Sea fog recognition method based on cloud height guidance and adaptive receptive field
By using cloud top height guidance and adaptive receptive field methods to dynamically adjust the receptive field range, and combining global context and local detail feature extraction, the problem of difficult differentiation of spectral features in sea fog identification is solved, achieving high-precision and high-reliability sea fog monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 天津海洋中心气象台(天津港航气象服务中心)
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-14
AI Technical Summary
Existing sea fog identification methods rely on spectral features, which makes it difficult to effectively distinguish sea fog from low clouds, resulting in a high false positive rate, especially with performance degradation during the transition from dawn to dusk and under complex cloud systems.
By employing cloud top height guidance and adaptive receptive field methods, a roughness map and an adaptive weight map are generated from the cloud top height map, and the receptive field range is dynamically adjusted. Combined with global context and local detail feature extraction, accurate differentiation between sea fog and low clouds can be achieved.
It improves the accuracy and reliability of sea fog recognition, especially in complex cloud systems and edge areas, and can take into account both large-scale connectivity and fine extraction of local details, thereby improving the completeness of recognition and boundary clarity.
Smart Images
Figure CN122391835A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of sea fog recognition technology, specifically to a sea fog recognition method based on cloud top height guidance and adaptive receptive field. Background Technology
[0002] Currently, sea fog is a condensation phenomenon occurring in the lower atmosphere of the ocean and coastal areas, resulting in a significant reduction in horizontal visibility. It not only seriously threatens maritime shipping traffic and operational safety but also has a significant impact on the throughput efficiency of coastal ports and military activities. Therefore, achieving all-weather, high-precision sea fog monitoring, especially accurately distinguishing sea fog from low clouds, is of significant strategic importance and practical value for marine meteorological early warning and disaster prevention and mitigation.
[0003] Traditional satellite remote sensing for sea fog monitoring primarily relies on physical threshold-based methods, the most representative of which is the Brightness Temperature Difference Method (BDM). This method utilizes the difference in emissivity of water cloud particles between the mid-infrared and thermal infrared channels to separate fog areas by setting static or dynamic thresholds. However, threshold methods heavily depend on expert experience and seasonal adjustments, and are highly susceptible to solar radiation contamination during the dawn-dusk transition. More importantly, sea fog and stratus clouds are highly similar in spectral characteristics, and relying solely on brightness temperature differences is often insufficient to effectively eliminate interference from lower-level cloud systems, resulting in a persistently high false positive rate.
[0004] With the advent of deep learning technology, semantic segmentation models based on Convolutional Neural Networks (CNNs) and Transformers (such as U-Net and DeepLab series) have gradually become mainstream. These methods significantly improve identification efficiency by learning a nonlinear mapping between multi-channel spectral data and fog region labels. However, existing data-driven methods still face the core bottleneck of confusion caused by spectral homogeneity when dealing with sea fog problems. Sea fog is essentially a low-lying stratocumulus cloud, and the two share almost identical spectral characteristics in the visible and infrared bands. Existing deep learning models mainly rely on spectral texture features, making them prone to misinterpreting clouds as fog. Summary of the Invention
[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.
[0006] This application provides a sea fog identification method based on cloud top height guidance and adaptive receptive field. Instead of simply relying on the stacking of spectral features or a general segmentation network, it uses the local statistical features of cloud top height as a physical "probe" to guide the receptive field to adapt and dynamically adjust the network's focus range. This enables the receptive field to distinguish between low clouds and ground-hugging sea fog at the pixel level, reducing or even avoiding the phenomenon of "mistaking clouds for fog".
[0007] In some embodiments, the sea fog identification method based on cloud top height guidance and adaptive receptive field includes: Project the cloud top height product into a cloud top height map that corresponds one-to-one with the coordinates of the original cloud fog image; Obtain the cloud top roughness map corresponding to the cloud top height map; the coordinates of the cloud top roughness map and the cloud top height map are in one-to-one correspondence. A spatial adaptive weight map is generated based on the cloud top roughness map; the coordinates of the spatial adaptive weight map and the cloud top roughness map are in one-to-one correspondence; the lower the roughness of any position in the cloud top roughness map, the greater the spatial adaptive weight of any position in the spatial adaptive weight map. The global context features and local detail features of the original disk image are extracted using the local detail path and the global context path, respectively; the receptive field used when extracting features using the global context path is larger than that used when extracting features using the local detail path. For any position, the spatial adaptive weight corresponding to any position in the spatial adaptive weight map is used as the first mixing coefficient of the global context feature corresponding to any position, and the complementary weight of the spatial adaptive weight corresponding to any position in the spatial adaptive weight map is used as the second mixing coefficient of the local detail feature corresponding to any position. The global context feature and local detail feature corresponding to any position are mixed with the first mixing coefficient and the second mixing coefficient to obtain the roughness-based adaptive feature corresponding to any position. After obtaining the adaptive features corresponding to all locations, the default sea fog recognition algorithm is used to process the adaptive features corresponding to all locations to obtain the sea fog recognition results.
