Satellite cloud picture automatic identification and analysis method and system based on GoogLeNet model
Through the automatic identification and analysis method based on the GoogleLeNet model, the multi-scale feature extraction is performed using median filtering and Inception modules, and combined with multiple auxiliary classifiers, the efficiency and accuracy problems in automatic interpretation and classification of satellite cloud maps are solved, and the efficiency and accuracy problems are realized. Efficient and accurate cloud map data processing.
Patent Information
- Application Number
- CN202411573659.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-06
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art is difficult to process large-scale, high-resolution data in the automatic interpretation and classification of satellite cloud maps, and is susceptible to noise and cloud complexity, resulting in insufficient efficiency and accuracy.
Using automatic identification and analysis method based on GoogleLeNet model, the median filtering process is used to reduce noise, use the Inception module to perform multi-scale feature extraction, and introduce multiple auxiliary classifiers to reduce overfitting, and build a deep convolutional neural network for feature extraction. and classification.
It significantly improves the automatic interpretation and classification efficiency of satellite cloud maps, enhances the recognition accuracy of complex weather systems, reduces manual intervention, and is suitable for cloud map data of multiple sources and different resolutions, improving the robustness and adaptability of the system. sex.
Smart Images

Figure CN119091323A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image recognition, and in particular to a method and system for automatically recognizing and analyzing satellite cloud images based on a GoogLeNet model. Background Art
[0002] With the rapid development of remote sensing technology, satellite cloud images have become an indispensable data source in the fields of weather forecasting, environmental monitoring and disaster warning. However, the acquisition and processing of traditional satellite cloud image data are highly dependent on manual interpretation, which is not only inefficient but also easily affected by subjective factors, resulting in uneven data quality. In recent years, the rise of computer vision and deep learning technology has provided a new way for the automatic interpretation of satellite cloud images, significantly improving the efficiency and accuracy of data acquisition.
[0003] At present, although some algorithms can realize the automatic interpretation of satellite cloud images, there are still challenges in efficiently obtaining high-quality cloud image data. On the one hand, the large amount of satellite cloud image data and high resolution put forward higher requirements for the processing speed and accuracy of the algorithm; on the other hand, cloud image data often contains noise, overlapping clouds and other complex situations, which increases the difficulty of automatic interpretation.
[0004] How to use advanced deep learning models to efficiently and accurately automatically interpret and classify large amounts of satellite cloud images has become a key issue that needs to be solved urgently. Summary of the invention
[0005] The technical problem to be solved and the technical task proposed by the present invention are to improve and perfect the existing technical solutions, and provide a method for automatic recognition and analysis of satellite cloud images based on the GoogLeNet model, so as to improve the efficiency and accuracy of interpretation and classification of satellite cloud image data. To this end, the present invention adopts the following technical solutions.
[0006] A method for automatically identifying and analyzing satellite cloud images based on the GoogLeNet model includes the following steps: A method for automatically identifying and analyzing satellite cloud images based on the GoogLeNet model includes the following steps: 1) Collect and integrate raw cloud image data from multiple satellite sources; 2) Preprocess the original cloud image data and use median filtering technology to smooth the data and reduce noise; 3) Segment the preprocessed satellite cloud image and unify the scale of each pixel. At the same time, normalize the size of the cloud image data to keep the size of all images consistent; normalize the pixel values of the image and scale the pixel values to between 0 and 1; 4) Constructing an automatic interpretation and classification model based on GoogLeNet, the model includes: Inception module, used for multi-scale feature extraction; The network structure for feature extraction and classification is established by convolutional layers, pooling layers, and fully connected layers; A main classifier and at least two auxiliary classifiers, the auxiliary classifiers are located at different depths of the network; the outputs of the auxiliary classifiers are added to the total loss during training to provide additional gradient paths and reduce overfitting by superimposing multiple classification losses; The preprocessed and normalized cloud image data is input into the automatic interpretation and classification model as a training set, and the model is trained; 5) After the model training is completed, the new satellite cloud image data is input into the trained automatic interpretation and classification model, and the output layer of the automatic interpretation and classification model generates the final qualitative analysis results.
