All-weather sea fog detection method based on multispectrum

Abstract

The application relates to the field of weather prediction, in particular to a full-time sea fog detection method based on multispectrum, which solves the problems of invalid detection algorithm in the morning and evening period, insufficient precision in distinguishing sea fog and low clouds and lack of systematicness in the current sea fog detection field. Specifically, the method comprises the following steps: acquiring satellite data and performing pretreatment, calculating the solar zenith angle (SZA) of each pixel; according to the value of the SZA, each pixel is assigned to a corresponding processing mode; based on the assigned processing mode, adaptive spectral preliminary screening of the corresponding mode is performed to generate a preliminary screening binary mask; based on the processed mode, regional texture consistency analysis is performed on the preliminary screening binary mask to effectively alleviate the low cloud interference; the retained sea fog area is post-processed and optimized to output the final sea fog detection result. The application can realize high-precision and high-stability detection of full-time sea fog, still maintains good performance in the morning and evening transition period, and can effectively distinguish sea fog and low clouds.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting, specifically to a multispectral, all-day sea fog detection method. Background Technology

[0002] Sea fog is a hazardous weather phenomenon that severely reduces visibility at sea, posing a continuous and serious safety threat to shipping, fisheries, and nearshore activities. Achieving high-precision, all-weather sea fog monitoring has become an urgent need for maritime safety early warning and emergency response. Satellite remote sensing technology, especially geostationary meteorological satellites with high temporal resolution (such as the Himawari-8, GOES, and FY-4 series), has become the primary means of sea fog monitoring due to its continuous observation capabilities.

[0003] Current sea fog monitoring technologies mainly focus on the spectral characteristics under different lighting conditions, specifically divided into nighttime detection, daytime detection, and detection during the twilight transition. Nighttime detection technology is relatively mature, generally employing the "dual-channel difference (DCD)" method, which identifies sea fog / low clouds based on the difference in radiative characteristics between the mid-infrared band (approximately 3.9 μm) and the thermal infrared band (approximately 11 μm). Daytime detection primarily utilizes the high reflectivity of the visible and near-infrared bands. Common methods include using reflectivity thresholds for the visible light channel, variants of the normalized differential vegetation index (NDVI), or combining the reflectivity ratio of the near-infrared and short-wave infrared channels (e.g., 1.6 μm) to distinguish fog from a bright background. However, these methods are susceptible to interference from mid- to high-level clouds and struggle to differentiate between sea fog and low clouds with similar spectral characteristics. Detection during the twilight transition is a weak point in existing technologies. During this period, the solar altitude angle is extremely low and changes rapidly, resulting in weak and unstable visible / near-infrared signals, rendering traditional methods based on fixed reflectivity thresholds completely ineffective. Although some studies have attempted to utilize the mid-infrared channel, its radiation characteristics undergo nonlinear changes around sunrise and sunset, making the fixed threshold method unreliable during this period, with high rates of missed detections and false detections.

[0004] Furthermore, a core challenge in all-day detection is the spectral similarity between sea fog and low-level cumulus clouds (low clouds). Both exhibit remarkably similar reflectance characteristics during the day and brightness-temperature difference characteristics at night. Existing techniques primarily employ texture analysis as an auxiliary criterion, such as calculating the gray-level co-occurrence matrix (GLCM) features of the image (e.g., contrast, entropy, homogeneity) or utilizing local standard deviation to quantify the roughness of cloud tops. Low clouds typically possess coarse textures and fragmented boundaries due to convection development; sea fog, on the other hand, forms under stable stratification, exhibiting uniform textures and continuous regions. However, existing techniques, when applying texture features, often employ globally fixed thresholds or independent judgments within a sliding window, failing to organically integrate texture features with the macroscopic spatial structure of the target (e.g., continuity and scale). This results in a still relatively high misclassification rate for sporadic low cloud or sea fog edges, failing to fundamentally solve the problem of fine-grained cloud and fog differentiation.

[0005] Therefore, existing sea fog detection methods have the following main shortcomings:

[0006] 1. Algorithm failure during sunrise and sunset: Existing methods lack adaptability to drastically changing lighting conditions. Around sunrise and sunset, when the solar zenith angle (SZA) is between approximately 75° and 88°, the signal-to-noise ratio of visible light signals is low, and the radiation characteristics of the mid-infrared channel undergo a transitional shift. Algorithms relying on fixed spectral thresholds struggle to adapt during this period, resulting in detection gaps or unreliable results, creating a break in the all-day monitoring chain.

[0007] 2. Insufficient accuracy in distinguishing between sea fog and low clouds: Although texture features are widely recognized as an effective distinguishing criterion, the application of existing methods is rather crude: local texture calculation based on sliding windows is easily affected by mixed pixels and strong gradients at cloud boundaries, leading to a large number of misclassifications; at the same time, the method lacks consideration for the overall consistency of the region, making it difficult to achieve robust region-level discrimination.

[0008] 3. Fragmented algorithm flow and lack of systematicity: Detection algorithms for daytime, nighttime, and dawn / dusk are often developed independently, using different feature and threshold systems, making it difficult to integrate them into a unified, adaptive processing framework. This increases the complexity of operational implementation.

