A cloud-aerosol hierarchical classification method based on semantic segmentation

Abstract

The application discloses a cloud-aerosol hierarchical classification method based on semantic segmentation. The semantic segmentation technology is introduced into the cloud and aerosol hierarchical classification of atmospheric remote sensing, and multiple channel observation results and texture information are combined to continuously classify original signals pixel by pixel. Firstly, the multi-channel attenuated backscatter coefficient of a spaceborne laser radar CALIOP is collected, preprocessed through denoising and credibility screening, and then the volume depolarization ratio and the attenuation color ratio are calculated as RGB three-channel values to form an image to obtain a training sample set, which is input into a constructed semantic segmentation neural network to be trained, so that a cloud-aerosol hierarchical classification model is obtained, and the hierarchical classification of clouds and aerosols in the atmosphere is realized. Compared with the CALIOP hierarchical classification result of the existing method, the method has the advantages of high accuracy, high resolution, rich hierarchy and the like, is suitable for multiple scenes, and is helpful for observing the distribution of clouds and aerosols, predicting global climate change and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of atmospheric environment remote sensing, and particularly relates to a cloud-aerosol hierarchical classification method based on semantic segmentation. BACKGROUND

[0002] Aerosol is an important component in the atmosphere, which is composed of small solid particles (such as dust) and liquid (such as water) suspended in the atmosphere. Atmospheric aerosols can be directly discharged into the atmosphere as particles (such as ash), or can be formed when the discharged gas undergoes complex chemical reactions and condenses into particles, and the particle radius can be as small as a few nanometers or as large as a few tens of microns.

[0003] The source of aerosol is very extensive, which can be further divided into natural aerosol and human-generated aerosol: natural aerosol includes desert dust, volatile organic compounds in vegetation, smoke generated by forest fires and volcanic ash, etc.; human-generated aerosol, i.e. aerosol generated by human activities, includes sulfate, nitrate, black carbon and organic carbon aerosol from the combustion of fossil fuels and the burning of agricultural waste. Although the proportion of aerosol in the whole atmospheric environment is relatively low, it plays a crucial role in human health and climate change research.

[0004] In order to better understand the formation, evolution and transmission of aerosol, and to explore how aerosol regulates the energy balance change of the atmospheric system, it is necessary to accurately grasp its three-dimensional distribution information in real time. Accurate and efficient three-dimensional detection of aerosol not only has great scientific significance for global ecological environment, but also has an important supporting role in improving the accuracy of climate prediction.

[0005] Because aerosol is widely distributed in space, it is necessary to obtain large-scale and high-quality observation data to meet the needs of scientific research. CALIOP, a spaceborne lidar, has the advantages of high temporal and spatial resolution, vertical distribution information, and all-day work, and has obvious advantages in atmospheric aerosol three-dimensional detection compared with passive remote sensing detection methods. Through the detected attenuated backscatter signal, combined with the hierarchical recognition algorithm, the three-dimensional structure information of the atmospheric aerosol can be further obtained.

[0006] The methods for hierarchical classification of aerosol mainly include slope method and threshold method. The slope method requires that the signal be strong enough and continuously increasing (or decreasing) to determine whether the slope is positive or negative. Otherwise, strong instrument noise will adversely affect the detection. Therefore, the slope method is not suitable for spaceborne lidar because of its low signal-to-noise ratio. The threshold method is more widely used in spaceborne lidar, which detects the signal strength in a certain height range and determines the layer type according to a series of thresholds set in advance.

[0007] The invention patent with the authorization announcement number CN 112698354 B discloses a method and system for identifying atmospheric aerosols and clouds, which comprises the following steps: acquiring laser radar original data; data correction; acquiring color ratio and attenuation backscattering coefficient, and comparing them with preset threshold conditions of color ratio and attenuation backscattering coefficient to obtain analysis results of aerosol and cloud types.

