Atmospheric river recognition method, system and equipment based on deep learning and medium

Abstract

The invention discloses an atmospheric river recognition method, system, equipment and medium based on deep learning, and belongs to the technical field of geophysics. The method comprises the following steps: judging whether an integrated water vapor conveying threshold value of each remote sensing image detection method exceeds a confidence coefficient threshold value requirement by analyzing the integrated water vapor conveying threshold value of each selected remote sensing image detection method; when a confidence threshold requirement is met, extracting a label of each remote sensing image detection method, and obtaining a region which is marked and filtered and is integrated with a water vapor conveying threshold as a high-credibility atmospheric river region; the atmospheric river distribution diagrams corresponding to all the remote sensing image detection methods are sequentially compared with the high-fidelity atmospheric river area distribution diagram, and when the atmospheric river distribution diagrams are overlapped with the high-fidelity atmospheric river area distribution diagram, the labels are updated to serve as a training set; and training the constructed deep learning neural network model by using the training set, and carrying out region identification on the original image data of the atmospheric river to be identified. The method can ensure the accuracy of model training, and the detection precision is higher.

Description

Technical Field

[0001] The present invention relates to the field of geophysical technology, and more specifically to an atmospheric river identification method, system, device and medium based on deep learning. Background Art

[0002] Atmospheric rivers are defined as long, extensive aerial water vapor transport pathways that are responsible for more than 90% of water vapor transport from mid-latitudes to polar regions. Studies have found that atmospheric rivers are responsible for more than 50% of precipitation in rainfall and floods in most mid-latitude regions. At the same time, if atmospheric rivers are missing, those regions that usually play an important role in annual precipitation will also experience severe drought disasters. Because atmospheric rivers have a significant impact on regional hydrological cycles, extreme precipitation and drought conditions, it is crucial to understand and detect atmospheric rivers.

[0003] Deep learning-based atmospheric river identification technology has been a hot topic in the fields of meteorology and hydrology in recent years. Traditionally, the identification of atmospheric rivers relies mainly on rule-based methods, which are usually defined using predetermined thresholds (such as the magnitude of water vapor flux) and geographical criteria (such as length and width). With the development of machine learning and deep learning technologies, researchers have begun to explore the use of these advanced algorithms to improve the detection and classification of atmospheric rivers. Deep learning models, especially convolutional neural networks (CNNs), are widely used in meteorological image analysis due to their powerful feature extraction capabilities and ability to automatically learn complex patterns. In addition, GAN and semi-supervised learning methods can effectively utilize a large amount of unlabeled meteorological data when data annotation is limited or the annotation cost is high. Through these methods, the model can learn to generate and identify the characteristics of atmospheric rivers with limited labeled data support, thereby improving the generalization ability and prediction accuracy of the model.

[0004] Traditional atmospheric river detection methods are mainly based on atmospheric river detection research based on different integrated water vapor transport thresholds (IVT). The existing technology only uses one remote sensing image detection method to select a threshold that is affected by multiple environmental factors, resulting in inaccurate selected thresholds. Even based on the water vapor concentration distribution map at the same time, different detection methods will obtain different atmospheric river distribution patterns, resulting in low confidence in the obtained atmospheric river area, affecting the detection accuracy of atmospheric rivers. Summary of the invention

[0005] In response to the problems existing in the above-mentioned fields, the present invention proposes an atmospheric river identification method, system, equipment and medium based on deep learning, which can solve the technical problem that the threshold selected by only one remote sensing image detection method in the prior art is affected by multiple environmental factors, resulting in inaccurate selected threshold. Even based on the water vapor concentration distribution map at the same time, different detection methods will obtain different atmospheric river distribution patterns, resulting in low confidence in the obtained atmospheric river area, affecting the detection accuracy of atmospheric rivers.

[0006] In order to solve the above technical problems, the present invention discloses an atmospheric river identification method based on deep learning, comprising the following steps: Obtaining original image data of the atmospheric river to be identified; Determine the confidence threshold of atmospheric river region detection, and judge whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection by analyzing the integrated water vapor transport threshold of each remote sensing image detection method selected. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as the atmospheric river area with high confidence. According to the high-confidence atmospheric river areas, the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method is collected; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; A deep learning neural network model is constructed and trained using the atmospheric river identification data training set to obtain an atmospheric river identification model. The atmospheric river identification model is used to perform regional recognition on the original image data of the atmospheric river to be identified.

