A ground-based cloud image classification and recognition method based on an improved convolutional neural network

Abstract

A ground-based cloud image classification and recognition method based on an improved convolutional neural network, S1, collects various types of ground-based cloud images to obtain an initial ground-based cloud image dataset; S2, the initial ground-based cloud image is pretreated to construct the ground-based cloud image dataset; S3, a convolutional neural network model is built; S4, the ground-based cloud image dataset is divided into a training set, a test set and a validation set in proportion, and the training set is used to train the convolutional neural network model; S5, the trained convolutional neural network model is verified by using the above-mentioned validation set, and the convolutional neural network model is adjusted according to the result; S6, the above-mentioned test set is identified by using the adjusted convolutional neural network model to obtain the identification result; in S3, the convolutional neural network model includes a feature extraction module and a main network module, the main network module is established based on a ResNet model and a DensNet model, and includes a BlockA module, a BlockB module and a BlockC module. The anti-overfitting property of the convolutional neural network model is improved, and the identification accuracy of the model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a ground-based cloud image classification and recognition method based on an improved convolutional neural network. BACKGROUND

[0002] The shape of the cloud has a significant indication effect on the current meteorological information and future gas phase change, and can be divided into high cloud family, medium cloud family and low cloud family according to the shape, structure, characteristics and cloud height. At present, the observation methods of the cloud are manual observation, satellite cloud observation and ground-based cloud observation. The manual observation requires the observer to have strong related knowledge and experience. However, the classification result is easily affected by the subjective factors of the person, and some small cloud changes are difficult to be recognized by manual observation, that is, it is easy to produce wrong judgment and misjudgment, and the recognition efficiency is relatively low. The satellite cloud observation is to observe the meteorological image taken from top to bottom by the observation satellite, which contains the shape information and movement of the cloud in a large range. However, due to the resolution, it cannot be used for local area, and the detection accuracy of low cloud and cirrus cloud is low, and the weather change condition of the local area cannot be accurately reflected in time. The ground-based cloud observation is mainly obtained by the all-sky imager, which has the advantages of low cost, high resolution and fixed angle compared with the satellite cloud observation. The background of the ground-based cloud is only the sky with single color, which greatly avoids the interference of the background to the cloud. Therefore, in recent years, the ground-based cloud has been paid more and more attention.

[0003] However, there are still many difficulties in the application of ground-based cloud observation. First, the convolutional neural network needs a large amount of labeled data for training. Second, the limited training data is easy to cause overfitting, so that the features cannot be well detected. SUMMARY

[0004] In order to solve the problem of overfitting in the training process of the convolutional neural network in the prior art, the present application provides a ground-based cloud image classification and recognition method based on an improved convolutional neural network, which fully utilizes the feature information in the ground-based cloud, perfects the comprehensiveness of the feature information, improves the anti-overfitting performance of the convolutional neural network model, and improves the recognition accuracy of the model.

[0005] In order to achieve the above purpose, the present application adopts the following specific scheme: a ground-based cloud image classification and recognition method based on an improved convolutional neural network, comprising the following steps:

[0006] S1, collecting a plurality of types of ground-based cloud images to obtain an initial ground-based cloud image data set;

[0007] S2, preprocessing the initial ground-based cloud image to construct a ground-based cloud image data set;

[0008] S3, building a convolutional neural network model;

[0009] S4, proportionally divide the ground cloud image dataset into a training set, a test set and a validation set, and use the training set to train the convolutional neural network model;

[0010] S5, use the above-mentioned validation set to verify the trained convolutional neural network model, and adjust the convolutional neural network model according to the result;

[0011] S6, use the adjusted convolutional neural network model to identify the above-mentioned test set, and obtain the identification result.

[0012] In the S3, the convolutional neural network model comprises a feature extraction module and a main network module, the main network module is established based on a ResNet model and a DensNet model, and comprises a BlockA module, a BlockB module and a BlockC module.

