North pole cloud detection method for remote sensing image of FY-3 satellite

Abstract

The invention relates to a north pole cloud detection method, in particular to a north pole cloud detection method for a remote sensing image of a FY-3D / MERS-II satellite, which comprises the following steps of: S1, acquiring all-year and all-day FY-3D / MERS-II data of a north pole region, and preprocessing the all-year and all-day FY-3D / MERS-II data; s2, selecting characteristic factors of a north pole cloud detection method, and synthesizing the characteristic factors into an FY-3D / MERS-II multispectral remote sensing image; s3, dividing the FY-3D / MERSI-II multispectral remote sensing image of the arctic region all year round into data of different seasons and different time periods; s4, constructing a Dense-Net cloud detection model based on a deep learning method, and obtaining cloud detection results of the arctic region in different seasons and different time periods; according to the invention, the accuracy and the automation degree of north pole cloud detection are improved; according to the method, FY-3D / MERS-II remote sensing image cloud detection of the arctic region in different seasons and different time periods is realized through a Dense-Net cloud detection model based on deep learning, and another research thought is provided for arctic cloud detection; the overall cloud detection precision is high.

Description

TECHNICAL FIELD

[0001] The application relates to a polar cloud detection method, in particular to a polar cloud detection method for Fengyun-3 satellite remote sensing images. BACKGROUND

[0002] Polar ground verification data are sparsely distributed in space and time, and optical satellite remote sensing technology can collect a large amount of effective information in a short time, and has become one of the main means for polar cloud detection. However, different waveband satellite remote sensing images have different response signals for clouds at different altitudes and types, and a single waveband image often only identifies clouds at a specific altitude, and in practice, it is necessary to combine multispectral remote sensing images to identify clouds at different altitudes and of different types and distinguish clouds from ice and snow underlying surfaces. In addition, the absence of optical bands during polar night will lead to a decrease in the overall accuracy of the cloud detection algorithm. Therefore, accurate polar cloud detection has been a difficulty and focus in the field of cloud detection.

[0003] Most of the existing polar cloud detection methods are traditional threshold methods. The polar cloud detection method based on the traditional threshold value finds the threshold value parameter with the highest recognition accuracy through spectral testing to distinguish cloud and non-cloud pixels, and realizes the detection of polar clouds. The polar cloud detection method based on the traditional threshold value has a certain subjectivity, and is sensitive to satellite load, season, underlying surface type, temperature, atmospheric humidity and the like, resulting in poor universality. SUMMARY

[0004] The application aims to provide a polar cloud detection method for Fengyun-3 satellite remote sensing images, which detects clouds in the FY-3D / MERSI-Ⅱ remote sensing images of Fengyun-3D satellite in the polar region through a deep learning method, utilizes a large amount of true value label data for driving, adopts a deep learning model with semantic information extraction and learning as an algorithm framework, does not need a large number of parameter adjustments and threshold selections, retains the detail information in the satellite remote sensing images, improves the accuracy and automation degree of the polar cloud detection, and provides a new method for the polar cloud detection.