[0008] Optionally, obtain the cloud top roughness map corresponding to the cloud top height map, including: Get a sliding window of a set size; Move the sliding window position by position on the cloud top height map; For any given location, the roughness is determined based on the difference between the cloud top height at that location and the average cloud top height within the sliding window; where the greater the difference, the higher the roughness. The roughness corresponding to all locations is determined as the cloud top roughness map.
[0009] Optionally, the roughness at any location is determined based on the difference between the cloud top height at any location and the average cloud top height within the sliding window, including: ; in, Let the coordinates be any position. coordinates The roughness corresponding to the cloud top roughness map Set the size for the sliding window. Represented by coordinates The central sliding window area, Coordinates in the sliding window area The height of the cloud top at that location This represents the average cloud top height within the sliding window.
[0010] Optionally, a spatially adaptive weighted map is generated based on the cloud top roughness map, including: ; in, Let the coordinates be any position. coordinates The corresponding spatial adaptive weights in the spatial adaptive weight graph for Sigmoid Activation function For learnable weight matrix, Indicates normalization, coordinates The roughness corresponding to the cloud top roughness map This is the linear adjustment coefficient.
[0011] Optionally, the global context features and local detail features corresponding to any position are mixed using a first mixing coefficient and a second mixing coefficient to obtain a roughness-based adaptive feature corresponding to any position, including: ; in, Let the coordinates be any position. For any given location, the roughness-based adaptive feature is... coordinates In the spatial adaptive weighting graph, the corresponding spatial adaptive weight represents the first mixing coefficient. coordinates Corresponding global context features The second mixing coefficient, coordinates Corresponding local detail features.
[0012] Optionally, during the training process of the system used to perform the sea fog recognition method, feature points that have been marked as "fog / cloud" in the sample images are identified as anchor points; Regions in the sample image whose similarity to the roughness at the anchor point is higher than or equal to the first similarity threshold and whose spectral similarity is higher than or equal to the second similarity threshold are considered as positive samples. Regions in the sample image whose similarity to the roughness at the anchor point is lower than the third similarity threshold and whose spectral similarity is higher than or equal to the fourth similarity threshold are considered negative samples. Among them, the third similarity threshold represents a similarity that is lower than or equal to the first similarity threshold, and the fourth similarity threshold represents a similarity that is higher than or equal to the second similarity threshold.
[0013] Optionally, during the training of the system used to perform the sea fog recognition method, the loss function includes a morphological contrast loss term: ; in, For positive and negative sample loss terms, Features of anchor points Features of positive samples Features of negative samples This is the temperature coefficient.
[0014] Optionally, the loss function may also include a segmentation loss term; The loss function is composed of a weighted sum of a morphological comparison loss term and a segmentation loss term.
[0015] Optionally, the original disc image is an L1 level product, and the cloud top height product is an L2 level product.
[0016] Optionally, the cloud top height product is projected as a cloud top height map with coordinates corresponding one-to-one with the sea fog disk image, including: Construct a standard latitude and longitude grid; The cloud top height product and sea fog disk image are projected onto a standard latitude and longitude grid using the satellite's built-in positioning lookup table and orbital parameters. The projection result is determined as a cloud top height map.
[0017] The sea fog recognition method based on cloud top height guidance and adaptive receptive field provided in this application embodiment can achieve the following technical effects: The coordinates in the original disk image, cloud top height map, cloud top roughness map, and adaptive weight map of the sea fog correspond one-to-one. This provides a coordinate basis for the adaptive weight pair in the adaptive weight map to guide the process of extracting a mixture of global context features and local detail features from the original disk image through the two paths of large and small receptive fields in the subsequent steps.
[0018] The receptive field used for extracting features in the global context path is larger than that used for extracting features in the local detail path. Then, based on the adaptive weights in the adaptive weight graph, the global context features and local detail features at each position are mixed. Since the second mixing coefficient is complementary to the first mixing coefficient, the larger the first mixing coefficient is, the smaller the second mixing coefficient is, and vice versa.
[0019] In this mixing process, the larger the adaptive weight in the adaptive weight map, the larger the first mixing coefficient and the smaller the second mixing coefficient. Consequently, the global context features contribute more to the adaptive features, while the local detail features contribute less. As a result, the adaptive features are more biased towards features extracted from a larger receptive field. That is, this recognition method is like "opening" a large receptive field, analyzing the original disk image from within a larger receptive field. Furthermore, because a larger first mixing coefficient results in a larger adaptive weight, the corresponding roughness in the cloud top roughness map is lower. That is, the cloud top height values in the cloud top height map are relatively smooth, and this area in the original disk image is more likely to be entirely covered by clouds or sea fog. Analyzing this situation with relatively smooth cloud top height values using a larger receptive field can aggregate global information to ensure the completeness of the fog area prediction results.
