A method and system for quickly and efficiently identifying mesoscale convective systems in the mid - low latitude regions

Abstract

The present invention belongs to the field of image processing, and discloses a method and system for quickly and efficiently identifying mesoscale convective systems in mid - low latitude regions, including obtaining cloud - top brightness temperature data; constructing a dataset for identifying mesoscale convective systems based on the cloud - top brightness temperature data; constructing a model for identifying mesoscale convective systems; training the model using the constructed training set; after reaching the specified number of training rounds, testing the model with the test set to obtain the final result of identifying mesoscale convective systems. The present invention designs a piece - by - piece learning strategy, enabling the model to simultaneously learn the characteristics of mesoscale convective systems in different regions; constructs a multi - scale feature extractor with semi - residuals to prevent the loss of small - scale features during the feature extraction process; designs a semantic consistency regularization to constrain redundant feature extraction and the feature asymmetry during the feature recovery process. The present invention solves the problem of slow identification speed using traditional threshold methods in mid - low latitude regions.

Description

Technical Field

[0001] The present invention belongs to, but is not limited to, the field of image processing technology, and particularly relates to a method and system for quickly and efficiently identifying mesoscale convective systems in mid - low latitude regions. Background Art

[0002] Deep moist convection is a common meteorological feature in many regions of the world and is usually formed by updrafts and downdrafts. Independent convective cells composed of these two airflows often interact with each other to form storm cells, storm lines, or storm clusters. This combination is called mesoscale convective systems (MCSs), which have a length scale greater than 100 kilometers in one or more directions and a duration of more than 3 hours. MCSs are the most important convective weather systems on Earth and are ubiquitous in tropical regions such as the western Pacific, tropical Africa, and the Amazon. They play an important role in the global energy and hydrological cycles because they vertically transport mass, momentum, and heat. Previous studies have shown that although they are only a small part of the wet convective system, they contribute 50% - 90% of the tropical annual precipitation. In addition, MCSs are often closely related to disastrous weather events such as thunderstorms, floods, and strong winds. MCSs bring essential precipitation to agricultural production throughout the tropical and mid - latitude regions. Changes in the frequency of MCSs can directly affect agricultural production: too few MCSs can lead to a dry environment, while too many MCSs can lead to extreme weather events such as strong winds and floods. Given the hydrological importance and disastrousness of MCSs, it is extremely important to understand the regional distribution and spatial climatology characteristics of MCSs, especially in the context of climate warming.

[0003] Due to the limitation of computing power, most of the previous studies on the identification of MCSs were regional or short-term. With the development of Earth-orbiting satellites in recent years, high spatio-temporal resolution observations of clouds have made it possible to continuously capture the evolution of MCSs globally. Researchers have done a lot of work on identifying MCSs using two widely used identification methods. The first is to use geostationary satellite data to identify MCSs based on the characteristics of low cloud-top brightness temperature. Generally, infrared satellite data in the 11-μm band is most commonly used to identify MCSs. Existing techniques identify MCSs in the tropical region (30°N - 30°S) based on temperature thresholds and area coverage thresholds of geostationary satellite data, and combine the area overlap method and the Kalman filtering method to track the evolution of MCSs. Further, existing techniques extend the identification area of MCSs to 60°N to 60°S, and use IMERG precipitation data to complete the identification and tracking of MCSs in the tropical and mid-latitude regions, supplementing the deficiency of convective internal structure information in geostationary satellite data. The focus of existing techniques is on the variation of MCS characteristics over land and ocean and in the tropical and mid-latitude regions. The second is to use long-term satellite precipitation data for additional measurements of cloud and precipitation structures to identify MCSs. A large number of previous studies have shown that the results of MCS identification and tracking obtained using GPM and IMERG precipitation data are consistent with radar-driven results. Existing techniques have identified and tracked MCSs in the IMERG precipitation field from 2014 to 2018, and tracked continuous regions larger than 1,000 km 2 as MCSs. In addition, existing techniques have used the Forward-in-Time algorithm (FiT) to track MCSs in the IMERG precipitation field in the global tropical region (30°N - 30°S) for 10 years (2011 - 2020).