[0005] The polar cloud detection method for Fengyun-3 satellite remote sensing images comprises the following steps: S1. Obtain FY-3D / MERSI-Ⅱ data of the polar region throughout the year and at all times and perform pretreatment; S1.1. Obtain FY-3D / MERSI-II L1 1000m resolution data of the Arctic region all year round and all day long. FY-3D / MERSI-II L1 data includes earth observation data and geographical positioning GEO data. FY-3D / MERSI-II L1 earth observation data file stores earth observation data after radiation calibration preprocessing, and GEO data file stores geographical positioning data after geographical positioning preprocessing, which is auxiliary to FY-3D / MERSI-II L1 earth observation data. The GEO data with 1000m spatial resolution is used to resample the FY-3D / MERSI-II L1 earth observation data of the Arctic region to 1000m, and the projection mode is polar projection; S1.2. Obtain FY-3D / MERSI-II CLM products of the Arctic region all year round and all day long. FY-3D / MERSI-II is equipped with 25 channels, and the spectral coverage range is 0.4μm-12.0μm. FY-3D / MERSI-II CLM product is a cloud detection result obtained based on multi-feature threshold method using spectral reflectance or brightness temperature of 9 channels of FY-3D / MERSI-II, including channel 3 (0.65μm), channel 4 (0.865μm), channel 6 (1.64μm), channel 7 (2.13μm), channel 19 (1.03μm), channel 20 (3.8μm), channel 21 (4.050μm), channel 24 (10.8μm) and channel 25 (12.0μm). The content includes cloud and clear sky identification mark and reliability information. The certain clear sky and possible clear sky in the content of FY-3D / MERSI-II CLM product are assigned as 0, and the rest is assigned as 1 to generate a cloud mask image of the Arctic region; S2. Select the characteristic factor composition of the Arctic cloud detection method for FY-3D / MERSI-II multi-spectral remote sensing image, including the Arctic cloud detection image obtained by the cloud index method based on traditional threshold; S2.1. Considering the consistency of the data, first, the spectral reflectance or brightness temperature of the same 9 channels as the FY-3D / MERSI-II CLM product generated by 9 channels is selected as the characteristic factor. In addition, channel 5 (1.38μm) of FY-3D / MERSI-II belongs to a strong water vapor absorption band. Water vapor has strong absorption capacity for solar radiation, and the radiation from the ground to the lower part of the troposphere is difficult to reach the optical satellite sensor, resulting in low reflectivity. Cirrus clouds are mostly located in the upper troposphere, and most of the water vapor in the atmosphere is concentrated below the cirrus cloud. Relatively speaking, the solar radiation intensity above the cirrus cloud is strong, and the reflectivity is high. The spectral reflectance of channel 5 (1.38μm) can be used to detect cirrus clouds. Therefore, the spectral reflectance of FY-3D / MERSI-II channel 5 (1.38μm) is selected as the characteristic factor; S2.2. The Arctic cloud detection images obtained by the two traditional threshold-based cloud index methods can also reflect the distribution of clouds in the Arctic region; select cloud index 1 C CI1 The obtained Arctic cloud detection image is used as a characteristic factor, and the calculation formula is: ; In the formula: C CI1 is the cloud index 1; R1 is the spectral reflectance of FY-3D / MERSI-II band 1 (0.47 μm); R2 is the spectral reflectance of FY-3D / MERSI-II band 2 (0.55 μm); R3 is the spectral reflectance of FY-3D / MERSI-II band 3 (0.65 μm); R4 is the spectral reflectance of FY-3D / MERSI-II band 4 (0.865 μm); R6 is the spectral reflectance of FY-3D / MERSI-II band 6 (1.64 μm); When |C CI1 -1| < 0.25, the pixel is identified as a cloud pixel, and a cloud detection image is obtained; S2.3. Select cloud index 2 C CI2 The obtained Arctic cloud detection image is used as a characteristic factor, and the calculation formula is: ; In the formula: C CI2 is the cloud index 2; R3 is the spectral reflectance of FY-3D / MERSI-II band 3 (0.65 μm); R6 is the spectral reflectance of FY-3D / MERSI-II band 6 (1.64 μm); When 0 < C CI2 < 0.4 and simultaneously satisfy R3 > 0.2, it is identified as a cloud pixel, and a cloud detection image is obtained; S2.4. The spectral reflectance or brightness temperature images of FY-3D / MERSI-II band 3 (0.65 μm), band 4 (0.865 μm), band 6 (1.64 μm), band 7 (2.13 μm), band 19 (1.03 μm), band 20 (3.8 μm), band 21 (4.050 μm), band 24 (10.8 μm), and band 25 (12.0 μm), the spectral reflectance image of band 5 (1.38 μm), and C CI1 and C CI2 The obtained Arctic cloud detection image is merged into a multi-spectral remote sensing image; S3. The FY-3D / MERSI-II multi-spectral remote sensing images of the Arctic region throughout the year and at all times are divided into data of different seasons and different time periods; S4. Construct a Dense-Net cloud detection model based on deep learning method to obtain cloud detection results in different seasons and at different time periods in the Arctic region; S4.1. Due to its unique geographical location at the northernmost tip of the Earth, the Arctic region exhibits significant polar day and polar night phenomena. Furthermore, at night, the spectral reflectance of the visible light band of optical satellite sensors cannot be used properly. To investigate the differences in the accuracy of cloud detection methods using satellite remote sensing imagery in different seasons and time periods in the Arctic, FY-3D / MERSI-II multispectral remote sensing imagery from different seasons and time periods was used to construct seasonal all-day, daytime, and nighttime cloud detection models. 70% of the FY-3D / MERSI-II multispectral remote sensing imagery from different seasons and time periods formed the training set samples for the cloud detection model, and 30% of the FY-3D / MERSI-II multispectral remote sensing imagery from different seasons and time periods formed the test set samples. S4.2. Use the mean square error (MSE) loss function and the Adam optimizer in training the deep learning-based Dense-Net cloud detection model. The root mean square error represents the average of the sum of squares of the differences between the detected value and the true value, and is calculated as follows: ; In the formula: MSE is the abbreviation of Mean Square Error, which stands for root mean square error; x i is the detection value; y i is the true value; N is the number of training samples; The Adam optimizer iterates the neuron weight parameters in a deep learning model based on the model network's backpropagation and gradient information, updating them toward minimizing the loss function. As an adaptive learning rate optimization algorithm, the Adam optimizer can dynamically adjust the learning rate of a deep learning model and is suitable for processing large-scale datasets and complex model network structures. S4.3. During the encoding process of the deep learning-based Dense-Net cloud detection model, a dense connection module and a transition module were introduced to build a Dense-Net cloud detection model for Arctic remote sensing imagery based on the deep learning-based U-Net model for remote sensing image feature extraction. The basic structure of the U-Net model consists of an input layer, a convolutional layer, a pooling layer, an activation function layer, a fully connected layer, and an output layer; The convolution layer is used to extract multi-level features from remote sensing images. Each element of the convolution kernel slides in the form of a sliding window. The convolution operation Conv is performed on the local area of ​​the input remote sensing image to complete the feature image acquisition of the remote sensing image. The convolution operation formula is as follows: ; In the formula: K represents the convolution kernel, I(m,n) represents the input image matrix, and H(i,j) represents the output image matrix; The two key parameters in the convolution operation are the size and number of convolution kernels. The convolution kernel size is usually 3×3. The number of convolution kernels can be adjusted according to the design and requirements of the network structure. The number of convolution kernels determines the number of channels in the output cloud detection image. The formula for calculating the size of the output cloud detection image after the input remote sensing image is convolved is as follows: ; In the formula: N represents the output image size, W represents the input image size, K represents the convolution kernel size, P is the padding, and S is the step size; The pooling layer is usually connected after the convolution layer. It performs sparse processing on the feature vectors, reduces the spatial dimension of the feature image, simplifies the output of the convolution layer, and thus improves computational efficiency and generalization performance. Common pooling operations include Max Pooling and Average Pooling. Max Pooling emphasizes the most significant feature vectors within the pooling window, which helps retain more useful information, while Average Pooling considers the average of all feature vectors within the pooling window, which helps smooth the feature vectors and reduce noise interference. Unlike convolutional layers and pooling layers, activation functions, as a nonlinear operation, enable the U-Net model to learn more complex feature vectors and improve the model's nonlinear mapping capabilities. Depending on the model network structure and classification objectives, commonly used activation functions include ReLU activation function and Sigmoid activation function. The specific calculation formula is as follows: ; ; In the formula, x represents the feature vector of the input activation function; The fully connected layer is used to integrate the feature vectors extracted by the forward network layer, which is equivalent to linear processing of the multi-dimensional information of the previous network layer; The U-Net model achieves pixel-level classification and detection through its relatively symmetrical encoder-decoder architecture and skip connections. The encoder gradually extracts high-level semantic features of remote sensing images through convolution and pooling operations, while the decoder maps the feature vector back to the original resolution cloud detection image through upsampling operations. The upsampling operation is equivalent to the transpose operation of the convolution operation. The calculation formula is as follows: ; Where K represents the convolution kernel, I(m,n) represents the input image matrix of the upsampling operation, and H(i,j) represents the output image matrix of the upsampling operation; S4.4. In the decoding process of the deep learning-based Dense-Net cloud detection model, after the combined operation of three dense connection modules and transition modules, the training set samples are restored to the original size through three upsampling operations and the corresponding jump connection in the encoding process, and the cloud detection result is obtained.

[0006] Further, the method for dividing the FY-3D / MERSI-II multi-spectral remote sensing images of the Arctic region throughout the year and throughout the day into data of different seasons in step S3 is: S3.1. The multi-spectral remote sensing images of March, April and May are defined as spring data, the multi-spectral remote sensing images of June, July and August are defined as summer data, the multi-spectral remote sensing images of September, October and November are defined as autumn data, and the multi-spectral remote sensing images of December, January and February are defined as winter data.

[0007] Further, the method for dividing the FY-3D / MERSI-II multi-spectral remote sensing images of the Arctic region throughout the year and throughout the day into data of different time periods in step S3 is: S3.2. According to the solar zenith angle 85° as the division standard, the solar zenith angle ≤ 85° is divided into daytime, and the rest is divided into night; for the data that has both daytime imaging and nighttime imaging, it is classified as nighttime data; the FY-3D / MERSI-II multi-spectral remote sensing images of the Arctic region throughout the day in different seasons are divided into daytime and nighttime data.