[0020] Correspondingly, the smaller the adaptive weight in the adaptive weight map, the smaller the first mixture weight and the larger the second mixture weight. Consequently, the global context features contribute less to the adaptive features, while local detail features contribute more to the adaptive weight. As a result, the adaptive features are more biased towards features extracted from a smaller receptive field. That is, this recognition method seems to "switch" to a smaller receptive field. Furthermore, because the smaller the first mixture coefficient, the smaller the adaptive weight, the higher the corresponding roughness in the cloud top roughness map. That is, the cloud top height values in the cloud top height map are relatively coarse. This area in the original disk image is more likely to be the boundary between clouds and sea fog. Analyzing this situation of relatively coarse cloud top height values with a smaller receptive field is equivalent to using high-frequency details to distinguish image differences, which is beneficial for accurately delineating the boundaries between clouds and fog.
[0021] In the aforementioned recognition process, the adaptive feature is a mixture of global contextual features from a large receptive field and local detail features from a small receptive field. Both the global contextual features and the local detail features are spectral features; that is, the adaptive feature is actually also a spectral feature. However, guided by the adaptive weight map, this adaptive feature exhibits the characteristic of adaptively changing the receptive field.
[0022] At the edge of clouds and under complex cloud systems, sea fog and low clouds with similar spectra but different physical surface morphologies (the former is smooth, the latter is rough) are easily found in the original disk image. In this case, the cloud top height map guides the adaptive change of the receptive field in the aforementioned manner, which can obtain adaptive features with a larger or smaller field of view. Then, based on the adaptive features, the identification of sea fog or cloud layers can be carried out, which can achieve a high recognition accuracy.
[0023] On the other hand, sea fog typically appears as a large, smoothly textured connected region. When the cloud top elevation map guides the adaptive change of the receptive field in the aforementioned manner, a larger receptive field will be obtained, which can maintain internal consistency. Fragmented clouds or cloud boundaries that accompany sea fog, on the other hand, appear as high-frequency, drastically changing textures. When the cloud top elevation map guides the adaptive change of the receptive field in the aforementioned manner, a smaller receptive field will be obtained, which can capture the details of fragmented clouds or cloud boundaries.
[0024] By using cloud top height to guide the adaptive changes of the receptive field, the morphological differences of different targets in sea fog monitoring can be taken into account. Accurate identification of sea fog areas requires capturing the overall connectivity of the large area and the global structure of the advection extension, while identifying cloud boundaries or fragmented cloud areas relies on the fine extraction of minute features such as local texture abrupt changes and edge details.
[0025] In summary, the technical solution of this application can adaptively adjust the focus of feature extraction according to the physical morphology of the cloud top, and still has good discrimination ability when facing complex scenarios with alternating distribution of large-scale advection fog areas and fragmented cloud systems.
[0026] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description
[0027] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrative descriptions and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are considered similar elements, and wherein: Figure 1 This is a flowchart illustrating a sea fog identification method based on cloud top height guidance and adaptive receptive field provided in an embodiment of this application; Figure 2 This is a flowchart illustrating another sea fog identification method based on cloud top height guidance and adaptive receptive field provided in an embodiment of this application; Figure 3 This is a schematic diagram of a sea fog recognition device based on cloud top height guidance and adaptive receptive field provided in an embodiment of this application. Detailed Implementation
[0028] To provide a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this application. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.
[0029] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.
[0030] Unless otherwise stated, the term "multiple" means two or more.
[0031] In this embodiment, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.
[0032] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.
[0033] Figure 1 This is a flowchart illustrating a sea fog identification method based on cloud top height guidance and adaptive receptive field provided in an embodiment of this application.
[0034] Combination Figure 1 As shown, the sea fog recognition method based on cloud top height guidance and adaptive receptive field includes: S101. Project the cloud top height product into a cloud top height map that corresponds one-to-one with the coordinates of the sea fog original disk image.
[0035] In some application scenarios, cloud top height products are L2 level products, while the original disk image is an L1 level product. Specifically, L1 level multi-channel data from the FY-4B geostationary meteorological satellite at the same time are used as spectral input, i.e., the original disk image. Simultaneously, the corresponding L2 level cloud top height products are acquired. A standard latitude and longitude grid is then constructed. Using the satellite's built-in positioning lookup table and orbital parameters, the FY-4B L1 level original disk image and L2 level cloud top height products are projected onto this standard grid, and the projection result is determined as the cloud top height map.
[0036] S102. Obtain the cloud top roughness map corresponding to the cloud top height map.
[0037] The higher the roughness in the cloud top roughness map, the more drastic the change in the cloud top height value at that location; the lower the roughness, the smoother the change in the cloud top height value at that location.
[0038] In some application scenarios, cloud top roughness maps are obtained in the following ways: To address the physical differences between the relatively smooth surface of sea fog and the relatively rough surface of convective / fragmented clouds, a local sliding window variance operator is designed.