[0008] At present, in the prior art, the cloud-aerosol hierarchical classification algorithm mainly performs hierarchical search on a single profile, relies on geographic spatial information and spectral resolution attributes, does not effectively utilize the spatial information within the layer, ignores the connection between adjacent profiles, and thus the continuity and accuracy of the hierarchical classification result are problematic. Therefore, it is necessary to construct a more accurate aerosol hierarchical identification method, which is conducive to constructing an atmospheric model and thus better understanding global climate change. SUMMARY

[0009] In order to solve the problem that the existing hierarchical classification algorithm is mostly directed to a single profile, easily ignores the connection between adjacent profiles, and identifies a certain layer as the same type when identifying the layer, thus easily causing edge misidentification, the present application provides a cloud-aerosol hierarchical classification method based on semantic segmentation, which takes the 532nm attenuation backscattering coefficient, bulk depolarization ratio and attenuation color ratio obtained by CALIOP as input, fully considers the observation data of multiple channels and the texture information of the laser radar image, and thus obtains more continuous and higher-resolution cloud-aerosol hierarchical classification results, which are suitable for various environments.

[0010] A cloud-aerosol hierarchical classification method based on semantic segmentation comprises the following steps:

[0011] (1) acquiring satellite-borne laser radar CALIOP remote sensing observation data, including acquiring level 1 calibrated three-channel attenuation backscattering coefficients, the three channels being a 532nm vertical polarization channel, a 532nm parallel polarization channel and a 1064nm channel, and simultaneously acquiring corresponding level 2 cloud-aerosol hierarchical classification results as subsequent training labels;

[0012] (2) dividing the data into daytime and nighttime data according to the data acquisition time, and removing noise from the daytime high-noise signals;

[0013] (3) screening the data reliability through a cloud-aerosol identification algorithm, separating the clouds and aerosols through a multidimensional histogram of scattering characteristics such as intensity and spectral dependence, and calculating a confidence score according to the overlapping area between the multidimensional histograms to select high-confidence profile samples;

[0014] (4) Obtain the three-channel attenuated backscatter coefficients of the samples after denoising and screening processing, calculate the bulk depolarization ratio and attenuation color ratio, and normalize the total attenuated backscatter coefficient of the 532nm channel, the bulk depolarization ratio and the attenuation color ratio to RGB three channels to form an image, unify the size, and then perform data enhancement on the image to expand the data and divide the data set into training, verification and testing sets;

[0015] (5) Based on the semantic segmentation neural network model, a cloud-aerosol hierarchical classification model is constructed, the training set data is used as input, and the level 2 cloud-aerosol classification result is used as output for model training.

[0016] (6) For the test set data, the model trained in step (5) is used to perform actual cloud-aerosol hierarchical classification, and the results are compared and analyzed with the existing classification results.

[0017] The original multi-channel backscatter signals are combined to form an image, and the cloud-aerosol hierarchical classification is realized based on the semantic segmentation network, which can effectively improve the accuracy and resolution of the classification.

[0018] Preferably, in step (1), the backscatter coefficients of the 532nm and 1064nm channels of the CALIOP level 1 and the calculated bulk depolarization ratio and attenuation color ratio are combined to form an image, and the cloud-aerosol hierarchical classification is performed for each pixel in the image. The bulk depolarization ratio is the ratio of the vertical component to the parallel component of the 532nm attenuated backscatter coefficient, and the smaller the value, the closer the shape of the scattering body to a sphere; the attenuation color ratio is the ratio of the 1064nm attenuated backscatter coefficient to the total attenuated backscatter coefficient of the 532nm channel, and the larger the value, the larger the particle size of the scattering body.

[0019] Preferably, in step (1), the cloud-aerosol hierarchical classification result includes six categories: cloud, aerosol, clear sky, ground, sub-surface and missing signal.

[0020] The signals obtained at different time periods differ greatly. During the day, the signal-to-noise ratio of the data obtained by the satellite sensor is low due to the influence of sunlight noise, while the signal-to-noise ratio of the data collected at night is generally good. If these two types of data are directly mixed to form a data set without preprocessing, the classification effect of the model obtained by subsequent model training will be poor, and a large number of misclassifications will occur. Therefore, in step (2), the data collected at different times need to be classified, and the high-noise attenuated backscatter signals during the day need to be denoised.

[0021] Preferably, in step (2), the BM3D denoising algorithm is used to remove the daytime noise for the original backscatter signal, and an identification algorithm is used to filter the misclassified samples. The BM3D denoising algorithm combines time domain filtering and transform domain filtering methods. First, similar matching is performed on the image blocks in the image to obtain a three-dimensional image block. After 3D transformation, the transform domain matrix is intervened to remove noise components by using hard threshold filtering. Then, an inverse 3D transformation is performed to obtain an evaluation image block. Finally, Wiener filtering is used instead of hard threshold filtering to search for similar patches in the filtered image to obtain the final denoised data.