[0013] As an optimization scheme of the above-mentioned ground cloud image classification and identification method based on the improved convolutional neural network, the S1 specifically comprises: collecting multiple types of ground cloud images by a full-sky imager and transmitting them back to an upper computer to obtain an initial ground cloud image dataset.

[0014] As another optimization scheme of the above-mentioned ground cloud image classification and identification method based on the improved convolutional neural network, the S2 comprises:

[0015] S201, repairing strong light of the initial ground cloud image to eliminate the interference of strong light of the sun on the ground cloud image identification, so that the ground cloud image is close to the real sky;

[0016] S202, manually labeling and classifying the ground cloud image after strong light repair to construct a basic dataset of the ground cloud image;

[0017] S203, identifying the basic dataset according to transfer learning and adding SoftMax output maximum probability, when the maximum probability is greater than 85%, it is considered that the ground cloud image classification is correct, otherwise re-labeling and classifying manually; repeat the above operation to finally construct a standard dataset;

[0018] S204, performing image size normalization processing and image enhancement on the standard dataset to obtain a ground cloud image dataset.

[0019] As another optimization scheme of the above-mentioned ground cloud image classification and identification method based on the improved convolutional neural network, the operation of strong light repair in the S201 is to set a clear sky coefficient threshold and calculate the clear sky coefficient:

[0020]

[0021] Wherein, I mTo measure the value of solar irradiance, I clr The theoretical value of the direct solar irradiance under clear sky conditions;

[0022] The clear sky coefficient k is compared with the clear sky coefficient threshold value. If the clear sky coefficient is greater than the clear sky coefficient threshold value, there is a strong light point around the sun. The RGB color map of the ground-based cloud image is converted into an HSV map, and then the brightness channel V in the HSV map is extracted.

[0023] A grayscale threshold value is set. When the grayscale value of the brightness channel V is greater than the grayscale threshold value, the pixel point is marked as a strong light point, and the brightness of the strong light point is set as the average brightness of the ground-based cloud image.

[0024] As another optimization scheme of the above-mentioned ground-based cloud image classification and recognition method based on the improved convolutional neural network, the S3 comprises:

[0025] S301, the feature extraction part of the DehazeNet defogging algorithm is used as a feature extraction module to obtain the fuzzy features of the ground-based cloud image, and the fuzzy features are maximized by the Maxout unit to generate a new feature map;

[0026] S302, the BlockA module and the BlockB module both pass through the BN layer and the ReLu activation function first, and then perform convolution operation;

[0027] S303, the BlockC module is used to extract the deep features of the ground-based cloud image.

[0028] As another optimization scheme of the above-mentioned ground-based cloud image classification and recognition method based on the improved convolutional neural network, the network structure function of the fuzzy features is:

[0029]

[0030] The activation function is the Maxout function, and the hidden layer node output expression is:

[0031]

[0032] Wherein, W is the network weight matrix, B is the bias matrix, j is the weight matrix, i is the hidden unit, f i，j (x) is the acquired visibility feature, I is the input feature vector, k is the number of virtual hidden layers, x T is the input feature matrix.

[0033] As another optimization scheme of the above-mentioned ground-based cloud image classification and recognition method based on the improved convolutional neural network, the feature extraction module is composed of the feature extraction part of the DehazeNet defogging algorithm, the original ground-based cloud image is taken as the input through the visibility estimation method, and the influence of aerosol diffuse reflection in the air is reduced.

[0034] As another optimization scheme of the above-mentioned ground-based cloud image classification and recognition method based on the improved convolutional neural network, the BlockA module and the BlockB module both introduce the SE module.

[0035] As another optimization scheme of the above-mentioned ground-based cloud image classification and recognition method based on the improved convolutional neural network, the S3 comprises:

[0036] S301, the ground-based cloud image dataset is divided into a training set, a test set and a validation set according to a ratio of 7:2:1, and the training set, the test set and the validation set all contain various types of ground-based cloud images;

[0037] S302, the Adam optimizer and the loss function are used to train the convolutional neural network model through the training set.