[0039] Specifically, obtaining the cloud top roughness map corresponding to the cloud top height map includes: Get a sliding window of a set size; Move the sliding window position by position on the cloud top height map; For any given location, the roughness is determined based on the difference between the cloud top height at that location and the average cloud top height within the sliding window; where the greater the difference, the higher the roughness. The roughness corresponding to all locations is determined as the cloud top roughness map.
[0040] Furthermore, the roughness at any given location is determined based on the difference between the cloud top height at any given location and the average cloud top height within the sliding window, including: ; in, Let the coordinates be any position. coordinates The roughness corresponding to the cloud top roughness map Set the size for the sliding window; that is, the sliding window is... Window size Represented by coordinates The central sliding window area, Coordinates in the sliding window area The height of the cloud top at that location This represents the average cloud top height within the sliding window.
[0041] In this step, pixel-level physical values are transformed into contextual texture features that guide changes in the receptive field in subsequent steps, in order to resolve the contradiction between satellite cloud physical products, which are pixel-level (point-wise) data, and fog area recognition algorithms that require context-wise information.
[0042] And, when When the value approaches 0, it indicates that the cloud top height in the area is uniform, and it is highly likely to be sea fog or stratus clouds; when... When the value is large or approaches infinity, it indicates that the cloud top in that region is highly undulating, representing cumulus, altoclouds, or cloud edges. Therefore, calculating the cloud top roughness map in this way is equivalent to introducing prior knowledge, making the process of guiding changes in the receptive field highly interpretable.
[0043] Furthermore, the coordinates of the cloud top roughness map and the cloud top height map are in one-to-one correspondence. This maintains coordinate consistency so that in subsequent steps, the adaptive weights at any position in the adaptive weight map can specifically guide the mixing of global context features and local detail features extracted at that position.
[0044] S103. Generate a spatial adaptive weight map based on the cloud top roughness map.
[0045] The coordinates of the spatial adaptive weight map and the cloud top roughness map are in one-to-one correspondence. Similarly, this coordinate consistency is maintained so that in subsequent steps, the adaptive weights at any position in the adaptive weight map can specifically guide the mixing of the global context features and local detail features extracted at that position.
[0046] The lower the roughness of any location in the cloud top roughness map, the greater the spatial adaptive weight of that location in the spatial adaptive weight map; conversely, the higher the roughness of any location in the cloud top roughness map, the smaller the spatial adaptive weight of that location in the spatial adaptive weight map.
[0047] Furthermore, a spatial adaptive weight map is generated based on the cloud top roughness map, including: ; in, Let the coordinates be any position. coordinates The corresponding spatial adaptive weights in the spatial adaptive weight graph for Sigmoid Activation function For learnable weight matrix, Indicates normalization, coordinates The roughness corresponding to the cloud top roughness map This is the linear adjustment coefficient.
[0048] The above learnable weight matrix Linear adjustment coefficient After training, the above mapping relationship acquires the following physical laws: when At lower (smooth) levels, Larger; when At higher (roughness) levels, Smaller.
[0049] The process described above for generating a spatially adaptive weighted map from the cloud top roughness map is a roughness-adaptive gating process. This process takes the cloud top roughness map as input and maps it to a spatially adaptive weighted map through a lightweight multi-layer perceptron (MLP).
[0050] The term "lightweight" is a common industry term, generally referring to a multilayer perceptron consisting of two, three, four, or five layers of convolutional neural networks.
[0051] S104. Extract the global context features and local detail features of the original disk image using the local detail path and the global context path, respectively.
[0052] Specifically, the receptive field used when extracting features in the global context path is larger than that used when extracting features in the local detail path.
[0053] The dual-path extraction network backbone in this step contains two parallel convolutional paths, each responsible for extracting features of different frequencies: Local Path of Detail ): Employs standard convolutional layers or convolutional layers with low dilation ratios (Dilation Rated=1,2). This path has a smaller receptive field, focusing on capturing high-frequency information such as cloud edges, fragmented cloud textures, and fine boundaries of sea fog.
[0054] Global Context Path ): Employs dilated convolutions or self-attention modules with large dilation rates (e.g., d=6, 12, 18). This path has a large receptive field, focusing on capturing low-frequency contextual information, such as the connectivity, uniformity, and large-scale spatial dependencies of sea fog over a wide area of the ocean.
[0055] S105. For any position, the spatial adaptive weight corresponding to any position in the spatial adaptive weight map is used as the first mixing coefficient of the global context feature corresponding to any position, and the complementary weight of the spatial adaptive weight corresponding to any position in the spatial adaptive weight map is used as the second mixing coefficient of the local detail feature corresponding to any position. The global context feature and local detail feature corresponding to any position are mixed with the first mixing coefficient and the second mixing coefficient to obtain the roughness-based adaptive feature corresponding to any position.