[0008] Further, according to the polar cloud detection method of the FY-3 satellite remote sensing image in claim 3, characterized in that: S4.3.1. A dense connection module is introduced, which is composed of three layers of dense connection operations, the convolution kernel size of the dense connection operation is 3x3, the step is 1, and a ReLU activation function is connected after each convolution layer; The dense connection module is composed of multiple dense connection operations, and the input of each layer in the dense connection operation includes the output of all previous layers. Each layer directly accesses the feature vectors of all previous layers, and the operation process is as follows: ; In the formula, x i represents the feature vector of the i-th layer, H L (•) is defined as the composite operation of convolution operation and activation function, and multiple feature vectors are connected to form a tensor for dense connection operation; S4.3.2. A transition module is introduced, which is composed of a convolution operation with a kernel size of 1*1 and a step of 1 and an average pooling operation, and the pooling window is 2*2 and the step is 2, and the combination of the dense connection module and the transition module is operated for three times.

[0009] An electronic device comprising a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to enable the processor to implement the polar cloud detection method of the FY-3 satellite remote sensing image.

[0010] A computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the polar cloud detection method of the FY-3 satellite remote sensing image.

[0011] The present application has the advantages that: the present application improves the accuracy and automation of polar cloud detection; the present application realizes FY-3D / MERSI-II remote sensing image cloud detection in different seasons and different time periods in the polar region through the Dense-Net cloud detection model based on deep learning, providing another research idea for polar cloud detection; the overall cloud detection precision is high, and the accuracy of cloud detection in spring, summer, autumn and winter is 84.88%, 91.35%, 80.13%, 86.33%, 91.52%, 83.86%, 83.25%, 90.28%, 79.74% and 80.57%, 87.51%, 79.20% respectively. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The flowchart of the present application is shown in the figure; Figure 2 The sample quantity in the first embodiment of the present application is shown in the figure; Figure 3 The network structure diagram of the Dense-Net cloud detection model based on deep learning in the present application is shown in the figure; Figure 4 The comparison chart of the polar cloud detection results of the first embodiment of the present application is shown in the figure; Figure 5 The polar cloud detection precision evaluation index chart of the first embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0013] As shown in the figure, Figure 1 The polar cloud detection method of the FY-3 satellite remote sensing image comprises the following steps: S1. Obtain FY-3D / MERSI-II data in the polar region all year round and all day long and perform preprocessing; S1.1. Obtain FY-3D / MERSI-II Level 1 L1 1000m resolution data of the Arctic region throughout the year and throughout the day. FY-3D / MERSI-II L1 data includes earth observation data and geographic positioning GEO data. FY-3D / MERSI-II L1 earth observation data files store earth observation data after radiation calibration preprocessing, and GEO data files store geographic positioning data after geographic positioning preprocessing, which is used to assist FY-3D / MERSI-II L1 earth observation data. The GEO data with a spatial resolution of 1000m is used to resample the FY-3D / MERSI-II L1 earth observation data of the Arctic region to 1000m, and the projection mode is polar projection; S1.2. Obtain FY-3D / MERSI-II CLM products of the Arctic region throughout the year and throughout the day. FY-3D / MERSI-II is equipped with 25 channels, and the spectral coverage ranges from 0.4μm to 12.0μm. FY-3D / MERSI-II CLM product is a cloud detection result obtained based on a multi-feature threshold method using the spectral reflectance or brightness temperature of 9 channels of FY-3D / MERSI-II, including channel 3 (0.65μm), channel 4 (0.865μm), channel 6 (1.64μm), channel 7 (2.13μm), channel 19 (1.03μm), channel 20 (3.8μm), channel 21 (4.050μm), channel 24 (10.8μm) and channel 25 (12.0μm). The content includes cloud and clear sky identification markers and reliability information. The certain clear sky and possible clear sky in the content of FY-3D / MERSI-II CLM product are assigned a value of 0, and the rest are assigned a value of 1 to generate a cloud mask image of the Arctic region; S2. Select the characteristic factor composition of the Arctic cloud detection method for FY-3D / MERSI-II multi-spectral remote sensing image, including the Arctic cloud detection image obtained by the cloud index method based on traditional threshold; S2.1. Considering the consistency of the data, first, select the spectral reflectance or brightness temperature of the same 9 channels as the characteristic factor as the FY-3D / MERSI-II CLM product generated by the 9 channels. In addition, channel 5 (1.38μm) of FY-3D / MERSI-II belongs to a strong water vapor absorption band. Water vapor has strong absorption ability to solar radiation, and the radiation from the ground to the lower part of the troposphere is difficult to reach the optical satellite sensor, resulting in low reflectivity. Cirrus clouds are mostly located in the upper troposphere, and most of the water vapor in the atmosphere is concentrated below the cirrus cloud. Relatively speaking, the solar radiation intensity above the cirrus cloud is strong, and the reflectivity is high. The spectral reflectance of channel 5 (1.38μm) can be used to detect cirrus clouds. Therefore, the spectral reflectance of FY-3D / MERSI-II channel 5 (1.38μm) is selected as the characteristic factor; S2.2. The Arctic cloud detection images obtained by the two traditional threshold-based cloud index methods can also reflect the distribution of clouds in the Arctic region; cloud index 1 C CI1 The obtained Arctic cloud detection image is taken as a characteristic factor, and the calculation formula is: ; In the formula, C CI1 is cloud index 1; R1 is the spectral reflectance of FY-3D / MERSI-II band 1 (0.47 μm); R2 is the spectral reflectance of FY-3D / MERSI-II band 2 (0.55 μm); R3 is the spectral reflectance of FY-3D / MERSI-II band 3 (0.65 μm); R4 is the spectral reflectance of FY-3D / MERSI-II band 4 (0.865 μm); and R6 is the spectral reflectance of FY-3D / MERSI-II band 6 (1.64 μm); When |C CI1 -1| < 0.25, the pixel is identified as a cloud pixel, and a cloud detection image is obtained. S2.3. Cloud index 2 C CI2 is selected, and the calculation formula is: ; In the formula, C CI2 is cloud index 2; R3 is the spectral reflectance of FY-3D / MERSI-II band 3 (0.65 μm); and R6 is the spectral reflectance of FY-3D / MERSI-II band 6 (1.64 μm); When 0 < C CI2 < 0.4 and R3 > 0.2 are simultaneously satisfied, the pixel is identified as a cloud pixel, and a cloud detection image is obtained. S2.4. The spectral reflectance images or brightness temperature images of FY-3D / MERSI-II bands 3 (0.65 μm), 4 (0.865 μm), 6 (1.64 μm), 7 (2.13 μm), 19 (1.03 μm), 20 (3.8 μm), 21 (4.050 μm), 24 (10.8 μm), and 25 (12.0 μm), the spectral reflectance image of band 5 (1.38 μm), and cloud index 1 C CI1 and cloud index 2 C CI2 are obtained, and the obtained Arctic cloud detection images are combined into a multi-spectral remote sensing image. The cloud index method is a cloud detection method based on traditional threshold values, and has high detection efficiency and is relatively easy to implement. The cloud index method comprehensively considers the spectral reflectance characteristics of clouds and underlying surfaces in visible light bands, near-infrared bands, and short-wave infrared bands, and finds the best band combination for detecting clouds. Clouds have strong spectral reflectivity in both the visible and infrared bands. In the visible band, clouds generally exhibit higher reflectivity, while in the infrared band, the temperature characteristics of clouds are more obvious. The cloud index 1 C is constructed using the ratio of the infrared band to the visible band. CI1 The method is used to measure the similarity of reflectance characteristics in the visible light band and the infrared band, which can effectively indicate cloud cover.