[0009] For example, the application published on May 28, 2019, with publication number CN109815968A, entitled "An Algorithm for Sea Fog Detection at Dawn and Dusk Based on Himawari-8 Satellite Data," dynamically adjusts the threshold by establishing the relationship between the identification index and SZA (Spectral Z-Aspect). However, it is still limited to the single scene of dawn and dusk and uses an empirical combination of thresholds, failing to solve the fundamental problem of spectral ambiguity in low clouds and sea fog. Another example is the application published on June 10, 2009, with publication number CN101452078A, entitled "Remote Sensing of Daytime and Dusk by Polar-orbiting Meteorological Satellites." The method for detecting sea fog at night relies on local statistics within a fixed window and preset thresholds to indirectly characterize consistency. However, it is susceptible to interference from cloud boundaries and mixed pixels, making it difficult to robustly distinguish between sea fog and low clouds. The application published on September 11, 2013, with publication number CN103293084A, entitled "All-day and All-weather Inversion Method for Sea Fog Based on Multispectral Meteorological Satellite Information," divides the time into dawn / dusk, daytime, and nighttime, ignoring the differences in illumination caused by seasonal changes, and fails to effectively distinguish between low clouds and sea fog with blurred spectral characteristics.

[0010] Therefore, there is an urgent need for a high-precision and high-stability all-day sea fog detection method that can effectively detect sea fog during dawn and dusk and accurately distinguish between sea fog and low clouds. Summary of the Invention

[0011] To address the problems of existing sea fog detection methods, such as algorithm failure during dawn and dusk, insufficient accuracy in distinguishing sea fog from low clouds, fragmented algorithm flow, and lack of systematic approach, this invention proposes a multispectral-based all-weather sea fog detection method, specifically including the following steps:

[0012] (1) Acquire satellite data and preprocess it to calculate the solar zenith angle SZA of each pixel;

[0013] (2) Assign each pixel to the corresponding processing mode according to the value of SZA;

[0014] (3) Based on the assigned processing mode, perform adaptive spectral screening of the corresponding mode to generate a preliminary screening binary mask;

[0015] (4) Based on the corresponding processing mode, perform regional texture consistency analysis on the initial screening binary mask to filter out low cloud interference;

[0016] (5) Perform post-processing optimization on the retained sea fog area and output the final sea fog detection results.

[0017] This invention divides the observation period into three modes: daytime, dawn / dusk, and nighttime, based on the solar zenith angle (SZA). It adaptively activates the corresponding spectral recognition strategy for initial screening of sea fog and further introduces a texture analysis module based on multi-scale Laplace response to effectively distinguish sea fog from low clouds, ultimately achieving all-weather, high-precision sea fog detection.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] This invention addresses the problem of low detection accuracy during sunrise and sunset by establishing a dynamic function model of the solar zenith angle and key spectral index thresholds. This enables the algorithm to intelligently adapt to the rapid nonlinear changes in illumination before and after sunrise and sunset, achieving stable and reliable detection of sea fog during this period.

[0020] This invention addresses the challenge of accurately distinguishing sea fog from low clouds due to their similar spectral characteristics. It combines connected component analysis with multi-scale regional texture statistics, elevating the discrimination scale from the noise-sensitive pixel level to the more physically meaningful object (region) level. This method effectively identifies and filters out coarse-textured, spatially isolated low cloud regions while maintaining the integrity of uniformly textured, continuous sea fog regions, significantly reducing the false alarm rate and solving a core challenge in sea fog detection.

[0021] This invention proposes a unified and robust all-day sea fog monitoring method: using the solar zenith angle as the core criterion, it organically integrates heterogeneous detection strategies for different time periods such as daytime, dawn / dusk, and nighttime into an adaptive framework; the method has clear logic, coordinated processes, and smooth transitions in output results at the boundaries of time periods, and has good potential for operational applications, making it suitable for continuous and real-time sea fog monitoring. Attached Figure Description

[0022] Figure 1 This is a general framework diagram of the all-weather sea fog detection system in this invention;

[0023] Figure 2 NDSI probability density distribution maps for sea fog / low clouds, clear-sky sea surface, and mid-to-high clouds at different solar zenith angles;

[0024] Figure 3 SFSI probability density distribution maps for sea fog / low clouds, clear-sky sea surface, and mid-to-high clouds at different solar zenith angles;

[0025] Figure 4 The probability density distribution of brightness temperature difference in the 3.9μm and 11.2μm channels for sea fog / low clouds, clear-sky sea surface, and mid-to-high clouds at different solar zenith angles;

[0026] Figure 5 A graph showing the trend of the 95th percentile of SFSI under different SZA values;

[0027] Figure 6 A graph showing the trend of NDSI at different SZA values;

[0028] Figure 7 This is an image showing the ablation experiment results of the texture feature module in this invention;

[0029] Figure 8 This is a comparison chart of the results of the daytime algorithm and the comparison algorithm of this invention;

[0030] Figure 9 This is a comparison chart of the results of the dawn / dusk time algorithm of this invention and the comparison algorithm. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0032] This embodiment uses AHI sensor data from the Japan Meteorological Agency's Himawari-8 satellite as an example to detect sea fog in the coastal waters of eastern China (28.2°–41.0°N, 117.5°–130.3°E). The following will refer to... Figure 1 The diagram shown illustrates the overall framework of the all-weather sea fog detection system, and introduces the multispectral-based all-weather sea fog detection method.