[0056] Optionally, the global context features and local detail features corresponding to any position are mixed using a first mixing coefficient and a second mixing coefficient to obtain a roughness-based adaptive feature corresponding to any position, including: ; in, Let the coordinates be any position. For any given location, the roughness-based adaptive feature is... coordinates In the spatial adaptive weighting graph, the corresponding spatial adaptive weight represents the first mixing coefficient. coordinates Corresponding global context features The second mixing coefficient, coordinates Corresponding local detail features.
[0057] In the central area of the sea fog (smooth) →1. The network automatically "activates" a large receptive field, aggregating global information to ensure the completeness of fog area prediction; in areas where clouds and fog meet or in fragmented cloud regions (coarse). →0, the network automatically “switch” to a local receptive field, using high-frequency details to accurately delineate the boundary.
[0058] The aforementioned S104 and S105 employ an asymmetric architecture of "physically guided branching + dual-path feature extraction," enabling the receptive field to simultaneously adapt to the large-scale connectivity of sea fog and the fine structure of cloud boundaries. This adaptive change of the receptive field is guided by prior knowledge, which on the one hand improves the interpretability of the model, and on the other hand, guides the mixing of global context features and local detail features with prior knowledge corresponding to cloud top roughness. Based on prior knowledge, the receptive field can be increased or decreased in a timely manner according to the actual situation, allowing the receptive field to simultaneously adapt to the large-scale connectivity of sea fog and the fine structure of cloud boundaries.
[0059] Through the aforementioned hybrid approach, the spatially adaptive weights (i.e., cloud top height) in the spatially adaptive weight map serve as prior knowledge, providing direct physical indicators for the sea fog identification method to distinguish between "smooth sea fog" and "coarse low clouds." This physical information, acting as a guiding signal, deeply participates in the fusion process of the sea fog identification method, overcoming the fundamental limitation of traditional methods that rely solely on spectral features, making it difficult to distinguish between "homogeneous but heterogeneous" targets. This significantly improves the physical interpretability of the model and the reliability of the discrimination criteria under complex cloud systems.
[0060] The spatially adaptive weights in the spatially adaptive weight map dynamically adjust the weights of the mixed global context features and local detail features based on the adaptive changes in cloud top roughness in different regions. For smooth fog areas, features with a large receptive field are automatically enhanced to ensure the connectivity and consistency of the recognition results; for rough cloud boundaries, local details are automatically focused to improve the precision of the segmentation boundaries. This "physical morphology-driven" receptive field adaptive mechanism enables this sea fog recognition method to intelligently match the feature extraction needs of different targets, improving its universality and adaptability to multi-scale cloud and fog morphologies.
[0061] Finally, the physical guidance branch and roughness adaptive gating process are designed to be lightweight, introducing only a small number of parameters and computational overhead, yet delivering significant performance improvements. The entire network requires only a single forward propagation during the inference phase, eliminating the need for multi-scale testing or complex post-processing, thus balancing recognition accuracy and computational efficiency. This makes it suitable for deployment in business early warning systems with real-time requirements.
[0062] S106. After obtaining the adaptive features corresponding to all locations, the default sea fog recognition algorithm is used to recognize the adaptive features corresponding to all locations in order to obtain the sea fog recognition results.
[0063] The default sea fog recognition algorithm is a relatively mature image recognition algorithm, such as U-Net and DeepLab series.
[0064] In the sea fog recognition method based on cloud top height guidance and adaptive receptive field provided in this application embodiment, the coordinates in the original disk image, cloud top height map, cloud top roughness map, and adaptive weight map of the sea fog correspond one-to-one. This provides a coordinate basis for the guiding role of the adaptive weight in the adaptive weight map in the subsequent steps in the process of extracting a mixture of global context features and local detail features from the original disk image by the two paths of large and small receptive fields.
[0065] The receptive field used for extracting features in the global context path is larger than that used for extracting features in the local detail path. Then, based on the adaptive weights in the adaptive weight graph, the global context features and local detail features at each position are mixed. Since the second mixing coefficient is complementary to the first mixing coefficient, the larger the first mixing coefficient is, the smaller the second mixing coefficient is, and vice versa.
[0066] In this mixing process, the larger the adaptive weight in the adaptive weight map, the larger the first mixing coefficient and the smaller the second mixing coefficient. Consequently, the global context features contribute more to the adaptive features, while the local detail features contribute less. As a result, the adaptive features are more biased towards features extracted from a larger receptive field. That is, this recognition method is like "opening" a large receptive field, analyzing the original disk image from within a larger receptive field. Furthermore, because a larger first mixing coefficient results in a larger adaptive weight, the corresponding roughness in the cloud top roughness map is lower. That is, the cloud top height values in the cloud top height map are relatively smooth, and this area in the original disk image is more likely to be entirely covered by clouds or sea fog. Analyzing this situation with relatively smooth cloud top height values using a larger receptive field can aggregate global information to ensure the completeness of the fog area prediction results.