[0007] This technical solution uses the GoogLeNet model and the Inception module in the GoogLeNet model to achieve automatic interpretation and classification of large-scale satellite cloud images, effectively reducing the workload of manual analysis and improving the efficiency of recognition and analysis. It is particularly suitable for processing tasks that require rapid processing, such as real-time weather forecasts and climate monitoring. Using the Inception module in GoogLeNet, multi-scale feature extraction of cloud image data can capture local and global features in the image, improve the recognition accuracy of complex weather systems, and adapt to the complex features of different types of clouds. In the preprocessing process, the median filtering technology is used to effectively smooth the cloud image data, reduce noise interference, enhance the robustness of the model, and make the classification results more accurate. By introducing auxiliary classifiers in the GoogLeNet model, additional gradient paths are provided during the training process, and the loss function is optimized to avoid overfitting of the model, thereby improving the generalization ability of the model on new data. A deep convolutional neural network is used to extract and classify features through convolutional layers, pooling layers, and fully connected layers, which can provide high-precision cloud image classification results, especially suitable for the fine classification of complex clouds. From data collection, preprocessing, interpretation to the output of classification results, the automation level of satellite cloud image interpretation has been significantly improved, the reliance on manual intervention has been reduced, and a large amount of real-time satellite data can be better processed. This technical solution can be applied to satellite cloud image data from multiple sources and to cloud image images of different resolutions and formats, and has strong versatility and adaptability.
[0008] As a preferred technical means: in step 2), the cloud image is smoothed by using median filtering technology, the window is set to 3×3, and the median filtering process is implemented by the following function:
[0009] Among them, G(x,y), f(x,y) are the processed image and the original image respectively, and W is the template window.
[0010] This technical solution uses median filtering technology to achieve efficient and accurate cloud image smoothing, laying the foundation for the performance improvement of subsequent interpretation and classification models. Median filtering can effectively remove random noise in images, especially strong interference noise such as "salt and pepper noise"; for satellite cloud images, since satellite images are susceptible to transmission noise and other interference, median filtering can better maintain the clarity and accuracy of cloud image data. Compared with other filtering methods (such as mean filtering), median filtering can better protect the edge and detail features in the image while removing noise. For key features such as cloud boundaries in satellite cloud images, median filtering can effectively retain edge information and avoid blurring that affects cloud interpretation. The median filter calculation is relatively simple. The use of a 3×3 window size means that each pixel only needs to be compared with the surrounding 8 pixels and take the median. The calculation complexity is low, which is suitable for the rapid processing of large-scale cloud image data, improves the processing speed of the system, and can cope with the real-time processing needs of large quantities of satellite data. The smoothed cloud image processed by median filtering is input into the subsequent GoogLeNet convolutional neural network, which can reduce the interference of image noise on model training and prediction, allowing the model to focus more on important features, thereby improving the accuracy of interpretation and classification. Median filtering technology can be applied to multi-source satellite cloud image data of different resolutions and formats, improving the robustness and reliability of the solution under diverse data conditions.
[0011] As a preferred technical means: in step 3), the filtered cloud image data is adjusted to the input size required by the automatic interpretation and classification model through the interpolation method, and the image data is normalized. The normalization formula is:
[0012] Among them, x is the original pixel value, is the mean, is the standard deviation, and x' is the normalized pixel value.
[0013] By adjusting the image size through interpolation methods, it can be ensured that all input images meet the input requirements of the automatic interpretation and classification model, thereby ensuring the accuracy of subsequent analysis and classification. Normalization processing scales pixel values to a fixed range, reduces numerical fluctuations during the calculation process, improves computational efficiency, and speeds up model training. Normalization processing can also reduce the impact of noise and outliers in image data on the model, making the model more robust, the data more numerically stable, and less susceptible to interference from extreme values or noise. Normalized image data is more consistent and comparable during feature extraction and classification, improving the classification performance of the model, eliminating numerical differences between different images, and reducing the complexity of the model.
[0014] As a preferred technical means: in step 4), a convolution operation is performed on the image using multiple convolution kernels of different sizes to extract local features. The formula of the convolution operation is:
[0015] Among them, y ij is the value of the i-th row and j-th column of the output feature map, M and N are the number of rows and columns of the convolution kernel, and w mn is the weight of the convolution kernel, x i+m-1,j+n-1 is the pixel value of the input image or the feature map of the previous layer, and b is the bias term.
[0016] This technical solution uses multiple convolution kernels of different sizes to capture features of different scales and shapes in the input image or feature map, enhance the model's adaptability to changes in input data, and improve the model's generalization ability. By changing the size and step size of the convolution kernel, the dimension of the input data can be reduced, thereby achieving feature dimensionality reduction, reducing computational complexity, and improving computational efficiency. The size of the feature map can be further reduced by stacking multiple convolution layers and using pooling operations, thereby accelerating the model's training and reasoning process. Using multiple convolution kernels of different sizes can increase the complexity of the model without adding too many parameters, preventing overfitting to a certain extent.