[0022] The aerosol classification results obtained by the threshold analysis method used by CALIOP officials also have some misclassification cases. Therefore, in step (3), the data needs to be further screened to remove misclassified samples, improve the accuracy of the training set, and ensure that the model trained by the data set can also achieve correct classification.

[0023] Preferably, in step (3), the cloud-aerosol identification CAD algorithm is used to screen the data reliability. The CAD algorithm separates clouds and aerosols by establishing a multi-dimensional histogram of scattering characteristics such as intensity and spectral dependence, and calculates the confidence score according to the overlap area between the multi-dimensional histograms. The score calculation formula is:

[0024]

[0025] where X is an input parameter array containing the backscatter coefficient of the 532 nm channel, the backscatter ratio of 1064 nm and 532 nm, and the altitude, p cloud and p aerosol are the probability density functions of clouds and aerosols, respectively, and k is the probability ratio of the occurrence of clouds and aerosols. When p cloud is greater than kp aerosol , the CAD score is positive, indicating that the probability of cloud characteristics is greater than that of aerosols, and vice versa. The absolute value represents the reliability, and the greater the value, the higher the reliability. The CAD score of each sample is calculated, and the samples with higher reliability are retained.

[0026] Preferably, in step (3), the profile samples with a confidence score absolute value between 70 and 100 are selected.

[0027] In step (4), the volume depolarization ratio and attenuation color ratio are further calculated based on the three-channel attenuated backscatter coefficients of the confidence profile samples selected in step (3), and the obtained values are normalized to ensure consistent ranges. Then, the attenuated backscatter coefficient, the volume depolarization ratio, and the attenuation color ratio are combined as a picture with red, green, and blue channels as intensities. The image contains three kinds of information, which are used as inputs for the semantic segmentation model for hierarchical classification.

[0028] Preferably, the three signals of the attenuation backscattering coefficient, the depolarization ratio and the attenuation color ratio of the 532 nm channel are normalized to the range of 0-1, and the normalization formula is:

[0029]

[0030] wherein x is the input sample data, x min is the maximum value of the sample data, x min is the minimum value of the sample data, x new is the new sample data obtained after normalization.

[0031] The processed samples are still not rich compared with the global, in order to improve the scene richness and sample balance, the image is subjected to data enhancement for data expansion.

[0032] The image data enhancement strategy specifically includes random cropping, rotation, shifting and other processing, which improves the robustness and generalization ability of the model by improving the data volume and data distribution. At the same time, the edges of each image are adjusted to overlap, so as to ensure the continuity of the data. Then the data set is divided for training, verification and testing.

[0033] Preferably, 60% of the data in the data set constitutes a training set for model training; 20% of the data constitutes a validation set for model performance testing and hyperparameter adjustment; and the remaining 20% of the data constitutes a test set for evaluating the actual performance of the model.

[0034] In step (5), a cloud-aerosol hierarchical classification model is constructed based on a semantic segmentation neural network model. Semantic segmentation is an application of deep learning in the field of computer vision, which can extract features in images and realize pixel-level classification.

[0035] Preferably, in step (5), the semantic segmentation neural network model is an encoder-decoder structure, the encoder extracts image features, and the decoder classifies according to the image features.

[0036] Preferably, the encoder is composed of a deep separable convolution network and a dilated convolution pyramid pooling-double channel attention parallel module, the decoder receives the output of the deep separable convolution network and the parallel module, and finally outputs the full image pixel cloud-aerosol prediction classification through multiple full convolution upsampling.

[0037] The input image of the encoder part is subjected to a deep separable convolution network for preliminary information extraction, and then subjected to a dilated convolution pyramid pooling layer for multi-scale deep feature extraction. The input of the decoder part includes two parts, which are the shallow features extracted by the deep separable convolution network and the output of the deep features after upsampling extracted by the pyramid pooling-attention parallel module, and then the prediction classification corresponding to each pixel in the output picture is obtained after multiple rounds of full convolution upsampling.