[0007] Preferably, the determination of the confidence threshold for atmospheric river region detection specifically includes: Among the selected multiple remote sensing image detection methods, the cloud diagrams of integrated water vapor transport thresholds of each remote sensing image detection method are analyzed respectively; The cloud maps of the integrated water vapor transport thresholds of each remote sensing image detection method are intersected, and the integrated water vapor transport thresholds of each remote sensing image detection method corresponding to the intersection area in the cloud map are used as the confidence threshold for atmospheric river area detection.

[0008] Preferably, the high-confidence atmospheric river region comprises the following steps: By judging whether the cloud map of the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river area detection, when the confidence threshold requirement is met, the original image data of the atmospheric river to be identified is marked and filtered, the labels of each remote sensing image detection method are extracted, and the area of ​​the integrated water vapor transport threshold after marking and filtering is obtained; When the region in the obtained labeled filtered integrated water vapor transport threshold cloud map is within the confidence threshold range of atmospheric river region detection, the region of the obtained labeled filtered integrated water vapor transport threshold is defined as a high-confidence atmospheric river region and is retained; When the cloud map of the integrated water vapor transport threshold obtained after marking and filtering is not within the confidence threshold range of the atmospheric river area detection, the area of ​​the integrated water vapor transport threshold obtained after marking and filtering is defined as a low-confidence atmospheric river area, and the low-confidence atmospheric river area is filtered out.

[0009] Preferably, the step of sequentially comparing the atmospheric river identification areas of each remote sensing image detection method with the atmospheric river areas with high credibility further comprises: When the atmospheric river distribution map corresponding to the remote sensing image detection method does not contain any verified high-fidelity atmospheric river regional distribution map, the label of the remote sensing image detection method is regarded as a low-confidence label and discarded.

[0010] Preferably, the building of a deep learning neural network model specifically includes: The deep learning neural network model utilizes two parallel CNN pathways to process two different enhancements of the same input image; With DeepLabV3 as the core architecture, it adopts an encoding path with multiple convolutional layers and gradually increasing filters. DeepLabV3 uses dilated convolution to capture multi-scale contextual information by aggregating multi-scale features of the input image. At the same time, ReLU activation and maximum pooling are used for downsampling, and the backbone network MobileNetV3 is used for feature extraction. The decoding path uses up-convolution to restore the spatial resolution, and the high-dimensional feature map is mapped to the required output category through 1×1 convolution; Using the training set, the constructed deep learning neural network model is trained through the focal loss function to process the unbalanced labels commonly seen in atmospheric river detection and obtain a deep learning neural network model for atmospheric river identification.

[0011] Preferably, the two parallel CNN paths include an upper part and a lower part, wherein the upper part is a fully supervised learning segment and the lower part uses a semi-supervised learning method, and the upper and lower paths are integrated through a discriminator to evaluate reconstruction and similarity losses.

[0012] Preferably, the use of the atmospheric river identification data training set to train the deep learning neural network model is to crop the input image and label into small blocks of 360×360 pixels, and use the small blocks of 360×360 pixels for semi-supervised learning training; At the end of training, the cropped patches are reassembled into a complete segmentation result, which serves as the atmospheric river region recognition result for the training evaluation procedure.

[0013] Preferably, it also includes an atmospheric river identification system based on deep learning, comprising: A data acquisition module, used for acquiring original image data of the atmospheric river to be identified; The data marking and filtering module is used to determine the confidence threshold of atmospheric river area detection. By analyzing the integrated water vapor transport threshold of each selected remote sensing image detection method, it is determined whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river area detection. When the confidence threshold requirement is met, the original image data of the atmospheric river to be identified is marked and filtered, the labels of each remote sensing image detection method are extracted, and the area of ​​the integrated water vapor transport threshold after marking and filtering is obtained as a high-confidence atmospheric river area; The training set construction module is used to collect the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method based on the high-confidence atmospheric river area; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; The region recognition module is used to build a deep learning neural network model. The deep learning neural network model is trained using the atmospheric river recognition data training set to obtain the atmospheric river recognition model. The atmospheric river recognition model is used to perform region recognition on the original image data of the atmospheric river to be identified.

[0014] Preferably, a computer device is also included, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor executes the following steps: Obtaining original image data of the atmospheric river to be identified; Determine the confidence threshold of atmospheric river region detection, and judge whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection by analyzing the integrated water vapor transport threshold of each remote sensing image detection method selected. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as the atmospheric river area with high confidence. According to the high-confidence atmospheric river areas, the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method is collected; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; A deep learning neural network model is constructed and trained using the atmospheric river identification data training set to obtain an atmospheric river identification model. The atmospheric river identification model is used to perform regional recognition on the original image data of the atmospheric river to be identified.