[0038] As another optimization scheme of the above-mentioned ground-based cloud image classification and recognition method based on the improved convolutional neural network, after the main network module extracts the ground-based cloud image features, the BlockA module, the BlockB module and the BlockC module are repeatedly stacked to obtain the feature map, and the feature map is globally averaged and pooled, and a Dropout layer is added in the full connection layer of the feature map, and the random inactivation ratio is 0.6.

[0039] Compared with the prior art, the present application has the following beneficial effects:

[0040] 1. The present application provides a ground-based cloud image classification and recognition method based on an improved convolutional neural network, the convolutional neural network model comprising a feature extraction module and a main network module, the feature extraction module being capable of extracting model features on the ground-based cloud image and performing a pixel-by-pixel maximization operation on the fuzzy features through a Maxout unit to generate a new feature map; the main network module combining a ResNet model and a DensNet model to propose a BlockA module, a BlockB module and a BlockC module, wherein the BlockA module and the BlockB module solve the problem of gradient disappearance through a skip connection, while obtaining more feature information, the BlockC module being used to extract deep features of the ground-based cloud image, fully utilizing the feature information in the ground-based cloud image, perfecting the comprehensiveness of the feature information, improving the anti-overfitting property of the convolutional neural network model and improving the recognition accuracy of the model.

[0041] 2. In the present application, the ground-based cloud image dataset is constructed through artificial labeling and transfer learning, the confidence of the ground-based cloud image dataset is improved, the subsequent training model is more accurate, and the time-consuming of constructing the ground-based cloud image dataset is greatly reduced.

[0042] 3、 The strong light of the ground-based cloud picture is repaired in the application, the interference of the strong light on the ground-based cloud picture is removed, and the recognition precision of the model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a convolutional neural network structure diagram in the application;

[0044] Figure 2 is a structure diagram of BlockA module;

[0045] Figure 3 is a structure diagram of BlockB module. DETAILED DESCRIPTION

[0046] The technical solutions of the application will be further described in detail below in combination with specific embodiments. Parts not described and disclosed in the following embodiments of the application should be understood as the prior art known or should be known by those skilled in the art.

[0047] A ground-based cloud picture classification and recognition method based on an improved convolutional neural network, comprising the following steps:

[0048] S1, collecting various types of ground-based cloud pictures to obtain an initial ground-based cloud picture data set, specifically, collecting various types of ground-based cloud pictures by a full-sky imager and returning to an upper computer to obtain an initial ground-based cloud picture data set.

[0049] S2, preprocessing the initial ground-based cloud picture to construct a ground-based cloud picture data set.

[0050] Specifically:

[0051] S201, repairing the initial ground-based cloud picture to eliminate the interference of the sun's strong light on the ground-based cloud picture recognition, so that the ground-based cloud picture is close to the real sky. The operation of strong light modification is to set a clear sky coefficient threshold, calculate the clear sky coefficient k:

[0052]

[0053] Wherein, I m is the measured solar irradiance value, I clr is the theoretical value of the direct solar irradiance under clear sky conditions.

[0054] Compare the clear sky coefficient k with the clear sky coefficient threshold. If the clear sky coefficient is greater than the clear sky coefficient threshold, there is a strong light point around the sun. Convert the RGB (red, green, blue) color chart of the ground-based cloud picture into an HSV (hue, saturation, brightness) chart, then extract the brightness channel V in the HSV chart. Set a gray scale threshold, compare the gray scale value of the brightness channel V with the gray scale threshold, and when the gray scale value of the brightness channel V is greater than the gray scale threshold, mark the pixel point as a strong light point, and set the brightness of the strong light point to the average brightness of the ground-based cloud picture.

[0055] S202, manually labeling and classifying the ground-based cloud image after strong light repair, specifically, carefully reading the China Meteorological Administration published "China Cloud Image" and manually labeling and classifying the collected ground-based cloud image under the guidance of professional meteorological instructors to form a basic data set including ten cloud types and no cloud.