[0017] As a preferred technical means: the output layer in step 4) includes the following: Main classifier: The height and width of each channel of the feature map are changed to 1 through the global average pooling layer, and then the final classification result is output through the fully connected layer. The fully connected layer has multiple output units corresponding to different categories; Auxiliary classifiers: Located at different depths of the network, including average pooling layers, 1x1 convolutional layers, fully connected layers, and softmax layers, the output of the auxiliary classifiers is added to the total loss during training to provide additional gradient paths and prevent overfitting.
[0018] The main classifier compresses the height and width of each channel of the feature map to 1 through the global average pooling layer, which helps to reduce the number of parameters, avoid overfitting, and retain global feature information. The fully connected layer then outputs the final classification result, and its multiple output units correspond to different categories, allowing the model to distinguish multiple categories and improve the accuracy of classification.
[0019] The auxiliary classifiers are located at different depths of the network and can capture feature information at different levels, which helps the model learn richer feature representations. The output of the auxiliary classifiers is added to the total loss during training, providing an additional gradient path for the model, which helps the model better adjust parameters during training, thereby enhancing the generalization ability of the model.
[0020] The introduction of auxiliary classifiers increases the complexity of the model, but also provides more regularization effects for the model. By optimizing the loss of the main classifier and the auxiliary classifier at the same time during the training process, the model can better learn the essential characteristics of the data and reduce the risk of overfitting. The use of auxiliary classifiers can be used as an effective regularization method to help the model reduce overfitting while maintaining high performance, so that the model can converge faster during training, and because they provide additional gradient paths, the model can find the optimal solution faster during training. In addition, auxiliary classifiers can also be used as a monitoring method during the model training process. By observing the output of the auxiliary classifier, the model's feature learning ability at different levels can be judged, so that the training strategy can be adjusted in time.
[0021] As a preferred technical means: Step 4) flatten the feature map into a one-dimensional vector, and perform further feature transformation and dimensionality reduction through a fully connected layer; convert the output of the fully connected layer into a probability distribution, and the probability of each category represents the possibility that the input image belongs to that category. The formula of the Softmax function is as follows:
[0022] Where p_k is the predicted probability of the kth category, z_k is the output of the fully connected layer for the kth category, and k is the total number of categories.
[0023] After the processing of the first few convolutional layers and pooling layers, the feature maps are gradually reduced and high-dimensional dependent features are extracted. These feature maps contain high-level information of the input image, such as edges, textures, and shapes. After being flattened into a one-dimensional vector, they can contain important information of the entire image. The flattened feature map undergoes further feature transformation and dimensionality reduction through the fully connected layer, compressing the high-dimensional features into a compact low-dimensional representation while retaining important feature information, which helps to reduce computational complexity and improve computational efficiency; the Softmax function normalizes the output of all categories into a probability distribution, which enables the model to compare the output values between different categories and further determine which category the image is most likely to belong to. The cross entropy loss function using the Softmax function gives a larger penalty for misclassification during training, which helps to have a clearer gradient signal, thereby increasing the convergence speed of training and improving the performance and stability of the model. It has the advantages of efficient feature extraction and information retention, helps to reduce computational complexity and improve computational efficiency, provides intuitive and easy-to-interpret probability output, and helps to optimize and train stability, and is suitable for multi-category classification tasks.
[0024] As a preferred technical means: During the training process, the outputs of the main classifier and the auxiliary classifier are both calculated for loss, and the network weights are updated through back propagation; the loss function uses cross entropy loss to calculate the correctness of the classification; the loss function is:
[0025] Among them, L main is the loss of the main classifier, L aux1 and L aux2 are the losses of the two auxiliary classifiers, and λ1 and λ2 are the weights of the auxiliary classifier losses.
[0026] By adding auxiliary classifiers to the middle layers of the network, loss signals can be introduced earlier in the network, thereby alleviating the gradient vanishing problem; the gradient information is not only propagated through the output of the main classifier, but also propagated through the auxiliary classifiers in the shallower layers of the network, which helps to update the weights more evenly. The auxiliary classifier provides additional feedback signals, allowing each layer of the network to participate in effective training earlier, thereby accelerating the convergence of the overall network. The network can learn useful features faster, thereby shortening the training time. By introducing auxiliary classifiers during the training process, additional regularization effects can be provided to the network to prevent the model from overfitting. In addition, the nature of multi-task learning enables the model to generalize better to unseen data. With the help of auxiliary classifiers, additional supervision signals can be provided at multiple levels, thereby enhancing the stability of training. By adjusting the weights of the auxiliary loss (λ1 and λ2), the proportion of the auxiliary classifier in the overall loss can be flexibly controlled, making the scheme highly adaptable to different tasks and data sets.