[0033] A multispectral-based all-weather sea fog detection method includes the following steps:

[0034] S1: Data Input and Preprocessing

[0035] We acquired L1 level full-disk data products from the Himawari-8 AHI. Data from seven key channels were selected: Channel 1 (0.47 μm), Channel 3 (0.64 μm), Channel 4 (0.86 μm), Channel 5 (1.6 μm), Channel 7 (3.9 μm), Channel 13 (10.4 μm), and Channel 14 (11.2 μm), with a unified spatial resolution of 2 km. Simultaneously, we acquired solar zenith angle (SZA) information for each pixel within the study area, matching the spatiotemporal data. Land masking was applied to all acquired files, and reflectivity enhancement was performed on channels 1, 3, and 4 to construct the VCI index. The specific enhancement methods are as follows:

[0036] The output reflectance is calculated using cubic spline interpolation within the range [0, 1] through six pairs of predefined input and output reflectance control points (input: [0.0, 0.12, 0.24, 0.47, 0.75, 1.0]; output: [0.0, 0.34, 0.63, 0.82, 0.94, 1.0]). The output reflectance remains unchanged when the input reflectance is 0 or 1.

[0037] S2: Select the appropriate processing mode based on the value of SZA.

[0038] To construct a sea fog detection method suitable for full-sunlight conditions, this invention proposes a layered detection strategy based on solar zenith angle (SZA): the SZA value corresponding to each pixel obtained in step S1 is divided into corresponding processing modes according to a preset threshold.

[0039] Daytime mode: When pixel SZA satisfies 0°≤SZA<75°;

[0040] Dusk / Sunset mode: When the pixel SZA satisfies 75°≤SZA<88°;

[0041] Night mode: When the pixel SZA satisfies SZA≥88°.

[0042] The aforementioned thresholds are based on a systematic analysis of the evolution of sea fog spectral characteristics under varying illumination conditions. To this end, this invention constructed a "Sea Fog Dedicated Dataset" (SFDD), encompassing 8 typical sea fog events and a total of 54 Himawari-8 satellite images. Through visual interpretation, combined with the spatiotemporal information from the *Marine Weather Review*, WS-SFDD, and the unique spectral characteristics of different targets, different surface features were identified and segmented. The interpreted feature masks were precisely registered with the satellite images, and a land mask was applied to exclude land areas, retaining only pixels with SZA ≥ 75° to focus on the spectral response characteristics during the transition from dawn to dusk. Based on this, multi-channel spectral information was extracted for three typical targets: sea fog / low clouds (1,160,758 pixels), clear-sky sea surface (5,767,665 pixels), and mid-to-high clouds (993,219 pixels).

[0043] To quantify spectral variations with illumination, the solar zenith angle range (75°–90°) was divided into 15 one-degree intervals. Subsequently, statistical analysis was performed on key parameters such as reflectance, brightness temperature, and normalized difference spectral index for each feature type within each interval. In the statistical analysis, the highest and lowest 2% values ​​for each feature type within each interval were removed to minimize the potential impact of anomalous pixels in visual interpretation. This process systematically revealed the evolution of spectral characteristics of sea fog and other targets with increasing solar zenith angle and enabled the calculation of key statistics. Ultimately, a quantitative characterization of the spectral properties of sea fog during dawn and dusk was achieved.

[0044] like Figure 2 and Figure 3As shown, within the solar zenith angle range of 75°–88°, both the Normalized Difference Snow Index (NDSI) and the Sea Fog-Sea Surface Separation Index (SFSI) can effectively distinguish clear-sky sea surfaces from sea fog / low clouds, and also show good separability from mid-to-high clouds. However, the applicability of these two discriminant indices is limited within the high solar zenith angle range (88°–90°). On the other hand, the analysis of the brightness temperature difference between the 3.9μm and 11.2μm channels (such as...) Figure 4 As shown in the figure, traditional nighttime methods become effective for detecting sea fog / low clouds when the solar zenith angle exceeds 88°.

[0045] Based on the above comprehensive analysis, this study proposes a hierarchical detection strategy based on solar zenith angle, specifically: a daytime algorithm suitable for solar zenith angles <75°, a twilight algorithm suitable for solar zenith angles between 75° and 88°, and a nighttime method suitable for solar zenith angles ≥88°. This hierarchical design fully considers the fundamental differences in the spectral response mechanisms of surface and cloud targets under different lighting conditions, effectively overcoming the problem of a sharp performance drop in a single algorithm during the low light period of dawn and dusk, thus achieving all-weather, highly robust remote sensing identification of sea fog.