[0056] S203, identifying the basic data set according to transfer learning and adding SoftMax function to output the maximum probability, when the maximum probability is greater than 85%, it is considered that the ground-based cloud image classification is correct, otherwise, re-labeling and classifying manually; repeat the above operation to finally build a standard data set. Specifically, select a suitable pre-training network model for transfer learning, in this embodiment, the pre-training model is ResNet50 network model, put the ground-based cloud image in the basic data set into the pre-trained ResNet50 network model for identification, and output the maximum probability of the ground-based cloud image through the SoftMax function. The confidence of the ground-based cloud image data set is improved, and the time-consuming of building the data set is reduced.

[0057] S204, image size normalization processing is performed on the standard data set, when the number of standard data sets is insufficient, the data amount between each category is unbalanced, and the standard data set needs to be directionally image enhanced by horizontal flip, vertical flip, fixed angle rotation, translation, random Gaussian noise processing and other methods to obtain the ground-based cloud image data set.

[0058] S3, build a convolutional neural network model; the convolutional neural network model includes a feature extraction module and a main network module, the main network module is established based on ResNet model and DensNet model, and includes BlockA module, BlockB module and BlockC module.

[0059] Specifically, S301, the feature extraction part of DehazeNet defogging algorithm is used as the feature extraction module to obtain the fuzzy features of the ground-based cloud image, and the Maxout unit is used to maximize the fuzzy features to generate a new feature map;

[0060] S302, the BlockA module and the BlockB module are first processed through the BN layer and the ReLu activation function, and then the convolution operation is performed;

[0061] S303, the BlockC module is used to extract the high-level features of the ground-based cloud image.

[0062] The feature extraction module includes:

[0063] In the ground-based cloud image, different cloud shapes have different structures and characteristics, and the convolutional neural network model can learn and identify these characteristics. However, when the sky is clear, the pipeline directly irradiates the all-sky imager, causing the sun's surrounding area to appear halos and flares in the photos taken by the all-sky imager, and causing the sun's surrounding area to appear white and the area far from the sun to appear black, ultimately making it difficult for the convolutional neural network to identify the ground-based cloud image. When cirrus or convective clouds appear in the sky, the sky can still be seen in the ground-based cloud image because the cloud body is thin, making it difficult to identify features in the ground-based cloud image. In summary, the particles in the air affect the light and thus affect the feature extraction of the ground-based cloud image, which is attributed to the problem of sky visibility.

[0064] Weather visibility is determined by an atmospheric scattering model, which describes the combined imaging mechanism of air particles and light:

[0065] I(x) = J(x)t(x) + a(1-(t(x))

[0066] where I(x) is the original image of the all-sky imager, J(x)t(x) is the brightness of the original image of the transmitted object, and a(1-(t(x)) is the atmospheric light of the particle scattering in the original image. J(x) is the reflection imaging of the object surface, t(x) is the medium transmittance, which describes the part of the light that is not scattered and directly reaches the imaging camera, and the last transmitted part only has part of the energy J(x)t(x) to reach the imaging camera. At the same time, the other sunlight that does not pass through the object surface after scattering reaches the imaging camera as a(1-(t(x)).

[0067] In this embodiment, the feature extraction module of the DehazeNet defogging algorithm is used to solve the influence of particles in the air on light, where the DehazeNet defogging algorithm is an end-to-end system that reduces the influence of aerosol diffuse reflection in the air by using the dark channel prior, maximum contrast prior, and color attenuation visibility estimation method to input the original ground-based cloud image, thereby extracting more effective features. The dark channel prior feature is based on the observation of low-visibility ground-based cloud images. In areas with strong visibility, the pixel brightness value of at least one color channel is very low, even zero. Define this channel D as the minimum value of the pixel color in the local area:

[0068]

[0069] where I c is a certain color channel, and Ω r (x) is a local window centered at x with a size of r x r.