[0004] Based on satellite data, both methods can accurately capture the evolution of short-term and local-area MCSs, including their germination, maturity, and demise. However, both methods are based on the threshold method, that is, by calculating the scale size and temperature of a single target to determine whether it is an MCS. When applied globally, especially in the tropical and mid-latitude regions, the time required for the threshold method to determine whether a target is an MCS is quite long. The huge time cost will lead to a slowdown in downstream tasks (tracking and forecasting), affecting the analysis of the climate characteristics of MCSs.

[0005] From the perspective of computer vision, the identification of mesoscale convective systems MCSs can be regarded as an image target recognition task. Visualize the cloud-top brightness temperature data in the mid- and low-latitude regions and the corresponding mesoscale convective system MCSs labels obtained based on the threshold method, such as Figure 1As shown. Target recognition is a fundamental task of identifying the regions of interest (foreground objects) in an image from the image itself (background). The recognition of mesoscale convective systems (MCSs) can be regarded as a recognition task that combines salient object detection and camouflaged object detection.

[0006] The task of salient object detection aims to identify the most important and informative objects in a scene and then segment them at the pixel level. It has been applied to many vision problems, including image segmentation, object recognition, and visual tracking. Early deep learning methods for the SOD task extracted local patches of an image from single or multiple scales and classified each image pixel one by one into salient or non-salient categories. These methods usually produced rough output results because the fully connected layers could not effectively utilize the spatial information of the image. With the development of the fully convolutional neural network (FCN), FCN-based models have made significant progress in the SOD task. The prior art introduced R-dropout to construct an uncertain set of internal feature units, which enabled the model to learn deep uncertain convolutional features (UCFs), thereby improving the robustness and accuracy of saliency detection. The prior art constructed a deep level set network to generate a compact and uniform saliency map, thus outputting a more precise boundary and a compact saliency. The prior art proposed a network that could incorporate saliency prior knowledge and used a recursive architecture to enable the network to refine the saliency map by correcting previous errors. The prior art developed a novel deep saliency network equipped with recursive aggregated deep features (RADF) to more accurately detect salient objects in an image by leveraging complementary salient information captured by different layers. SAC-Net achieved salient object detection by fusing local and global image contexts inside, around, and outside the salient object. The prior art proposed a novel progressive structure with a knowledge review network (PA-KRN) to simulate the global localization and local segmentation processes in human biology.

[0007] Different from the salient object detection task, the goal of camouflaged object detection is to identify objects from highly similar environments. The latest research has made significant progress in identifying camouflaged objects in a data-driven manner by leveraging the complex representation capabilities of deep learning models. The prior art has proposed a two-branch (probability branch and main branch) network for camouflaged object detection, named Anabranch, which fuses the probability of camouflaged objects into the main branch for segmentation to improve segmentation accuracy. The prior art has proposed a new framework that makes full use of multiple visual cues (i.e., saliency and edges) to refine the prediction of camouflaged objects. The prior art has established a two-stage model for camouflaged object detection, first finding the approximate area where camouflaged objects exist through a localization model, and then precisely segmenting the selected area through a segmentation model. The prior art has proposed a new type of camouflaged object detection framework called CubeNet, which effectively fuses multi-layer features by introducing X connections. To address the problems of camouflaged objects with different scales and blurred appearances, the prior art has proposed a hybrid-scale triple network - ZoomNet, which mimics the behavior of humans when observing blurred images. The prior art has constructed a deep gradient network (DGNet), which is a new type of deep framework for camouflaged object detection using object gradient supervision. The prior art has proposed a feature decomposition and edge reconstruction (FEDER) model for camouflaged object detection, which solves the problem of the inherent similarity between the foreground and the background by decomposing features into different frequency bands using learnable wavelets. The prior art has proposed a novel learnable and separable frequency-aware mechanism to distinguish camouflaged objects from the background.