[0046] S3: Adaptive Spectral Screening

[0047] According to the mode assigned in step S2, the corresponding spectral discrimination process is executed in parallel to generate a preliminary binary mask (the initial sea fog / low cloud pixels are 1, and others are 0).

[0048] 1. Preliminary screening using daytime mode spectra:

[0049] The adaptive spectral screening in daytime mode uses two parallel paths to identify sea fog and low cloud areas. The results of the two paths are cross-checked, which can effectively filter out misjudged pixels caused by feature limitations of a single path, thereby improving detection accuracy.

[0050] The first path includes:

[0051] (1) Sea surface filtration:

[0052] In daytime mode, existing research shows that relying solely on the Visible Cloud Index (VCI) for cloud detection can easily misclassify clear-sky sea surfaces as clouds. To reduce such misclassifications, this invention employs a conservative VCI threshold strategy and introduces thermal infrared brightness temperature difference as an auxiliary criterion to construct a joint cloud monitoring mechanism based on VCI and BTD.

[0053] VCI is calculated using the following formula:

[0054] in, , , These represent the enhanced reflectivity of channels 1, 3, and 4, respectively.

[0055] Simultaneously, define the thermal infrared brightness temperature difference:

[0056] = ,

[0057] in, and These are the brightness temperatures for channels 7 and 13, respectively.

[0058] If a pixel satisfies VCI<28 or Pixels >9K are identified as cloud candidate pixels; otherwise, they are identified as clear sky / sea surface and discarded. By combining VCI and BTD criteria, clouds and clear sky / sea surface can be effectively distinguished, achieving high-precision initial cloud screening.

[0059] (2) Filtering out mid-to-high cloud cover:

[0060] The study area was divided into several sub-regions according to latitude. For each sub-region, the number of pixels identified as clear sky over sea was counted, and an appropriate mid-to-high cloud removal method was selected accordingly.

[0061] If the number of pixels over clear sky and sea in the sub-region is less than 1000, then a simple brightness temperature threshold method is used: for cloud candidate pixels, if the brightness temperature of its channel 14 (11.2μm) is less than 1000, then the threshold method is used: If the value is ≤270K, it is classified as a medium-high cloud and removed.

[0062] If the number of pixels over clear sky and sea within a sub-region is ≥1000, then the cloud height estimation method is used:

[0063] First, calculate the average brightness temperature of the clear sky sea surface pixels within the sub-region in the 11.2μm channel, denoted as . ;

[0064] For each cloud candidate pixel, with its 11.2μm brightness temperature

[0065] H= ,

[0066] in, The vertical temperature lapse rate is 0.6 K / 100m.

[0067] If the estimated cloud height H ≥ 2000 m, it is determined to be a medium-high cloud. For the pixel region determined to be a medium-high cloud, a 3×3 pixel morphological dilation operation is performed to ensure the spatial integrity of the medium-high cloud region, and the dilated region is removed from the initial screening mask.

[0068] The second path includes:

[0069] Calculate the Normalized Difference Snow Index (NDSI):

[0070] NDSI= ,

[0071] in, and These are the reflectivities of channels 3 and 5, respectively.

[0072] If its NDSI < 0.1, it is determined to be a preliminary screening pixel of sea fog / low cloud and marked as 1 in the mask.

[0073] This invention employs two independent and parallel spectral screening paths in daytime mode: the first path fuses VCI and BTD to eliminate clear-sky sea surface and combines a regional adaptive strategy to filter out mid-to-high clouds; the second path constructs the Normalized Differential Snow Index (NDSI) to screen out sea fog and low cloud pixels. The final screening results are determined by dual-path cross-validation. This mechanism can effectively suppress isolated false alarms and significantly improve the accuracy and robustness of sea fog and low cloud detection.

[0074] 2. Preliminary spectral screening using the dawn-dusk mode:

[0075] The adaptive spectral screening of the dawn-dusk mode uses two parallel paths to identify sea fog and low cloud regions. The results of the two paths are cross-validated, which can effectively filter out pixels that are misidentified as sea fog by a single path, thereby improving detection accuracy. At the same time, the sea fog-sea surface separation index (SFSI) and normalized differential snow index (NDSI) used are both in ratio form, which have inherent robustness to drastic changes in solar irradiance.

[0076] The first path includes:

[0077] (1) Separation of sea surface and clouds:

[0078] The sea fog-sea surface separation index (SFSI) was used to effectively separate the sea surface from clouds during dawn and dusk. SFSI is defined as the ratio of the reflectance of AHI channel 1 (0.47 μm, blue light) to channel 5 (1.6 μm, shortwave infrared).

[0079] SFSI= ,

[0080] in, and These are the reflectivities of channels 1 and 5, respectively.

[0081] The effectiveness of this index stems from the significant spectral differences between sea surface and sea fog in the two wavelength bands: clear seawater has high reflectivity in the blue light band, but extremely low reflectivity in the shortwave infrared band due to strong absorption, resulting in a significantly higher SFSI value; while sea fog or low clouds, affected by Mie scattering from liquid water droplets, have similar reflectivity in both wavelength bands, with SFSI values ​​close to 1. Therefore, SFSI can effectively distinguish between sea fog / low clouds and clear sea surface.

