Lightweight satellite cloud image neural network training method and rainfall detection method

Abstract

The present invention discloses a lightweight satellite cloud image neural network training method and a rainfall detection method, including a satellite cloud image preprocessing process and a neural network training process; the satellite cloud image preprocessing process includes: inputting a satellite cloud image photo into a satellite cloud image photo collection; using a GPM precipitation distribution map at the same time as the satellite cloud image; marking all precipitation areas in the satellite cloud image photo with a frame; performing an image threshold binarization operation on the marked satellite cloud image; the neural network training process includes: performing image enhancement processing on the satellite cloud image photo; establishing a lightweight target detection algorithm M-YOLOv4; obtaining a corresponding loss function; inputting the satellite cloud image into the neural network to obtain the result after the algorithm inference; calculating the function value of the corresponding loss function to obtain a corresponding weight value. The present invention uses deep learning technology for image training to obtain precipitation detection areas in the satellite cloud image.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a lightweight satellite cloud image neural network training method and a rainfall detection method. Background Art

[0002] Satellite cloud images are a type of remote sensing imagery captured by meteorological satellites. These images reveal that, on average, nearly half of Earth's surface is covered by clouds each day. In some areas, the clouds are thicker, while in others, they are very thin, allowing for a glimpse of the underlying features. Clouds are crucial players in weather processes, forming due to atmospheric motion and reflecting the physical characteristics of the air. Clouds of various types are generally composed of water, ice crystals, or a mixture of these. They are not in contact with the ground and must have a certain thickness. Clouds are classified into three groups, 10 genera, and 29 categories based on their formation, physical properties, and their appearance, color, and cloud base height. The appearance of different clouds often indicates different atmospheric conditions and provides a strong indicator of weather changes. This is why there's a saying that "reading clouds can predict the weather," with certain weather phenomena often being associated with certain types of clouds. Satellite cloud images offer advantages such as a wide observation range, high density, and all-weather coverage. After years of development, they have become increasingly important in atmospheric science research, natural disaster observation, and weather forecasting.

[0003] Currently, there is little research on satellite cloud image detection. Most of the research is only on cloud image recognition and cannot accurately locate a specific cloud cluster. A small number of cloud image detection studies cannot be applied on edge devices due to the limited memory and computing power of edge devices.

[0004] Deep learning involves learning the inherent patterns and representational hierarchies of sample data. During this learning process, machines gradually acquire human-like analytical and learning capabilities. Deep learning is an artificial neural network composed of many layers. Higher layers in the network can acquire more abstract features based on input from lower layers. Its primary capability is automatic feature extraction. Currently, most satellite cloud image precipitation detection methods still rely on machine learning algorithms, with only a small number using deep learning algorithms. Machine learning algorithms are used to manually filter out cloud cluster features from satellite cloud images and then feed these features into the machine learning algorithm. However, these methods have the following drawbacks: 1. Manual feature selection is prohibitively expensive; 2. Feature selection takes a long time. Existing models using deep learning algorithms for satellite cloud image precipitation detection suffer from a large number of model parameters and slow detection speed.

[0005] Therefore, the defect of the existing technology is the lack of a lightweight satellite cloud image rainfall detection algorithm, which uses deep learning technology for image training to obtain precipitation detection areas in satellite cloud images. Summary of the Invention

[0006] In view of at least one defect of the prior art, the object of the present invention is to provide a lightweight satellite cloud image neural network training method and a rainfall detection method. The algorithm uses deep learning technology for image training to obtain precipitation detection areas in satellite cloud images.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: a lightweight satellite cloud image neural network training method, including a satellite cloud image preprocessing process and a neural network training process;

[0008] The satellite cloud image preprocessing process includes the following steps:

[0009] Step A1: establishing a satellite cloud image photo collection, inputting the satellite cloud image photos into the satellite cloud image photo collection; establishing a corresponding precipitation distribution atlas, and storing the corresponding GPM precipitation distribution map into the precipitation distribution atlas;

