A convection cloud detection method based on a double-branch fusion network

Abstract

The application discloses a convection cloud detection method based on a double-branch fusion network, and comprises the following steps: acquiring visible light data and infrared data, which are denoted as a set X; processing infrared data of different wave bands to obtain BTD data, which are denoted as a set Y; pre-processing data in the set X and the set Y; constructing a DBDCN model, training the DBDCN model by using a training set, and denoting the DBDCN model after training as a convection cloud detection model; judging whether the brightness temperature of a pixel point to be detected of a cloud image block is greater than a predefined threshold value, and if yes, directly identifying the cloud image block as a non-convection cloud; otherwise, detecting the pixel point to be detected of the cloud image block by using the convection cloud detection model, so that high-precision convection cloud detection is realized.

Description

Technical Field

[0001] This invention relates to intelligent meteorological detection technology, and in particular to a method for detecting convective clouds based on a dual-branch fusion network. Background Technology

[0002] Convective clouds are a special type of cloud that plays a crucial role in the global water cycle and the balance of radiative energy on local or global scales. These clouds often organize into storm complexes, forming mesoscale convective systems. Furthermore, the formation of convective clouds is frequently accompanied by extreme weather events such as heavy rainfall and hail, making effective detection of convective clouds essential.

[0003] Based on the low cloud top temperature of convective clouds, the brightness temperature (BT) thresholding method using infrared data is widely used due to its simplicity and efficiency. Besides BT thresholding, brightness temperature difference (BTD) data between infrared channels is also widely used because it is sensitive to certain types of clouds. Compared to traditional thresholding methods, deep learning, as a data-driven approach, does not require an understanding of atmospheric physics processes and has therefore attracted much attention in the field of cloud detection. This method integrates multi-band data, facilitating the comprehensive extraction of multi-scale features and enhancing the ability to identify cloud distributions. Although deep learning-based methods introduce a new approach to cloud detection, limitations still exist in practical applications. First, most existing methods only use visible light and infrared data as input. They distinguish convective clouds by utilizing changes in cloud top BT observed in the infrared channels (the BT of convective clouds is often lower than that of non-convective clouds). However, this method is prone to misclassifying clouds with lower brightness temperatures surrounding convective clouds as convective clouds, leading to false positives. Furthermore, traditional convolutional networks have kernels of fixed shapes, but convective clouds vary in shape and size, and traditional convolution can easily cause problems such as blurred cloud boundaries and noise.

[0004] One of the most commonly used methods in convective cloud detection is the fixed brightness temperature threshold method. Since strong convective clouds have low brightness temperatures, an appropriate threshold can be set for infrared data with a center wavelength close to 11 micrometers to detect them. Brightness temperature-based thresholding can quickly detect convective cloud regions during both day and night, but it is not universally applicable due to the influence of atmospheric conditions, cloud state, and other microscopic physical characteristics. Brightness temperature thresholding (BTD) reveals cloud top features to detect convective clouds by calculating the difference between the infrared and water vapor bands. However, factors such as spatial and spectral resolution, as well as signal-to-noise ratio, can affect the judgment of the BTD thresholding method. Furthermore, the final detection result is largely influenced by the choice of threshold, and a significant amount of time is usually required in experiments to select an appropriate threshold. Summary of the Invention

[0005] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a convective cloud detection method based on a dual-branch fusion network.

[0006] Technical solution: The present invention provides a convective cloud detection method based on a dual-branch fusion network, comprising the following steps:

[0007] Step 1: Obtain visible light data and infrared data, denoted as set X;

[0008] Step 2: Process the infrared data of different bands to obtain BTD data, denoted as set Y;

[0009] Step 3: Preprocess the data in set X and set Y to obtain image patch x and image patch y, and divide image patch x and image patch y into training set and test set according to the ratio;

[0010] Step 4: Construct the DBDCN model. Train the DBDCN model using the training set. The trained DBDCN model is denoted as the convective cloud detection model. Image patch x and image patch y serve as the input terms of the two branches in the DBDCN model.

[0011] Step 5: Determine whether the brightness temperature of the pixel to be tested in the cloud image block is greater than a predefined threshold. If it is, it is directly identified as a non-convective cloud; otherwise, the convective cloud detection model is used to detect convective clouds in the pixel to be tested in the cloud image block.