[0082] To adapt to the drastic fluctuations in light caused by the rapid changes in the solar zenith angle (SZA) during the twilight period, the present invention introduces a dynamic threshold function DT1(SZA):

[0083] ,

[0084] This function is obtained by statistical fitting based on a large number of observations and can effectively track the changing trend of the 95% quantile of SFSI under different SZAs (as shown in Figure 5 ), thereby adaptively alleviating the impact of light attenuation on the stability of the criterion.

[0085] If a certain pixel satisfies SFSI < DT1(SZA), it is determined as a pixel待定 for sea fog and enters the next judgment.

[0086] (2) Preliminary filtering of middle and high clouds:

[0087] Although the visible light signal significantly weakens during the twilight period, the thermal infrared brightness temperature can still stably reflect the thermodynamic characteristics of the cloud top; based on the negative correlation that the higher the cloud top in the troposphere, the lower the brightness temperature, the present invention utilizes this characteristic to effectively pre-screen middle and high clouds. Specifically, for candidate pixels, if the brightness temperature of their channel 14 (11.2 μm) ≤ 270 K, it is preliminarily determined as middle and high clouds. For the region of the preliminarily determined middle and high cloud pixels, two consecutive 3×3 pixel morphological dilation operations are performed, and the final dilated region is completely removed from the pre-screening mask.

[0088] The second path includes:

[0089] Calculate the Normalized Difference Snow Index NDSI:

[0090] NDSI = ,

[0091] where and are the reflectances of channels 3 and 5 respectively.

[0092] To adapt to the impact of light attenuation caused by the change in the solar zenith angle (SZA) during the twilight period on the spectral index, the present invention introduces a dynamic threshold function DT2(SZA):

[0093] ,

[0094] This function is obtained by statistical fitting based on a large number of sea fog observation samples and can effectively track the changing trend of the 95% quantile of NDSI under different SZAs (as shown in Figure 6 ), thereby adaptively distinguishing sea fog / low clouds from background interference under low light conditions.

[0095] For pixels that meet the SFSI and brightness temperature conditions, if their NDSI < DT2(SZA), they are determined as initially screened sea fog / low cloud pixels and marked as 1 in the mask.

[0096] 3. Initial Screening of Nighttime Mode Spectra:

[0097] Identification Based on the Dual-Channel Difference (DCD) Method:

[0098] Calculate the dual-channel brightness temperature difference: = , where and are the brightness temperatures of channels 7 and 14 respectively.

[0099] If the value of a certain pixel is within the interval [-7.5K, -2.0K], it is determined as an initially screened sea fog / low cloud pixel and marked as 1 in the mask. Pixels not within this interval (such as clear sea surfaces, medium and high clouds, etc.) are all excluded.

[0100] S4: Low Cloud Removal - Regional Texture Consistency Analysis

[0101] After obtaining the candidate regions of sea fog and low clouds, in view of the significant differences in the spatial texture structures between sea fog and low clouds, the present invention filters out the rough and isolated low clouds from the initially screened mask through multi-scale texture consistency analysis and retains the uniform and continuous sea fog regions, specifically including the following steps:

[0102] (1) Selection of Source Images: According to the time period to which the pixels of the initially screened mask belong, select the corresponding single-channel source images for calculation. If it is a daytime period, select the reflectance image of channel 3 (0.64μm), because it has higher spatial resolution and rich texture information and can sensitively reflect physical properties such as cloud optical thickness; if it is a twilight or nighttime period, select the brightness temperature image of channel 13 (10.4μm) and use the brightness temperature texture differences caused by the undulations of cloud top heights for identification. This strategy effectively addresses the challenge of the drastic changes in spectral response during the twilight period and overcomes the limitation of the failure of the visible light channel at night, thereby ensuring the continuity and robustness of sea fog detection under all-day illumination conditions.

[0103] (2) Image Normalization and Texture Feature Extraction: Perform min-max normalization on the selected source images to linearly scale the pixel values to the interval [0, 1].

[0104] Extract texture features using a multi-scale Laplacian operator. Set four scale parameters : 0.3, 0.7, 1.2, 2.0.

[0105] For each scale, first perform Gaussian filtering on the normalized image (standard deviation is Then, the Laplacian response of the filtered image is calculated.

[0106] Calculate the multi-scale Laplacian response mean map (ML): For each pixel (x, y), take the absolute values ​​of its Laplacian response at all four scales, and then average them. The calculation formula is as follows:

[0107]

[0108] (3) Connectivity analysis: Perform connectivity component analysis (4-domain connectivity) on the initial binary mask obtained in step S3 to identify all spatially independent candidate regions, denoted as... .

[0109] (4) Calculation of average texture in regions: For each connected region Calculate the average value of all pixels within the region in the ML map, and use it as a measure of texture uniformity for that region. The calculation formula is as follows:

[0110] ,

[0111] in, For the region The area (number of pixels).

