Deep learning sea fog detection method based on geographically weighted regression

Abstract

The invention relates to the field of sea fog detection, and discloses a deep learning sea fog detection method based on geographically weighted regression. The invention aims to improve the accuracy of sea fog identification by combining geographically weighted regression and deep learning methods. According to the main scheme, the method comprises the steps of preprocessing, obtaining and preprocessing satellite data and sea fog related marine meteorological element data, and integrating to form a comprehensive data set. Next, fitting is carried out by utilizing a geographically weighted regression model, collinear variables are removed through VIF detection, visual local parameters are analyzed, and a combination with good goodness of fit and few independent variables is selected; and finally, designing a GeoSatFogNet network, inputting data such as regression coefficients and marine environment elements, obtaining meteorological features through an attention mechanism, fusing the features, and inputting the fused features into an image segmentation network to realize sea fog identification. Through training, the sea fog identification accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of sea fog detection, and provides a deep learning sea fog detection method based on geographically weighted regression. Background Art

[0002] Sea fog is a serious marine disaster and a common disastrous weather phenomenon in which water vapor near the sea surface condenses to form suspended water droplets. In foggy sea areas, the horizontal visibility is less than 1 km. Since sea fog greatly reduces visibility, it has a negative impact on maritime activities such as marine transportation, fishing and aquaculture, marine oil and gas exploration, and military activities. Therefore, the accurate identification of sea fog is a task with high requirements and great significance, and sea fog identification is of great importance for maintaining maritime production activities.

[0003] In the current research field, deep learning models, with their powerful feature extraction capabilities, deeply learn and accurately identify the sea fog band features in sea fog-related data. By training on a large number of remote sensing images containing sea fog information, deep learning models can capture the unique spectral features presented by sea fog in specific bands. While this commonly adopted technical path focuses on mining the sea fog band features, it largely ignores the marine environmental factors that affect the formation of sea fog. Marine environmental factors play a crucial role in the generation, development, and dissipation of sea fog. Marine environmental factors such as sea surface temperature, wind speed, wind direction, and humidity interact with and influence each other, jointly constituting the complex environmental background for the formation of sea fog. This limitation may lead to insufficient understanding of the complex physical process of sea fog formation during the sea fog identification process. As a result, when facing some special sea fog scenarios or situations affected by a combination of various marine environmental factors, the identification accuracy decreases, and it is impossible to accurately and comprehensively identify and monitor sea fog.

[0004] At present, the analysis methods for sea fog and marine environmental elements are mainly divided into three categories, namely synoptic analysis, numerical analysis, and objective analysis. Synoptic analysis only analyzes the changes in the synoptic situation and marine hydrological conditions that cause sea fog. This method only simply statistically analyzes the changes in driving factors during the occurrence of sea fog. Numerical analysis is proposed from the perspective of physical mechanisms, which requires a large amount of computing resources and time, and its calculation process cannot be combined with existing sea fog detection technologies. Objective analysis establishes a statistical model between sea fog and various driving factors based on a large number of historical statistical analyses. Previous studies have analyzed the general relationship between different marine sea fogs and their related marine meteorological environmental elements, without considering the different relationships between sea fogs in different regions and their marine meteorological environmental elements. Because of the differences in geographical location, topography, and ocean circulation in different sea areas, there are obvious spatial differences in the responses of sea fog to environmental elements such as sea surface temperature, air-sea temperature difference, and relative humidity. Traditional statistical analysis methods are difficult to fully capture the complex non-linear relationship between sea fog and environmental elements.

[0005] To improve the accuracy of sea fog identification, this study proposes a deep learning sea fog detection method based on geographically weighted regression. Geographically Weighted Regression (GWR) is a regression method based on modeling spatial-varying relationships. It generates a regression model describing local relationships at each location in the study area, thus being able to well explain the local spatial relationships and spatial heterogeneity of variables. On this basis, a network is constructed. The features after processing the meteorological elements are concatenated with the satellite channel data in the channel dimension to form the total features, integrating meteorological and satellite observation information. The total features are input into the semantic segmentation network to achieve precise monitoring and identification of sea fog. Summary of the Invention

[0006] The main purpose of this method is to conduct spatial heterogeneity analysis on the marine meteorological environmental elements affecting the occurrence of sea fog, and combine the fusion features of meteorology and satellite with deep learning to improve the accuracy of sea fog identification.