[0070] The maximum contrast prior is based on atmospheric scattering, and the contrast of the ground-based cloud image decreases with the decrease of the visibility transmission, which is obtained as:

[0071]

[0072] where t is the transmittance, I(x) is the original image, and J(x) is the reflected light image. In a local window region Ω centered at x with size r×r, the maximum contrast is defined as the variance of the pixel intensity I(z) with respect to the center pixel I(y) in a window of size s×s centered at y:

[0073]

[0074] where I(z) is the pixel intensity, I(y) is the center pixel, and Ω is the window region.

[0075] Color attenuation refers to the fact that the saturation I s (x) of a local region of the ground cloud image will decrease sharply under the influence of air particles, and the brightness I v (x) will increase, and at the same time, the entire scene will fade, resulting in a high difference value, so the difference between the brightness and the saturation is used to estimate the visibility of the ground cloud image:

[0076] A(x) = I v (x) - I s (x)

[0077] The hue difference between the original ground cloud image and its semi-inverse ground cloud image is also used to detect the visibility of the ground cloud image, and the semi-inverse ground cloud image of the original ground cloud image I(x) is defined as:

[0078] I st (x) = max[I c (x), 1 - I c (x)]

[0079] where c is a color channel component, and c ∈ (r, g, b).

[0080] The pixel values of the three channels of the semi-inverse ground cloud image with high visibility do not flip simultaneously, resulting in a large hue change between I st (x) and I(x); the pixel values of the three channels of the semi-inverse ground cloud image with low visibility need to flip simultaneously. The hue difference feature is defined as:

[0081]

[0082] In this embodiment, the feature extraction part of the DehazeNet defogging algorithm is used as the feature extraction module of the convolutional neural network model, and the fuzzy features of the ground cloud image are obtained through the feature extraction module, and the network structure function is:

[0083]

[0084] wherein, W is a network weight matrix, B is a bias matrix, j is a weight matrix, i is a hidden unit, f i，j (x) is an acquired visibility feature, I is an input feature vector, and k is the number of virtual hidden layers.

[0085] The activation function is a Maxout function, which has strong fitting ability and can fit any convex function, and the output expression of the hidden layer node is:

[0086]

[0087] wherein, W is a network weight matrix, B is a bias matrix. When W is a reverse filter, the Maxout function output is equivalent to the dark channel prior feature, when W is a ring filter, the Maxout function output is the maximum contrast prior feature, and when W contains both the reaction and all-pass filters, the Maxout function output is equivalent to the color attenuation feature. In this embodiment, W is a network weight matrix, b is a bias matrix, k is the number of virtual hidden layers, x T is an input feature matrix, j is a weight matrix, and i is a hidden unit.

[0088] Main network module:

[0089] The traditional convolutional neural network model mainly extracts deep features through the number of convolutional layers. However, as the depth of the convolutional layer increases, fewer and fewer features can be obtained, and even the gradient vanishes and the model degrades. Therefore, the main network module is constructed based on the residual module of the ResNet model and the cascade module of the DensNet model, and the main network module includes BlockA module, BlockB module and BlockC module. Among them, BlockA and BlockB first pass through the BN layer and the ReLu activation function, and then perform convolution operation. In the residual module of the ResNet model, the BN layer performs batch normalization operation on the features after convolution. However, due to the connection of the skip connection, the fused features do not have the normalization characteristic. Therefore, BlockA and BlockB pass through the BN layer before convolution, so that the results after convolution and the skip connection are normalized, thereby making the regularization ability of the convolutional neural network model stronger. At the same time, BlockA and BlockB introduce the SE module, which gives different weights to the positions in the ground-based cloud image from the channel domain through a weight matrix, so as to obtain more important feature information. Specifically, the SE module uses a new neural network to acquire the importance of each channel of the feature map in an automatic learning manner, and then gives each feature a weight value according to the importance, so that the neural network focuses on the feature channels with high importance, enhances the feature map channels useful for the current task, and always pays less attention to the feature channels less useful for the current task.