[0027] Another object of the present invention is to provide a satellite cloud image automatic recognition and analysis system based on the GoogLeNet model, a satellite cloud image automatic recognition and analysis system based on the GoogLeNet model comprises: Data collection and integration module, used to collect and integrate raw cloud image data from multiple source satellites; Data preprocessing module, used to preprocess the original cloud image data, including using median filtering technology to smooth the data and reduce noise; Image segmentation and normalization module, which is used to segment the preprocessed satellite cloud image, unify the scale of each pixel, normalize the size of the cloud image data, and scale the pixel value of the image to between 0 and 1; Automatic interpretation and classification module, including: A model building unit, used to build an automatic interpretation and classification model based on GoogLeNet, which includes an Inception module, a multi-level convolutional layer, a pooling layer and a fully connected layer, as well as a main classifier and at least two auxiliary classifiers located at different depths; A model training unit, which is used to input the preprocessed and normalized cloud image data as a training set into the automatic interpretation and classification model for training, wherein the output of the auxiliary classifier is added to the total loss during the training process to provide an additional gradient path and reduce overfitting; Analysis result generation module: After the model training is completed, it is used to input the new satellite cloud image data into the trained automatic interpretation and classification model. The output layer of the automatic interpretation and classification model generates the final qualitative analysis results.
[0028] The data collection and integration module can collect and integrate raw cloud image data from multiple satellite sources. This integration of multi-source data can provide more comprehensive and richer cloud image information, thereby improving the accuracy of subsequent analysis and interpretation.
[0029] The data preprocessing module uses median filtering technology to smooth data and reduce noise, which helps to improve the quality of cloud image data and lay a good foundation for subsequent image segmentation and normalization.
[0030] The image segmentation and normalization module segments the preprocessed satellite cloud images, unifies the scale of each pixel, and normalizes the size of the cloud image data. This standardization process enables the model to process cloud image data of different sizes and resolutions, improving the generalization ability of the model.
[0031] The automatic interpretation and classification module uses a model architecture based on GoogLeNet, which includes an Inception module, multi-level convolutional layers, pooling layers, and fully connected layers, as well as a main classifier and at least two auxiliary classifiers at different depths. This complex model architecture can capture multi-level features in cloud image data and improve classification accuracy.
[0032] The output of the auxiliary classifier is added to the total loss during training to provide additional gradient paths and reduce overfitting. This strategy helps the model better learn data features during training while avoiding overfitting the training data, thereby improving the generalization performance of the model.
[0033] As a preferred technical means: In the data preprocessing module, the median filter processing technology is used to smooth the cloud image, the window is set to 3×3, and the median filter processing is implemented by the following function:
[0034] Among them, G(x,y), f(x,y) are the processed image and the original image respectively, and W is the template window.
[0035] As a preferred technical means: in the automatic interpretation and classification module, the automatic interpretation and classification model constructed by the model construction unit: The main classifier outputs the final classification result through the global average pooling layer and the fully connected layer. The fully connected layer has multiple output units corresponding to different categories. Auxiliary classifiers are located at different depths of the network, including average pooling layers, 1x1 convolutional layers, fully connected layers, and softmax layers, to provide additional gradient paths during training and prevent overfitting; During the training process, the cross entropy loss function is used to calculate the loss of the main classifier and the auxiliary classifier, and the network weights are updated by back propagation. The loss function is:
[0036] Among them, L main is the loss of the main classifier, L aux1 and L aux2 are the losses of the two auxiliary classifiers, and λ1 and λ2 are the weights of the auxiliary classifier losses.
[0037] Beneficial effects: This technical solution significantly improves the efficiency and quality of data acquisition, and quickly obtains high-quality cloud images through deep learning automatic interpretation. At the same time, it strengthens the pre-processing process, effectively removes noise and filters, and reduces interference. Combines data mining and machine learning to optimize data and improve prediction accuracy. Overall, the present invention optimizes the processing flow, improves the accuracy of pre-judgment, and provides stronger technical support for meteorological forecasting. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of the process of the present invention.
[0039] Figure 2 It is a schematic diagram of the Inception module structure of the present invention. DETAILED DESCRIPTION
[0040] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.
[0041] Embodiment 1: Figure 1 As shown, the present invention comprises the following steps: S1: Collect and integrate raw cloud image data from multiple source satellites; Satellite cloud images can be divided into many types according to different observation bands and imaging technologies. Common satellite cloud images include: visible light cloud images, infrared cloud images, water vapor cloud images, microwave cloud images, multi-spectral cloud images and radar cloud images.
