CN104504389B - A kind of satellite cloudiness computational methods based on convolutional neural networks - Google Patents

Abstract

The invention discloses a method for calculating satellite cloud amount based on a convolutional neural network. Firstly, a satellite cloud image training sample containing 6000~8000 training samples is established, and manually mark thick clouds and thin clouds of 2000~3000 samples in the satellite cloud image. The cloud and clear sky cloud tiles are used as the training samples of the convolutional neural network; then the training samples and satellite cloud images are preprocessed as the data input of the convolutional neural network, and then the convolutional neural network detection is performed to detect each image in the cloud image. The location of thick cloud, thin cloud and clear sky area; finally, according to the positions of thick cloud, thin cloud and clear sky in the cloud image, calculate their gray value respectively, and calculate the cloud amount of satellite cloud image according to their gray value. The invention can directly use the satellite cloud image as the input of CNN, and integrates the feature extraction function into the neural network to implicitly extract the features of the image, which is more convenient and accurate than the prior art, and has important application value.

Description

A satellite cloud calculation method based on convolutional neural network

technical field

The invention relates to the field of meteorological detection, in particular to a method for calculating satellite cloudiness based on a convolutional neural network.

Background technique

Cloud is one of the most important factors in weather and climate. On the one hand, it regulates the internal radiation balance of the earth's atmosphere, and on the other hand, it plays an important role in the water cycle. Therefore, cloud observation plays an important role. The long-term reliance on manual visual observation has become the bottleneck of automatic forecasting of meteorological satellites, and automatic recognition of cloud images has become an urgent need.

Cloud detection, cloud classification and calculation of cloud amount based on satellite images are the main ways to obtain global cloud amount distribution. At present, the international satellite cloud calculation methods mainly include the ISCCP method. Through the ISCCP multi-threshold cloud detection method, the pixels are divided into two types: clear sky and cloudy; there is the CLAVR-1 method, which divides the pixels into clear sky, mixed and cloudy. There are three types of clouds; there is the CLAVR-X method, which divides the pixels into four types: full cloud, mixed cloud, mixed clear sky, and situation; there is also the MODIS method, which divides the pixels into four types: definite cloud, possible cloud, likely clear sky, and definite clear sky class; there are also methods such as UW HIRS, NIR/VIS, etc. The above-mentioned cloud amount calculation methods can be roughly divided into two categories: one is to calculate cloud amount based on the ratio of cloudy pixels to total pixel points in the area; the other is to calculate equivalent cloud amount based on pixel radiation/reflectivity. The first type of method is simple to operate, but it cannot analyze the sub-pixel cloud amount, which often leads to high calculation results; the second type of method solves the problem of sub-pixel cloud amount to a certain extent, but it is not suitable for the case of multi-layer cloud and drastic changes in the surface type. Not very applicable. Regardless of the calculation method, its accuracy depends on the accuracy of cloud detection results.

At present, research on cloud detection at home and abroad mainly includes threshold method and neural network, and the recognition accuracy of neural network is generally considered to be higher than that of other classifiers. Although the neural network classification method has unique advantages among many methods, there are also some problems. The traditional neural network adopts the error feedback gradient learning method (BP), which has the disadvantages of slow learning speed, too many iterations, and the solution is easy to fall into local minimum. These shortcomings seriously affect the application of neural network in cloud classification.

Deep learning combines low-level features to form more abstract high-level representation attribute categories or features to discover distributed feature representations of data. It can effectively solve the deficiencies in the existing methods. The classifier is the core of the cloud class recognition model, and the effectiveness of the model directly affects the intelligent analysis results of the cloud image. Due to the self-adaptive, self-learning and non-linear approximation capabilities of convolutional neural network, it has more advantages than some other algorithms in the process of realizing cloud classification.

Contents of the invention

Purpose of the invention: The present invention aims at various shortcomings of current cloud image detection classifiers and technical deficiencies of low cloud cover detection accuracy, and provides a satellite cloud cover calculation method based on convolutional neural network through a large number of experimental studies.