[0038] Since aerosols and clouds in the atmosphere do not have a fixed shape and the size is not completely uniform, when segmenting them, not only the feature extraction of different scales needs to be considered, but also the context information is very important. It is generally believed that there is a correlation between different feature channels, so a connection needs to be established between multiple feature channels. Then, using this connection, the feature mapping information is highlighted, and the semantic segmentation accuracy is improved.

[0039] Preferably, the double-channel attention parallel module is a position attention and a channel attention module, both of which adaptively combine local features with global features.

[0040] In the position attention module, assuming that a local feature A with a dimension size of CxHxW is input into a convolution layer to obtain two new features B and C, and the dimension of B is converted to CxN through dimension conversion, where N is the number of feature image pixels. The transpose B T and C are subjected to matrix multiplication operation, and the spatial attention feature S can be calculated by applying a softmax layer:

[0041]

[0042] where s ji represents the influence of position i on position j. Further, the new feature D obtained by convolution of the feature A is subjected to matrix multiplication with S, and the dimension is reset to CxHxW. The output E is obtained by adding the original feature A:

[0043]

[0044] where a is a scale parameter, which can be learned in model training. The feature of each position is the weighted sum of the original feature and the global position feature, realizing the aggregation of global semantic information.

[0045] In the channel attention module, the original feature A and the transpose A T are directly subjected to matrix multiplication operation, and then a softmax operation is applied to obtain the channel attention feature X:

[0046]

[0047] x ji is the influence of the i-th channel on the j-th channel, in addition, the transpose of the matrix X and the matrix A are subjected to matrix multiplication operation and the dimension is converted to CxHxW, then a scale parameter β is multiplied by the result, a pixel-level operation is performed, and a final output E is obtained:

[0048]

[0049] β is similar to α, which is a scale parameter of the channel attention module. Both parameters are initialized to 0 and can be learned to appropriate values during model training. As can be seen from the formula, the features of each channel aggregate the weight information of all channel features, which is equivalent to establishing a relationship between different channels. Through learning this relationship, the discriminative ability of the model can be effectively improved.

[0050] Preferably, in step (5), after the cloud-aerosol hierarchical classification model is constructed, the training set data obtained in step (4) is input, and the corresponding level 2 cloud-aerosol classification result is output for model training, and the model effect is optimized through multiple iterations.

[0051] Preferably, in step (5), the model training parameters are: 100 iterations, an initial learning rate of 0.01, a single training batch size of 8, a random gradient descent method, a momentum of 0.9, and a loss function of Focal loss function. The calculation formula of the Focal loss function Focal loss

[0052]

[0053] Where N is all pixels, p t is the actual prediction probability, p t is 1 minus the prediction probability, w t is the class weight, which suppresses the number imbalance problem of different classes, and γ is a hyperparameter, which is increased to reduce the loss of high confidence samples, so that the model training focuses on difficult samples. In the model, the value is set to 2.

[0054] In step (6), the semantic segmentation model trained is applied to the test set image to obtain the predicted cloud-aerosol classification of the pixels, and the official classification result is compared to calculate the average intersection over union and the classification accuracy.

[0055] Compared with the prior art, the present application has the beneficial effects that:

[0056] ​(1) The cloud-aerosol hierarchical classification algorithm based on the semantic segmentation network of the present application fully considers the observation data of multiple channels and the texture information of the laser radar image, and obtains more continuous and higher resolution hierarchical recognition results.

[0057] (2) The cloud-aerosol hierarchical classification algorithm based on the semantic segmentation network of the present application has a classification accuracy of more than 90% for the data categories of clear sky, ground, sub-surface and signal loss, an aerosol classification accuracy of 85%, a cloud classification accuracy of 81%, an average intersection over union of 78%, and an overall accuracy of 96%.

[0058] (3) Compared with the CALIOP hierarchical classification results of the existing method, the method of the present application has the advantages of high accuracy, high resolution and rich hierarchy, and is suitable for multiple scenes, which is helpful for observing the distribution of clouds and aerosols and predicting global climate change. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 FIG. 1 is a flowchart of the cloud-aerosol hierarchical classification method based on semantic segmentation in the present embodiment.

[0060] Figure 2 FIG. 2 is a schematic diagram of the construction of the semantic segmentation neural network model in the present embodiment.

[0061] Figure 3 (a) is an RGB image composed of the data of the backscattering coefficient, the color ratio and the volume depolarization ratio in the present embodiment; Figure 3 (b) is the hierarchical classification estimation result of the semantic segmentation neural network model.