[0008] From the perspective of computer vision, the identification of MCSs in meteorology can be regarded as an image binary classification task, that is, separating MCSs from cloud top brightness temperature data. However, the identification of MCSs globally is not a simple binary classification task. Specifically, for low-latitude regions (near the equator), MCSs have clearer contours and are well distinguishable from the background. We can consider the identification of MCSs in low-latitude regions as an SOD task. While MCSs in mid-latitude regions have blurred contours and low distinguishability. We can regard the identification of MCSs in mid-latitude regions as a COD task. In addition, most MCSs in low-latitude regions are relatively scattered, and the distance between adjacent MCSs is very close. While the monitoring in mid-latitude regions appears as a whole. Globally, MCSs in both mid-latitude and low-latitude regions have the characteristics of large-scale variations.

[0009] In view of the above analysis, the technical problems urgently needed to be solved in the prior art are:

[0010] (1) Based on satellite data, traditional methods can accurately capture the evolution of MCSs in short-term and local regions, including their germination, maturity, and extinction. However, these methods are all based on thresholds, that is, by calculating the scale size and temperature of a single target to determine whether it is an MCS. When applied globally, especially in tropical and mid-latitude regions, the time required for the threshold method to determine whether a target is an MCS is quite long. The huge time cost will lead to slow progress in downstream tasks (tracking and prediction), affecting the analysis of the climatological characteristics of MCSs.

[0011] (2) Deep learning models have made significant progress in the field of target recognition, while the research in the field of mesoscale convective identification is relatively blank. MCSs identification is a mixed task, and most existing deep learning models are only for single significant target recognition or camouflaged target recognition tasks. Using a single significant target recognition (camouflaged target recognition) model to identify MCSs will lead to unstable gradient backpropagation. How to use a single model to quickly and accurately identify MCSs is the primary problem. In addition, the characteristics of large scale differences and close distances of MCSs also pose great challenges to identification. Summary of the Invention

[0012] Aiming at the problems existing in the prior art, the present invention provides a method and system for quickly and efficiently identifying mesoscale convective systems in the mid-low latitude region.

[0013] The present invention is implemented as follows. A method for quickly and efficiently identifying mesoscale convective systems in the mid-low latitude region is characterized in that the method specifically includes:

[0014] S1. Obtain the cloud top brightness temperature data of the mid-low latitude region;

[0015] S2. Construct a mesoscale convective system identification data set for the cloud top brightness temperature data, and the data set includes a training set and a test set;

[0016] S3. Construct a mesoscale convective system identification model;

[0017] S4. Use the training set to train the mesoscale convective system identification model;

[0018] S5. Based on the trained mesoscale convective system identification model, obtain the mesoscale convective system identification result;

[0019] S6. After S4 and S5 reach the specified number of training times in sequence, input the test set into the trained mesoscale convective system identification model to determine whether the current mesoscale convective system identification model meets the index requirements. If so, save the parameters of the current model and enter S7; otherwise, return to S4;

[0020] S7. Input the test set into the mesoscale convective system recognition model that meets the index requirements to obtain the final mesoscale convective system recognition result.

[0021] Further, S2 includes the following steps:

[0022] S201. Set the temperature threshold of the mesoscale convective system MCSs, and mark the areas smaller than the temperature threshold from the original cloud top brightness temperature data;

[0023] S202. Set the regional coverage threshold of the mesoscale convective system MCSs, and further screen the areas marked in S201 using the regional coverage threshold. Select the areas in S201 with an area larger than the regional coverage threshold, and mark the places that meet the requirements of both thresholds as the mesoscale convective system MCSs areas. Among them, the mesoscale convective system MCSs areas are marked as 1, and the non-mesoscale convective system MCSs are marked as 0. After the above processing, the mesoscale convective system MCSs label is obtained, and the label size is 5143×1715;

[0024] S203. Fill the null values in the cloud top brightness temperature data using the interpolation method. In addition, process the outliers in the data. Specifically, the values greater than 345 in the cloud top brightness temperature data are replaced with 345, and the values less than 170 are replaced with 170. The obtained cloud top brightness temperature data size is 5143×1715;

[0025] S204. Perform edge padding on the mesoscale convective system MCSs label obtained in S202 and the cloud top brightness temperature data obtained in S203. Specifically, fill the cloud top brightness temperature data with a value of 300, and fill the label data with a value of 0. The sizes of both after padding are 5632×2048.