[0067] Correspondingly, the smaller the adaptive weight in the adaptive weight map, the smaller the first mixture weight and the larger the second mixture weight. Consequently, the global context features contribute less to the adaptive features, while local detail features contribute more to the adaptive weight. As a result, the adaptive features are more biased towards features extracted from a smaller receptive field. That is, this recognition method seems to "switch" to a smaller receptive field. Furthermore, because the smaller the first mixture coefficient, the smaller the adaptive weight, the higher the corresponding roughness in the cloud top roughness map. That is, the cloud top height values in the cloud top height map are relatively coarse. This area in the original disk image is more likely to be the boundary between clouds and sea fog. Analyzing this situation of relatively coarse cloud top height values with a smaller receptive field is equivalent to using high-frequency details to distinguish image differences, which is beneficial for accurately delineating the boundaries between clouds and fog.
[0068] In the aforementioned recognition process, the adaptive feature is a mixture of global contextual features from a large receptive field and local detail features from a small receptive field. Both the global contextual features and the local detail features are spectral features; that is, the adaptive feature is actually also a spectral feature. However, guided by the adaptive weight map, this adaptive feature exhibits the characteristic of adaptively changing the receptive field.
[0069] At the edge of clouds and under complex cloud systems, sea fog and low clouds with similar spectra but different physical surface morphologies (the former is smooth, the latter is rough) are easily found in the original disk image. In this case, the cloud top height map guides the adaptive change of the receptive field in the aforementioned manner, which can obtain adaptive features with a larger or smaller field of view. Then, based on the adaptive features, the identification of sea fog or cloud layers can be carried out, which can achieve a high recognition accuracy.
[0070] On the other hand, sea fog typically appears as a large, smoothly textured, connected region. When the cloud top height map guides the adaptive change of the receptive field in the aforementioned manner, a larger receptive field is obtained, maintaining internal consistency. Conversely, fragmented clouds or cloud boundaries associated with sea fog exhibit high-frequency, drastically changing textures. When the cloud top height map guides the adaptive change of the receptive field in the aforementioned manner, a smaller receptive field is obtained, capable of capturing details of fragmented clouds or cloud boundaries. Therefore, using cloud top height to guide the adaptive change of the receptive field helps improve the accuracy of the recognition results.
[0071] Especially in scenarios where the performance of traditional methods degrades, such as at the edge of sea fog, during the transition between dawn and dusk, and when there is interference from fragmented cloud systems, the sea fog recognition results output by this method have better spatial continuity, internal consistency, and boundary clarity, providing an effective technical solution for high-precision and high-reliability operational sea fog monitoring.
[0072] In summary, the technical solution of this application can take into account the morphological differences of different targets in sea fog monitoring. Accurately identifying sea fog areas requires capturing their large-scale connectivity and global structure of advection extension, while identifying cloud system boundaries or fragmented cloud areas relies on the fine extraction of minute features such as local texture abrupt changes and edge details.
[0073] The technical solution of this application can adaptively adjust the focus of feature extraction according to the physical morphology of the cloud top, and still has good discrimination ability when facing complex scenes with alternating distribution of large-scale advection fog areas and broken cloud systems.
[0074] Figure 2 This is a flowchart illustrating another sea fog identification method based on cloud top height guidance and adaptive receptive field provided in an embodiment of this application.
[0075] Combination Figure 2 As shown, the sea fog recognition method based on cloud top height guidance and adaptive receptive field includes: S201. Obtain satellite spectral data of sea fog.
[0076] S202. Preprocess the satellite spectral data to obtain the original disk data.
[0077] S203. Extract local detailed features from the original disk data.
[0078] S204. Extract global context features from the original disk data.
[0079] S205, Obtain Cloud Top Height Product.
[0080] S206. Preprocess the cloud top height product to obtain a cloud top height map with coordinates consistent with the original disk data.
[0081] S207. Obtain the cloud top roughness map based on the cloud top height map.
[0082] S208. Obtain the spatial adaptive weight map based on the cloud top roughness map.
[0083] S209. Obtain adaptive features by mixing local detail features and global context features based on the spatial adaptive weight map.
[0084] S210. Determine whether the model is in the training phase. If yes, execute S211; otherwise, execute S212.
[0085] S211, Conduct joint training.
[0086] S212. Optimize the model based on the loss function and continue with S210.
[0087] S213. Use the default sea fog recognition algorithm to identify adaptive features.
[0088] S214. Obtain the identification results of sea fog.
[0089] The foregoing embodiments have provided an exemplary description of the specific working process of the sea fog recognition method based on cloud top height guidance and adaptive receptive field. The training process of the model used in this sea fog recognition method will be described exemplarily below.
[0090] Optionally, during the training process of the system used to perform the fog recognition method, feature points that have been labeled as "fog / cloud" in the sample images are identified as anchor points.