[0042] Collecting and integrating multi-source satellite cloud images can make up for the time and space deficiencies of single satellite observations and provide more continuous and comprehensive data coverage. It greatly expands the capacity and scope of meteorological observations, helps to provide more accurate and comprehensive meteorological information, and improves the quality and reliability of weather forecasts and climate research.
[0043] In this embodiment, three different types of satellite cloud images are collected, namely, visible light satellite cloud images, infrared satellite cloud images and water vapor satellite cloud images.
[0044] S2: Preprocess the original cloud image data and use median filtering technology to smooth the data and reduce noise; The cloud image is smoothed using median filtering technology. The window is set to 3×3 and the median filtering is implemented using the following function:
[0045] Among them, G(x,y), f(x,y) are the processed image and the original image respectively, and W is the template window.
[0046] By adopting median filtering technology, efficient and accurate cloud image smoothing processing is achieved, laying the foundation for improving the performance of subsequent interpretation and classification models.
[0047] S3: Segment the preprocessed satellite cloud image, unify the scale of each pixel, and normalize the size of the cloud image data to keep the size of all images consistent; normalize the pixel values of the image and scale the pixel values to between 0 and 1; form a satellite cloud image data set.
[0048] Among them, the filtered cloud image data is adjusted to the input size required by the automatic interpretation and classification model through the interpolation method, and the image data is normalized. The normalization formula is:
[0049] Among them, x is the original pixel value, is the mean, is the standard deviation, and x' is the normalized pixel value.
[0050] Adjusting the image size and normalizing it through interpolation methods is beneficial to improving model compatibility, computational efficiency, robustness, classification performance and simplifying the model, providing strong support for the automatic recognition and analysis of satellite cloud images.
[0051] In this embodiment, various satellite cloud images are segmented and classified, and the pixel ratio is unified to 50*50.
[0052] S4: Construct an automatic interpretation and classification model based on GoogLeNet, the model includes: Inception module, such as Figure 2 As shown, it is used for multi-scale feature extraction; The network structure for feature extraction and classification is established by convolutional layers, pooling layers, and fully connected layers; A main classifier and at least two auxiliary classifiers, the auxiliary classifiers are located at different depths of the network; the outputs of the auxiliary classifiers are added to the total loss during training to provide additional gradient paths and reduce overfitting by superimposing multiple classification losses; The preprocessed and normalized cloud image data is input into the automatic interpretation and classification model as a training set, and the model is trained; The high-level features of the image are gradually abstracted through structures such as convolution layers and pooling layers. Multiple convolution kernels of different sizes are used to perform convolution operations on the image to extract local features. The formula for the convolution operation is:
[0053] Among them, y ij is the value of the i-th row and j-th column of the output feature map, M and N are the number of rows and columns of the convolution kernel, and w mn is the weight of the convolution kernel, x i+m-1,j+n-1 is the pixel value of the input image or the feature map of the previous layer, and b is the bias term.
[0054] The local features in the cloud image are extracted through the convolution operation, and the pooling layer is used for downsampling to reduce the feature dimension. At the same time, the feature detection is made more robust, and multi-scale feature extraction is performed on the cloud image. The multi-scale feature extraction can be cloud height, cloud thickness, cloud type, cloud brightness temperature value, cloud brightness temperature difference, cloud color; it can also extract information such as the area and thickness of cloud coverage from visible light satellite cloud images, extract information such as cloud height and cloud top temperature from infrared satellite cloud images, and extract information such as water vapor distribution and humidity from water vapor satellite cloud images.
[0055] The output layer of the network consists of the following: Main classifier: The height and width of each channel of the feature map are changed to 1 through the global average pooling layer, and then the final classification result is output through the fully connected layer. The fully connected layer has multiple output units corresponding to different categories; Two auxiliary classifiers: Located at different depths of the network, including average pooling layer, 1x1 convolution layer, fully connected layer and softmax layer, the output of the auxiliary classifier is added to the total loss during training to provide additional gradient paths and prevent overfitting.
[0056] The main classifier compresses the height and width of each channel of the feature map to 1 through the global average pooling layer, which helps to reduce the number of parameters, avoid overfitting, and retain global feature information. The fully connected layer then outputs the final classification result, and its multiple output units correspond to different categories, allowing the model to distinguish multiple categories and improve the accuracy of classification.
[0057] The auxiliary classifiers are located at different depths of the network and can capture feature information at different levels, which helps the model learn richer feature representations. The output of the auxiliary classifiers is added to the total loss during training, providing an additional gradient path for the model, which helps the model better adjust parameters during training, thereby enhancing the generalization ability of the model.