[0026] S205. Divide the padded mesoscale convective system MCSs label and cloud top brightness temperature data obtained in S204 into a training set and a test set. Among them, both the training set and the test set include cloud top brightness temperature data and mesoscale convective system MCSs labels, and the storage format is numpy.

[0027] Further, S3 includes the following steps:

[0028] S301. Process the cloud top brightness temperature data obtained in S204 using maximum-minimum normalization and convert it to the tensor format, where the maximum value is taken as 345 and the minimum value is taken as 170. The MCS label data obtained in S204 remains unchanged;

[0029] S302. Construct a mesoscale convective system recognition model according to the normalization processing result of S301.

[0030] Furthermore, S4 includes the following steps:

[0031] S401. For the cloud top bright temperature data after normalization in S301, the size is 5632×2048. First, the data is split into 11×4 data with a size of 512×512. Every 11 groups of data are divided into a batch, and a total of 4 batches of data are obtained and denoted as B i . Among them, the first batch B1 and the fourth batch B4 are cloud top bright temperature data in the mid-latitude region, and the second B2 and the third batch B3 are data in the low-latitude region;

[0032] S402. Input the first batch B1 into the semi-residual multi-scale feature extractor SRFE to extract hierarchical multi-scale pyramid features, and output the feature map Ei of each layer of the encoder;

[0033] S403. In the decoder part, the feature map Ei output by the encoder and the feature map Di output by the decoder are concatenated in the channel dimension and then input into the convolution and upsampled to obtain Di, where Di represents the feature recovered by the i-th layer of the decoder;

[0034] S404. Input the output of the last layer of the decoder into the convective recognition head SegHead to obtain the recognition result of the mesoscale convective system;

[0035] S405. When optimizing the model by backpropagation, we use semantic consistency regularization to optimize the model;

[0036] S406. Repeat S401 - S405 until all 4 batches of data B i are input into the model.

[0037] Furthermore, the model includes:

[0038] An encoder for extracting hierarchical pyramid features after normalization in S301. Among them, the encoder includes four semi-residual multi-scale feature extractors SRFE; the semi-residual multi-scale feature extractor SRFE is used to extract multi-scale features to prevent small-scale features from being lost during downsampling. The semi-residual design can prevent the forward propagation of redundant features;

[0039] A decoder for decoding and recovering pyramid features at different levels. Among them, the decoder includes a convolution and an upsampling module.

[0040] Furthermore, S402 includes the following steps:

[0041] S4021. Input the first batch B1 into the first semi-residual multi-scale feature extractor SRFE of the encoder. First, duplicate B1 four times in the channel dimension. Use a 1×1 convolution to extract features from the first copy to obtain the first branch feature map F1, use a 3×3 convolution to extract features from the second copy to obtain the second branch feature map F2, use a 5×5 convolution to extract features from the third copy to obtain the third branch feature map F3, and use a 7×7 convolution to extract features from the fourth copy to obtain the fourth branch feature map F4;

[0042] S4022. Concatenate the four feature maps F1, F2, F3, and F4 obtained in S4021 in the channel dimension and then use a 1*1 convolution for multi-scale feature fusion to obtain the preliminary fusion feature map M1;

[0043] S4023. Split the preliminary fusion feature map M1 into four parts in the channel dimension. Use a 1×1 convolution to extract features from the first part to obtain the first branch feature map F 11 , use a 3×3 convolution to extract features from the second part to obtain the second branch feature map F 22 , use a 5×5 convolution to extract features from the third part to obtain the third branch feature map F 33 , use a 7×7 convolution to extract features from the fourth part to obtain the fourth branch feature map F 44 ;

[0044] S4024. Concatenate the four feature maps F 11 , F 22 , F 33 , F 44 obtained in S4023 in the channel dimension and then use a 1*1 convolution for multi-scale feature fusion to obtain the final fusion feature map M 11 , and add M 11 to M1 to obtain the final output E1 of the first semi-residual multi-scale feature extractor SRFE;

[0045] S4025. Input the final output E1 of the first semi-residual multi-scale feature extractor SRFE into the next semi-residual multi-scale feature extractor SRFE and perform the operations of S4021 - S4025 to obtain the final output E2 of the second semi-residual multi-scale feature extractor SRFE.