[0091] Regions in the sample image whose roughness similarity to the anchor point is higher than or equal to a first similarity threshold and whose spectral similarity is higher than or equal to a second similarity threshold are considered positive samples; that is, positive samples have the same physical morphology as the anchor point at the cloud top roughness level and are spectrally similar. For example, two different smooth sea fog regions can be considered positive samples for each other.
[0092] Regions in the sample image whose roughness similarity to the anchor point is below the third similarity threshold and whose spectral similarity is above or equal to the fourth similarity threshold are considered negative samples. That is, negative samples are spectrally similar to the anchor point and are difficult samples, but the two have significant differences in physical morphology at the cloud top roughness level. For example, smooth sea fog and low-level fragmented clouds with similar spectra but rough surfaces are negative samples for each other.
[0093] Among them, the third similarity threshold represents a similarity that is lower than or equal to the first similarity threshold, indicating that the negative sample is extremely similar to the anchor point spectrum and is a difficult sample for spectral recognition; the fourth similarity threshold represents a similarity that is higher than or equal to the second similarity threshold, indicating that the negative sample can actually be distinguished by the cloud top height.
[0094] Furthermore, during the training of the system used to perform the sea fog recognition method, the loss function includes a morphology-aware contrastive loss term: ; in, For positive and negative sample loss terms, Features of anchor points Features of positive samples Features of negative samples This is the temperature coefficient.
[0095] Based on the aforementioned positive and negative sample selection method, this morphological comparison loss term can bring similar morphological features closer together and push different morphological features further apart.
[0096] Furthermore, this morphological contrast loss term forces the network to classify features not only based on the spectrum, but also in conjunction with physical morphology.
[0097] In addition, the loss function also includes a segmentation loss term; the loss function is composed of a weighted sum of the morphological comparison loss term and the segmentation loss term.
[0098] Specifically, the loss function during model training is: ; in, The loss function in the modeling process, The segmentation loss term is specifically the cross-entropy loss term. For positive and negative sample loss terms, For positive and negative sample loss terms The weight.
[0099] After joint training based on the aforementioned loss function, the model can memorize high-frequency edge textures dependent on the cross-entropy loss term. Since the selection process for positive and negative samples introduces the physical property of cloud top roughness, the positive and negative sample loss term based on these samples enables the model to understand this physical property. The model can clearly distinguish the boundaries between spectrally similar features belonging to "fog" and features belonging to "low clouds." During the inference phase, even without explicit labels, the model can automatically separate easily confused low cloud features from sea fog features in the feature space based on the input physical roughness map, significantly reducing the misclassification rate. Furthermore, because the model learns the physical property of cloud top roughness, it exhibits better generalization ability and robustness when facing complex samples not present in the training set.
[0100] In some embodiments, the cloud top height-guided and adaptive receptive field-based sea fog identification device includes a processor and a memory storing program instructions. The processor is configured to execute the cloud top height-guided and adaptive receptive field-based sea fog identification method provided in the foregoing embodiments when executing the program instructions.
[0101] Figure 3 This is a schematic diagram of a sea fog recognition device based on cloud top height guidance and adaptive receptive field provided in an embodiment of this application.
[0102] Combination Figure 3 As shown, the sea fog recognition device based on cloud top height guidance and adaptive receptive field includes: The processor 31 and memory 32 may also include a communication interface 33 and a bus 34. The processor 31, communication interface 33, and memory 32 can communicate with each other via the bus 34. The communication interface 33 can be used for information transmission. The processor 31 can call logical instructions in the memory 32 to execute the sea fog identification method based on cloud top height guidance and adaptive receptive field provided in the foregoing embodiments.
[0103] Furthermore, the logic instructions in the aforementioned memory 32 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.
[0104] The memory 32, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this application. The processor 31 executes functional applications and data processing by running the software programs, instructions, and modules stored in the memory 32, thereby implementing the methods in the above-described method embodiments.
[0105] The memory 32 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 32 may include high-speed random access memory and may also include non-volatile memory.
[0106] This application provides a computer-readable storage medium storing computer-executable instructions configured to execute the sea fog recognition method based on cloud top height guidance and adaptive receptive field provided in the foregoing embodiments.
[0107] This application provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer, cause the computer to perform the sea fog recognition method based on cloud top height guidance and adaptive receptive field provided in the aforementioned embodiment.
[0108] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.
[0109] The technical solutions of this application embodiment can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in this application embodiment. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code; it can also be a transient storage medium.
[0110] The foregoing description and accompanying drawings fully illustrate embodiments of this application to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operations may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Additionally, when used in this application, the terms “comprise” and its variations “comprises” and / or “comprising” refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Unless otherwise specified, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes that element. In this document, each embodiment may focus on describing the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, then the relevant parts can be referred to the description of the method section.