[0062] Figure 4 FIG. 4 is a confusion matrix for comparing the recognition and classification results of the cloud-aerosol in the present embodiment with the official classification results of CALIOP. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0064] The present embodiment provides a neural network-based cloud-aerosol hierarchical classification algorithm, as shown in FIG. 1, which includes six steps, which will be described in detail below for each step: Figure 1

[0065] (1) Profile data acquisition

[0066] ​The CALIOP original data is acquired, and the time range is selected as three months from May 2015 to July 2015, containing 1154 orbit data. The main data products include the attenuation backscatter coefficients of 532nm and 1064nm channels at level 1, wherein the 532nm channel includes vertical and parallel channels, and the level 2 product of hierarchical classification results is downloaded as a prediction label.

[0067] (2) Daytime signal denoising

[0068] Since the influence of solar noise is great during the day, the BM3D denoising algorithm is applied to the attenuation backscatter signal, and a large standard deviation is selected for denoising to better remove noise. The daytime signal after denoising is input into the semantic segmentation network to obtain better hierarchical recognition results.

[0069] (3) Data credibility screening

[0070] The data credibility is screened by the cloud-aerosol hierarchical classification algorithm. The CAD algorithm calculates the CAD score of the feature according to the scattering characteristics such as intensity and spectral dependence to separate the cloud and aerosol. If the overlapping or intersecting area is small, it is determined that the confidence of classifying the feature is high. In order to train the model with high-quality data, the profiles with absolute CAD score between 70 and 100 are selected, and the classification results of these profiles are relatively more accurate and have higher credibility.

[0071] (4) Normalization and data enhancement to construct training data set

[0072] The attenuation backscatter coefficients of the denoised and screened samples are obtained, the depolarization ratio and attenuation color ratio are calculated, and the attenuation backscatter coefficients of the 532nm channel, the depolarization ratio and the attenuation color ratio are normalized to the range of 0-1. The three kinds of data are respectively taken as the values of the RGB three channels of the image, and are adjusted to a uniform size (545 (height pixel size) x 3000 (pixel size along the track) x 3 (spectrum size)). The edges of each adjusted image have a part of the area overlapping to ensure the continuity of the data, and data enhancement is adopted. Finally, 25120 sample image of the training set is obtained, which is used as the input of the model training. In the whole sample, 60% of the data is randomly selected for training the classification model, 20% of the data is used for verification and optimization, and the remaining 20% of the data is used for final model test and evaluation.

[0073] (5) Constructing cloud-aerosol hierarchical classification model for training

[0074] The semantic segmentation neural network model is constructed, and the training set data is taken as input. After training iteration, the model with the best effect on the validation set is obtained. The model training hyperparameters are set as follows: 100 iterations, initial learning rate of 0.01, batch size of 8 for single training, random gradient descent method, momentum of 0.9, and focal loss function.

[0075] Figure 2 A schematic diagram of constructing a semantic segmentation neural network model for the present embodiment is shown. The semantic segmentation neural network model can be divided into an encoder and a decoder. The encoder takes the processed RGB image as input, first extracts features through a depth separable convolutional neural network, and inputs the features to the decoder. Meanwhile, the deep feature information is further extracted through the parallel module of the atrous convolution pyramid pooling and the double-channel attention. After upsampling, the deep feature information is fused with the shallow feature in the full convolution network to restore the boundary information, and finally the high-precision cloud-aerosol hierarchical classification result is output through upsampling.

[0076] (6) Classification result verification and evaluation

[0077] The trained model is used for actual cloud-aerosol hierarchical classification. Figure 3 (a) is an RGB image combined with the data of the attenuation backscatter coefficient, the attenuation color ratio and the bulk depolarization ratio. After processing, there is still noise, and the naked eye cannot directly distinguish the atmospheric classification. Figure 3 (b) is the hierarchical classification estimation result of the semantic segmentation neural network model. Each pixel point in the figure is classified, and the classification result is clear and explicit, continuous in distribution, and has no void inside the hierarchy, with excellent classification effect.