[0015] Preferably, it further comprises a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the following steps: Obtaining original image data of the atmospheric river to be identified; Determine the confidence threshold of atmospheric river region detection, and judge whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection by analyzing the integrated water vapor transport threshold of each remote sensing image detection method selected. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as the atmospheric river area with high confidence. According to the high-confidence atmospheric river areas, the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method is collected; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; A deep learning neural network model is constructed and trained using the atmospheric river identification data training set to obtain an atmospheric river identification model. The atmospheric river identification model is used to perform regional recognition on the original image data of the atmospheric river to be identified.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes an atmospheric river identification method based on deep learning. By defining the area of ​​the integrated water vapor transport threshold after the acquired marker filtering as a high-credibility atmospheric river area, the method can collect the high-fidelity atmospheric river area distribution map corresponding to each time slice to extract the high-fidelity label of each remote sensing image detection method. The atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn. When the atmospheric river distribution map corresponding to the remote sensing image detection method does not contain any verified high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is regarded as a low-credibility label and discarded; when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the extracted labels of each remote sensing image detection method are updated and used as the atmospheric river identification data training set. Compared with the traditional atmospheric river identification algorithm using deep learning, the atmospheric river identification area obtained by the method proposed by the present invention has a higher credibility. Using it as a training set to train the constructed deep learning neural network model can ensure the accuracy of model training, thereby improving the accuracy of atmospheric river detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the atmospheric river identification method based on deep learning proposed by the present invention; Figure 2 This is a data set labeling and filtering framework diagram of the method proposed in the present invention; Figure 3 A deep learning neural network model training architecture constructed for the present invention; Figure 4 A distribution diagram of global IVT intensity identified by an embodiment of the present invention; Figure 5 This is a result map of atmospheric rivers identified by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the attached embodiment of the present invention Figure 1-Figure 5 , the technical solutions in the embodiments of the present invention are clearly and completely described. It should be understood that the terms described in the present invention are only used to describe specific implementation methods and are not used to limit the present invention.

[0019] like Figure 1 As shown, the present invention proposes an atmospheric river identification method based on deep learning, comprising the following steps: S1: Obtain the original image data of the atmospheric river to be identified; S2: Determine the confidence threshold of atmospheric river region detection. By analyzing the integrated water vapor transport threshold of each selected remote sensing image detection method, determine whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as a high-confidence atmospheric river area. S3: Based on the high-confidence atmospheric river areas, collect the high-fidelity atmospheric river area distribution maps corresponding to each time slice of each remote sensing image detection method; compare the atmospheric river distribution maps corresponding to all remote sensing image detection methods with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, update the extracted label of the remote sensing image detection method as the atmospheric river identification data training set; S4: Construct a deep learning neural network model, use the atmospheric river identification data training set to train the deep learning neural network model, and obtain an atmospheric river identification model; through the atmospheric river identification model, perform regional recognition on the original image data of the atmospheric river to be identified.

[0020] Specifically, the confidence threshold for atmospheric river area detection is determined, including: Among the selected multiple remote sensing image detection methods, the cloud diagrams of integrated water vapor transport thresholds of each remote sensing image detection method are analyzed respectively; The cloud maps of the integrated water vapor transport thresholds of each remote sensing image detection method are intersected, and the integrated water vapor transport thresholds of each remote sensing image detection method corresponding to the intersection area in the cloud map are used as the confidence threshold for atmospheric river area detection.

[0021] As a high-confidence atmospheric river region, the following steps are included: By judging whether the cloud map of the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river area detection, when the confidence threshold requirement is met, the original image data of the atmospheric river to be identified is marked and filtered, the labels of each remote sensing image detection method are extracted, and the area of ​​the integrated water vapor transport threshold after marking and filtering is obtained; When the region in the obtained labeled filtered integrated water vapor transport threshold cloud map is within the confidence threshold range of atmospheric river region detection, the region of the obtained labeled filtered integrated water vapor transport threshold is defined as a high-confidence atmospheric river region and is retained; When the cloud map of the integrated water vapor transport threshold obtained after marking and filtering is not within the confidence threshold range of the atmospheric river area detection, the area of ​​the integrated water vapor transport threshold obtained after marking and filtering is defined as a low-confidence atmospheric river area, and the low-confidence atmospheric river area is filtered out.