[0010] Step A2: extracting satellite cloud image photos from the satellite cloud image photo collection in sequence;

[0011] Step A3: Retrieving the corresponding GPM precipitation distribution map, which has the same date as the satellite cloud image; wherein the GPM precipitation distribution map is provided by the Meteorological Bureau;

[0012] Step A4: Find the precipitation area in the GPM precipitation distribution map and manually mark all precipitation clouds in the corresponding satellite cloud image using the marking tool;

[0013] Step A5: Using the image threshold binarization operation on the marked satellite cloud image to obtain the pre-processed satellite cloud image;

[0014] Step A6: Determine whether there are any unlabeled satellite cloud images. If so, proceed to step A2. If not, end.

[0015] The training process of a neural network includes the following steps:

[0016] Step B1: Obtain a set of labeled satellite cloud image photos;

[0017] Step B2: performing image enhancement processing on the satellite cloud image photos in the satellite cloud image photo set to obtain an enhanced satellite cloud image;

[0018] Step B3: Build a custom M-YOLOv4 neural network model;

[0019] Step B4: Obtain the loss function corresponding to the M-YOLOv4 neural network model;

[0020] Step B5: Determine the threshold of the number of iterations of the loss function;

[0021] Step B6: Extract satellite cloud image images from the satellite cloud image collection in sequence, input the satellite cloud image images into the M-YOLOv4 neural network model, and obtain prediction results;

[0022] Step B7: Calculate the corresponding loss function value based on the manual labeling results and the corresponding prediction results, train the M-YOLOv4 neural network model, and obtain the corresponding weight value;

[0023] Step B8: Determine whether the number of iterations of the loss function is greater than the iteration threshold; if so, go to step B9; otherwise, go to step B6;

[0024] Step B9: Save the weight value as the final weight value.

[0025] The lightweight satellite cloud image neural network training method, step A5 includes:

[0026] Step A5: Use formula (1) to perform image threshold binarization operation on the marked satellite cloud image photos;

[0027]

[0028] Among them, pre_pixel is the original pixel value in the satellite cloud image, M is the pixel threshold, and M can be 195; New_pixel is the pixel value after the image threshold binarization operation.

[0029] Since the most important feature of precipitation clouds is the thickness of the cloud layer, we can use image threshold binarization to perform image preprocessing here. The image is composed of RGB. Where the cloud layer is thick, the image pixel value is correspondingly larger. Image threshold binarization refers to setting a pixel threshold. When the pixel value is greater than the pixel threshold, the pixel value remains unchanged. When the pixel value is less than the pixel threshold, the pixel value is set to 0.

[0030] The lightweight satellite cloud image neural network training method, step B2 includes:

[0031] Step B21: horizontally flipping the satellite cloud image in the satellite cloud image photo set;

[0032] Step B22: Mix the horizontally flipped and non-horizontally flipped satellite cloud image photos, randomly select 4 images from them, and randomly crop and splice the 4 images at random positions to form a new image.

[0033] The lightweight satellite cloud image neural network training method, the customized M-YOLOv4 neural network model in step B3 includes:

[0034] Input module, Backbone module, Neck module and Head module;

[0035] The Input module is used to obtain satellite cloud imagery; the Backbone module is used to extract shallow image features from the satellite cloud imagery input by the Input module; the Neck module enhances the shallow image features extracted by the Backbone module; the Head module obtains the image features enhanced by the Neck module and decodes these features to obtain the output results; among them, the Backbone module uses the MobileNetv1 network as the backbone feature extraction network; the Neck module uses the depthwise separable convolution DW-PANet as the feature fusion module, and the Neck module also uses the spp layer as an additional module.

