physical problem of "same temperature, different height" under complex cloud phases, resulting in higher inversion accuracy.

[0122] The following table compares the present invention with conventional technology, the closest prior art, and the present invention:

[0123]

[0124] In a second aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of the method described in any of the preceding claims.

[0125] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-supervised cloud top altitude inversion method based on binary star asymmetric spectral residuals, characterized in that, Includes the following steps: S1. Acquire primary radiometer observation data and secondary cloud classification products from FY-4A and FY-4B satellites at the same time, and perform spatiotemporal matching to obtain standardized channel data, cloud detection, and cloud phase classification labels with pixel alignment within the shared field of view; S2. Based on asymmetric difference and A gated spectral residual fusion network processes the standardized channel data to generate the initial cloud top height; The asymmetric difference and Gated spectral residual fusion networks include: The first N channels of FY-4A data are used as the input base stream to extract baseline features and output the baseline cloud top height; The correction stream is input using the differential data of corresponding channels of FY-4B and FY-4A, and the Mth channel data of FY-4B, to extract differential features and The channel's high-level semantic features are analyzed, and the cloud height correction and dynamic gating coefficient are output. The fusion module is used to calculate and output the initial cloud top height based on the base cloud top height, cloud height correction amount and dynamic gating coefficient through weighted residual calculation; S3. Process the standardized channel data, cloud classification labels, and the initial cloud top height based on a spectral-semantic dual-stream adaptive network to generate the final corrected cloud top height; The spectral-semantic dual-stream adaptive network includes: Semantic streams are used to encode and embed high-dimensional semantics into the cloud classification labels of FY-4A and FY-4B, generating semantic difference features. The physical stream is used to extract spectral features from at least two specified channels of FY-4B data and output a cirrus confidence feature map. The arbitration fusion module is used to generate a conflict region mask based on the semantic difference features, fuse the semantic difference features and the cirrus confidence feature map within the conflict region to generate arbitration weights, and use the arbitration weights to perform weighted compensation on the initial cloud top height difference between FY-4A and FY-4B, and output the final corrected cloud top height.

2. The self-supervised cloud top altitude inversion method based on binary star asymmetric spectral residuals according to claim 1, characterized in that, Step S1 includes: Acquire primary multichannel radiometer observation data and secondary cloud classification products at the same time of FY-4A and FY-4B; the primary multichannel radiometer observation data includes visible light channels, near-infrared channels, and infrared channels; the secondary cloud classification products include cloud detection and cloud phase classification labels. The observation data from the two satellites are reprojected onto the same standard latitude and longitude grid, and the shared viewing area is cropped to ensure that the pixels correspond one-to-one. The observation data from the first-level multichannel radiometer were Z-score standardized to have a mean of 0 and a variance of 1.

3. The self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals according to claim 1, characterized in that, In step S2, the correction stream includes: The differential encoder is used to process the differential data of the corresponding channels of FY-4B and FY-4A and extract differential features that characterize the systematic deviation between the two satellites. The channel semantic encoder is used to process the Mth channel data of FY-4B and extract high-level semantic features. The feature fusion layer is used to concatenate the differential features with high-level semantic features; A correction predictor is used to predict cloud height correction based on the output of the feature fusion layer. A gating controller is used to generate dynamic gating coefficients with values ​​ranging from 0 to 1 based on high-level semantic features.

4. The self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals according to claim 3, characterized in that, In step S2, the formula used by the fusion module to calculate the initial cloud top height is: ， in, Baseline predictions from FY-4A, Cloud height correction derived from channel differential features and Ch15. High cloud confidence weights from Ch15.

5. The self-supervised cloud top altitude inversion method based on binary star asymmetric spectral residuals according to claim 1, characterized in that, In step S3, the semantic stream includes: One-Hot encoding units are used to convert the cloud classification label maps of FY-4A and FY-4B into multi-channel sparse tensors, respectively; The semantic embedding unit is used to map multi-channel sparse tensors to dense semantic feature vectors through learnable convolutional layers and to calculate the difference between two semantic feature vectors to generate semantic difference features.

6. The self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals according to claim 1, characterized in that, In step S3, the physical flow is a lightweight convolutional neural network. Its input is the 4th and 15th channel data of FY-4B. It extracts spectral physical features through several convolutional layers and outputs a cirrus confidence feature map with a value range of 0 to 1 through the Sigmoid activation function.

7. The self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals according to claim 1, characterized in that, In step S3, the arbitration fusion module includes: The conflict mask generation unit is used to identify pixel regions with inconsistent cloud classification labels between FY-4A and FY-4B based on semantic difference features, and generate conflict region masks. The arbitration weight calculation unit is used to stitch together semantic difference features and cirrus confidence feature maps within the conflict region mask, and calculate the arbitration weight coefficients through a process that includes weighting, biasing and sigmoid activation functions. The cloud height correction unit is used to calculate the final corrected cloud top height based on the arbitration weight coefficient and a relative correction strategy.

8. The self-supervised cloud top altitude inversion method based on binary asymmetric spectral residuals according to claim 1, characterized in that, The formula used by the cloud height correction unit to calculate the final corrected cloud top height is: ， ， ， in, It is a semantic difference feature conflict mask. It is a cirrus cloud feature map. This is the arbitration weighting coefficient corrected by YunGao. and It is a Fengyun Level 2 Cloud High-end product. This is the adaptively corrected cloud height label.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images