[0007] To achieve the above object, the present invention adopts the following technical means. A deep learning sea fog detection method based on geographically weighted regression, characterized by comprising the following steps:

[0008] I. Pretreatment

[0009] Step 1: Acquisition and pretreatment of satellite data. Obtain satellite data at 9:00 am every day for five years, use the mature sea fog detection method to identify sea fog, divide it into regular hexagons, and count the sea fog occurrence frequency within the grid as the dependent variable for model fitting;

[0010] Step 2: Acquisition and preprocessing of sea fog related marine meteorological elements. According to the existing sea fog prediction research, collect the marine meteorological element data with the same time and space as the sea fog data, preprocess the collected meteorological data to ensure the integrity of the data, and use the acquired marine meteorological elements as independent variables for model fitting;

[0011] Step 3: Integrate data. After dividing the regular hexagonal grid, match the calculated sea fog frequency in the grid with the sea fog-related marine meteorological elements to make them have the same resolution. Integrate the data of the same time and space and the same resolution to form a comprehensive data set, including sea fog frequency and related meteorological data;

[0012] 2. Geographically Weighted Regression Model Fitting and Parameter Optimization

[0013] Step 4: Model fitting. Considering the spatial heterogeneity and data characteristics of the data, a geographically weighted regression model was selected. Before model fitting, global regression was used, sample data of the comprehensive data set was selected for model fitting, VIF was used to detect multicollinearity of independent variables, and variables with collinearity were eliminated;

[0014] Step 5: Parameter optimization. Visualize the local parameters of the independent variables, analyze the independent variables that affect the occurrence of sea fog, check the model goodness of fit, and select independent variable combinations with better goodness of fit and relatively fewer independent variables to enhance model performance and stability;

[0015] 3. Establishing GeoSatFogNet Sea Fog Detection Network

[0016] Step 6: Input features. Input the regression coefficients of each marine environmental element obtained by geographical weighted regression in each grid, as well as the marine environmental element data corresponding to the detection. For each environmental element in each grid, combine its actual value with the local regression coefficient to obtain the weighted environmental element and then combine it with the satellite channel data;

[0017] Step 7: Acquisition of meteorological features. The channel attention mechanism and the spatial attention mechanism are added. The channel attention mechanism performs global average pooling and global maximum pooling on the input marine environmental element data, and then processes and fuses them through a multi-layer perceptron to obtain the attention weights of each channel. Based on the channel attention output, the spatial attention mechanism performs average pooling and maximum pooling along the channel dimension, concatenates the results, and compresses the channel dimension through a convolutional layer to generate spatial attention weights. Finally, the channel attention and spatial attention weights act on the input features in turn, focusing on key channels and spatial areas;

[0018] Step 8: Applying the fused features to the segmentation network. The meteorological element features processed by the meteorological element branch are concatenated with the satellite channel data in the channel dimension to form a total feature. This combination method fuses the weighted information of meteorological elements and the observation information of multiple satellite channels, providing a more comprehensive data basis for subsequent semantic segmentation. The fused total feature is input into the image segmentation network for sea fog recognition.

[0019] In the above technical solution, step 2 includes the following steps:

[0020] (1) Selection of sea fog independent variables. According to existing sea fog prediction studies, the selected sea fog driving factors are sea surface temperature (sst), 2m temperature (2t), 2m dew point temperature (2dt), 10m wind U (u), 10m wind V (v), total cloud cover (tcc), relative humidity (rh), geopotential height (gt), and vertical velocity (vv). The sea-air temperature difference (TS) is obtained using the 2m temperature and the mean sea surface temperature, the air temperature dew point difference (TD) is obtained using the 2m dew point temperature and the 2m temperature, and the wind speed (WSD) and wind direction (WDR) are obtained using the 10m wind U and the 10m wind V.