[0058] The feature map is flattened into a one-dimensional vector, and further feature transformation and dimensionality reduction are performed through a fully connected layer. The output of the fully connected layer is converted into a probability distribution. The probability of each category represents the possibility that the input image belongs to that category. The formula of the Softmax function is as follows:
[0059] Where p_k is the predicted probability of the kth category, z_k is the output of the fully connected layer for the kth category, and k is the total number of categories.
[0060] Loss function calculation: During the training process, the output of the main classifier and the auxiliary classifier will calculate the loss, and the network weights will be updated through back propagation. In this embodiment, the loss function uses cross-entropy loss to calculate the correctness of the classification. For the main classifier and the auxiliary classifier, the loss function can be expressed as:
[0061] Among them, L main is the loss of the main classifier, L aux1 and L aux2 are the losses of the two auxiliary classifiers, and λ1 and λ2 are the weights of the auxiliary classifier losses.
[0062] S5: After the model training is completed, the new satellite cloud image data is input into the trained automatic interpretation and classification model, and the output layer of the automatic interpretation and classification model generates the final qualitative analysis results.
[0063] The qualitative results of the output are shown in the following table:
[0064] Note: Clear (0-1), few clouds (2-3), cloudy (4-7), cloudy to overcast (8-9), cloudy with rain and fog (10); clear sky (A), stratocumulus (B), cumulonimbus (C), dense stratus (D), cirrostratus (E); It can be seen that compared with the actual situation, the accuracy of the qualitative information of the cloud map finally output by this method can reach more than 95%.
[0065] Of course, as satellite images continue to change, in order to maintain the accuracy of the model, the performance of the model can be further evaluated and optimized to ensure the accuracy and reliability of the results. Based on the evaluation results, the model can be further optimized and adjusted to improve its performance.
[0066] This technical solution uses the GoogLeNet model and the Inception module in the GoogLeNet model to achieve automatic interpretation and classification of large-scale satellite cloud images, effectively reducing the workload of manual analysis and improving the efficiency of recognition and analysis. It is particularly suitable for processing tasks that require rapid processing, such as real-time weather forecasts and climate monitoring. Using the Inception module in GoogLeNet, multi-scale feature extraction of cloud image data can capture local and global features in the image, improve the recognition accuracy of complex weather systems, and adapt to the complex features of different types of clouds. In the preprocessing process, the median filtering technology is used to effectively smooth the cloud image data, reduce noise interference, enhance the robustness of the model, and make the classification results more accurate. By introducing auxiliary classifiers in the GoogLeNet model, additional gradient paths are provided during the training process, and the loss function is optimized to avoid overfitting of the model, thereby improving the generalization ability of the model on new data. A deep convolutional neural network is used to extract and classify features through convolutional layers, pooling layers, and fully connected layers, which can provide high-precision cloud image classification results, especially suitable for the fine classification of complex clouds. From data collection, preprocessing, interpretation to the output of classification results, the automation level of satellite cloud image interpretation has been significantly improved, the reliance on manual intervention has been reduced, and a large amount of real-time satellite data can be better processed. This technical solution can be applied to satellite cloud image data from multiple sources and to cloud image images of different resolutions and formats, and has strong versatility and adaptability.
[0067] Embodiment 2: A satellite cloud image automatic recognition and analysis system based on the GoogLeNet model, a satellite cloud image automatic recognition and analysis system based on the GoogLeNet model includes: Data collection and integration module, used to collect and integrate raw cloud image data from multiple source satellites; Data preprocessing module, used to preprocess the original cloud image data, including using median filtering technology to smooth the data and reduce noise; Image segmentation and normalization module, which is used to segment the preprocessed satellite cloud image, unify the scale of each pixel, normalize the size of the cloud image data, and scale the pixel value of the image to between 0 and 1; Automatic interpretation and classification module, including: A model building unit, used to build an automatic interpretation and classification model based on GoogLeNet, which includes an Inception module, a multi-level convolutional layer, a pooling layer and a fully connected layer, as well as a main classifier and at least two auxiliary classifiers located at different depths; A model training unit, which is used to input the preprocessed and normalized cloud image data as a training set into the automatic interpretation and classification model for training, wherein the output of the auxiliary classifier is added to the total loss during the training process to provide an additional gradient path and reduce overfitting; Analysis result generation module: After the model training is completed, it is used to input the new satellite cloud image data into the trained automatic interpretation and classification model. The output layer of the automatic interpretation and classification model generates the final qualitative analysis results.