[0014] Cloud Index 2 C CI2 The method uses visible light band and shortwave infrared band to distinguish clouds from underlying surface, which can highlight cloud information and facilitate cloud detection. The purpose of the cloud index in the present invention is to obtain a cloud detection image based on the calculated cloud index and its threshold value, which serves as one of the characteristic factors of the cloud detection model; S3. Divide the FY-3D / MERSI-II multispectral remote sensing images of the Arctic region throughout the year into data for different seasons and time periods; S3.1. Define the multispectral remote sensing images of March, April, and May as spring data, the multispectral remote sensing images of June, July, and August as summer data, the multispectral remote sensing images of September, October, and November as autumn data, and the multispectral remote sensing images of December, January, and February as winter data. S3.2. Based on the solar zenith angle of 85°, data with a solar zenith angle ≤ 85° are classified as daytime data, and data with other solar zenith angles are classified as nighttime data. Data with both daytime and nighttime imaging are classified as nighttime data. FY-3D / MERSI-II multispectral remote sensing images of the Arctic region during all daytime and in different seasons are classified as daytime and nighttime data. The present invention constructs cloud detection models for all day, daytime, and nighttime in different seasons, totaling 12 models. FY-3D is a solar polar orbit satellite with a scanning time resolution of 5 minutes in the same orbit and 85 minutes in different orbits. The time resolution of the band spectral reflectance or brightness temperature obtained by FY-3D is uneven.

[0015] S4. Build a Dense-Net cloud detection model based on deep learning methods to obtain cloud detection results in different seasons and time periods in the Arctic region; S4.1. Due to the unique geographical location of the Arctic region, the region presents a significant phenomenon of polar day and night, and the spectral reflectance of the visible light band of the optical satellite sensor cannot be normally used at night; in order to study the differences in the accuracy of cloud detection methods of satellite remote sensing images in different seasons and different time periods in the Arctic region, different seasons and different time periods of FY-3D / MERSI-Ⅱ multispectral remote sensing images in the Arctic region are used to construct different seasonal all-day, daytime and nighttime cloud detection models, wherein 70% of the time of the FY-3D / MERSI-Ⅱ multispectral remote sensing images in different seasons and different time periods in the Arctic region are selected to constitute the training set samples of the cloud detection model, and 30% of the time is selected to constitute the test set samples; S4.2. In the training of the Dense-Net cloud detection model based on deep learning, the Mean Square Error, MSE loss function and Adam optimizer are used; The mean square error represents the average value of the square sum of the difference between the detected value and the true value, and the calculation formula is as follows: ; In the formula, MSE is the abbreviation of Mean Square Error, which represents the mean square error; x i is the detected value; y i is the true value; N is the number of training samples; The Adam optimizer updates the neuron weight parameters in the iterative deep learning model according to the model network back propagation and gradient information, and updates the neuron weight parameters towards the direction of minimizing the loss function; the Adam optimizer is an adaptive learning rate optimization algorithm, which can dynamically adjust the learning rate of the deep learning model, and is suitable for processing large-scale data sets and complex model network structure problems; S4.3. In the encoding process of the Dense-Net cloud detection model based on deep learning, the dense connection module and the transition module are introduced based on the U-Net model based on deep learning to construct the Arctic remote sensing image cloud detection model Dense-Net, and the remote sensing image feature extraction is carried out; The basic structure of the U-Net model is composed of an input layer, a convolution layer, a pooling layer, an activation function layer, a full connection layer and an output layer; The convolution layer is used to extract multi-level features in the remote sensing image, each element of the convolution kernel slides in the form of a sliding window, and the feature image of the remote sensing image is obtained by convolution operation Conv between the convolution kernel and the local region of the input remote sensing image, and the operation formula of the convolution operation is as follows: ; In the formula: K represents the convolution kernel, I(m, n) represents the input image matrix, and H(i, j) represents the output image matrix; Two key parameters in the convolution operation are the size and number of the convolution kernel; the size of the convolution kernel is usually 3x3, and the number of the convolution kernel can be adjusted according to the design and requirements of the network structure, and the number of the convolution kernel determines the channel number of the output cloud detection image; the size of the output cloud detection image after the convolution operation of the input remote sensing image is calculated according to the following formula: ; In the formula: N represents the output image size, W represents the input image size, K represents the convolution kernel size, P is the padding, and S is the step; The pooling layer is usually connected after the convolution layer, and through sparse processing of the feature vector, the spatial dimension of the feature image is reduced, the output of the convolution layer is simplified, and the calculation efficiency and generalization performance are improved; common pooling operations include Max Pooling and Average Pooling, Max Pooling emphasizes the most significant feature vector in the pooling window, which is beneficial to retain more useful information, while Average Pooling considers the average value of all feature vectors in the pooling window, which is beneficial to smooth the feature vector and reduce noise interference; Unlike the convolution layer and the pooling layer, the activation function as a nonlinear operation can make the U-Net model learn more complex feature vectors and improve the nonlinear mapping ability of the model; according to the different network structures of the model and classification targets, the commonly used activation functions mainly include ReLU activation function and Sigmoid activation function, and the specific calculation formulas are as follows: ; ; In the formula, x represents the feature vector input into the activation function; The fully connected layer is used to integrate the feature vectors extracted by the forward network layer, which is equivalent to linear processing of the multi-dimensional information of the previous network layer; The U-Net model realizes pixel-level classification and detection through its relatively symmetrical encoder-decoder architecture and skip connection; among them, the encoder gradually extracts high-level semantic features of the remote sensing image through convolution operation and pooling operation, while the decoder maps the feature vector back to the cloud detection image of the original resolution through up-sampling operation, and the up-sampling operation is equivalent to the transpose operation of the convolution operation, and the calculation formula is as follows: ; Among them, K represents the convolution kernel, I(m, n) represents the input image matrix of the up-sampling operation, and H(i, j) represents the output image matrix of the up-sampling operation; S4.3.1. Introducing a dense connection module, the dense connection module is composed of three layers of dense connection operations, the convolution kernel size of the dense connection operation is 3x3, the step is 1, and a ReLU activation function is connected after each convolution layer; The dense connection module is composed of a plurality of dense connection operations, the input of each layer in the dense connection operation includes the output of all previous layers, each layer directly accesses the feature vectors of all previous layers, and the operation process is as follows: ; In the formula, x i represents the feature vector of the i-th layer, H L (•) is defined as a composite operation of convolution operation and activation function, a plurality of feature vectors are connected to form a tensor for dense connection operation; S4.3.2. Introducing a transition module, the transition module is composed of a convolution operation with a convolution kernel size of 1x1 and a step of 1 and an average pooling operation, the pooling window is 2x2, the step is 2, and the combination of the dense connection module and the transition module is operated three times in the encoding process. S4.4. In the decoding process of the Dense-Net cloud detection model based on deep learning, after the combination operation of the dense connection module and the transition module three times, the training set sample is restored to the original size through three upsampling operations and the corresponding jump connection in the encoding process, and the cloud detection result is obtained.