[0021] (2) Preprocessing the data. Spatial interpolation is used to preprocess the data to ensure its integrity.

[0022] In the above technical solution, step 6 includes the following steps:

[0023] (1) Collect the actual marine environmental element data corresponding to each grid during detection to ensure that the data is accurately matched with the grid and time point corresponding to the regression coefficient.

[0024] (2) Normalization operation. For each type of marine environmental element data within each grid, calculations are performed according to the selected normalization method.

[0025] (3) Combine the actual values of the marine environmental elements within each grid after normalization with the local regression coefficients obtained from the geographically weighted regression corresponding to the grid. The normalized marine environmental element data is \(x_{ij}\), where \(i\) represents the grid number and \(j\) represents the element number. Let the local regression coefficient obtained from the geographically weighted regression be \(w_i\). For the environmental element \(x_{ij}\) within each grid \(i\), the weighted environmental element \(y_{ij}\) is generated through the following formula: ij where \(i\) represents the grid number, \(j\) represents the element number. Let the local regression coefficient obtained from the geographically weighted regression be \(w_i\). For the environmental element \(x_{ij}\) within each grid \(i\), the weighted environmental element \(y_{ij}\) is generated through the following formula: ij For each grid \(i\) of environmental element \(x_{ij}\) ij The weighted environmental element \(y_{ij}\) is generated through the following formula: ij . The formula is:

[0026]

[0027] In this way, the weighted marine environmental element data of each type within each grid is obtained, highlighting the relative importance of different environmental elements in each local area for the sea fog frequency.

[0028] Beneficial effects

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] In the field of sea fog research, many deep learning algorithms have been used for sea fog detection. For example, although the TransUnet model performs well in sea fog recognition, like many traditional methods, it fails to fully consider the marine meteorological elements affecting sea fog. This method uniquely adopts geographically weighted regression analysis to conduct in-depth spatial heterogeneity analysis of the marine environmental elements affecting sea fog. Currently, when most sea fog recognition models are constructed, they often do not fully consider the key role played by the marine underlying surface. As an important basis for the formation of sea fog, the temperature, humidity and other factors of the marine underlying surface environment have a profound impact on the generation, development and dissipation of sea fog. Moreover, due to the significant differences in geographical environment, the sea fog driving factors in different regions show obvious spatial heterogeneity. Traditional methods fail to effectively capture this spatial variation feature, resulting in limitations in the understanding of the sea fog formation mechanism.

[0031] By using geographically weighted regression analysis, this method can analyze the differences in the impact of marine environmental elements on sea fog at different geographical locations, opening up a new perspective for sea fog research. This analysis method not only reveals the spatial variation laws of each element, but also provides more detailed and comprehensive information for understanding the sea fog formation mechanism. On this basis, this method further uses fusion features to detect sea fog, integrating information from multiple sources and giving full play to the advantages of different data. This innovative fusion strategy comprehensively considers the spectral characteristics of sea fog and the closely related marine environmental elements, providing stronger support for the accurate detection of sea fog and reaching a new depth in the understanding of the sea fog formation mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is the overall flow chart of this method

[0033] Figure 2 is the geographically weighted regression flow chart of this method

[0034] Figure 3 is the sea fog detection flow chart of this method DETAILED DESCRIPTION OF THE INVENTION

[0035] For the convenience of those of ordinary skill in the art to understand and implement the present invention, and to make the purpose, content and advantages of this method clearer, the following further describes this method in detail with reference to the drawings and embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain this method and are not used to limit this method.

[0036] The overall flowchart of a deep learning sea fog detection method based on geographically weighted regression given by this method is as follows Figure 1 as shown.