[0068] The data collection and integration module can collect and integrate raw cloud image data from multiple satellite sources. This integration of multi-source data can provide more comprehensive and richer cloud image information, thereby improving the accuracy of subsequent analysis and interpretation.
[0069] The data preprocessing module uses median filtering technology to smooth data and reduce noise, which helps to improve the quality of cloud image data and lay a good foundation for subsequent image segmentation and normalization.
[0070] The image segmentation and normalization module segments the preprocessed satellite cloud images, unifies the scale of each pixel, and normalizes the size of the cloud image data. This standardization process enables the model to process cloud image data of different sizes and resolutions, improving the generalization ability of the model.
[0071] The automatic interpretation and classification module uses a model architecture based on GoogLeNet, which includes an Inception module, multi-level convolutional layers, pooling layers, and fully connected layers, as well as a main classifier and at least two auxiliary classifiers at different depths. This complex model architecture can capture multi-level features in cloud image data and improve classification accuracy.
[0072] The output of the auxiliary classifier is added to the total loss during training to provide additional gradient paths and reduce overfitting. This strategy helps the model better learn data features during training while avoiding overfitting the training data, thereby improving the generalization performance of the model.
[0073] In the data preprocessing module, the median filter processing technology is used to smooth the cloud image. The window is set to 3×3, and the median filter processing is implemented by the following function:
[0074] Among them, G(x,y), f(x,y) are the processed image and the original image respectively, and W is the template window.
[0075] In the automatic interpretation and classification module, the automatic interpretation and classification model constructed by the model building unit: The main classifier outputs the final classification result through the global average pooling layer and the fully connected layer. The fully connected layer has multiple output units corresponding to different categories. Auxiliary classifiers are located at different depths of the network, including average pooling layers, 1x1 convolutional layers, fully connected layers, and softmax layers, to provide additional gradient paths during training and prevent overfitting; During the training process, the cross entropy loss function is used to calculate the loss of the main classifier and the auxiliary classifier, and the network weights are updated by back propagation. The loss function is:
[0076] Among them, L main is the loss of the main classifier, L aux1 and L aux2 are the losses of the two auxiliary classifiers, and λ1 and λ2 are the weights of the auxiliary classifier losses.
[0077] It is understandable that the detailed functional implementation of each of the above modules can be found in the introduction of the aforementioned method embodiment, and no further elaboration is made here.
[0078] The satellite cloud image automatic recognition and analysis method and system based on the GoogLeNet model shown above are specific embodiments of the present invention, which have embodied the substantial characteristics and progress of the present invention. According to actual use needs, under the guidance of the present invention, equivalent modifications can be made to them, which are all within the protection scope of this solution.
Claims
1. A method for automatic recognition and analysis of satellite cloud images based on the GoogLeNet model, characterized in that The following steps are involved: 1) Collect and integrate raw cloud image data from multiple satellite sources; 2) Preprocess the original cloud image data and use median filtering technology to smooth the data and reduce noise; 3) Segment the preprocessed satellite cloud image and unify the scale of each pixel. At the same time, normalize the size of the cloud image data to keep the size of all images consistent; Normalize the pixel values of the image and scale the pixel values to between 0 and 1; 4) Constructing an automatic interpretation and classification model based on GoogLeNet, the model includes: Inception module, used for multi-scale feature extraction; The network structure for feature extraction and classification is established by convolutional layers, pooling layers, and fully connected layers; A main classifier and at least two auxiliary classifiers, the auxiliary classifiers are located at different depths of the network; the outputs of the auxiliary classifiers are added to the total loss during training to provide additional gradient paths and reduce overfitting by superimposing multiple classification losses; The preprocessed and normalized cloud image data is input into the automatic interpretation and classification model as a training set, and the model is trained; 5) After the model training is completed, the new satellite cloud image data is input into the trained automatic interpretation and classification model, and the output layer of the automatic interpretation and classification model generates the final qualitative analysis results.
2. The method for automatic satellite cloud image recognition and analysis based on the GoogLeNet model according to claim 1, characterized in that: In step 2), the cloud image is smoothed using median filtering technology, the window is set to 3×3, and the median filtering process is implemented by the following function: Among them, G(x,y), f(x,y) are the processed image and the original image respectively, and W is the template window.
3. The method for automatic satellite cloud image recognition and analysis based on the GoogLeNet model according to claim 2, characterized in that: In step 3), the filtered cloud image data is adjusted to the input size required by the automatic interpretation and classification model through interpolation method, and the image data is normalized. The normalization formula is: Among them, x is the original pixel value, is the mean, is the standard deviation, and x' is the normalized pixel value.