[0112] (5) Region filtering based on texture threshold: filter each region Compare with preset time period / channel related thresholds.

[0113] The texture threshold is established based on statistical analysis of the training samples. When using channel 3, if the average feature response of the region satisfies... If the value is less than 0.0145, the region is identified as sea fog. When using channel 13, higher feature response values ​​often appear at the boundary between sea fog and mid-to-high clouds. When mid-to-high clouds are highly fragmented and partially obscure the sea fog region, the average response level of the region is further increased. To address this issue, a second 3×3 morphological dilation operation is performed on the already dilated mid-to-high cloud region during dawn and dusk to extract the sea fog region. To ensure sufficient detection recall, a... A threshold of <0.0043 was used as the criterion for sea fog. Although channel 13 was used for texture analysis during both nighttime and twilight periods, the average regional feature response value was much lower at night. This is mainly attributed to the fact that the multi-scale Laplacian feature response value within the sea fog region is inherently lower at night than during twilight, meaning that the cloud tops of the sea fog region are more uniform at night. Therefore, the texture discrimination threshold for nighttime was set to 0.0022. The combination of multi-scale analysis and regional statistics enables this method to effectively filter out low clouds with coarse textures and spatial isolation, thereby improving the accuracy of sea fog detection.

[0114] In summary, the region filtering rules based on texture thresholds are as follows:

[0115] For areas using channel 3 (daytime): If If the value is less than 0.0145, it is identified as a sea fog area and retained; otherwise, it is identified as a low cloud area and filtered out from the final result.

[0116] For areas using channel 13 (dawn / dusk): If If the value is less than 0.0043, it is identified as a sea fog area and retained; otherwise, it is filtered out.

[0117] For areas using channel 13 (nighttime): If If the value is less than 0.0022, it is identified as a sea fog area and retained; otherwise, it is filtered out.

[0118] The aforementioned region filtering mechanism based on multi-scale texture consistency effectively utilizes the physical differences between sea fog cloud tops, which are smooth and uniform, and low clouds, which have fragmented and coarse structures. Through time-adaptive source image selection and dynamic texture thresholding strategies, it can robustly distinguish between sea fog and non-target low clouds during the day, dawn, dusk, and night, significantly suppressing isolated false alarms, improving the spatial continuity and classification accuracy of detection results, and providing crucial post-processing support for all-day sea fog remote sensing.

[0119] S5: Result Optimization and Output

[0120] After completing the initial screening of sea fog and the filtering of low clouds, the resulting mask may still contain a small number of isolated noise pixels or tiny false targets. To improve the accuracy of the results, morphological post-processing is performed on the final mask: for the binary image of the sea fog region retained after filtering in step S4 (results for daytime and twilight periods), a morphological opening operation (erosion followed by dilation) is performed using a 3×3 square structuring element to remove small discrete noise points and smooth the region boundaries.

[0121] The optimized binary image is output as the final sea fog detection product. Pixels with a value of 1 represent detected sea fog, while pixels with a value of 0 represent non-sea fog (including clear sea surface, low clouds, mid-to-high clouds, etc.).

[0122] To comprehensively evaluate the performance of the proposed sea fog recognition algorithm, this invention uses six core evaluation indicators: accuracy, precision, recall, false alarm rate (FAR), F1 score, and intersection-over-union ratio (IoU) to comprehensively verify its superiority from multiple dimensions such as overall classification accuracy, reliability of positive sample recognition, and accuracy of negative sample discrimination.

[0123] (1) Ablation experiment of texture feature module under different lighting periods

[0124] Given the significant differences in spatial texture structure between sea fog and low clouds—low clouds are typically coarse and isolated, while sea fog regions exhibit uniform and continuous texture—this invention introduces a texture analysis module based on existing spectral criteria. To evaluate the contribution of texture features to sea fog detection, ablation experiments were conducted during daytime, dawn / dusk, and nighttime, testing the performance of using only spectral features versus fusing spectral and texture features. All experiments were quantitatively evaluated based on the WS-SFDD test set and its extended samples after applying a land mask. For ease of evaluation, entire labels in the WS-SFDD dataset with sea surface pixels having a solar zenith angle ≥75° were used for evaluation at dawn / dusk. The results are shown in Table 1.

[0125] Table 1 shows the ablation experimental results of spectral and texture features under different illumination periods.

[0126]

[0127] Table 1 summarizes the results of ablation experiments on texture feature modules using the proposed algorithm under daytime, nighttime, and twilight conditions. The results show that the algorithm maintains a high recall rate in all lighting scenarios (87.97% during the day, 94.85% at night, and 89.05% at twilight), demonstrating its ability to identify most real sea fog pixels with a low false negative probability. This characteristic is particularly valuable in applications where completeness of detection is paramount, such as maritime safety and weather warning.