[0022] The atmospheric river identification areas of each remote sensing image detection method are compared with the atmospheric river areas with high confidence, including: When the atmospheric river distribution map corresponding to the remote sensing image detection method does not contain any verified high-fidelity atmospheric river regional distribution map, the label of the remote sensing image detection method is regarded as a low-confidence label and discarded.

[0023] Build a deep learning neural network model, including: The deep learning neural network model uses two parallel CNN paths to process two different enhancements of the same input image; With DeepLabV3 as the core architecture, it adopts an encoding path with multiple convolutional layers and gradually increasing filters. DeepLabV3 uses dilated convolution to capture multi-scale contextual information and outlines the complex structure in the input image by aggregating the multi-scale features of the input image.

[0024] At the same time, ReLU activation and maximum pooling are used for downsampling, and the backbone network MobileNetV3 is used for feature extraction. The decoding path uses up-convolution to restore the spatial resolution, and the high-dimensional feature map is mapped to the required output category through 1×1 convolution.

[0025] The constructed deep learning neural network model is trained through the focal loss function to handle the unbalanced labels commonly seen in atmospheric river detection, and a deep learning neural network model for atmospheric river identification is obtained.

[0026] The two parallel CNN paths include an upper part and a lower part, where the upper part is a fully supervised learning segment and the lower part uses a semi-supervised learning method. The upper and lower paths are integrated through a discriminator to evaluate the reconstruction and similarity losses to optimize the model performance.

[0027] The deep learning neural network model is trained using the atmospheric river identification data training set by cropping the input image and label into small patches of 360×360 pixels, and using small patches of 360×360 pixels for semi-supervised learning training.

[0028] At the end of training, these cropped patches are reassembled into a complete segmentation result as the atmospheric river identification result for the training evaluation procedure.

[0029] The present invention also proposes an atmospheric river identification system based on deep learning, comprising: A data acquisition module, used for acquiring original image data of the atmospheric river to be identified; The data marking and filtering module is used to determine the confidence threshold of atmospheric river area detection. By analyzing the integrated water vapor transport threshold of each selected remote sensing image detection method, it is determined whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river area detection. When the confidence threshold requirement is met, the original image data of the atmospheric river to be identified is marked and filtered, the labels of each remote sensing image detection method are extracted, and the area of ​​the integrated water vapor transport threshold after marking and filtering is obtained as a high-confidence atmospheric river area; The training set construction module is used to collect the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method based on the high-confidence atmospheric river area; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; The region recognition module is used to build a deep learning neural network model. The deep learning neural network model is trained using the atmospheric river recognition data training set to obtain the atmospheric river recognition model. The atmospheric river recognition model is used to perform region recognition on the original image data of the atmospheric river to be identified.

[0030] Embodiment 1 In order to verify the feasibility of the atmospheric river identification method based on deep learning proposed in this paper, Figure 2 As shown in the figure, five currently selected remote sensing image detection methods are given, namely AR connect, Guan, Lora, Mundhenk and TempestLR, among which: AR connect is a possible connection to augmented reality or algorithm Augmented Reality Applications: When AR connect refers to the application of augmented reality technology in remote sensing image processing, it may involve combining remote sensing images with real-time data or three-dimensional models to provide richer spatial information.

[0031] Algorithm connection: Another possibility, AR connect, refers to the connection or integration between different remote sensing image processing algorithms to achieve more complex or more precise analysis tasks.

[0032] Guan may be associated with spectral analysis or specific algorithms Spectral analysis: In remote sensing image processing, spectral analysis is an important means of extracting information about objects. When Guan refers to a spectral analysis method, it may use the reflectance characteristics of objects in different bands to identify or classify targets.

[0033] Specific algorithm: Guan may also refer to a specific remote sensing image processing algorithm that may have unique data processing and analysis capabilities.

[0034] Lora may be associated with deep learning or long short-term memory networks Deep learning: In the field of remote sensing image processing, deep learning technology has been widely used. When Lora refers to a deep learning model, it may use a convolutional neural network (CNN) or other deep learning architectures to automatically extract and analyze features in remote sensing images.

[0035] Long Short-Term Memory Network LSTM: Although LSTM is usually used to process sequence data, in some cases, it may also be used for time series analysis or change detection of remote sensing images. If Lora is associated with LSTM, it may involve the use of time series remote sensing images to monitor and analyze surface changes.