[0037] This method includes the following steps:

[0038] Step 1: First, obtain satellite data at 9 am every day for five years, use a mature sea fog detection method to detect sea fog, divide appropriate grids, and count the sea fog occurrence frequency within the grids as the dependent variable for model fitting. Obtain Himawari-8 data of sea fog at 9 am from 2020 to 2024. Process the Himawari-8 data to convert nc data to tif data and clip out the Yellow Sea and Bohai Sea regions. Use the SVM algorithm to identify sea fog for the daytime Himawari-8 data, divide grids for the Yellow Sea and Bohai Sea regions, calculate the sea fog occurrence frequency by counting the sea fog values and the total number of pixels within the grids over five years, and the statistical results are used as the dependent variable for model fitting.

[0039] Step 2: Acquisition and preprocessing of sea fog-related marine meteorological elements. According to existing sea fog prediction research, collect marine meteorological element data with the same spatio-temporal as the sea fog data, preprocess the collected meteorological data to ensure data integrity, and the obtained marine meteorological elements are used as independent variables for model fitting; use ERA5 reanalysis data, and select the sea fog driving factors as sea surface temperature (sst), 2m temperature (2t), 2m dew point temperature (2dt), 10m wind U (u), 10m wind V (v), total cloud cover (tcc). Obtain the sea-air temperature difference (TS) from the 2m temperature and the sea surface temperature, the air temperature dew point difference (TD) from the 2m dew point temperature and the 2m temperature, and the wind speed (WSD) and wind direction (WDR) from the 10m wind U and 10m wind V. The detailed description is shown in Table 1:

[0040] Table 1:

[0041] Variable Description Unit Sea surface temperature Sea water temperature near the surface K 2m dew point temperature Dew point temperature 2m above the Earth's surface K 2m temperature Air temperature at 2m above the surface of land, ocean or inland water K 10m u-component of wind u-component of wind at 10m above the Earth's surface m / s 10m v-component of wind v-component of wind at 10m above the Earth's surface m / s Total cloud cover Proportion of the grid box covered by clouds Dimensionless

[0042] The calculation formula for the sea-air temperature difference (TS) is as follows:

[0043] TS = T 2t - T sst

[0044] The calculation formula for the air temperature dew point difference (TD) is as follows:

[0045] TD = T 2dt - T sst

[0046] The calculation formula for the wind speed (WSD) is as follows:

[0047]

[0048] The wind speed (WDR) calculation formula is as follows:

[0049] WDR=mod(180+arctan2(u,v),360)

[0050] Using ERA5 reanalysis data, the driving factors relative humidity (rh), geopotential height (gt) and vertical velocity (vv) were selected at 600, 650, 700, 750, 800, 850, 900, 950 and 1000 hPa respectively. A detailed description is shown in Table 2:

[0051] Table 2:

[0052]

[0053] The missing values ​​of the collected meteorological data are preprocessed to ensure the integrity of the data.

[0054] Step 3: Data fusion. After dividing the regular hexagonal grid, the calculated sea fog frequency in the grid is matched with the sea fog-related marine meteorological elements to make them have the same resolution. The data of the same time and space and the same resolution are integrated to form a comprehensive data set, including sea fog frequency and related meteorological data.

[0055] This method provides a geographically weighted regression flow chart of a deep learning sea fog detection method based on geographically weighted regression. Figure 2 shown.

[0056] Step 4: Model fitting. Considering the spatial heterogeneity and data characteristics of the data, the geographically weighted regression model is selected. First, a global regression is performed on the independent variables and the dependent variables. The multicollinearity problem of the independent variables is detected based on the VIF value calculated by regression, and the variables with collinearity are eliminated. The geographically weighted regression model is fitted to the comprehensive data set.

[0057] Step 5: Result analysis. Visualize the parameters of each independent variable of the geographically weighted regression, analyze the impact of each independent variable on the occurrence and non-occurrence of sea fog by comparing the current value of the independent variable, and quantitatively analyze the changes in the driving factors of sea fog. Compare the R values ​​of global regression and geographically weighted regression. 2 The values ​​of χ2 and AICc are used to select a combination of independent variables with better goodness of fit and relatively fewer independent variables, which enhances the performance and stability of the model. It can be seen that the geographically weighted regression has a better fitting effect than the global regression.