[0016] An electronic device comprising a memory and a processor, the memory storing a computer program, and the computer program being executed by the processor to enable the processor to implement the polar cloud detection method of the FY-3 satellite remote sensing image.

[0017] A computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the polar cloud detection method of the FY-3 satellite remote sensing image.

[0018] Embodiment one: According to the FY-3D / MERSI-II data of the polar region in 2022, the polar cloud detection method of the FY-3 satellite remote sensing image is run.

[0019] S1. Obtain FY-3D / MERSI-II data of the polar region in 2022 and perform preprocessing; S1.1. Obtain FY-3D / MERSI-II Level 1 L1 1000m resolution data of the Arctic region in 2022 all day all year round. FY-3D / MERSI-II L1 data includes earth observation data and geographic positioning GEO data. FY-3D / MERSI-II L1 earth observation data file stores earth observation data after radiation calibration preprocessing, and GEO data file stores geographic positioning data after geographic positioning preprocessing, which is used to assist FY-3D / MERSI-II L1 earth observation data. The GEO data with a spatial resolution of 1000m is used to resample the FY-3D / MERSI-II L1 earth observation data of the Arctic region to 1000m, and the projection mode is polar projection; S1.2. Obtain FY-3D / MERSI-II CLM products of the Arctic region all day all year round; FY-3D / MERSI-II is equipped with 25 channels, and the spectral coverage range is 0.4-12.0 μm. FY-3D / MERSI-II CLM product is a cloud detection result obtained based on multi-feature threshold method using 9 channels of spectral reflectance or brightness temperature of FY-3D / MERSI-II, including channel 3 (0.65 μm), channel 4 (0.865 μm), channel 6 (1.64 μm), channel 7 (2.13 μm), channel 19 (1.03 μm), channel 20 (3.8 μm), channel 21 (4.050 μm), channel 24 (10.8 μm) and channel 25 (12.0 μm). The content includes cloud and clear sky identification mark and confidence information; the certain clear sky and possible clear sky in the content of FY-3D / MERSI-II CLM product are assigned as 0, and the rest is assigned as 1, to generate a cloud mask image of the Arctic region; S2. Select the characteristic factor composition of the Arctic cloud detection method for FY-3D / MERSI-II multispectral remote sensing image, including the Arctic cloud detection image obtained by the cloud index method based on traditional threshold; S2.1. Considering the consistency of the data, first, select the spectral reflectance or brightness temperature of the same 9 channels as the characteristic factor as the characteristic factor as the characteristic factor of the FY-3D / MERSI-II CLM product generated by 9 channels; In addition, FY-3D / MERSI-II channel 5 (1.38 μm) belongs to a strong water vapor absorption band, and the water vapor has strong absorption capacity for solar radiation. The radiation from the ground to the lower part of the troposphere is difficult to reach the optical satellite sensor, resulting in low reflectivity; Cirrus clouds are mostly located in the upper troposphere, and most of the water vapor in the atmosphere is concentrated below the cirrus cloud. Relatively speaking, the solar radiation intensity above the cirrus cloud is strong, and the reflectivity is high. The spectral reflectance of channel 5 (1.38 μm) can be used to detect cirrus clouds; therefore, the spectral reflectance of FY-3D / MERSI-II channel 5 (1.38 μm) is selected as the characteristic factor; S2.2. The Arctic cloud detection images obtained by the two traditional threshold-based cloud index methods can also reflect the distribution of clouds in the Arctic region; select cloud index 1 C CI1 The obtained Arctic cloud detection image is taken as a characteristic factor, and the calculation formula is: ; In the formula, C CI1 is the cloud index 1; R1 is the spectral reflectance of FY-3D / MERSI-II band 1 (0.47 μm); R2 is the spectral reflectance of FY-3D / MERSI-II band 2 (0.55 μm); R3 is the spectral reflectance of FY-3D / MERSI-II band 3 (0.65 μm); R4 is the spectral reflectance of FY-3D / MERSI-II band 4 (0.865 μm); and R6 is the spectral reflectance of FY-3D / MERSI-II band 6 (1.64 μm); When |C CI1 -1| < 0.25, the pixel is identified as a cloud pixel, and a cloud detection image is obtained. S2.3. Select cloud index 2 C CI2 The obtained Arctic cloud detection image is taken as a characteristic factor, and the calculation formula is: ; In the formula, C CI2 is the cloud index 2; R3 is the spectral reflectance of FY-3D / MERSI-II band 3 (0.65 μm); and R6 is the spectral reflectance of FY-3D / MERSI-II band 6 (1.64 μm); When 0 < C CI2 < 0.4 and R3 > 0.2 at the same time, the pixel is identified as a cloud pixel, and a cloud detection image is obtained. S2.4. The spectral reflectance images or brightness temperature images of FY-3D / MERSI-II bands 3 (0.65 μm), 4 (0.865 μm), 6 (1.64 μm), 7 (2.13 μm), 19 (1.03 μm), 20 (3.8 μm), 21 (4.050 μm), 24 (10.8 μm) and 25 (12.0 μm), the spectral reflectance image of band 5 (1.38 μm), and C CI1 and C CI2 The obtained Arctic cloud detection images are merged into a multi-spectral remote sensing image. S3. The FY-3D / MERSI-II multi-spectral remote sensing images of the Arctic region throughout the year and at all times are divided into data of different seasons and different time periods. S3.1. Define the multispectral remote sensing images of March, April and May as spring data, the multispectral remote sensing images of June, July and August as summer data, the multispectral remote sensing images of September, October and November as autumn data, and the multispectral remote sensing images of December, January and February as winter data; S3.2. According to the solar zenith angle of 85° as the division standard, the solar zenith angle ≤ 85° is divided into daytime, and the rest is divided into night. For data with both daytime and nighttime imaging, it is classified as nighttime data. The FY-3D / MERSI-II multispectral remote sensing images of the Arctic region in different seasons and at all times are divided into daytime and nighttime data; S4. Construct a Dense-Net cloud detection model based on deep learning method to obtain the cloud detection results of the Arctic region in different seasons and at different times; S4.1. Since the Arctic region is located at the northernmost end of the earth, its unique geographical location makes the region exhibit significant polar day and polar night phenomena, and in the night, the spectral reflectance of the visible light band of the optical satellite sensor cannot be normally used. In order to study the accuracy differences of satellite remote sensing image cloud detection methods in different seasons and at different times in the Arctic region, FY-3D / MERSI-II multispectral remote sensing images of the Arctic region in different seasons and at different times are used to construct daytime, nighttime and daytime cloud detection models in different seasons, respectively. Among them, 70% of the time of FY-3D / MERSI-II multispectral remote sensing images of the Arctic region in different seasons and at different times are selected to constitute the training set samples of the cloud detection model, and 30% of the time is selected to constitute the test set samples; Figure 2 The number of FY-3D / MERSI-II multispectral remote sensing image dataset samples is intuitively shown when constructing a Dense-Net cloud detection model based on deep learning; S4.2. In the training of the Dense-Net cloud detection model based on deep learning, the Mean Square Error, MSE loss function and Adam optimizer are used; The mean square error represents the average value of the square sum of the difference between the detection value and