[0090] In this embodiment, in the BlockA module, the input x is subjected to two 3x3 convolutions to obtain a corresponding feature map F(x), the SE module obtains a corresponding weight parameter, the obtained weight is fused with F(x), and finally added with the input x. In the BlockB module, 1x1 convolution is first performed to reduce the channel number of the input ground cloud matrix, so as to reduce the calculation amount of subsequent operations; then, 3x3 convolution is performed to extract the detailed features of the ground cloud; finally, 1x1 convolution is performed to expand the number of channels. That is, the BlockB module strengthens the connection between the internal feature channels of the residual module, and enhances the ability of the convolutional neural network model to extract deep key features.

[0091] The BlockC module is used to extract high-level effective feature information of the ground cloud, realizes comprehensive extraction of feature information of the ground cloud, and improves the precision of the convolutional neural network model. Moreover, the BlockC module can densely connect all the preceding convolutional layers and the subsequent convolutional layers, and realizes feature reuse through connection of features in channels.

[0092] The BlockA, BlockB and BlockC modules are repeatedly stacked to obtain a feature map, global average pooling is performed on the feature map, and a Dropout layer is added in the full connection layer of the feature map, and the random inactivation ratio is 0.6, which reduces network overfitting, increases the generalization ability of the network, and reduces the training parameters of the network. The parameters of the convolutional neural network model are shown in Table 1.

[0093] Table 1

[0094]

[0095]

[0096] S4, the ground cloud data set is divided into training set, test set and validation set in proportion, and the training set is used to train the convolutional neural network model. The proportion of the training set, the test set and the validation set is 7:2:1, and the training set, the test set and the validation set all contain various types of ground clouds. The Adam optimizer and the cross-entropy loss function are selected for training. The Adan optimizer is an extension of the stochastic gradient descent method, which considers the first and second moment estimates of the gradient, has the advantages of simplicity and high efficiency, and the replacement of parameters is not affected by the gradient scaling transformation. The cross-entropy loss function measures the gap between the predicted results and the actual results of the convolutional neural network model:

[0097]

[0098] wherein x i represents the i-th element of the true label, y iTo predict the probability that x belongs to the i-th class by the convolutional neural network model, C is the total number of classes.

[0099] The essence of cross-entropy loss is to measure the distance between two probability distributions, which are the real label y distribution and the probability distribution y predicted by the convolutional neural network model * . For each class i, y i represents the probability that the model predicts x belongs to the i-th class, y * represents the probability that the model predicts x belongs to the i-th class. When the two probability distributions are closer, the cross-entropy loss is smaller, indicating that the model prediction result is more accurate.

[0100] S5, verify the trained convolutional neural network model by using the above-mentioned verification set, and adjust the convolutional neural network model according to the result.

[0101] S6, use the adjusted convolutional neural network model to identify the above-mentioned test set, and obtain the identification result.

[0102] The above description of the disclosed embodiments enables those skilled in the art to carry out or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ground-based cloud classification recognition method based on an improved convolutional neural network, comprising the following steps: S1, collecting multiple types of ground-based cloud images to obtain an initial ground-based cloud image dataset; S2, preprocessing the initial ground-based cloud image to construct a ground-based cloud image dataset; S3, building a convolutional neural network model; S4, dividing the ground-based cloud image dataset into a training set, a test set and a validation set in proportion, and training the convolutional neural network model using the training set; S5, verifying the trained convolutional neural network model using the above-mentioned validation set, and adjusting the convolutional neural network model according to the result; S6, identifying the above-mentioned test set using the adjusted convolutional neural network model to obtain an identification result; characterized in that, in S3, the convolutional neural network model comprises a feature extraction module and a main network module, the main network module is established based on a ResNet model and a DensNet model, and comprises a BlockA module, a BlockB module and a BlockC module; the BlockA module and the BlockB module both introduce an SE module; S301, using the feature extraction part of the DehazeNet defogging algorithm as the feature extraction module to obtain the fuzzy features of the ground-based cloud image, and performing a maximization operation on the fuzzy features through a Maxout unit to generate a new feature map; S302, in the BlockA module, input x is subjected to two 3x3 convolutions to obtain a corresponding feature map F(x), an SE module obtains corresponding weight parameters, the obtained weight is fused with F(x), and finally added with input x; in the BlockB module, 1x1 convolution is first performed to reduce the channel number of the input ground-based cloud image matrix, 3x3 convolution is then performed to extract the detailed features of the ground-based cloud image; finally, 1x1 convolution is performed to expand the number of channels; the BlockA module and the BlockB module both first pass through a BN layer and a ReLu activation function, and then perform convolution operation; S303, using the BlockC module to extract deep features of the ground-based cloud image; The BlockA, BlockB and BlockC modules are repeatedly stacked to obtain a feature map, and the feature map is subjected to global average pooling and a Dropout layer is added in the full connection layer of the feature map. The specific process of S1 is that multiple types of ground-based cloud images are collected by an all-sky imager and transmitted back to an upper computer to obtain an initial ground-based cloud image dataset.