[0058] The GeoSatFogNet network design is as follows: Figure 3 , the specific steps are as follows:

[0059] Step 6: Input features. Input the regression coefficients of sea fog frequency obtained from geographically weighted regression in each grid and the ocean environmental factors, as well as the corresponding ocean environmental factor data at the time of detection. For the environmental factors in each grid, combine their actual values with the local regression coefficients to obtain the combination of weighted environmental factors and satellite channel data.

[0060] Step 7: Add the CBAM module. CBAM consists of a channel attention mechanism and a spatial attention mechanism. The channel attention mechanism performs global average pooling and global maximum pooling on the input ocean environmental factor data, and then processes and fuses them through a multi-layer perceptron to obtain the attention weights for each channel. This can highlight the channels that have an important impact on the sea fog formation mechanism, thereby reflecting the correlation between different combinations of independent variables (ocean environmental factors) and sea fog at the channel level, making the model's understanding of the sea fog formation mechanism more in-depth and comprehensive. The spatial attention mechanism focuses on the key regions in space, performs average pooling and maximum pooling operations to obtain the feature information in the spatial dimension. Then, these two pooling results are concatenated in the channel dimension to form a feature map containing rich spatial features. Next, pass this concatenated feature map through a convolutional layer. The convolutional kernels in the convolutional layer can perform convolutional operations on the feature map to extract local features in space.

[0061] (1) Collect the actual ocean environmental factor data corresponding to each grid at the time of detection to ensure that the data is precisely matched with the grid and time point corresponding to the regression coefficients.

[0062] (2) Normalization operation. Calculate the data of each type of ocean environmental factor in each grid according to the selected normalization method.

[0063] (3) Combine the actual values of the ocean environmental factors in each grid after normalization with the local regression coefficients obtained from geographically weighted regression for that grid. The normalized ocean environmental factor data is \(x_{ij}\), where \(i\) represents the grid number and \(j\) represents the factor number. Let the local regression coefficient obtained from geographically weighted regression be \(w_{ij}\). For the environmental factor \(x_{ij}\) in each grid \(i\), generate the weighted environmental factor \(y_{ij}\) through the following formula. The formula is: ij , where \(i\) represents the grid number and \(j\) represents the factor number. Let the local regression coefficient obtained from geographically weighted regression be \(w_{ij}\). ij For the environmental factor \(x_{ij}\) in each grid \(i\), ij generate the weighted environmental factor \(y_{ij}\) through the following formula. ij . The formula is:

[0064]

[0065] Through the above method, obtain the weighted ocean environmental factor data of various types in each grid, highlighting the relative importance of different environmental factors in each local area for sea fog frequency.

[0066] Step 8: Feature fusion. The obtained meteorological element features are concatenated with the satellite channel data in the channel dimension to form a total feature. This combination method fuses the weighted information of meteorological elements and the observation information of multiple satellite channels, providing a more comprehensive data basis for subsequent semantic segmentation.

[0067] Step 9: Application of the segmentation network. The fused total feature is input into the TransUNet network. This network combines the advantages of convolutional neural networks and Transformers. The CNN part is responsible for capturing local detailed information and extracting features of different scales in the image through convolutional operations at different levels. The Transformer module uses the self-attention mechanism to learn global context information and effectively handle long-range dependencies. Finally, through operations such as upsampling, the feature map is restored to the original image size for semantic segmentation, and the probability that each pixel belongs to sea fog or non-sea fog is output, thereby achieving precise monitoring and recognition of sea fog.