the true value, and the calculation formula is as follows: ; In the formula, MSE is the abbreviation of Mean Square Error, which represents the mean square error; x i is the detection value; y i is the true value; N is the number of training samples; The Adam optimizer updates the neuron weight parameters in the iterative deep learning model according to the model network back propagation and gradient information, and updates the neuron weight parameters towards the direction of minimizing the loss function; the Adam optimizer is an adaptive learning rate optimization algorithm, which can dynamically adjust the learning rate of the deep learning model, and is suitable for processing large-scale data sets and complex model network structure problems; In the encoding process of the deep learning-based Dense-Net cloud detection model, on the basis of the deep learning-based U-Net model, a dense connection module and a transition module are introduced to construct an Arctic remote sensing image cloud detection model Dense-Net for remote sensing image feature extraction; The basic structure of the U-Net model is composed of an input layer, a convolutional layer, a pooling layer, an activation function layer, a full connection layer and an output layer; The convolutional layer is used to extract multi-level features in the remote sensing image, each element of the convolutional kernel slides in the form of a sliding window, and the feature image of the remote sensing image is obtained by convolution operation Conv between the convolutional kernel and the local region of the input remote sensing image, and the convolution operation formula is as follows: ; In the formula, K represents the convolutional kernel, I(m, n) represents the input image matrix, and H(i, j) represents the output image matrix; The two key parameters in the convolution operation are the size and number of the convolutional kernel; the size of the convolutional kernel is usually 3x3, and the number of the convolutional kernel can be adjusted according to the design and requirements of the network structure, and the number of the convolutional kernel determines the channel number of the output cloud detection image; the size calculation formula of the output cloud detection image after the input remote sensing image is subjected to the convolution operation is as follows: ; In the formula, N represents the output image size, W represents the input image size, K represents the convolutional kernel size, P is the padding, and S is the step; The pooling layer is usually connected after the convolutional layer, and the spatial dimension of the feature image is reduced by sparse processing of the feature vector, which simplifies the output of the convolutional layer, thereby improving the calculation efficiency and generalization performance; common pooling operations include maximum pooling Max Pooling and average pooling Average Pooling, maximum pooling emphasizes the most significant feature vector in the pooling window, which is beneficial to retaining more useful information, while average pooling considers the average value of all feature vectors in the pooling window, which is beneficial to smoothing the feature vector and reducing noise interference; Different from the convolutional layer and the pooling layer, the activation function as a nonlinear operation can make the U-Net model learn more complex feature vectors and improve the nonlinear mapping ability of the model; according to the different network structures and classification targets of the model, the commonly used activation functions mainly include the ReLU activation function and the Sigmoid activation function, and the specific calculation formulas are as follows: ; ; In the formula, x represents the feature vector input to the activation function; The fully connected layer is used to integrate the feature vectors extracted by the forward network layer, which is equivalent to linear processing of the multi-dimensional information of the previous network layer; The U-Net model realizes pixel-level classification and detection through its relatively symmetrical encoder-decoder architecture and skip connection; among them, the encoder extracts high-level semantic features of remote sensing images through convolution operation and pooling operation, while the decoder maps the feature vectors back to the original resolution cloud detection image through up-sampling operation, and the up-sampling operation is equivalent to the transposed operation of the convolution operation, and the calculation formula is as follows: ; Among them, K represents the convolution kernel, I(m, n) represents the input image matrix of the up-sampling operation, and H(i, j) represents the output image matrix of the up-sampling operation; S4.3.1. Introduce the dense connection module, which is composed of three layers of dense connection operations, the convolution kernel size of the dense connection operation is 3x3, the step is 1, and the ReLU activation function is connected after each convolution layer; The dense connection module is composed of multiple dense connection operations, and the input of each layer in the dense connection operation includes the output of all previous layers. Each layer directly accesses the feature vectors of all previous layers, and the operation process is as follows: ; In the formula, x i represents the feature vector of the i-th layer, and H L (•) is defined as the composite operation of convolution operation and activation function, which connects multiple feature vectors to form a tensor for dense connection operation; S4.3.2. Introduce the transition module, which is composed of convolution operation with convolution kernel size of 1x1 and step of 1 and average pooling operation, and the pooling window is 2x2 and the step is 2. The combination of the dense connection module and the transition module is operated three times in the encoding process; S4.4. In the decoding process of the Dense-Net cloud detection model based on deep learning, after the combination operation of three dense connection modules and transition modules, the training set samples are restored to the original size through three upsampling operations and the corresponding jump connection in the encoding process, and the cloud detection result is obtained; Figure 3 The network structure diagram of the Dense-Net cloud detection model based on deep learning is shown.

[0020] The cloud detection result of Example 1 is evaluated for precision, Figure 4 The comparison diagram of the cloud detection result of the Arctic region in different seasons in 2022 and the FY-3D / MERSI-Ⅱ RGB true color image is shown.

[0021] Through the FY-3 satellite remote sensing image Arctic cloud detection method, the approximate outline and coverage range of the cloud can be accurately identified, and the cloud and clear sky classification effect is good.

[0022] The FY-3 satellite remote sensing image Arctic cloud detection method can identify the cloud in the Arctic region in different seasons. In summer, due to the melting of ice and snow in the Arctic region, the surface reflectivity is reduced, and the spectral feature difference between the cloud and the underlying surface of ice and snow is more obvious. The detection effect of the Dense-Net cloud detection model based on deep learning is good, and most of the cloud pixels can be identified. In winter, the Dense-Net cloud detection model based on deep learning has a small amount of false detection. In the transition season of spring and autumn, the Dense-Net cloud detection model based on deep learning shows good stability, can adapt to the change of surface reflectivity and illumination condition, and better describes the outline and coverage range of the cloud.