[0012] Furthermore, step 2 includes:

[0013] Infrared channel cloud image data at 3.9μm, 6.2μm, 7.3μm, 8.6μm, 10.4μm, 11.2μm, and 12.4μm were acquired and denoted as BT3.9, BT6.2, BT7.3, BT8.6, BT10.4, BT11.2, and BT12.4. The cloud image data of different bands were subtracted, where the first BTD data was BT7.3-BT11.2, the second BTD data was BT11.2-BT3.9, the third BTD data was BT11.2-BT12.4, the fourth BTD data was BT8.6-BT11.2, the fifth BTD data was BT6.2-BT10.4, and the sixth BTD data was BT10.4-BT12.4.

[0014] Furthermore, step 3 includes:

[0015] Cropping is performed on each pixel in sets X and Y, centered at the same spatial location, to obtain image patches x∈R. p×p×C and image patch y∈R p×p×C′ , where p is the size of the image patch, C represents the number of channels in set X, and C′ represents the number of classes in the BTD data.

[0016] Furthermore, step 4 includes:

[0017] Image patch x is input into the first branch of the DBDCN model. After passing through a convolutional layer, image patch x is input into the first fusion unit, and the output feature is denoted as x′. Feature x′ is input into the second fusion unit, and the output feature is denoted as x″. Feature x″ is input into the third fusion unit, and the output feature is denoted as F4.

[0018] Image patch y is input into the second branch of the DBDCN model. After passing through a convolutional layer, image patch y is input into the first fusion unit, and the output feature is denoted as y′. Feature y′ is input into the second fusion unit, and the output feature is denoted as y″. Feature y″ is input into the third fusion unit, and the output feature is denoted as F4′.

[0019] Features F4 and F4′ are concatenated along the channel dimension and input into a fully connected layer to obtain cloud detection results.

[0020] Furthermore, step 4 includes:

[0021] Image patch x is fed into a deformable convolutional layer after passing through a convolutional layer, and the output feature is denoted as F1; image patch y is fed into a deformable convolutional layer after passing through a convolutional layer, and the output feature is denoted as F1′; features F1 and F1′ are fed into a cross-branch fusion module, and the output features of the cross-branch fusion module are multiplied and added pointwise with features F1 and F1′ respectively, and then the outputs are fed into the residual module respectively, and the output features of the residual module in the two branches are x′ and y′ respectively;

[0022] Feature x′ is input into a deformable convolutional layer, and the output feature is denoted as F2; ​​feature y′ is input into a deformable convolutional layer, and the output feature is denoted as F2′; features F2 and F2′ are input into a cross-branch fusion module, and the output features of the cross-branch fusion module are multiplied and added pointwise with features F2 and F2′ respectively. Then the outputs are input into the residual module, and the output features of the residual module in the two branches are x″ and y″ respectively.

[0023] Feature x″ is input into a deformable convolutional layer, and the output feature is denoted as F3; feature y″ is input into a deformable convolutional layer, and the output feature is denoted as F3′; features F3 and F3′ are input into a cross-branch fusion module, and the output features of the cross-branch fusion module are multiplied and added point by point with features F3 and F3′ respectively. Then the outputs are input into the residual module respectively, and the output features of the residual module in the two branches are F4 and F4′ respectively.

[0024] Furthermore, the cross-branch fusion module in each fusion unit includes:

[0025] Let F represent the features generated by the i-th deformable convolutional layer in both branches. i ∈Rh×w×c and F i ′∈R h×w×c , will F i and F i The two features are concatenated along the channel dimension, and then a 1×1 convolutional layer is used to reduce the channel dimension to obtain feature Z. i The mathematical model represents:

[0026] Z i =Conv([F i ,F i ′])

[0027] Where [.,.] represents a channel concatenation operation, Conv() represents a 1×1 convolution operation, and Z i ∈R h×w×c ;

[0028] Feature Z i The shape is a feature A∈R n×c and Z i Transpose integer as feature B∈R c×n Where n represents h×w; matrix multiplication is performed on the features and features, and the channel attention map C∈R is obtained through a Sigmoid function layer. c×c The mathematical model is:

[0029] C = σ(BA).

[0030] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0031] 1. In order to effectively detect convective clouds, this invention collects BTD data, which is sensitive to cloud analysis, as a supplementary input to visible light and infrared data, and constructs a dual-branch convolutional network to fully fuse these data.