[0036] The lightweight satellite cloud image neural network training method, the Conv_block module of the Backbone module is composed of 3×3 convolution, Batch Normalization layer and Relu6 activation function; the DW_block module of the Backbone module is composed of depthwise separable convolution Dwconv, Batch Normalization layer, Relu6 activation function, 1×1 convolution, BatchNormalization layer and Relu6 activation function;

[0037] The DBM module in the backbone module consists of a depthwise separable convolution (Dwconv), a batch normalization layer, and a Mish activation function. The enhanced feature extraction network uses a depthwise separable convolution (DW-PANet) and a spp layer to further increase the receptive field. In the later stages, we adopted the spp layer and customized a new DW-PANet to further increase the receptive field and enhance the feature extraction capability.

[0038] The lightweight satellite cloud image neural network training method, the loss function of the M-YOLOv4 neural network model includes a category loss function, a confidence loss function and a position loss function; the category loss function and the confidence loss function adopt a binary cross entropy loss function L;

[0039]

[0040] Where N represents the number of samples; i = 1-N; y i Indicates the label of sample i, the positive class is 1 and the negative class is 0; p i Indicates the probability of sample i predicting the positive class;

[0041] Position loss function L cIou Use CIOU loss function;

[0042]

[0043] IoU represents the intersection-over-union ratio between the predicted box and the real box, ρ 2 (b,b gt ) represents the Euclidean distance between the two center points, c represents the diagonal distance of the minimum closure area that can contain both the predicted box and the true box, α is the weight, and v is used to measure the similarity of the aspect ratio. Definition

[0044] Among them, b is the prediction box, b gt is the real box; w and h are the width and height of the predicted box respectively; w gt and h gt are the width and height of the ground truth box respectively.

[0045] In the lightweight satellite cloud image neural network training method, the iteration number threshold is 3000 times.

[0046] A lightweight satellite cloud image rainfall detection method;

[0047] The steps include:

[0048] Step C1: Obtain satellite cloud image photos;

[0049] Step C2: performing an image threshold binarization operation on the satellite cloud image;

[0050] Step C3: Processing the satellite cloud image using the M-YOLOv4 neural network model and the final weight value of claims 1-8 to generate corresponding precipitation areas;

[0051] Step C4: Display the precipitation area on the corresponding unprocessed satellite cloud image.

[0052] By taking out the satellite cloud image and GPM precipitation distribution map with the same date, finding the precipitation area in the GPM precipitation distribution map, and corresponding it to the satellite cloud image photo, a marked satellite cloud image photo is obtained.

[0053] Through the neural network training process, deep recognition is performed based on the labeled satellite cloud images, and the deep neural network algorithm is used to generate precipitation areas. The final weight values of the M-YOLOv4 neural network model are obtained.

[0054] The lightweight satellite cloud image detection algorithm, step C3 adopts the innovative lightweight target detection algorithm M-YOLOv4; the number of parameters is lower and the detection speed is faster.

[0055] Significant effect: The present invention provides a lightweight satellite cloud image neural network training method and a rainfall detection method, which uses deep learning technology for image training to obtain precipitation detection areas in satellite cloud images. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart for satellite cloud image annotation and image preprocessing;

[0057] Figure 2 This is the training flow chart of the M-YOLOv4 neural network model;

[0058] Figure 3 This is a flow chart of the method for detecting rainfall in satellite cloud images;

[0059] Figure 4 A schematic diagram of a specific use of this method;

[0060] Figure 5 This is a schematic diagram of the horizontal flip of a satellite cloud image;

[0061] Figure 6 This is the network structure diagram of the M-YOLOv4 neural network model. DETAILED DESCRIPTION

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] like Figures 1-6 As shown in the figure, a lightweight satellite cloud image neural network training method includes a satellite cloud image annotation and image preprocessing process and an M-YOLOv4 neural network model training process;

[0064] The satellite cloud map annotation and image preprocessing process includes the following steps:

[0065] Step A1: Create a satellite cloud image photo collection and input satellite cloud image photos into the satellite cloud image photo collection;

[0066] Step A2: extracting satellite cloud image photos from the satellite cloud image photo collection in sequence;

[0067] Step A3: Get the GPM precipitation distribution map with the same time as the satellite cloud map in step A2;

[0068] Step A4: Use the marking tool to manually find the precipitation areas in the GPM precipitation distribution map, map the precipitation areas to the satellite image, and mark them;

[0069] Step A5: Use the image threshold binarization operation to perform image preprocessing on the satellite cloud image.