[0128] Figure 7The ablation experiment results of the texture feature module are shown. All samples in the figure are divided into three categories according to the scene: daytime (top two groups), dawn and dusk (middle two groups), and nighttime (bottom two groups). Comparing the method using only spectral features in the second column with the method fusing spectral and texture features in the third column, it can be seen that introducing texture features effectively alleviates low cloud interference, further improving the overall sea fog detection performance while maintaining high recall. While maintaining high recall, the addition of texture features further improves the overall detection performance. The most significant improvement occurs under daytime conditions, where the fusion of texture features significantly increases precision by 11.86% (from 65.00% to 76.86%), while significantly reducing the false alarm rate by 2.30% (from 4.91% to 2.61%). Compared with nighttime and dawn / dusk periods, the improvement in precision and false alarm rate is more pronounced during the day. Furthermore, the F1 score and intersection-over-union ratio also significantly improve during the day by 7.77% and 9.47%, respectively, highlighting a comprehensive improvement in detection accuracy and segmentation precision under daylight conditions. For nighttime conditions, the fusion of texture features yielded meaningful performance improvements, reducing the false alarm rate by 2.21% (from 6.22% to 4.01%) and improving precision by 9.16% (from 68.72% to 77.88%). Simultaneously, the F1 score gained by 6.18%, the cross-union ratio improved by 8.24%, and the recall was the highest across all time periods. Similarly, twilight conditions also benefited from texture features, achieving the highest final precision (84.68%) and the lowest final false alarm rate (1.87%) across all time periods. These results were accompanied by a 7.05% improvement in the F1 score and the highest absolute cross-union ratio gain (9.55%). Even using only spectral features, the twilight period exhibited superior baseline performance, with the highest precision (73.69%) and the lowest false alarm rate (3.54%) across all conditions. These results clearly demonstrate the effectiveness of the proposed dynamic thresholding method based on solar zenith angle in handling the challenging illumination conditions of these transitional periods.

[0129] The experimental results above demonstrate that multi-scale texture information, as an effective discriminative feature, can effectively filter out coarse-textured low clouds that are not connected to sea fog. Introducing multi-scale texture features significantly improves the accuracy of the sea fog detection model, especially during the day. This is mainly attributed to the rich texture features provided by the visible light channel reflectance. In contrast, texture information relying on thermal infrared brightness temperature during nighttime and twilight periods offers relatively smaller performance improvements due to its limited spatial detail.

[0130] (2) Comparative experiment with other methods

[0131] Currently, research specifically targeting sea fog detection during dawn and dusk is relatively limited. In existing research, the algorithm developed by Ryu and Hong (2020) based on NDSI (hereinafter referred to as the comparison algorithm) has been proven effective during dusk, and the satellite data and study area used are highly consistent with this work. Given the direct comparability of this algorithm, this invention selects it as the primary comparison method for daytime and dawn / dusk periods.

[0132] The algorithm constructs two discriminant indices using channels 2 and 5 of the AHI: NDSI_obs and NDSI_obs-NDSI_cal. Considering the consistency of the study area and validation data, this invention directly adopts the parameters from the original paper. For NDSI_obs, the original paper does not provide a specific discrimination threshold, only mentioning a range of [-0.029, 0.29]. Preliminary experiments show that using -0.029 as the lower threshold leads to a large number of missed detections of sea fog pixels, resulting in a significantly low recall rate. Further testing shows that the NDSI_obs value of sea fog is usually lower than that of other ground features. Therefore, this invention does not set a lower discrimination threshold, but only adopts the upper limit of 0.29 mentioned in the original paper. The final implemented criterion is: NDSI_obs < 0.29 and NDSI_obs-NDSI_cal < 0.076.

[0133] After applying the above algorithm and masking land pixels, the detection results were validated using the WS-SFDD test set and its supplementary dataset. The results are shown in Table 2.

[0134] Table 2 shows the performance comparison between the present invention and existing algorithms.

[0135]

[0136] As shown in Table 2, under daytime conditions, the algorithm proposed in this invention outperforms the comparative methods in all key metrics. It achieves significantly higher scores in accuracy (96.79% vs. 88.28%), F1 score (80.66% vs. 57.93%), and intersection-over-union ratio (69.37% vs. 43.79%). While the recall rate is slightly lower (87.97% vs. 90.63%), the precision is significantly improved (76.86% vs. 46.73%), greatly enhancing the reliability of the detection results. Therefore, the false alarm rate is significantly reduced to 2.61%, an improvement of nearly 10 percentage points compared to the comparative method's 12.29%. Figure 8 As shown, the comparison algorithm struggles to effectively separate sea fog from low clouds, occasionally misclassifying mid-to-high clouds and sea surface as sea fog. In contrast, the proposed algorithm effectively identifies spatially isolated, coarse-textured low clouds and reduces misclassified pixels.

[0137] For the twilight period, the proposed algorithm also demonstrates a significant advantage over the comparison method in terms of overall performance indicators. It achieved higher accuracy (97.56% vs. 86.46%), F1 score (85.23% vs. 54.87%), and intersection-over-union ratio (76.24% vs. 41.25%). The comparison method achieved extremely high recall (97.83%), but its false alarm rate increased significantly (14.51%) and its precision was significantly lower (42.00%). Analysis of the results shows that as the solar zenith angle increases, the comparison method's ability to reliably distinguish between mid-to-high clouds and low clouds, as well as sea fog, decreases. At extreme solar zenith angles close to the terminator, its discrimination ability severely degrades or even fails. Figure 9 This is intuitively demonstrated. In contrast, the proposed algorithm maintains robust performance with higher accuracy (84.68%) and a significantly lower false alarm rate (1.87%), proving its effectiveness and stability under challenging twilight conditions. Furthermore, even the version of the algorithm using only spectral features shown in (1) outperforms the comparative methods in terms of accuracy, cross-union ratio, and F1 score, further validating the core reliability of the proposed twilight sea fog detection algorithm.