2. The ground-based cloud image classification and recognition method based on the improved convolutional neural network according to claim 1, characterized in that: S2 3. The ground-based cloud image classification and recognition method based on an improved convolutional neural network according to claim 1, characterized in that: comprises: S201, repairing strong light on the initial ground-based cloud image to eliminate the interference of strong sunlight on the ground-based cloud image recognition, so that the ground-based cloud image approaches the real sky; S202, manually labeling and classifying the ground-based cloud image after strong light repair to construct a basic dataset of ground-based cloud images; S203, identifying the basic dataset according to transfer learning and adding a SoftMax output maximum probability; when the maximum probability is greater than 85%, it is considered that the ground-based cloud classification is correct, otherwise the manual labeling and classification is performed again; the above operation is repeated to finally construct a standard dataset; S204, performing image size normalization processing and image enhancement on the standard dataset to obtain a ground-based cloud image dataset. ​ 4. The ground-based cloud image classification and recognition method based on the improved convolutional neural network according to claim 3, characterized in that: The operation of the strong light repair in S201 is to set a clear sky coefficient threshold and calculate the clear sky coefficient: where I m is the measured solar irradiance value, I clr is the theoretical value of the direct solar irradiance in clear sky conditions; Compare the contrast clear sky coefficient k with the clear sky coefficient threshold, if the clear sky coefficient is greater than the clear sky coefficient threshold, there is a strong light point around the sun, convert the RGB color map of the ground-based cloud image into an HSV map, then extract the brightness channel V in the HSV map; Set a grayscale threshold, when the grayscale value of the brightness channel V is greater than the grayscale threshold, mark the pixel point as a strong light point, and set the brightness of the strong light point to the average brightness of the ground-based cloud image.

5. The ground-based cloud image classification and recognition method based on an improved convolutional neural network according to claim 1, characterized in that: The network structure function of the fuzzy feature is: The activation function is a Maxout function, and the hidden layer node output expression is: where W is a network weight matrix, B is a bias matrix, j is a weight matrix, i is a hidden unit, f i,j (x) is an acquired visibility feature, I is an input feature vector, k is a number of virtual hidden layers, x T is an input feature matrix.

6. The ground-based cloud image classification and recognition method based on an improved convolutional neural network according to claim 1, characterized in that: The feature extraction module is composed of the feature extraction part of the DehazeNet dehazing algorithm, the original ground-based cloud image is taken as the input through the visibility estimation method, and the influence of aerosol diffuse reflection in the air is reduced.

7. The ground-based cloud image classification and recognition method based on an improved convolutional neural network according to claim 1, characterized in that: S3 includes: The ground-based cloud image dataset is divided into a training set, a test set and a validation set in a ratio of 7:2:1, and the training set, the test set and the validation set all contain various types of ground-based cloud images; The Adam optimizer and the loss function are used to train the convolutional neural network model through the training set.

8. The ground-based cloud image classification and identification method based on an improved convolutional neural network according to claim 1, characterized in that: The random inactivation ratio is 0.6.

Images