[0036] Mundhenk may be associated with image processing experts or algorithms Image processing expert: In some cases, Mundhenk may refer to an expert or researcher who has made significant contributions to the field of remote sensing image processing, and whose research results or algorithms may have been widely used in this field.

[0037] Specific Algorithm: Mundhenk may also refer to a specific remote sensing image processing algorithm that may have unique data processing and analysis capabilities.

[0038] TempestLR is a technology that may be associated with advanced image processing techniques or long-range remote sensing Advanced image processing technology: TempestLR may refer to an advanced image processing technology that combines multiple algorithms and methods to provide more accurate and reliable remote sensing image analysis results.

[0039] Long-range remote sensing: TempestLR is related to long-range remote sensing, which may involve the use of high-resolution remote sensing imagery to monitor and analyze objects or phenomena far from the ground.

[0040] Since the names of AR connect, Guan, Lora, Mundhenk and TempestLR are not standard professional terms, there may be some uncertainty in the above explanations and speculations. In practical applications, it is recommended to select appropriate algorithms and methods based on specific remote sensing image processing requirements and data characteristics. At the same time, you can also refer to the latest research results and technological advances in the field of remote sensing image processing to obtain more accurate and comprehensive information.

[0041] By analyzing the cloud maps of the integrated water vapor transport threshold of each remote sensing image detection method, whether there are three or more cloud maps of the integrated water vapor transport threshold of remote sensing image detection methods included in the atmospheric river area, when there are three or more cloud maps of the integrated water vapor transport threshold of remote sensing image detection methods included in the atmospheric river area, the atmospheric river identification area with a confidence level of more than 60% is determined and retained.

[0042] The specific contents of the five remote sensing image detection methods selected in this embodiment are shown in Table 1.

[0043] Table 1 Detailed information of the five selected remote sensing image detection methods

[0044] Specifically, among the five selected remote sensing image detection methods, the cloud maps of the integrated water vapor transport threshold of each remote sensing image detection method are analyzed respectively.

[0045] The original year label comes from Figure 2 In the five original atmospheric river labeled datasets shown in the leftmost part of , in order to extract high-fidelity labels, the goal of the label filtering step is to identify high-fidelity label areas.

[0046] like Figure 2 As shown in the label filtering step section, when high-value areas in the cloud map of the integrated water vapor transport threshold are included in the atmospheric river region in three or more of the five remote sensing image detection methods, then these areas will be considered as high-confidence atmospheric river areas and retained in the preprocessing algorithm. If the areas are only identified by one or two of the five remote sensing image detection methods, then these areas are considered as low-confidence atmospheric river areas and filtered out.

[0047] After completing the first step, the distribution of high-fidelity atmospheric river areas corresponding to each time slice can be collected.

[0048] The atmospheric river distribution maps corresponding to all the above-mentioned remote sensing image detection methods will be used to compare with the obtained high-fidelity atmospheric river distribution maps in turn. If the atmospheric river labels identified by the remote sensing image detection method for comparison do not contain any verified high-fidelity atmospheric river areas, then the labels of the remote sensing image detection method for comparison will be regarded as low-confidence labels and discarded.

[0049] According to the five updated atmospheric river detection labels collected in this process, when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river regional distribution map, such as Figure 2 As shown in the Updated Labeling section, the labels of each remote sensing image detection method should be updated.

[0050] Therefore, the filtered atmospheric rivers obtained by this method can ensure the accuracy of subsequent DL training. Finally, these labels from 2000 to 2019 will be used as the input dataset for subsequent SSPD network training.

[0051] A deep learning neural network model is constructed to achieve atmospheric river identification. A dual-path architecture is usually used for image segmentation to enhance the learning process by utilizing labeled and unlabeled data. When labels from various remote sensing image detection methods are used simultaneously in a deep learning method, fully supervised learning tends to absorb all features derived from these diverse labels, and several small errors existing in the above remote sensing image detection methods may be learned. In contrast, the semi-supervised learning adopted by the present invention first trains the network using labeled data, classifies the unlabeled data using a specific classification algorithm, and through a confidence assessment cycle, the unlabeled data accurately classified by the remote sensing image detection method is included in the training set. In each iteration, as the newly added data is considered relatively reliable, the model becomes more robust. Therefore, through continuous iterations, the newly extracted features will behave more consistently with the previous features, thereby achieving better performance in atmospheric river detection.