[0032] 2. Considering that the shape of convective clouds is often variable, this invention uses deformable convolutional layers to extract the features of convective clouds; based on the extracted features, this invention constructs a cross-branch fusion module to fuse the features of different branches;

[0033] 3. Through quantitative and qualitative comparative analysis, the convective cloud detection model proposed in this invention has achieved better results in detection accuracy than other existing models. Attached Figure Description

[0034] Figure 1 This is a flowchart of a convective cloud detection method based on a dual-branch fusion network in an embodiment;

[0035] Figure 2 Example diagrams for labels in the embodiments;

[0036] Figure 3This is a flowchart illustrating the workflow of the convective cloud detection model in this embodiment.

[0037] Figure 4 This is a schematic diagram of the residual module structure in the embodiment;

[0038] Figure 5 This is a schematic diagram illustrating the working principle of the cross-branch fusion module in the embodiment;

[0039] Figure 6 This is a schematic diagram illustrating the working principle of the deformable convolutional layer in the embodiment;

[0040] Figure 7 This is an example diagram of a deformable convolutional layer in the embodiment;

[0041] Figure 8 This is a comparison chart of convective cloud detection results in the example. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0043] The flowchart of the convective cloud detection method based on a dual-branch fusion network described in this embodiment is as follows: Figure 1 As shown, the convective cloud detection method includes the following steps:

[0044] Step 1: Obtain visible light data and infrared data, denoted as set X;

[0045] Step 2: Process the infrared data of different bands to obtain BTD data, denoted as set Y;

[0046] Step 3: Preprocess the data in set X and set Y to obtain image patch x and image patch y, and divide image patch x and image patch y into training set and test set according to the ratio;

[0047] Step 4: Construct the DBDCN model. Train the DBDCN model using the training set. The trained DBDCN model is denoted as the convective cloud detection model. Image patch x and image patch y serve as the input terms of the two branches in the DBDCN model.

[0048] Step 5: Determine whether the brightness temperature of the pixel to be tested in the cloud image block is greater than a predefined threshold. If it is, it is directly identified as a non-convective cloud; otherwise, the convective cloud detection model is used to detect convective clouds in the pixel to be tested in the cloud image block.

[0049] In one embodiment, China is influenced by the summer monsoon, with most of its annual precipitation concentrated in the summer and autumn, particularly from May to September. During this period, precipitation in South China generally exceeds that in the north. Analysis of climate bulletins published by the National Climate Center shows that the 2020 rainy season and precipitation in the middle and lower reaches of the Yangtze River reached record highs. Specifically, precipitation in most areas of the middle and lower reaches of the Yangtze River and southwestern Yunnan ranged from 1200 mm to 2000 mm, with vigorous convective cloud development. Although precipitation in 2021 was slightly lower than in 2020, it was still 6.7% higher than the average. Therefore, this embodiment selects most of southern China, between 97°E and 122°E and between 20°N and 35°N, as the study area, and selects May to September of 2020 and 2021 as the study period.

[0050] The Japan Meteorological Agency's Satellite Meteorological Center (MSC) has introduced a cloud classification product called High Resolution Cloud Analysis Information (HCAI), which can effectively distinguish typical convective clouds—cumulonimbus clouds—from other clouds. Therefore, this product was chosen as the initial convective cloud label. However, due to the strict threshold selection during HCAI generation, the resulting cumulonimbus cloud labels often contain a lot of noise and holes. Figure 2 Example images are created for convective cloud labels, where Figure (a) shows visible light data, Figure (b) shows infrared data, Figures (c) and (d) show HCAI cloud classification product samples and their magnified partial results, respectively, and Figures (e) and (f) show the results of manual fine-tuning and partial magnification, respectively. Furthermore, HCAI can ignore other types of convective clouds. To address these issues, this embodiment utilizes visible light and infrared channel cloud image data to manually fine-tune the labels through visual interpretation. Figure 2 As shown in (e), more uniform convective clouds were labeled, and small-scale convective clouds that are easily overlooked by HCAI were also correctly labeled as convective cloud labels X∈R. H×W×C This will be input into the first branch of the DBDCN model, where H, W, and C represent the height, width, and number of channels of the cloud map, respectively.