[0023] The Dense-Net cloud detection model based on deep learning can also identify the cloud in the Arctic region at different times. The cloud detection effect is good in the daytime. Compared with the all-day and daytime cloud detection effect, the night cloud detection effect is poor, which may be due to the fact that the FY-3 satellite remote sensing image Arctic cloud detection method cannot fully use the visible light band, resulting in the loss of feature information of the FY-3D / MERSI-Ⅱ multi-spectral remote sensing image in the Arctic region at night.

[0024] The accuracy rate ( A Acc ), precision rate ( P Pre ), recall rate ( R Rec ) and F1 score ( F F1scoreThe four precision evaluation indexes are used to quantitatively evaluate the Arctic cloud detection method. Among them, the accuracy, precision, recall and F1 score represent the detection accuracy. The higher the accuracy, precision, recall and F1 score values are, the higher the algorithm accuracy is. The calculation formulas are as follows: Table 1 Precision evaluation index calculation matrix: .

[0025] ; ; ; ; In the formula, T P represents the number of pixels whose detection result and true value label are both cloud; T N represents the number of pixels whose detection result and true value label are both non-cloud; F P represents the number of misdetected pixels, i.e. the number of pixels whose detection result is cloud while the true value label is non-cloud; conversely, F N represents the number of missed pixels, i.e. the number of pixels whose detection result is non-cloud while the true value label is cloud.

[0026] Figure 5 The precision evaluation indexes of cloud detection in different seasons of the Arctic region in 2022 are shown.

[0027] The accuracy of cloud detection in the Arctic region in spring, summer, autumn and winter is 84.88%, 91.35% and 80.13% for all-day, daytime and nighttime, respectively; 86.33%, 91.52% and 83.86% for all-day, daytime and nighttime, respectively; 83.25%, 90.28% and 79.74% for all-day, daytime and nighttime, respectively; and 80.57%, 87.51% and 79.20% for all-day, daytime and nighttime, respectively. The Dense-Net cloud detection model based on deep learning can automatically learn the complex features of clouds without the need for manual setting of complex threshold rules, and has stronger adaptability and generalization ability, so it has good cloud detection effect in the Arctic region.

Claims

1. A method for detecting Arctic clouds using Fengyun-3 satellite remote sensing images, characterized by: The following steps are included: S1. Acquire and preprocess FY-3D / MERSI-II data for the Arctic region throughout the year and all day long. S1.

1. Acquire FY-3D / MERSI-II Level 1 L1 1000m resolution data for the Arctic region throughout the entire day and year. FY-3D / MERSI-II L1 data includes Earth observation data and geolocated GEO data. The FY-3D / MERSI-II L1 Earth observation data files store Earth observation data that has undergone radiometric calibration preprocessing, while the GEO data files store geolocated data that has undergone geolocation preprocessing. These files assist in the use of FY-3D / MERSI-II L1 Earth observation data. The FY-3D / MERSI-II L1 Earth observation data for the Arctic region are resampled to 1000m spatial resolution using 1000m spatial resolution GEO data, using a polar projection. S1.

2. Obtain FY-3D / MERSI-II CLM products for the Arctic region year-round and all day long. FY-3D / MERSI-II is equipped with 25 channels, covering the spectral range of 0.4 μm to 12.0 μm. The FY-3D / MERSI-II CLM product uses a multi-feature threshold method to generate cloud detection results using the spectral reflectance or brightness temperature of nine FY-3D / MERSI-II channels: Bands 3 (0.65μm), 4 (0.865μm), 6 (1.64μm), 7 (2.13μm), 19 (1.03μm), 20 (3.8μm), 21 (4.050μm), 24 (10.8μm), and 25 (12.0μm). This product includes cloud and clear sky identification features and reliability information. Confirmed clear and possible clear sky values ​​in the FY-3D / MERSI-II CLM product are assigned a value of 0, while all other conditions are assigned a value of 1. This creates a cloud mask image for the Arctic region. S2. Select characteristic factors of Arctic cloud detection methods and synthesize them into FY-3D / MERSI-II multispectral remote sensing images, including Arctic cloud detection images obtained by the traditional threshold-based cloud index method; S2.

1. Considering data consistency, the spectral reflectance or brightness temperature of the same nine channels used in the FY-3D / MERSI-II CLM product are selected as characteristic factors. Furthermore, FY-3D / MERSI-II Band 5 (1.38 μm) is a band with strong water vapor absorption. Water vapor strongly absorbs solar radiation, making it difficult for radiation from the ground to the middle and lower troposphere to reach the optical satellite sensor, resulting in low reflectance. Cirrus clouds are mostly located in the upper troposphere, and most of the water vapor in the atmosphere is concentrated below them. Relatively speaking, the solar radiation intensity above cirrus clouds is strong and the reflectivity is high. Therefore, the spectral reflectance of Band 5 (1.38 μm) can be used to detect cirrus clouds. Therefore, the spectral reflectance of FY-3D / MERSI-Ⅱ Band 5 (1.38 μm) is selected as the characteristic factor. S2.

2. The Arctic cloud detection images obtained by two cloud index methods based on traditional thresholds can also reflect the distribution of clouds in the Arctic region; the cloud index 1 C CI1 The obtained Arctic cloud detection image is used as the characteristic factor, and the calculation formula is: ； In the formula: C CI1 is the cloud index 1; R1 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 1 (0.47μm); R2 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 2 (0.55μm); R3 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 3 (0.65μm); R4 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 4 (0.865μm); R6 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 6 (1.64μm); When | C CI1 When -1|<0.25, the pixel is identified as a cloud pixel and a cloud detection image is obtained; S2.

3. Select Cloud Index 2 C CI2 The obtained Arctic cloud detection image is used as the characteristic factor, and the calculation formula is: ； In the formula: C CI2 is the cloud index 2; R3 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 3 (0.65μm); R6 is the spectral reflectance of FY-3D / MERSI-Ⅱ band 6 (1.64μm); When 0<C CI2 When R<0.4 and R3>0.2, it is identified as a cloud pixel and a cloud detection image is obtained; S2.

4. Use the spectral reflectance or brightness temperature images of FY-3D / MERSI-Ⅱ bands 3 (0.65μm), 4 (0.865μm), 6 (1.64μm), 7 (2.13μm), 19 (1.03μm), 20 (3.8μm), 21 (4.050μm), 24 (10.8μm), and 25 (12.0μm), the spectral reflectance image of band 5 (1.38μm), and C CI1 and C CI2 The obtained Arctic cloud detection images are merged into multispectral remote sensing images; S3. Divide the FY-3D / MERSI-II multispectral remote sensing images of the Arctic region throughout the year into data for different seasons and time periods; S4. Build a Dense-Net cloud detection model based on deep learning methods to obtain cloud detection results in different seasons and time periods in the Arctic region; S4.