[0052] The deep learning neural network model uses two parallel CNN paths to process two different enhancements of the same input image. By adopting image enhancement techniques, such as modifications of brightness and contrast, this approach enhances the generalization ability of the model in a wide range of visual environments and more accurately simulates changes in the integrated water vapor transport threshold profile. This dual-path approach not only maximizes the use of data, but also enriches the training set without the need for additional labeled samples.

[0053] The core of this deep learning neural network model is the DeepLabV3 architecture, which was selected for its excellent performance in high-resolution semantic segmentation tasks. DeepLabV3 uses an encoding path with multiple convolutional layers and gradually increasing filters, which enhances the model's ability to capture complex features and details. Although not explicitly shown in the provided code, DeepLabV3 uses dilated convolutions to capture multi-scale contextual information without losing resolution, which is critical for segmenting details in atmospheric river detection. This approach enables the network to effectively aggregate multi-scale features and outline complex structures in the input image.

[0054] The DeepLabV3 architecture uses ReLU activation and max pooling for efficient downsampling and a powerful backbone network MobileNetV3 for powerful feature extraction, which is efficient in both computation and storage and suitable for high-resolution input images. The decoding path uses up-convolution to restore the spatial resolution, and finally maps the high-dimensional feature map to the desired output category through 1×1 convolution. At the same time, a carefully designed focal loss function is included in the training method to handle the unbalanced labels commonly found in atmospheric river detection. The focal loss is particularly effective when dynamically scaling the cross entropy loss based on prediction confidence, thereby focusing learning on difficult-to-classify examples.

[0055] Based on these foundations, the model was implemented and optimized in PyTorch. This seamless integration extends to rigorous training evaluation of the U-Net and discriminator components using key performance metrics such as Intersection over Union (IoU). This thorough validation not only confirms the reliability of the model, but also its adaptability, which is critical for producing high-quality atmospheric river detection segmentations based on integrated water vapor transport threshold profiles.

[0056] Since the images in the training dataset have high resolution, the images and labels are cropped into small patches for training, e.g. Figure 3 The prediction results are shown in the prediction results section. Due to the uneven distribution of atmospheric rivers over the entire region, if a smaller size (such as 240×240 pixels or smaller) is used, there will be too many cropped images without atmospheric river labels, which may reduce the prediction accuracy. After testing with different sizes, it is recommended to use cropped images of 360×360 pixels. Figure 3 The model is trained on the sub-images in . Note that the same cropping operation is applied to the image and its corresponding label to maintain the alignment required for the segmentation task. Based on this crop size, each original image is segmented into 8 sub-images for semi-supervised model training. At the end of training, these sub-images are reassembled into a complete segmentation result as the basic atmospheric river recognition result for subsequent training and evaluation procedures. This approach improves training efficiency and prevents memory overload. The system architecture of the entire semi-supervised model is shown in Figure 2. Figure 3 shown.

[0057] The deep learning neural network model constructed in the present invention is based on five widely used remote sensing image detection methods. The atmospheric river areas recognized by each remote sensing image detection method and recognized by each other are extracted as deep learning data sets for training. Therefore, compared with the traditional atmospheric river recognition algorithm using deep learning, it has higher credibility.

[0058] In order to better learn the features of multiple data sources to train the deep learning neural network model, when using labels from various remote sensing image detection methods, fully supervised learning tends to absorb all features derived from these different labels. Some small errors existing in the above algorithm may be learned. In contrast, the semi-supervised learning algorithm used in the present invention first trains the network using labeled data, and then uses a classification method to classify the unlabeled data. Through a cycle of accuracy evaluation, unlabeled data that is considered to be accurately classified will be included in the training set. In each iteration, as the newly added data is considered to be relatively reliable, the model becomes more robust. Therefore, through continuous cycles, the newly extracted features are more consistent with previous features, thereby achieving better performance in atmospheric river detection.

[0059] Embodiment 2 The AR recognition process of the present invention includes: after the entire neural network training is completed, it is saved on the server and a software model is constructed.

[0060] The current server has stored all ERA5 atmospheric river data from 1980 to 2024, and uses the trained neural network for prediction. The data is authentic and reliable, and is the mainstream meteorological analysis data. Among them, the calculation area of ​​the main interface of the constructed software model can select the calculation area and calculation time on the map, and the output time interval ranges from 1 hour to 24 hours. After selecting the appropriate time and space area and time resolution, click the "OK" button to jump to the login interface.

[0061] Taking into account the huge amount of meteorological data required for atmospheric river identification, the present invention uploads all required data to the online database in advance. The user only needs to enter the corresponding time range and calculation area to start identifying the atmospheric river within the selected time and space area, which greatly saves the time for carrying out the corresponding work.