[0051] In one example, step 2 includes:

[0052] Infrared channel cloud image data at 3.9μm, 6.2μm, 7.3μm, 8.6μm, 10.4μm, 11.2μm, and 12.4μm were obtained through infrared data analysis, denoted as BT3.9, BT6.2, BT7.3, BT8.6, BT10.4, BT11.2, and BT12.4. The cloud image data from different bands were subtracted to construct the BTD data Y∈R. H×W×C′The expression C′ represents the number of BTD data types. The first BTD data set, BT7.3-BT11.2, helps detect the distribution of water vapor content in clouds; a larger difference indicates higher water vapor content, suggesting the cloud is more likely to be a convective cloud. The second BTD data set, BT11.2-BT3.9, uses BT11.2 to observe high-altitude clouds and BT3.9 primarily for low-altitude clouds; their difference helps detect cloud thickness, with a larger difference indicating a thicker cloud. The third BTD data set, BT11.2-BT12.4, detects cloud top height; a larger difference indicates a higher probability of the cloud being a convective cloud. The fourth BTD data set, BT8.6-BT11.2, helps detect ice clouds; a larger value indicates a higher probability of the cloud being an ice cloud. The fifth BTD data set, BT6.2-BT10.4, uses BT6.2 and BT10.4 to observe cloud base and cloud top brightness temperatures, respectively; their difference represents the cloud's vertical structure information, with a large difference indicating a highly developed cloud. The sixth BTD data point is BT10.4-BT12.4, which reflects the water vapor content in the cloud. The larger the difference, the more likely the cloud is to be a convective cloud.

[0053] In one example, step 3 includes:

[0054] Cropping is performed on each pixel in sets X and Y, centered at the same spatial location, to obtain image patches x∈R. p×p×C and image patch y∈R p×p×C′ Where p is the size of the image patch, C represents the number of channels in set X, and C′ represents the number of classes in the BTD data. Empirically, p can be set to 7.

[0055] Specifically, step 4 includes:

[0056] Image patch x is input into the first branch of the DDBDN model. After passing through a convolutional layer, image patch x is input into the first fusion unit, and the output feature is denoted as x′. The kernel size of the convolutional layer is 5×5. An initial feature is extracted from the input data using the convolutional layer. Feature x′ is input into the second fusion unit, and the output feature is denoted as x″. Feature x″ is input into the third fusion unit, and the output feature is denoted as F4.

[0057] Image patch y is input into the second branch of the DBDCN model. After passing through a convolutional layer, image patch y is input into the first fusion unit, and the output feature is denoted as y′. Feature y′ is input into the second fusion unit, and the output feature is denoted as y″. Feature y″ is input into the third fusion unit, and the output feature is denoted as F4′.

[0058] Features F4 and F4′ are concatenated along the channel dimension and input into a fully connected layer to obtain cloud detection results.

[0059] like Figure 3 The flowchart of the convective cloud detection model shown herein further includes step 4, which includes:

[0060] Image patch x is passed through a convolutional layer and then input into a deformable convolutional layer, with the output feature denoted as F1; image patch y is passed through a convolutional layer and then input into a deformable convolutional layer, with the output feature denoted as F1′; features F1 and F1′ are input into a cross-branch fusion module, and the output features of the cross-branch fusion module are multiplied and added pointwise with features F1 and F1′ respectively, and then the outputs are input into the residual module respectively. The output features of the residual module in the two branches are x′ and y′ respectively (not shown in the figure);

[0061] Feature x′ is input into a deformable convolutional layer, and the output feature is denoted as F2; ​​feature y′ is input into a deformable convolutional layer, and the output feature is denoted as F2′; features F2 and F2′ are input into a cross-branch fusion module, and the output features of the cross-branch fusion module are multiplied and added pointwise with features F2 and F2′ respectively. Then the outputs are input into the residual module, and the output features of the residual module in the two branches are x″ and y″ respectively.

[0062] Feature x″ is input into a deformable convolutional layer, and the output feature is denoted as F3; feature y″ is input into a deformable convolutional layer, and the output feature is denoted as F3′; features F3 and F3′ are input into a cross-branch fusion module, and the output features of the cross-branch fusion module are multiplied and added point by point with features F3 and F3′ respectively. Then the outputs are input into the residual module respectively, and the output features of the residual module in the two branches are F4 and F4′ respectively.

[0063] like Figure 4 The diagram shows the structure of the residual module, which further enhances the discriminative ability of features in the two branches. The residual module consists of two convolutional layers (Conv), with a kernel size of 3×3. A ReLU layer is also connected after each convolutional layer (Conv).

[0064] Figure 5 The flowchart illustrating the cross-branch fusion module is shown. The cross-branch fusion module in each fusion unit includes:

[0065] Let F represent the features generated by the i-th deformable convolutional layer in both branches. i ∈R h×w×c and F i ′∈R h×w×c , will F i and F i The two features are concatenated along the channel dimension, and then a 1×1 convolutional layer is used to reduce the channel dimension to obtain feature Z. i The mathematical model represents:

[0066] Z i =Conv([F i ,F i ′])

[0067] Where [.,.] represents a channel concatenation operation, Conv() represents a 1×1 convolution operation, and Z i ∈R h×w×c ;

[0068] Feature Z i The shape is a feature A∈R n×c and Z i Transpose integer as feature B∈R c×n Where n represents h×w; matrix multiplication is performed on the features and features, and the channel attention map C∈R is obtained through a Sigmoid function layer. c×c The mathematical model is:

[0069] C = σ(BA).