1. Due to its unique geographical location at the northernmost tip of the Earth, the Arctic region exhibits significant polar day and polar night phenomena. Furthermore, at night, the spectral reflectance of the visible light band of optical satellite sensors cannot be used properly. To investigate the differences in the accuracy of cloud detection methods using satellite remote sensing imagery in different seasons and time periods in the Arctic, FY-3D / MERSI-II multispectral remote sensing imagery from different seasons and time periods was used to construct seasonal all-day, daytime, and nighttime cloud detection models. 70% of the FY-3D / MERSI-II multispectral remote sensing imagery from different seasons and time periods formed the training set samples for the cloud detection model, and 30% of the FY-3D / MERSI-II multispectral remote sensing imagery from different seasons and time periods formed the test set samples. S4.

2. Use the root mean square error (MSE) loss function and the Adam optimizer in training the deep learning-based Dense-Net cloud detection model. The root mean square error represents the average of the sum of squares of the differences between the detected value and the true value, and is calculated as follows: ； In the formula: MSE is the abbreviation of Mean Square Error, which stands for root mean square error; x i is the detection value; y i is the true value; N is the number of training samples; The Adam optimizer updates the neuron weight parameters in the iterative deep learning model based on the model network backpropagation and gradient information, and updates the neuron weight parameters in the direction of minimizing the loss function; S4.

3. During the encoding process of the deep learning-based Dense-Net cloud detection model, a dense connection module and a transition module were introduced to build a Dense-Net cloud detection model for Arctic remote sensing imagery based on the deep learning-based U-Net model for remote sensing image feature extraction. The basic structure of the U-Net model consists of an input layer, a convolutional layer, a pooling layer, an activation function layer, a fully connected layer, and an output layer; The convolution layer is used to extract multi-level features from remote sensing images. Each element of the convolution kernel slides in the form of a sliding window. The convolution operation Conv is performed on the local area of ​​the input remote sensing image to complete the feature image acquisition of the remote sensing image. The convolution operation formula is as follows: ； In the formula: K represents the convolution kernel, I(m,n) represents the input image matrix, and H(i,j) represents the output image matrix; The two key parameters in the convolution operation are the size and number of convolution kernels. The convolution kernel size is usually 3×3. The number of convolution kernels can be adjusted according to the design and requirements of the network structure. The number of convolution kernels determines the number of channels in the output cloud detection image. The formula for calculating the size of the output cloud detection image after the input remote sensing image is convolved is as follows: ； In the formula: N represents the output image size, W represents the input image size, K represents the convolution kernel size, P is the padding, and S is the step size; The pooling layer is usually connected after the convolution layer. It performs sparse processing on the feature vectors, reduces the spatial dimension of the feature image, simplifies the output of the convolution layer, and thus improves computational efficiency and generalization performance. Common pooling operations include Max Pooling and Average Pooling. Max Pooling emphasizes the most significant feature vectors within the pooling window, which helps retain more useful information, while Average Pooling considers the average of all feature vectors within the pooling window, which helps smooth the feature vectors and reduce noise interference. Unlike convolutional layers and pooling layers, activation functions, as a nonlinear operation, enable the U-Net model to learn more complex feature vectors and improve the model's nonlinear mapping capabilities. Depending on the model network structure and classification objectives, commonly used activation functions include ReLU activation function and Sigmoid activation function. The specific calculation formula is as follows: ； ； In the formula, x represents the feature vector of the input activation function; The fully connected layer is used to integrate the feature vectors extracted by the forward network layer, which is equivalent to linear processing of the multi-dimensional information of the previous network layer; The U-Net model achieves pixel-level classification and detection through its relatively symmetrical encoder-decoder architecture and skip connections. The encoder gradually extracts high-level semantic features of remote sensing images through convolution and pooling operations, while the decoder maps the feature vector back to the original resolution cloud detection image through upsampling operations. The upsampling operation is equivalent to the transpose operation of the convolution operation. The calculation formula is as follows: ； Where K represents the convolution kernel, I(m,n) represents the input image matrix of the upsampling operation, and H(i,j) represents the output image matrix of the upsampling operation; S4.

4. During the decoding process of the deep learning-based Dense-Net cloud detection model, after three runs of the dense connection module and the transition module, the training set samples are restored to their original size through three upsampling operations and the corresponding skip connections in the encoding process to obtain the cloud detection results.

2. The method for detecting Arctic clouds using Fengyun-3 satellite remote sensing images according to claim 1, characterized in that: The method for dividing the FY-3D / MERSI-II multispectral remote sensing images of the Arctic region throughout the year and all day into data of different seasons in step S3 is as follows: S3.

1. Multispectral remote sensing images of March, April, and May are defined as spring data; multispectral remote sensing images of June, July, and August are defined as summer data; multispectral remote sensing images of September, October, and November are defined as autumn data; and multispectral remote sensing images of December, January, and February are defined as winter data.

3. The method for detecting Arctic clouds using FY-3 satellite remote sensing images according to claim 2, characterized in that: The method for dividing the FY-3D / MERSI-II multispectral remote sensing images of the Arctic region throughout the year and all day into data of different time periods in step S3 is: S3.

2. Based on the solar zenith angle of 85° as the classification standard, data with a solar zenith angle ≤ 85° are classified as daytime data, and the rest are classified as nighttime data; data with both daytime and nighttime imaging are classified as nighttime data; FY-3D / MERSI-Ⅱ multispectral remote sensing images of the Arctic region throughout the day in different seasons are divided into daytime and nighttime data.

4. The method for detecting Arctic clouds using Fengyun-3 satellite remote sensing images according to claim 3, characterized in that: The step S4.3 also includes the following steps: S4.3.

1. Introduce the dense connection module. The dense connection module consists of three layers of dense connection operations. The convolution kernel size of the dense connection operation is 3×3, the stride is 1, and each convolution layer is connected to the ReLU activation function. The dense connection module consists of multiple dense connection operations. The input of each layer in the dense connection operation includes the output of all previous layers. Each layer directly accesses the feature vectors of all previous layers. The operation process is as follows: ； In the formula, x i represents the feature vector of layer i, H L (•) Defined as a composite operation of convolution and activation function, multiple feature vectors are connected to form a tensor for dense connection operation; S4.3.

2. Introduce the transition module. The transition module consists of a convolution operation with a convolution kernel size of 1×1 and a stride of 1 and an average pooling operation. The pooling window is 2×2 and the stride is 2. The encoding process performs a combination of three dense connection modules and transition modules.

5. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the computer program is executed by the processor, the processor implements the method for detecting Arctic clouds using Fengyun-3 satellite remote sensing images as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting Arctic clouds using Fengyun-3 satellite remote sensing images as described in any one of claims 1 to 4 is implemented.

Images