[0046] S4026. Repeat S4025 twice to obtain the final outputs E3 and E4 of the third and fourth semi-residual multi-scale feature extractors SRFE;

[0047] Further, the S405 includes the following steps:

[0048] S4051. For the output Ei of each layer of the encoder, calculate the L2 norm of each channel, reorder Ei by channel according to the value size, and represent the reordered result as Es_i;

[0049] S4052. Divide the reordered result Es_i into two parts on the channel. The first half is the part with large feature differences, and the second half is the part with small feature differences. By calculating the losses of these two parts, the model is constrained to extract more diverse features. The formula is as follows;

[0050]

[0051] Among them, represents the intra-class semantic consistency regularization loss, represents the result of the first half of the features extracted by the i-th layer decoder after reordering, represents the result of the second half of the features extracted by the i-th layer decoder after reordering, and N represents the total number of layers of the decoder.

[0052] S4053. Downsample the last layer of features D4 restored by the S403 decoder so that the scale of D4 is the same as the scale of the features Ei extracted by each layer of the encoder. The four downsampled features are respectively represented as d1, d2, d3, d4;

[0053] S4054. For the features Ei extracted by each layer of the encoder, perform random channel selection so that the number of channels of Ei is the same as that of D4. The features after random channel selection are represented as e1, e2, e3, e4;

[0054] S4055. After S4053 and S4054, d1, d2, d3, d4 and e1, e2, e3, e4 with the same spatial dimension and channel dimension will be obtained. By calculating the mean square error of their corresponding positions, the feature symmetry between the features extracted by the model and the features in the decoding process is constrained. The formula is as follows;

[0055]

[0056] Among them, represents the encoder-decoder semantic consistency regularization loss, represents the result of the output of each layer of the encoder after random channel selection, represents the results of D4 after being downsampled four times respectively, and N represents the total number of layers of the encoder.

[0057] Furthermore, the expression of the loss function of the model is as follows:

[0058]

[0059] Among them, represents the total loss, represents the regular loss of codec semantic consistency, represents the regular loss of intra-class semantic consistency, represents the weighted binary cross-entropy loss function, represents the dice coefficient loss function, Output represents the recognition result of the mesoscale convective image, and mask represents the label data.

[0060] Another object of the present invention is to provide a fast and efficient mesoscale convective system recognition system in the mid-low latitude region. The system specifically includes:

[0061] An input module for obtaining cloud top brightness temperature data in the mid-low latitude region;

[0062] A preprocessing module for constructing a mesoscale convective system recognition data set and preprocessing it;

[0063] A model construction module for constructing a mesoscale convective system recognition model;

[0064] A training module for training a mesoscale convective system recognition model;

[0065] An output module for inputting a test set into a mesoscale convective system recognition model that meets the index requirements to obtain the final mesoscale convective system recognition result.

[0066] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are:

[0067] First, a fast and efficient mesoscale convective system recognition method provided by the present invention solves the problems of large computational resource consumption, slow recognition speed, and small recognition area in the recognition of mesoscale convective systems MCSs by traditional methods, and solves the problem that most existing deep learning models only target single significant target recognition or camouflage target recognition tasks.