[0111] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0112] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed units may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, i.e., they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to implement this embodiment according to actual needs. Furthermore, the functional units in the embodiments of this application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0113] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
Claims
1. A sea fog identification method based on cloud top height guidance and adaptive receptive field, characterized in that, include: Project the cloud top height product into a cloud top height map that corresponds one-to-one with the coordinates of the original cloud fog image; Obtain the cloud top roughness map corresponding to the cloud top height map; the coordinates of the cloud top roughness map and the cloud top height map are in one-to-one correspondence. A spatial adaptive weight map is generated based on the cloud top roughness map; the coordinates of the spatial adaptive weight map and the cloud top roughness map are in one-to-one correspondence; the lower the roughness of any position in the cloud top roughness map, the greater the spatial adaptive weight of that position in the spatial adaptive weight map. The global context features and local detail features of the original disk image are extracted using local detail paths and global context paths, respectively; wherein, the receptive field used when extracting features using the global context path is larger than the receptive field used when extracting features using the local detail path; For any given position, the spatial adaptive weight corresponding to the given position in the spatial adaptive weight map is used as the first mixing coefficient of the global context feature corresponding to the given position. The complementary weight of the spatial adaptive weight corresponding to the given position in the spatial adaptive weight map is used as the second mixing coefficient of the local detail feature corresponding to the given position. The global context feature and the local detail feature corresponding to the given position are mixed using the first mixing coefficient and the second mixing coefficient to obtain the roughness-based adaptive feature corresponding to the given position. After obtaining the adaptive features corresponding to all locations, the default sea fog recognition algorithm is used to process the adaptive features corresponding to all locations to obtain the sea fog recognition results.
2. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to claim 1, characterized in that, Obtaining the cloud top roughness map corresponding to the cloud top height map includes: Get a sliding window of a set size; The sliding window is moved position by position in the cloud top height map; For any given location, the roughness of that location is determined based on the difference between the cloud top height at that location and the average cloud top height within the sliding window; wherein, the greater the difference, the higher the roughness. The roughness corresponding to all locations is determined as the cloud top roughness map.
3. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to claim 2, characterized in that, Determining the roughness at any given location based on the difference between the cloud top height at any given location and the average cloud top height within the sliding window includes: ； in, Let the coordinates of any of the positions be . coordinates The roughness corresponding to the roughness in the cloud top roughness map. The set size for the sliding window. Represented by coordinates The central sliding window area, Coordinates in the sliding window area The height of the cloud top at that location The average cloud top height within the sliding window.
4. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to claim 1, characterized in that, Generate a spatial adaptive weight map based on the cloud top roughness map, including: ； in, Let the coordinates of any of the positions be . coordinates The corresponding spatial adaptive weights in the spatial adaptive weight graph, for Sigmoid Activation function For learnable weight matrix, Indicates normalization, coordinates The roughness corresponding to the roughness in the cloud top roughness map. This is the linear adjustment coefficient.
5. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to claim 1, characterized in that, The global context features and local detail features corresponding to any given location are mixed using the first mixing coefficient and the second mixing coefficient to obtain the roughness-based adaptive features corresponding to any given location, including: ； in, Let the coordinates of any of the positions be . For any given position, the roughness-based adaptive feature is... coordinates The spatial adaptive weights corresponding to the spatial adaptive weights in the spatial adaptive weights graph represent the first mixing coefficients. coordinates Corresponding global context features This is the second mixing coefficient. coordinates Corresponding local detail features.
6. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to any one of claims 1 to 5, characterized in that, During the training process of the system used to perform the sea fog recognition method, feature points that have been marked as "fog / cloud" in the sample images are identified as anchor points; Regions in the sample image whose roughness similarity to the anchor point is higher than or equal to a first similarity threshold and whose spectral similarity is higher than or equal to a second similarity threshold are considered as positive samples. Regions in the sample image whose roughness similarity to the anchor point is lower than the third similarity threshold and whose spectral similarity is higher than or equal to the fourth similarity threshold are considered negative samples. Wherein, the similarity represented by the third similarity threshold is lower than or equal to the first similarity threshold, and the similarity represented by the fourth similarity threshold is higher than or equal to the second similarity threshold.
7. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to claim 6, characterized in that, During the training of the system used to perform the sea fog recognition method, the loss function includes a morphological contrast loss term: ； in, For positive and negative sample loss terms, The characteristics of the anchor point, The features of the positive samples, The features of the negative samples, This is the temperature coefficient.
8. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to claim 7, characterized in that, The loss function also includes a segmentation loss term; The loss function is composed of a weighted sum of the morphological comparison loss term and the segmentation loss term.
9. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to any one of claims 1 to 5, characterized in that, The original disk image is an L1 level product, and the cloud top height product is an L2 level product.
10. The sea fog identification method based on cloud top height guidance and adaptive receptive field according to any one of claims 1 to 5, characterized in that, Projecting the cloud top height product into a cloud top height map that corresponds one-to-one with the coordinates of the sea fog original disk image, including: Construct a standard latitude and longitude grid; The cloud top height product and the sea fog disk image are projected onto the standard latitude and longitude grid using the satellite's built-in positioning lookup table and orbital parameters. The projection result is determined as the cloud top height map.