4. The method for automatic satellite cloud image recognition and analysis based on the GoogLeNet model according to claim 3, characterized in that: In step 4), a convolution operation is performed on the image using multiple convolution kernels of different sizes to extract local features. The formula for the convolution operation is: Among them, y ij is the value of the i-th row and j-th column of the output feature map, M and N are the number of rows and columns of the convolution kernel, and w mn is the weight of the convolution kernel, x i+m-1,j+n-1 is the pixel value of the input image or the feature map of the previous layer, and b is the bias term.
5. The method for automatic satellite cloud image recognition and analysis based on the GoogLeNet model according to claim 4, characterized in that: The output layer in step 4) consists of the following: Main classifier: The height and width of each channel of the feature map are changed to 1 through the global average pooling layer, and then the final classification result is output through the fully connected layer. The fully connected layer has multiple output units corresponding to different categories; Auxiliary classifiers: Located at different depths of the network, including average pooling layers, 1x1 convolutional layers, fully connected layers, and softmax layers, the output of the auxiliary classifiers is added to the total loss during training to provide additional gradient paths and prevent overfitting.
6. The method for automatic satellite cloud image recognition and analysis based on the GoogLeNet model according to claim 5, characterized in that: Step 4) Flatten the feature map into a one-dimensional vector and perform further feature transformation and dimensionality reduction through a fully connected layer; convert the output of the fully connected layer into a probability distribution. The probability of each category represents the possibility that the input image belongs to that category. The formula of the Softmax function is as follows: Where p_k is the predicted probability of the kth category, z_k is the output of the fully connected layer for the kth category, and k is the total number of categories.
7. The method for automatic satellite cloud image recognition and analysis based on the GoogLeNet model according to claim 6, characterized in that: During the training process, the outputs of the main classifier and the auxiliary classifier are both subjected to loss calculation, and the network weights are updated through back propagation; the loss function uses cross entropy loss to calculate the correctness of the classification; the loss function is: Among them, L main is the loss of the main classifier, L aux1 and L aux2 are the losses of the two auxiliary classifiers, and λ1 and λ2 are the weights of the auxiliary classifier losses.
8. A satellite cloud image automatic recognition and analysis system based on the GoogLeNet model, characterized by include: Data collection and integration module, used to collect and integrate raw cloud image data from multiple source satellites; Data preprocessing module, used to preprocess the original cloud image data, including using median filtering technology to smooth the data and reduce noise; Image segmentation and normalization module, which is used to segment the preprocessed satellite cloud image, unify the scale of each pixel, normalize the size of the cloud image data, and scale the pixel value of the image to between 0 and 1; Automatic interpretation and classification module, including: A model building unit, used to build an automatic interpretation and classification model based on GoogLeNet, which includes an Inception module, a multi-level convolutional layer, a pooling layer and a fully connected layer, as well as a main classifier and at least two auxiliary classifiers located at different depths; A model training unit, which is used to input the preprocessed and normalized cloud image data as a training set into the automatic interpretation and classification model for training, wherein the output of the auxiliary classifier is added to the total loss during the training process to provide an additional gradient path and reduce overfitting; Analysis result generation module: After the model training is completed, it is used to input the new satellite cloud image data into the trained automatic interpretation and classification model. The output layer of the automatic interpretation and classification model generates the final qualitative analysis results.
9. The satellite cloud image automatic recognition and analysis system based on the GoogLeNet model according to claim 8, characterized in that: In the data preprocessing module, the median filter processing technology is used to smooth the cloud image. The window is set to 3×3, and the median filter processing is implemented by the following function: Among them, G(x,y), f(x,y) are the processed image and the original image respectively, and W is the template window.
10. The satellite cloud image automatic recognition and analysis system based on the GoogLeNet model according to claim 8, characterized in that: In the automatic interpretation and classification module, the automatic interpretation and classification model constructed by the model building unit: The main classifier outputs the final classification result through the global average pooling layer and the fully connected layer. The fully connected layer has multiple output units corresponding to different categories. Auxiliary classifiers are located at different depths of the network, including average pooling layers, 1x1 convolutional layers, fully connected layers, and softmax layers, to provide additional gradient paths during training and prevent overfitting; During the training process, the cross entropy loss function is used to calculate the loss of the main classifier and the auxiliary classifier, and the network weights are updated by back propagation. The loss function is: Among them, L main is the loss of the main classifier, L aux1 and L aux2 are the losses of the two auxiliary classifiers, and λ1 and λ2 are the weights of the auxiliary classifier losses.
Citation Information
Patent Citations
Foundation cloud atlas cloud identification method based on convolutional neural network
CN110929602A