[0068] The present invention constructs a fast and efficient mesoscale convective system recognition model for the mid - low latitude region, and uses the constructed mesoscale convective system MCSs recognition dataset as the training set to train the model. First, a multi - scale feature extractor with four semi - residuals SRFE is used to extract hierarchical multi - scale features, which improves the model's perception ability of MCSs at different scales with fewer parameters, prevents the loss of small - scale features during downsampling, and provides sufficient guidance for accurately identifying MCSs at different scales. In addition, the semi - residual design can prevent the forward propagation of redundant features. The present invention proposes a piece - wise learning strategy. By segmenting the MCSs in the mid - latitude and low - latitude regions and inputting the segments into the model in sequence for learning, the model can learn the features of MCSs in the mid - latitude (low - latitude) region without being affected by the features of MCSs in the low - latitude (mid - latitude) region. In this way, the model can stably learn the feature expressions of MCSs in different regions. Finally, we propose semantic consistency regularization to constrain the features of the model in terms of the diversity of feature extraction, and to constrain the model to extract features symmetric to those obtained in the feature extraction and decoding processes.

[0069] Second, as auxiliary evidence of the positive effects of the present invention, it is also reflected in the following important aspects:

[0070] (1) The technical solution of the present invention fills the technical gaps at home and abroad in the industry: Currently, most methods in the field of MCSs recognition are based on thresholds, and the threshold method is slow in recognizing MCSs. The present invention uses deep learning technology for MCSs recognition, greatly improving the speed of MCSs recognition.

[0071] (2) The technical solution of the present invention solves the technical problems that people have always been eager to solve but have never succeeded in: Limited by computing resources and the time cost of MCSs recognition, most existing MCSs recognitions are limited to local areas. There is less research on the recognition of MCSs in a large range (mid - low latitude region). The technical solution of the present invention can meet the needs of MCSs recognition in a large range. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a schematic diagram of the mid - low latitude mesoscale convective system MCSs recognition dataset provided by the embodiment of the present invention;

[0073] Figure 2 is a schematic diagram of the network structure provided by the embodiment of the present invention;

[0074] Figure 3 is a schematic diagram of the structure of the multi - scale feature extractor with semi - residuals SRFE provided by the embodiment of the present invention;

[0075] Figure 4 is a flowchart of the fast and efficient method for recognizing mesoscale convective systems in the mid - low latitude region provided by the embodiment of the present invention;

[0076] Figure 5 is a schematic structural diagram of a decoder provided by an embodiment of the present invention;

[0077] Figure 6 is a schematic structural diagram of a calculation method for semantic consistency regularization provided by an embodiment of the present invention;

[0078] Figure 7 is a module diagram of a system for quickly and efficiently identifying mesoscale convective systems in the mid - low latitude regions provided by an embodiment of the present invention;

[0079] Figure 8 is a precision - recall curve and a threshold - F - measure curve on a dataset for identifying mesoscale convective systems (MCSs) in the mid - low latitude regions provided by an embodiment of the present invention;

[0080] Figure 9 is a qualitative result graph on a dataset for identifying mesoscale convective systems (MCSs) in the mid - low latitude regions provided by an embodiment of the present invention;

[0081] Figure 10 is a spatio - temporal qualitative result graph on a dataset for identifying mesoscale convective systems (MCSs) in a local area provided by an embodiment of the present invention;

[0082] Figure 11 is another spatio - temporal qualitative result graph on a dataset for identifying mesoscale convective systems (MCSs) in a local area provided by an embodiment of the present invention. Detailed implementation manners

[0083] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0084] Before explaining the present invention, the following terms are explained:

[0085] MCSs: Mesoscale Convective Systems, mesoscale convective systems.

[0086] SRFE: Semi - residual Feature Extractor, multi - scale feature extractor with semi - residuals.