[0070] Input a satellite cloud image, use the GPM precipitation distribution map, compare and mark the two, find the precipitation area on the satellite cloud image, and give the corresponding precipitation cloud label; the XML file example of its data is as follows:

[0071]

[0072]

[0073] Among them, the XML file is the generated marking file, and the results of our cloud map marking are stored in the XML file; after marking a satellite cloud map photo, the coordinate points of the marking box and the category information are stored in the XML file; object represents a precipitation area marked in a satellite cloud map photo; name represents the marked category, here the precipitation cloud is marked as cloud, and the four values under bndbox correspond to the coordinate points of the upper left corner and the lower right corner of the marking box respectively.

[0074] Satellite cloud images are from photos provided by the National Meteorological Administration;

[0075] Among them, step A6: determine whether there are any unlabeled satellite cloud map photos. If so, go to step A2; if not, end.

[0076] The training process of the M-YOLOv4 neural network model includes the following steps:

[0077] Step B1: Obtain a set of labeled satellite cloud image photos;

[0078] Step B2: performing image enhancement processing on the satellite cloud image photos in the satellite cloud image photo set to obtain an enhanced satellite cloud image;

[0079] Step B3: Build a custom M-YOLOv4 neural network model;

[0080] Step B4: Obtain the loss function corresponding to the M-YOLOv4 neural network model;

[0081] Step B5: Determine the threshold of the number of iterations for training the neural network model;

[0082] Step B6: Extract satellite cloud imagery from the satellite cloud imagery collection in sequence, input the satellite cloud imagery into the M-YOLOv4 neural network model, and obtain precipitation area results;

[0083] Step B7: Calculate the corresponding loss function value based on the manually marked precipitation areas and the precipitation areas obtained by model inference, train the M-YOLOv4 neural network model, and obtain the corresponding weight values;

[0084] Step B9: Determine whether the number of iterations of the neural network model training is greater than the iteration number threshold; if yes, go to step B10; otherwise, go to step B6;

[0085] Step B10: Save the weight value as the final weight value.

[0086] The key to the lightweight satellite cloud image precipitation detection method is that:

[0087] The image threshold binarization image preprocessing operation in step A5 takes advantage of the fact that precipitation clouds have thick layers and large pixel values, and sets the pixel values in non-precipitation cloud areas to 0.

[0088] By setting the above method, the precipitation cloud area is preserved and the non-precipitation cloud area is removed, which greatly eliminates the interference of the non-precipitation cloud area;

[0089] The key to the lightweight satellite cloud image precipitation detection method is that step B2 includes:

[0090] Step B21: horizontally flipping the satellite cloud image in the satellite cloud image photo set;

[0091] The satellite cloud map photos are centered on the vertical midline and swapped left and right. The purpose of this is to enlarge the photo collection.

[0092] Step B22: Randomly crop a portion of each horizontally flipped satellite cloud image to obtain the image sub-block, splice the four randomly cropped image sub-blocks to obtain a new satellite cloud image, and add the new satellite cloud image to the satellite cloud image collection.

[0093] This process is called data augmentation, and the enhanced satellite cloud image is added to the satellite cloud image collection; the purpose of this step is to overcome the problem of insufficient images in our image library.

[0094] The key to the lightweight satellite cloud image precipitation detection method is that the termination iteration threshold is 3000 times.

[0095] Continuously take enhanced images from the satellite cloud image collection for iteration. If the number of iterations reaches 3000, it will be terminated and the weight will be saved.