Claims

1. A deep learning sea fog detection method based on geographically weighted regression, characterized in that: The following steps are involved:

1. Preprocessing Step 1: Acquisition and preprocessing of satellite data. Obtain satellite data at 9:00 a.m. every day for five years, use the mature sea fog detection method to identify sea fog, divide it into regular hexagons, and count the frequency of sea fog in the grid as the dependent variable of model fitting; Step 2: Acquisition and preprocessing of sea fog related marine meteorological elements. According to the existing sea fog prediction research, collect the marine meteorological element data with the same time and space as the sea fog data, preprocess the collected meteorological data to ensure the integrity of the data, and use the acquired marine meteorological elements as independent variables for model fitting; Step 3: Integrate data. After dividing the regular hexagonal grid, match the calculated sea fog frequency in the grid with the sea fog-related marine meteorological elements to make them have the same resolution. Integrate the data of the same time and space and the same resolution to form a comprehensive data set, including sea fog frequency and related meteorological data; 2. Geographically Weighted Regression Model Fitting and Parameter Optimization Step 4: Model fitting. Considering the spatial heterogeneity and data characteristics of the data, a geographically weighted regression model was selected. Before model fitting, global regression was used, sample data of the comprehensive data set was selected for model fitting, VIF was used to detect multicollinearity of independent variables, and variables with collinearity were eliminated; Step 5: Parameter optimization. Visualize the local parameters of the independent variables, analyze the independent variables that affect the occurrence of sea fog, check the model goodness of fit, and select independent variable combinations with better goodness of fit and relatively fewer independent variables to enhance model performance and stability; 3. Establishing GeoSatFogNet Sea Fog Detection Network Step 6: Input features. Input the regression coefficients of each marine environmental element obtained by geographical weighted regression in each grid, as well as the marine environmental element data corresponding to the detection. For each environmental element in each grid, combine its actual value with the local regression coefficient to obtain the weighted environmental element and then combine it with the satellite channel data; Step 7: Acquisition of meteorological features. The channel attention mechanism and the spatial attention mechanism are added. The channel attention mechanism performs global average pooling and global maximum pooling on the input marine environmental element data, and then processes and fuses them through a multi-layer perceptron to obtain the attention weights of each channel. Based on the channel attention output, the spatial attention mechanism performs average pooling and maximum pooling along the channel dimension, concatenates the results, and compresses the channel dimension through a convolutional layer to generate spatial attention weights. Finally, the channel attention and spatial attention weights act on the input features in turn, focusing on key channels and spatial areas; Step 8: Apply the fused features to the segmentation network. The meteorological element features processed by the meteorological element branches are spliced ​​with the satellite channel data in the channel dimension to form a total feature. This combination combines the weighted information of meteorological elements and the observation information of satellite multi-channels, providing a more comprehensive data foundation for subsequent semantic segmentation. The fused total feature is input into the image segmentation network for sea fog recognition.

2. The deep learning sea fog detection method based on geographically weighted regression according to claim 1 is characterized in that: Step 2 includes the following steps: (1) Selection of sea fog independent variables. According to existing sea fog prediction research, the driving factors of sea fog are selected as mean sea surface temperature (sst), 2m temperature (2t), 2m dew point temperature (2dt), 10m wind U (u), 10m wind V (v), total cloud cover (tcc), relative humidity (rh), geopotential height (gt) and vertical velocity (vv). The air-sea temperature difference (TS) is obtained by using 2m temperature and mean sea surface temperature, the air-dew point difference (TD) is obtained by using 2m dew point temperature and 2m temperature, and the wind speed (WSD) and wind direction (WDR) are obtained by using 10m wind U and 10m wind V. (2) Preprocess the data. Use spatial interpolation to ensure data integrity.

3. The deep learning sea fog detection method based on geographically weighted regression according to claim 1 is characterized in that: Step 6 includes the following steps: (1) Collect the actual ocean environment element data corresponding to each grid during detection to ensure that the data accurately matches the grid and time point corresponding to the regression coefficient. (2) Normalization operation: For each type of marine environmental element data in each grid, the selected normalization method is used for calculation. (3) The actual value of the marine environmental elements in each grid after normalization is combined with the local regression coefficient obtained by the geographically weighted regression corresponding to the grid. ij , where i represents the grid number and j represents the element number. Let the local regression coefficient obtained by geographical weighted regression be w ij , for each environmental element x in grid i ij , the weighted environmental factor y is generated by the following formula ij The formula is: In this way, various weighted marine environmental factor data within each grid are obtained, highlighting the relative importance of different environmental factors to the frequency of sea fog in each local area.

Images