[0138] Overall, the proposed algorithm demonstrates significant advantages over the baseline in several key metrics. Its consistent performance under daytime and twilight conditions reflects strong adaptability and robustness, supporting its applicability to continuous 24 / 7 sea fog monitoring. These results indicate that the proposed method shows promise for high-precision, low-false-alarm-rate sea fog detection and provides a feasible technical path for achieving operational all-weather monitoring.

[0139] In summary, this invention divides the observation period into three modes—daytime, dawn / dusk, and nighttime—based on the solar zenith angle (SZA), adaptively activates the corresponding spectral recognition strategies for initial screening of sea fog, and further introduces a texture analysis module based on multi-scale Laplace response to effectively alleviate the problem of ambiguity between the spectral features of sea fog and low clouds, ultimately achieving all-weather, high-precision sea fog detection.

[0140] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A multispectral-based all-weather sea fog detection method, characterized in that, Includes the following steps: S1: Acquire satellite data and preprocess it to calculate the solar zenith angle SZA for each pixel; S2: Assign each pixel to the corresponding processing mode according to the value of SZA; S3: Based on the assigned processing mode, perform adaptive spectral screening for the corresponding mode and generate a preliminary screening binary mask; S4: Based on the corresponding processing mode, perform regional texture consistency analysis on the initial screening binary mask to filter out low cloud interference; The regional texture consistency analysis includes: S4.1: Select the corresponding single-channel source image based on the time period to which the initial binary mask pixels belong; S4.2: Normalize the single-channel source image and extract multi-scale texture features to generate a texture feature map aligned with the initial screening binary mask space; S4.3: Perform connected component analysis on the initial screening binary mask generated in S3 to obtain several candidate regions; S4.4: Calculate the average texture response value of each candidate region in the texture feature map; S4.5: Compare the average texture response value with the preset threshold for the time period to filter out low cloud areas with coarse texture; The selection of the corresponding single-channel source image according to the time period in S4.1 includes: selecting the reflectance image of channel 3 during the daytime; selecting the brightness temperature image of channel 13 during dawn, dusk or nighttime. S5: Perform post-processing optimization on the preserved sea fog areas and output the final sea fog detection results.

2. The method for all-weather sea fog detection based on multispectral imaging according to claim 1, characterized in that, In S2, the rules for dividing the processing modes are as follows: Daytime mode: When the pixel SZA satisfies greater than or equal to 0° and less than 75°; Dusk / Sunset mode: When pixel SZA satisfies greater than or equal to 75° and less than 88°; Night mode: When the pixel SZA is greater than or equal to 88°.

3. The all-weather sea fog detection method based on multispectral imaging according to claim 2, characterized in that, The adaptive spectral screening in the daytime mode uses two parallel paths to identify sea fog and low cloud regions: The first path includes: (1) removing sea surface pixels based on the joint criteria of visible cloud index (VCI) and brightness temperature difference (BTD); (2) identifying and removing mid-to-high cloud regions using a latitude partitioning strategy. The second approach involves calculating the Normalized Differential Snow Index (NDSI) and classifying pixels with index values ​​less than a set threshold as areas of sea fog and low clouds.

4. The all-weather sea fog detection method based on multispectral imaging according to claim 2, characterized in that, The adaptive spectral screening of the dawn-dusk mode uses two parallel paths to identify sea fog and low cloud regions: The first path includes: (1) Calculating the sea fog-sea surface separation index SFSI and comparing it with the value of the first dynamic threshold function DT1 under the current solar zenith angle; when SFSI is less than DT1, it is determined to be a sea surface pixel and is removed; (2) Using the brightness temperature of the thermal infrared channel to determine the mid-to-high cloud region and remove it. The second path includes: calculating the Normalized Differential Snow Index (NDSI) and comparing it with the value of the second dynamic threshold function at the current solar zenith angle; when the NDSI is less than DT2, it is determined to be a sea fog and low cloud region.

5. The all-weather sea fog detection method based on multispectral imaging according to claim 2, characterized in that, The adaptive spectral screening of the night mode uses a dual-channel brightness temperature difference DCD to identify sea fog and low cloud regions.

6. The all-weather sea fog detection method based on multispectral imaging according to claim 1, characterized in that, In step S4.5, the preset threshold is adaptively set according to the selected channel: For the area using channel 3 during the daytime, the threshold is 0.0145; For the region using channel 13 during the dawn and dusk hours, the threshold is 0.0043; For the area using channel 13 during the nighttime period, the threshold is 0.0022.

Images