[0070] Existing feature fusion methods mostly involve directly concatenating or weighting features across channels. This approach is simple and straightforward, requiring only that different types of features have the same dimension. However, these methods are generally suitable for situations where there is no significant correlation between features, or the correlation between features is weak. In this embodiment, BTD data is used as complementary information to infrared data and input into the second branch of the DBDCN model, which can provide the distribution of convective clouds. There is a strong correlation between the features of the two branches. Therefore, this embodiment uses a cross-branch feature fusion module to fully fuse features from different branches by utilizing the dependencies between different channels and weighting them accordingly.

[0071] In the above process, considering the frequent changes in the shape of different convective clouds, a deformable convolutional layer with a kernel size of 3×3 is used to capture various shape information. Traditional convolutions have fixed kernels, making it difficult to learn the diverse shape information of convective clouds. In contrast, deformable convolutional layers can adaptively adjust the shape of the kernel, making them more suitable for extracting features from convective clouds. Figure 6 As shown, in this embodiment, a deformable convolutional layer is used to learn an offset for each sampling location using a single-layer convolution. This offset is used to adjust the sampling position of the convolutional kernel on the input sample point. For each sampling point, an offset is typically generated in both the horizontal and vertical directions. Then, the learned offset is used to calculate the actual sampling position of the sample point on the input feature map. Assuming a sampling point on the input feature map is (e, f), and the corresponding offset is (Δe, Δf), then the new sampling point position is (e + Δe, f + Δf). Therefore, the deformable convolutional layer can autonomously adjust the receptive field according to the shape of the convective cloud. Figure 7An example is presented to compare the differences between traditional convolution and deformable convolution, where Figure (a) shows traditional convolution and Figure (b) shows deformable convolution. From Figure 7 It can be clearly observed that deformable convolution can better fit the shape of convective clouds.

[0072] This invention provides a dual-branch deformable convolutional neural network to extract convective cloud features, achieving high-precision detection of convective clouds. To verify the effectiveness of this invention, the detection results are compared with those of HCAI cloud classification products, dual-stream fully convolutional networks (DSFC), convective cloud detection U-shaped networks (Convection-UNet), CNN-based cloud detection algorithms (CDNet), and deformable context boundary weighted networks (DABNet).

[0073] Figure 8 The images show the detection results of the method of this invention, along with HCAI cloud classification products, Dual-Stream Fully Convolutional Network (DSFC), Convection-UNet for Flow Cloud Detection, CNN-based Cloud Detection Algorithm (CDNet), and Deformable Context Boundary Weighted Network (DABNet), on infrared datasets. Figure (a) shows the AHI data, Figure (b) shows the label database, and Figures (c)-(h) show the detection results of HCAI, DSFC, Convection-UNet, CDNet, DABNet, and DDBCN, respectively. Figure 8 As can be seen from the table, compared with other methods, the present invention can provide more accurate results on convective cloud distribution and has higher detection results on cloud boundaries. The present method can obtain more accurate convective cloud detection results. Table 1 shows a comparison of the detection accuracy of six different methods on the infrared dataset.

[0074] Table 1 Comparison of convective cloud detection accuracy in infrared datasets

[0075]

[0076]

[0077] Note: An upward arrow indicates that the higher the better, and vice versa.

[0078] As can be seen from Table 1, the hit rate (POD), false alarm rate (MAR), and critical success index (CSI) of the method of this invention are significantly higher than those of the other five classification methods. Among them, DABNet focuses on cloud boundary detection through deformable convolution and has better segmentation accuracy compared with the other four methods. DDBCN has a slightly lower false alarm rate than DABNet, but overall, our detection results are better than those of DABNet.

[0079] In summary, compared with HCAI cloud classification products, dual-stream fully convolutional networks (DSFC), convection-UNet, CNN-based cloud detection algorithms (CDNet), and deformable context boundary weighted networks (DABNet), the method of this invention has a significant advantage in the accuracy of convection cloud detection.