[0062] After entering the preset username and password in the login interface, you can access the resources on the server. If you need to register a new account, you can send an email to Professor Zhao Yang of Ocean University of China in advance to apply for it. In this way, you can obtain relevant usage permissions. You can also use the default account in the figure to log in for atmospheric river analysis.

[0063] After clicking the "Login" button, you will enter the SSHCommandForm interface. First, the Output window shows the time range and space area selected by the user, as well as the time interval entered. At the same time, users with proficient Linux theoretical knowledge and deep learning application experience can also enter content through the command window below and click the "Submit" button to submit it to the server. When it displays "File copied from XXXXX", it means that the data that meets the time and space range has been copied to a special directory, and training can begin. Click the "Start Training" button to call the previously trained deep learning network to start prediction. The prediction results will be saved in txt format in sequence, and can be downloaded to the local computer by clicking the "Download Results" button, and click the "Data Arrangement" button to organize it into the desired format.

[0064] Finally, click the "Visualize" button to open a previously trained result and draw it into a corresponding atmospheric river image, thus providing an intuitive visualization effect. Figure 4 and Figure 5 As shown, it is the visualization result of atmospheric river identification, where: Figure 4 To identify the distribution map of global IVT intensity, Figure 5 For the identified atmospheric river result map, users can also draw the map based on the exported data to get the expected effect.

[0065] Considering the huge computing resources and high requirements for graphics card devices required for deep learning, the software model provided by the present invention can provide a trained semi-supervised learning network. At the same time, users do not need to master complex artificial intelligence knowledge and Linux system operations, and can integrate relevant operations into the software model. They only need to click the corresponding button according to the operation process to perform the atmospheric river identification training task and download the corresponding data.

[0066] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

[0067] In addition, unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as commonly understood by those skilled in the art to which the present invention belongs. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of any conflict with any incorporated document, the content of this specification shall prevail.

Claims

1. A deep learning-based atmospheric river identification method, characterized in that: The following steps are involved: Obtaining original image data of the atmospheric river to be identified; Determine the confidence threshold of atmospheric river region detection, and judge whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection by analyzing the integrated water vapor transport threshold of each remote sensing image detection method selected. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as the atmospheric river area with high confidence. According to the high-confidence atmospheric river areas, the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method is collected; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; A deep learning neural network model is constructed and trained using the atmospheric river identification data training set to obtain an atmospheric river identification model. The atmospheric river identification model is used to perform regional recognition on the original image data of the atmospheric river to be identified.

2. The atmospheric river identification method based on deep learning according to claim 1, characterized in that: Determining the confidence threshold for atmospheric river region detection specifically includes: Among the selected multiple remote sensing image detection methods, the cloud diagrams of integrated water vapor transport thresholds of each remote sensing image detection method are analyzed respectively; The cloud maps of the integrated water vapor transport thresholds of each remote sensing image detection method are intersected, and the integrated water vapor transport thresholds of each remote sensing image detection method corresponding to the intersection area in the cloud map are used as the confidence threshold for atmospheric river area detection.

3. The atmospheric river identification method based on deep learning according to claim 2 is characterized in that: The method of identifying an atmospheric river region with high confidence includes the following steps: By judging whether the cloud map of the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river area detection, when the confidence threshold requirement is met, the original image data of the atmospheric river to be identified is marked and filtered, the labels of each remote sensing image detection method are extracted, and the area of ​​the integrated water vapor transport threshold after marking and filtering is obtained; When the region in the obtained labeled filtered integrated water vapor transport threshold cloud map is within the confidence threshold range of atmospheric river region detection, the region of the obtained labeled filtered integrated water vapor transport threshold is defined as a high-confidence atmospheric river region and is retained; When the cloud map of the integrated water vapor transport threshold obtained after marking and filtering is not within the confidence threshold range of the atmospheric river area detection, the area of ​​the integrated water vapor transport threshold obtained after marking and filtering is defined as a low-confidence atmospheric river area, and the low-confidence atmospheric river area is filtered out.

4. The atmospheric river identification method based on deep learning according to claim 3 is characterized in that: The step of sequentially comparing the atmospheric river identification areas of each remote sensing image detection method with the atmospheric river areas with high credibility also includes: When the atmospheric river distribution map corresponding to the remote sensing image detection method does not contain any verified high-fidelity atmospheric river regional distribution map, the label of the remote sensing image detection method is regarded as a low-confidence label and discarded.