[0096] In this invention, we use a custom lightweight neural network model M-YOLOv4, whose backbone network selects MobileNetv1, and replaces all ordinary convolutions in the enhanced feature extraction network PANet with depthwise separable convolutions. The new enhanced feature extraction network is DW-PANet, which can provide better accuracy and faster speed; the final display of precipitation areas on the satellite cloud image is obtained by displaying the results of the M-YOLOv4 model inference on the unprocessed satellite cloud image.

[0097] The network structure of the M-YOLOv4 neural network model is as follows Figure 6 shown.

[0098] The M-YOLOv4 network architecture consists of three main components: the backbone, the neck, and the head. The backbone of M-YOLOv4 uses MobileNetV1 as the backbone network to extract features from satellite cloud images. It consists of a Conv_block and a DW_block. The Conv_block consists of a 3x3 convolution, a BN layer, and a Relu6 activation function. The DW_block consists of a 3x3 depthwise convolution, a BN layer, a Relu6, a 1x1 regular convolution, a BN layer, and a Relu6 activation function. The Neck part is composed of DW-PANet and SPP layers, which can enhance feature extraction. The SPP layer is processed by maximum pooling of four different scales. The pooling kernel sizes of maximum pooling are 13×13, 9×9, 5×5, and 1×1 (1×1 means no processing). DW-PANet consists of a series of convolution, pooling and upsampling layers. Its very important feature is the repeated feature extraction. The head uses YOLOv3head, which decodes the extracted features to obtain the predicted results. YOLOv3head consists of 3x3 convolution and 1x1 convolution. The function of 3x3 convolution is feature integration, and the function of 1x1 convolution is to adjust the number of channels.

[0099] The Conv_block consists of a 3×3 convolution, a Batch Normalization layer, and a Relu6 activation function; the DBM consists of a depthwise separable convolution (DWconv), a Batch Normalization layer, and a Mish activation function; and the DW_block consists of a depthwise separable convolution (DWconv), a Batch Normalization layer, a Relu6 activation function, a 1×1 convolution, a Batch Normalization layer, and a Relu6 activation function. The M-YOLOv4 neural network model initially uses the MobileNetv1 network as the backbone for feature extraction. Later, we adopted the SPP layer and customized the new DW-PANet to further increase the receptive field and enhance feature extraction capabilities.

[0100] The key to the lightweight satellite cloud image precipitation detection method is that the corresponding loss function of the M-YOLOv4 neural network model consists of three parts: category loss function, confidence loss function, and position loss function. The category loss function and confidence loss function use binary cross entropy loss function, and the position loss function uses CIOU loss function, as shown in formulas (1) and (2) respectively:

[0101]

[0102] y i Indicates the label of sample i, the positive class is 1 and the negative class is 0;

[0103] p i Indicates the probability of sample i predicting the positive class;

[0104]

[0105] IoU represents the intersection-over-union ratio between the predicted box and the real box, ρ 2 (b,b gt ) represents the Euclidean distance between the two center points, c represents the diagonal distance of the minimum closure area that can contain both the predicted box and the true box, α is the weight, and v is used to measure the similarity of the aspect ratio. Definition

[0106] Using this M-YOLOv4 neural network model and loss function, we can train the model using the augmented image input and the corresponding manually annotated label file, performing reverse derivation and continuous approximation to obtain the corresponding weight values. Our initial learning rate is 0.01, and with every 20% of the iterations, the learning rate decreases tenfold, reaching 0.0001 and then stops decreasing. This approach allows for faster and more accurate learning.

[0107] like Figure 3As shown in the figure, a lightweight satellite cloud image precipitation detection method is proposed. The key lies in:

[0108] The steps include:

[0109] Step C1: Obtain new satellite cloud image photos;

[0110] Step C2: performing an image threshold binarization operation on the satellite cloud image;

[0111] Step C3: Use the M-YOLOv4 neural network model and the final weight value to process the new satellite cloud image to generate the corresponding precipitation area;

[0112] Step C4: Display the precipitation area on the unprocessed satellite cloud image.