Claims

1. A method for detecting convective clouds based on a dual-branch fusion network, characterized in that, The steps include the following: Step 1: Acquire visible light data and infrared data, denoted as set X; Step 2: Process the infrared data of different bands to obtain BTD data, denoted as set Y; Step 3: Preprocess the data in sets X and Y to obtain image patches. and image blocks , image blocks and image blocks The dataset is divided into a training set and a test set according to a certain ratio. include: Each pixel in set X and set Y is cropped with the same spatial location as the center to obtain image patches. and image blocks ,in The size of the image patch. This represents the number of channels in set X. Indicates the number of classes in the BTD data; Step 4: Construct the DBDCN model. Train the DBDCN model using the training set, and denote the trained DBDCN model as the convective cloud detection model; where image patches and image blocks As input items for two branches in the DBDCN model; including: Image blocks The image patch is input into the DBDCN model as the first branch of the DBDCN model. After passing through the convolutional layer, the input is fed into the first fusion unit, and the output feature is denoted as... ; Features The input is fed into the second fusion unit, and the output feature is denoted as... ; Features The input is fed into the third fusion unit, and the output feature is denoted as... ; Image blocks As the input to the second branch of the DBDCN model, the image patch is input into the DBDCN model. After passing through the convolutional layer, the input is fed into the first fusion unit, and the output feature is denoted as... ; Features The input is fed into the second fusion unit, and the output feature is denoted as... ; Features The input is fed into the third fusion unit, and the output feature is denoted as... ; Image Patch After passing through the convolutional layer, the input is fed into a deformable convolutional layer, and the output feature is denoted as... Image blocks After passing through the convolutional layer, the input is fed into a deformable convolutional layer, and the output feature is denoted as... ; Features and characteristics The input is fed into the cross-branch fusion module, and the output features of the cross-branch fusion module are respectively compared with the feature... and characteristics Perform point-by-point multiplication and addition, and then input the outputs into the residual module. The output characteristics of the residual module in the two branches are as follows: and ; Features The input is fed into a deformable convolutional layer, and the output feature is denoted as... ; Features The input is fed into a deformable convolutional layer, and the output feature is denoted as... ; Features and characteristics The input is fed into the cross-branch fusion module, and the output features of the cross-branch fusion module are respectively compared with the feature... and characteristics Perform point-by-point multiplication and addition, and then input the outputs into the residual module. The output characteristics of the residual module in the two branches are as follows: and ; Features The input is fed into a deformable convolutional layer, and the output feature is denoted as... ; Features The input is fed into a deformable convolutional layer, and the output feature is denoted as... ; Features and characteristics The input is fed into the cross-branch fusion module, and the output features of the cross-branch fusion module are respectively compared with the feature... and characteristics Perform point-by-point multiplication and addition, and then input the outputs into the residual module. The output characteristics of the residual module in the two branches are as follows: and ; Features and characteristics The data are concatenated along the channel dimension and input into a fully connected layer to obtain cloud detection results; Step 5: Determine whether the brightness temperature of the pixel to be tested in the cloud image block is greater than a predefined threshold. If it is, it is directly identified as a non-convective cloud; otherwise, the convective cloud detection model is used to detect convective clouds in the pixel to be tested in the cloud image block.

2. The convective cloud detection method based on a dual-branch fusion network according to claim 1, characterized in that, Step 2 includes: Infrared channel cloud image data at 3.9μm, 6.2μm, 7.3μm, 8.6μm, 10.4μm, 11.2μm, and 12.4μm were acquired and denoted as BT3.9, BT6.2, BT7.3, BT8.6, BT10.4, BT11.2, and BT12.

4. The cloud image data of different bands were subtracted, where the first BTD data was BT7.3-BT11.2, the second BTD data was BT11.2-BT3.9, the third BTD data was BT11.2-BT12.4, the fourth BTD data was BT8.6-BT11.2, the fifth BTD data was BT6.2-BT10.4, and the sixth BTD data was BT10.4-BT12.

4.

3. The convective cloud detection method based on a dual-branch fusion network according to claim 1, characterized in that, The cross-branch fusion module in each fusion unit includes: The first of the two branches The features generated by a deformable convolutional layer are represented as follows: and ,Will and The two features are concatenated along the channel dimension, and then utilized Convolutional layers perform channel dimensionality reduction to obtain features The mathematical model represents: ， in, This indicates a channel serialization operation. express Convolution operation, ; Features Plastic surgery as a feature and will Characterized by transpose shaping ,in represent ; Perform matrix multiplication on features and obtain channel attention maps through a Sigmoid function layer. The mathematical model is as follows: 。

Images