Technical scheme: in order to achieve the above object, the technical scheme that the present invention takes is:

The steps of a cloud computing method based on a convolutional neural network are as follows:

(1) Establish a satellite cloud image training sample containing 6000-8000 samples, manually mark thick cloud, thin cloud and clear sky cloud tiles of 2000-3000 samples in the satellite cloud image, and use this as the training sample of the convolutional neural network,

(2) Preprocessing the training sample and the satellite cloud image is used as the data input of the convolutional neural network, and then the convolutional neural network is detected, so as to detect the positions of each thick cloud, thin cloud and clear sky area in the cloud image;

(3) According to the positions of thick cloud, thin cloud and clear sky in the cloud image after detection, calculate their gray value respectively, and calculate the cloud amount of the satellite cloud image according to their gray value.

As a preferred version, the above-mentioned a kind of satellite cloud amount calculation method based on convolutional neural network, said step (1) specifically includes:

1) Download the required HJ-1A/1B satellite cloud image data from the China Resources Satellite Center;

2) Use the collector to collect 39*39 pixels of thick clouds, 2000 pieces of thin clouds and 2000 clear cloud blocks on the HJ-1A/1B satellite cloud image, and uniformly zoom to 32*32 pixels.

As a preferred solution, the above-mentioned satellite cloud amount calculation method based on convolutional neural network, the preprocessing required for step (2) is to convert the entire satellite cloud image format into a data form of 32*32*X, where X is the quantity, and then the training samples are input as the convolutional neural network together;

The convolutional neural network includes 7 layers, the first layer is the input layer, and the second layer is the volume base layer, also known as the feature mapping layer, which extracts different features of the picture through multiple 5*5 convolution kernels, including 12 A 28*28 feature map. The third layer is the downsampling layer, also known as the feature extraction layer. It is composed of 12 feature maps with a size of 14*14. Each neuron in the feature map is related to the 2 neurons of the first layer. *2 fields are connected, the fourth layer is a convolutional layer composed of 16 10*10 feature maps, the fifth layer is a downsampling layer like the third layer, and contains 16 5*5 feature maps, and the sixth layer is The fully connected layer has a total of 400 connection points, and the last layer is the output layer, which has 3 nodes representing thick clouds, thin clouds and clear sky.

The calculation method of the convolutional layer is as follows:

Where l is the number of layers of the network, K is the convolution kernel, and M _j represents the set of input maps.

The downsampling layer is calculated as follows:

Among them, down() represents the downsampling function, and β and b correspond to the feature map of each output respectively.

The detected satellite cloud images are represented by red, green, blue and grayscale images respectively;

According to the detected color map, the minimum gray value of the thick cloud in the red area, the highest gray value of the thin cloud in the green area and the average gray value of the clear sky in the blue area are calculated respectively, and the cloud amount is calculated according to the cloud amount calculation formula.

Beneficial effect: compared with the prior art, the present invention has the following beneficial effects:

The present invention screens through a large number of experiments, and designs a satellite cloud calculation method based on a convolutional neural network. As a convolutional neural network for deep learning, the satellite cloud image can be directly used as the input of CNN, and the feature extraction function is integrated into the neural network. The implicit extraction of image features is more convenient and accurate than manual extraction. Weight sharing reduces the training parameters of the network, reduces the complexity of the neural network, and adapts to the current demand for large amounts of data.

The detection of clouds in satellite cloud images is the premise of satellite image interpretation. Aiming at the deficiency of threshold method cloud detection, the present invention utilizes the global and local characteristics of clouds, and carries out cloud semantic feature learning and cloud classification based on convolutional neural networks. On the basis of improving the albedo-based spatial correlation method to calculate the total cloud amount, improving and perfecting the cloud classification algorithm and cloud amount calculation, laying a solid theoretical foundation for the comprehensive automatic detection of satellite cloud images.