5. The atmospheric river identification method based on deep learning according to claim 4 is characterized in that: The construction of the deep learning neural network model specifically includes: The deep learning neural network model utilizes two parallel CNN pathways to process two different enhancements of the same input image; With DeepLabV3 as the core architecture, it adopts an encoding path with multiple convolutional layers and gradually increasing filters. DeepLabV3 uses dilated convolution to capture multi-scale contextual information by aggregating multi-scale features of the input image. At the same time, ReLU activation and maximum pooling are used for downsampling, and the backbone network MobileNetV3 is used for feature extraction. The decoding path uses up-convolution to restore the spatial resolution, and the high-dimensional feature map is mapped to the required output category through 1×1 convolution; Using the training set, the constructed deep learning neural network model is trained through the focal loss function to process the unbalanced labels commonly seen in atmospheric river detection and obtain a deep learning neural network model for atmospheric river identification.

6. The atmospheric river identification method based on deep learning according to claim 5, characterized in that: The two parallel CNN paths include an upper part and a lower part, wherein the upper part is a fully supervised learning segment and the lower part uses a semi-supervised learning method, and the upper and lower paths are integrated through a discriminator to evaluate reconstruction and similarity losses.

7. The atmospheric river identification method based on deep learning according to claim 6, characterized in that: The deep learning neural network model is trained using the atmospheric river identification data training set by cropping the input image and label into small blocks of 360×360 pixels, and using the small blocks of 360×360 pixels for semi-supervised learning training; At the end of training, the cropped patches are reassembled into a complete segmentation result, which serves as the atmospheric river region recognition result for the training evaluation procedure.

8. An atmospheric river identification system based on deep learning, characterized in that: include: A data acquisition module, used for acquiring original image data of the atmospheric river to be identified; The data marking and filtering module is used to determine the confidence threshold of atmospheric river area detection. By analyzing the integrated water vapor transport threshold of each selected remote sensing image detection method, it is determined whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river area detection. When the confidence threshold requirement is met, the original image data of the atmospheric river to be identified is marked and filtered, the labels of each remote sensing image detection method are extracted, and the area of ​​the integrated water vapor transport threshold after marking and filtering is obtained as a high-confidence atmospheric river area; The training set construction module is used to collect the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method based on the high-confidence atmospheric river area; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; The region recognition module is used to build a deep learning neural network model. The deep learning neural network model is trained using the atmospheric river recognition data training set to obtain the atmospheric river recognition model. The atmospheric river recognition model is used to perform region recognition on the original image data of the atmospheric river to be identified.

9. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps: Obtaining original image data of the atmospheric river to be identified; Determine the confidence threshold of atmospheric river region detection, and judge whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection by analyzing the integrated water vapor transport threshold of each remote sensing image detection method selected. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as the atmospheric river area with high confidence. According to the high-confidence atmospheric river areas, the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method is collected; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; A deep learning neural network model is constructed and trained using the atmospheric river identification data training set to obtain an atmospheric river identification model. The atmospheric river identification model is used to perform regional recognition on the original image data of the atmospheric river to be identified.

10. A computer-readable storage medium, characterized in that: The computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor performs the following steps: Obtaining original image data of the atmospheric river to be identified; Determine the confidence threshold of atmospheric river region detection, and judge whether the integrated water vapor transport threshold of each remote sensing image detection method exceeds the confidence threshold requirement of atmospheric river region detection by analyzing the integrated water vapor transport threshold of each remote sensing image detection method selected. When the confidence threshold requirement is met, mark and filter the original image data of the atmospheric river to be identified, extract the labels of each remote sensing image detection method, and obtain the area of ​​the integrated water vapor transport threshold after marking and filtering as the atmospheric river area with high confidence. According to the high-confidence atmospheric river areas, the high-fidelity atmospheric river area distribution map corresponding to each time slice of each remote sensing image detection method is collected; the atmospheric river distribution maps corresponding to all remote sensing image detection methods are compared with the high-fidelity atmospheric river area distribution map in turn, and when the atmospheric river distribution map corresponding to the remote sensing image detection method overlaps with the high-fidelity atmospheric river area distribution map, the label of the remote sensing image detection method is updated and extracted as the atmospheric river identification data training set; A deep learning neural network model is constructed and trained using the atmospheric river identification data training set to obtain an atmospheric river identification model. The atmospheric river identification model is used to perform regional recognition on the original image data of the atmospheric river to be identified.

Images