[0113] Using the above detection method settings, we perform a threshold binarization operation on the new satellite cloud image to highlight precipitation areas in the satellite cloud image and remove the interference of non-precipitating clouds. In this method, we use a custom lightweight neural network model M-YOLOv4 to achieve faster results and provide the precipitation areas to be detected in the satellite cloud image at one time.

[0114] like Figure 4 The figure shows a specific usage diagram of this method. Step 1 is to obtain a satellite cloud image dataset, step 2 is to perform an image threshold binarization operation on the satellite cloud image, step 3 is to perform image enhancement processing on the satellite cloud image dataset, step 4 is to use the custom lightweight neural network model M-YOLOv4 to infer the enhanced image, and step 5 is to display the inference result on the unprocessed satellite cloud image.

[0115] Due to limited storage and computing resources, applying neural network models to edge devices remains a significant challenge. To better utilize neural network models on edge devices, a growing interest is being placed on lightweight neural networks. MobileNetv1 reduces both parameter count and computational complexity by using depthwise separable convolutions and replacing Maxpooling and convolutional layers with convolutions with a stride of 2. This reduces computational complexity by approximately one-quarter and eliminates the need for Maxpooling. Using a 3×3 convolution kernel significantly reduces parameter count and significantly improves computational speed, enabling better application on edge devices.

[0116] The present invention has the following characteristics:

[0117] Feature 1: Data annotation of satellite cloud images through GPM precipitation distribution map;

[0118] Feature 2: We use image threshold binarization to pre-process the labeled satellite cloud images, so that our satellite cloud images become mostly precipitation clouds.

[0119] Feature 3: In the first stage, a neural network model is trained based on the satellite cloud image after image preprocessing, and the trained custom lightweight object detection algorithm M-YOLOv4 is used for inference to generate precipitation areas;

[0120] Feature 4: In this method, we use a custom lightweight neural network model, M-YOLOv4, to achieve faster results. Furthermore, we simultaneously display precipitation areas on satellite cloud images, improving accuracy.

[0121] Finally, it should be noted that the above examples are only specific implementation examples of the present invention. Of course, those skilled in the art can make changes and modifications to the present invention. If these modifications and variations fall within the scope of the claims of the present invention and their equivalent technologies, they should be considered to be within the scope of protection of the present invention.