Description of drawings

Fig. 1 is a kind of satellite cloud cover calculation method schematic diagram based on convolutional neural network of the present invention;

Fig. 2 is a schematic diagram of the convolutional neural network structure of the present invention;

Fig. 3 is a schematic diagram of convolutional neural network detection according to the present invention;

Fig. 4 is a schematic diagram of convolution and downsampling in the convolutional neural network process of the present invention;

Detailed ways

Further illustrate technical scheme of the present invention below in conjunction with accompanying drawing;

As shown in Figure 1, a satellite cloud calculation method based on convolutional neural network, including sample acquisition and processing, convolutional neural network training, satellite cloud detection and cloud calculation four stages, the sample acquisition and processing Include the following steps:

(1) The satellite cloud image sample is obtained from the CCD channel of China Satellite Resource Satellite HJ-1A/1B satellite;

(2) Collect 6000 samples of 39*39 pixel cloud blocks from the satellite cloud image through the collector, 2000 blocks each for thick clouds, thin clouds and clear sky;

(3) Scale the 6000 samples uniformly to 32*32 pixels, and convert the entire satellite cloud image format into a 32*32*X data format, where X is the quantity;

Convolutional neural network training of the present invention comprises the following steps:

(1) 4200 samples out of 6000 cloud samples are used as training samples, 1400 are thick clouds, thin clouds and clear sky respectively, and 1800 test samples are thick clouds. Thin clouds and clear sky are 600 yuan each;

(2) Training is the structure of the convolutional neural network as shown in Figure 2. Through continuous training and testing of cloud block samples, the parameters in the convolutional neural network are constantly adjusted, and the parameters are determined based on the cloud block detection rate. Satellite cloud image detection as a foreshadowing;

The satellite cloud image detection of the present invention comprises the following steps:

(1) The entire satellite cloud image after preprocessing is used as the data input of the convolutional neural network to perform cloud detection, wherein the process of cloud block detection is shown in Figure 3;

(2) The convolutional neural network consists of 7 layers, the first layer is the input layer, and the second layer is the volume base layer, also known as the feature mapping layer, which extracts different features of the picture through multiple 5*5 convolution kernels, including 12 A 28*28 feature map. The third layer is the downsampling layer, also known as the feature extraction layer. It is composed of 12 feature maps with a size of 14*14. Each neuron in the feature map is related to the 2 neurons of the first layer. *2 fields are connected, the fourth layer is a convolutional layer composed of 16 10*10 feature maps, the fifth layer is a downsampling layer like the third layer, and contains 16 5*5 feature maps, and the sixth layer is The fully connected layer has a total of 400 connection points, and the last layer is the output layer, which has 3 nodes representing thick clouds, thin clouds and clear sky.

(3) The learning process of the convolutional neural network is shown in Figure 4. The convolution base layer is the feature mapping layer, and the sub-sampling layer is the feature extraction layer. A trainable filter f _x is used to deconvolute the global and local features of the satellite cloud image. , and then add a paranoid b _x to obtain the volume base layer C _x . The calculation form of the convolutional layer is as follows:

Where l is the number of layers of the network, K is the convolution kernel, and M _j represents the set of input maps.

The subsampling process consists of summing per domain, weighting by W _x+1 , adding bias b _x+1 , and then passing a sigmoid

The activation function, which produces a reduced feature map layer S _x+1 , is calculated as:

Among them, down() represents the downsampling function, and β and b correspond to the feature map of each output respectively.

(4) The results of cloud detection are distinguished by red, green and blue, where red represents thick clouds, green represents thin clouds, and blue represents clear sky.

The detection result cloud amount calculation of the present invention comprises the following steps:

(1) In order to solve the problem of partial cloud cover, this study uses an improved "spatial correlation method" based on reflectance to calculate the total cloud cover. The basic principle of the "spatial correlation method" is to obtain the total cloud amount of a single pixel based on the detection of the radiation amount of a single pixel and the radiation amount in the case of clear sky and thick clouds. The calculation formula is as follows:

I=(1-A _c )I _clr +A _c I _cld

In the formula, I is the radiation amount received by the pixel; A _c represents the total cloud amount of the pixel; I _clr and I _cld represent the radiation amount of the highest gray value of clear sky and the lowest gray value of thick cloud, respectively. Thus, the conversion can be obtained, and the total cloud amount is:

(2) Cloud amount calculation, the cloud amount of each pixel is between 0 and 1, 1 represents thick cloud, and 0 represents clear sky. The present invention uses 200 randomly selected pictures for testing, and checks in the way of experts' marking. The cloud amount calculation accuracy rate of this method reaches 84.3%. In the same situation, the accuracy rate obtained by the traditional threshold method is 75.4%. It shows that this method is feasible in practical application, and the accuracy rate is higher than that of the existing technology, which has important application value.

The above descriptions are only part of the embodiments of the present invention. It should be pointed out that those skilled in the art can make some improvements and modifications without departing from the principles of the present invention. It should be regarded as the protection scope of the present invention.

Claims (2)

1. based on convolutional neural networks it is satellite cloudiness computational methods a kind of, which is characterized in that it includes the following steps：

(1) the satellite cloud picture training sample for including 6000~8000 training samples is established, is marked out in satellite cloud picture manually each Spissatus, Bao Yun and the clear sky cloud atlas block, in this, as the training sample of convolutional neural networks of 2000~3000 samples；

(2) training sample and satellite cloud picture are carried out pretreatment to input as the data of convolutional neural networks, then carries out convolution Neural network detects, and each spissatus in cloud atlas, Bao Yun and clear sky region position are detected with this；

(3) according to the position of spissatus, Bao Yun and clear sky in cloud atlas after detection, calculate separately its gray value, according to its gray value come The cloud amount for carrying out satellite cloud picture calculates；

The step (1) specifically includes：

1) the HJ-1A/1B satellite cloud picture data needed for being downloaded in Chinese Resources satellite hub；

2) collector is utilized to acquire the spissatus of 39*39 pixels on HJ-1A/1B satellite cloud pictures respectively, Bao Yun and clear sky cloud mass are each 2000 pieces, uniformly it is scaled 32*32 pixels；

The step (2) specifically includes：

1) it is that 32*32 pixel forms and training sample are defeated together as convolutional neural networks by whole picture satellite cloud picture format conversion Enter source；

2) convolutional neural networks include 7 layers, and the 1st layer is input layer, and the 2nd layer is volume base, is extracted by the convolution kernel of multiple 5*5 The different characteristic of picture, includes the Feature Mapping figure of 12 28*28, and the 3rd layer is down-sampling layer, be by 12 sizes is 14*14 Characteristic pattern composition, each neuron of characteristic pattern is connected with the 1st layer of the fields 2*2, the 4th layer of characteristic pattern group by 16 10*10 At convolutional layer, the 5th layer is down-sampling layer as the 3rd layer, includes the characteristic pattern of 16 5*5, and the 6th layer is full articulamentum, altogether It is output layer to have 400 tie points, last layer, has 3 nodes, respectively represents spissatus, Bao Yun and clear sky；

3) satellite cloud picture after detection is showed with redgreenblue and gray-scale map respectively；

Cloud amount computational methods described in step (3) are：

The basic principle of " Spatial coherence method " be based on to single pixel amount of radiation and clear sky and it is spissatus in the case of amount of radiation Detection, obtains the total amount of cloud of single pixel, calculation formula is as follows：

I=(1-A_c)I_clr+A_cI_cld

In formula, I is the amount of radiation that pixel receives；A_cIndicate pixel total amount of cloud；I_clr、I_cldThe highest ash of clear sky/blue is indicated respectively Angle value, spissatus/red lowest gray value amount of radiation；Total amount of cloud calculation formula is：

I is the amount of radiation that pixel receives；A_cIndicate pixel total amount of cloud；I_clr、I_cldRespectively indicate clear sky/blue highest gray value, Spissatus/red lowest gray value amount of radiation.

2. it is according to claim 1 it is a kind of based on convolutional neural networks be satellite cloudiness computational methods, which is characterized in that The satellite cloud picture training sample for including 6000 training samples is established in step (1).