[0078] The classification model is used for cloud-aerosol hierarchical classification of the validation set image, and the classification result is compared with the official classification result of CALIOP to calculate the confusion matrix (such as Figure 4 ), and the classification effect is further analyzed. Figure 4 The diagonal data of the confusion matrix represents the classification accuracy of each category. The larger the value, the greater the predicted classification probability. It can be seen that the classification accuracy of all categories is more than 80%, among which the classification accuracy of the clear sky, ground, sub-surface and signal missing data categories is more than 90%, the aerosol classification accuracy reaches 85%, and the cloud classification accuracy reaches 81%. The main misidentification occurs in identifying clouds and aerosols as clear skies, and the main reason is that the signal-to-noise ratio of the original data is too low, and the classification provided by CALIOP is not completely accurate. The average intersection over union of the classification is calculated to be 78%, and the overall accuracy is 96%, and the hierarchical edge recognition is accurate and continuous, proving that the method can more truly calculate each hierarchical category.

Claims

1. A cloud-aerosol hierarchical classification method based on semantic segmentation, characterized in that, The method comprises the following steps: (1) obtaining CALIOP satellite-borne lidar remote sensing observation data, including obtaining level 1 calibrated three-channel attenuated backscatter coefficients, the three channels being a 532 nm vertical polarization channel, a 532 nm parallel polarization channel and a 1064 nm channel, and simultaneously obtaining corresponding level 2 cloud-aerosol hierarchical classification results as subsequent training labels; (2) dividing the data into daytime and nighttime data according to the data collection time, removing noise from the daytime high-noise signal; for the original backscatter signal, using the BM3D denoising algorithm to remove daytime noise, and using an identification algorithm to filter false classification samples; (3) screening the data reliability through a cloud-aerosol identification algorithm, separating clouds and aerosols through a multi-dimensional histogram of scattering characteristics, and calculating a confidence score according to the overlapping area between the multi-dimensional histograms to screen the confidence profile samples; (4) obtaining the three-channel attenuated backscatter coefficients of the denoised and screened samples, calculating the volume depolarization ratio and the attenuation color ratio, and normalizing the 532 nm channel total attenuated backscatter coefficient, the volume depolarization ratio and the attenuation color ratio three signals to match the RGB three channels to form an image, unify the size, and then perform data enhancement on the image to expand the data and divide the data set for training, verification and testing; (5) constructing a cloud-aerosol hierarchical classification model based on a semantic segmentation neural network model, taking the training set data as input and the level 2 cloud-aerosol classification results as output to train the model; The semantic segmentation neural network model is constructed as an encoder-decoder structure, the encoder extracts image features, and the decoder classifies according to the features; The encoder is composed of a depth separable convolution network and a dilated convolution pyramid pooling-double-channel attention parallel module, the decoder receives the output of the depth separable convolution network and the parallel module, and finally outputs the full image pixel cloud-aerosol prediction classification through multiple full convolution upsampling; (6) for the test set data, using the model trained in step (5) to perform actual cloud-aerosol hierarchical classification and comparing with the existing classification results.

2. The method of claim 1, wherein the cloud-aerosol hierarchical classification based on semantic segmentation is characterized by, In step (1), the 532 nm and 1064 nm channel backscatter coefficients of the CALIOP level 1 and the calculated volume depolarization ratio and attenuation color ratio are combined to form an image, and cloud-aerosol hierarchical classification is performed for each pixel in the image.

3. The method of claim 1, wherein the cloud-aerosol hierarchical classification based on semantic segmentation is characterized by, In step (1), the cloud-aerosol hierarchical classification results include cloud, aerosol, clear sky, ground, sub-surface and missing signal.

4. The method of claim 1, wherein the cloud-aerosol hierarchical classification based on semantic segmentation is characterized by, In step (3), the profile samples with a confidence score absolute value between 70 and 100 are screened.

5. The method of claim 1, wherein the cloud-aerosol hierarchical classification based on semantic segmentation is characterized by, In step (4), the data augmentation method for the image includes random cropping, rotation and shifting processing, and adjusting each image so that the edges overlap.

6. The method of claim 1, wherein the cloud-aerosol hierarchical classification based on semantic segmentation is characterized by, The double-channel attention parallel module is a position attention module and a channel attention module.

7. The method of claim 1, wherein the cloud-aerosol hierarchical classification based on semantic segmentation is performed by a convolutional neural network (CNN) trained on a dataset of images of clouds and aerosols. In step (6), the average intersection over union and the classification accuracy of the classification are calculated based on the results of the comparative analysis.

Images