Claims

1. A lightweight satellite cloud image neural network training method, characterized in that: Including satellite cloud image preprocessing process and neural network training process; The satellite cloud image preprocessing process includes the following steps: Step A1: establishing a satellite cloud image photo collection, inputting the satellite cloud image photos into the satellite cloud image photo collection; establishing a corresponding precipitation distribution atlas, and storing the corresponding GPM precipitation distribution map into the precipitation distribution atlas; Step A2: extracting satellite cloud image photos from the satellite cloud image photo collection in sequence; Step A3: Retrieve the corresponding GPM precipitation distribution map, which has the same date as the satellite cloud image; Step A4: Find the precipitation area in the GPM precipitation distribution map and manually mark all precipitation clouds in the corresponding satellite cloud image using the marking tool; Step A5: Using the image threshold binarization operation on the marked satellite cloud image to obtain the pre-processed satellite cloud image; Step A6: Determine whether there are any unlabeled satellite cloud images. If so, proceed to step A2. If not, end. The training process of a neural network includes the following steps: Step B1: Obtain a set of labeled satellite cloud image photos; Step B2: performing image enhancement processing on the satellite cloud image photos in the satellite cloud image photo set to obtain an enhanced satellite cloud image; Step B3: Build a custom M-YOLOv4 neural network model; Step B4: Obtain the loss function corresponding to the M-YOLOv4 neural network model; Step B5: Determine the threshold of the number of iterations of the loss function; Step B6: Extract satellite cloud image images from the satellite cloud image collection in sequence, input the satellite cloud image images into the M-YOLOv4 neural network model, and obtain prediction results; Step B7: Calculate the corresponding loss function value based on the manual labeling results and the corresponding prediction results, train the M-YOLOv4 neural network model, and obtain the corresponding weight value; Step B8: Determine whether the number of iterations of the loss function is greater than the iteration number threshold; If yes, go to step B9, otherwise go to step B6; Step B9: Save the weight value as the final weight value; The customized M-YOLOv4 neural network model in step B3 includes: Input module, Backbone module, Neck module and Head module; The Input module is used to obtain satellite cloud images; the Backbone module is used to extract shallow image features from the satellite cloud images input by the Input module; the Neck module enhances the shallow image features extracted by the Backbone module; the Head module obtains the image features enhanced by the Neck module and decodes these features to obtain the output results. Among them, the Backbone module uses the MobileNetv1 network as the backbone feature extraction network; the Neck module uses the depthwise separable convolution DW-PANet as the feature fusion module, and the Neck module also uses the SPP layer as an additional module; The Conv_block module of the Backbone module is composed of a 3×3 convolution, a Batch Normalization layer, and a Relu6 activation function; the DW_block module of the Backbone module is composed of a depthwise separable convolution Dwconv, a BatchNormalization layer, a Relu6 activation function, a 1×1 convolution, a Batch Normalization layer, and a Relu6 activation function; The DBM module of the Backbone module consists of a depthwise separable convolution Dwconv, a Batch Normalization layer, and a Mish activation function; The enhanced feature extraction network uses deep separable convolution DW-PANet and spp layers to further increase the receptive field.

2. The lightweight satellite cloud image neural network training method according to claim 1 is characterized in that: Step A5 includes: Formula (1) is used to perform image threshold binarization operation on the marked satellite cloud image photos; ； in, is the original pixel value in the satellite cloud image, is the pixel threshold; It is the pixel value after the image threshold binarization operation.

3. The lightweight satellite cloud image neural network training method according to claim 1 is characterized in that: The step B2 includes: Step B21: horizontally flipping the satellite cloud image in the satellite cloud image photo set; Step B22: Mix the horizontally flipped and non-horizontally flipped satellite cloud image photos, randomly select 4 images from them, and randomly crop and splice the 4 images at random positions to form a new image.

4. The lightweight satellite cloud image neural network training method according to claim 1, characterized in that: The loss function of the M-YOLOv4 neural network model includes a category loss function, a confidence loss function, and a position loss function; the category loss function and the confidence loss function adopt a binary cross entropy loss function. ; ； Where N represents the number of samples; =1-N; Represents a sample The label of the positive class is 1 and the negative class is 0; Represents a sample Predict the probability of the positive class; Position loss function Use CIOU loss function; ； Represents the intersection-over-union ratio of the predicted box and the true box, represents the Euclidean distance between two center points, Represents the diagonal distance of the minimum closure area that can contain both the predicted box and the true box. is the weight, v is used to measure the similarity of aspect ratio, and is defined ; Among them, b is the prediction box, b gt is the real frame; and h are the width and height of the prediction box respectively; and are the width and height of the ground truth box respectively.

5. The lightweight satellite cloud image neural network training method according to claim 1 is characterized in that: The iteration number threshold is 3000 times.

6. A lightweight satellite cloud image rainfall detection method, characterized by: The steps include: Step C1: Obtain satellite cloud image photos; Step C2: performing an image threshold binarization operation on the satellite cloud image; Step C3: using the M-YOLOv4 neural network model and final weight values obtained by the lightweight satellite cloud image neural network training method according to any one of claims 1 to 5 to process the satellite cloud image photos and generate corresponding precipitation areas; Step C4: Display the precipitation area on the corresponding unprocessed satellite cloud image.

Images