[0087] The present invention proposes a mesoscale convective system identification model (MCSeg), Figure 2Schematic diagram of the MCSeg network structure proposed by the present invention. Its input data is Input, and the output is the recognition result map Output. Among them, the resolution of Input is C×W×H, where C represents the number of channels, H represents the height of the image, and W represents the width of the image. The resolution sizes of the dataset are: 1×5143×1715. The present invention proposes a piecewise learning strategy, a semi-residual multi-scale feature extractor SRFE, and semantic consistency regularization. The piecewise learning strategy slices the MCSs in the middle and low latitudes, and inputs the slices into the model in sequence for learning, so that the model can learn the features of the middle (low) latitude MCSs without being affected by the features of the low (middle) latitude MCSs during the learning process. In this way, the model can stably learn the feature expressions of different regions of MCSs, improve the stability during the model training process, and accelerate the model convergence. The semi-residual multi-scale feature extraction SRFE ( Figure 3 ) extracts hierarchical multi-scale features, improves the model's perception ability of MCSs at different scales with fewer parameters, prevents small-scale features from being lost during downsampling, and provides sufficient guidance for accurately identifying MCSs at different scales. Semantic consistency regularization is used to constrain the features of the model in terms of feature extraction diversity, and to constrain the model to extract features symmetric to those obtained in the decoding process.

[0088] As Figure 4 shown, an embodiment of the present invention provides a method for quickly and efficiently identifying mesoscale convective systems in the middle and low latitude regions. The method specifically includes:

[0089] S1. Obtain the cloud top brightness temperature data of the middle and low latitude regions;

[0090] S2. Construct a mesoscale convective system recognition dataset for the cloud top brightness temperature data. The dataset includes a training set and a test set;

[0091] S3. Construct a mesoscale convective system recognition model;

[0092] S4. Use the training set to train the mesoscale convective system recognition model;

[0093] S5. Based on the trained mesoscale convective system recognition model, obtain the mesoscale convective system recognition result;

[0094] S6. Based on S4 and S5, when the input data volume reaches 1 / 5 of the total data volume, input the test set into the mesoscale convective system recognition model for testing, and determine whether the current Dice index of the network test is the highest. If so, save the current model parameters and enter S7. Otherwise, return to S4;

[0095] S7. Input the test set into the mesoscale convective system recognition model that meets the index requirements to obtain the final mesoscale convective system recognition result.

[0096] Retrieve the cloud top brightness temperature data of the mid - low latitude region through meteorological satellite remote sensing technology. The cloud top brightness temperature reflects the intensity and scope of the convective system. After processing the spatial and temporal resolutions of these data, they are used to construct an identification dataset for mesoscale convective systems. The dataset is divided into a training set and a test set for model training and validation respectively.

[0097] Construct a deep learning model suitable for identifying mesoscale convective systems, such as a convolutional neural network (CNN). This model can extract spatial and texture features from the input cloud top brightness temperature data. Use the data in the training set to train the model. The model continuously adjusts its parameters through the backpropagation algorithm to minimize the loss function. During the training process, the model learns how to distinguish mesoscale convective systems from other weather systems.

[0098] During the model training process, whenever the input data volume reaches 1 / 5 of the total data volume, input the test set into the model for testing and calculate the Dice metric. The Dice metric is used to evaluate the accuracy and consistency of the model in segmenting mesoscale convective systems. If the Dice metric of the current test reaches the highest value, save the current model parameters; otherwise, return to continue training and optimize the model parameters to ensure the continuous improvement of the model's performance on the test set.

[0099] When the Dice metric of the model on the test set meets the requirements, input the entire test set into the trained identification model for comprehensive identification of mesoscale convective systems. The model accurately identifies the location and scope of mesoscale convective systems by extracting features and classifying the input data, and generates an identification result map. The final result provides an efficient and accurate analysis tool for meteorological research and disaster warning.

[0100] For S1, obtain the original cloud top brightness temperature data. In this embodiment, obtain the original global cloud top brightness temperature data (European Union Cloud Archive User Service (CLAUS) project dataset), and the data range is (60°S–60°N, 180°W–180°E).

[0101] S2 includes:

[0102] S201: Set the temperature threshold for mesoscale convective systems MCSs, and mark the areas with temperatures lower than the threshold from the original cloud top brightness temperature data.

[0103] S202. Set the regional coverage threshold for mesoscale convective systems (MCSs). For the regions marked in S201, further screening is carried out using the regional coverage threshold. Select the regions in S201 with an area larger than the regional coverage threshold, and mark the places that meet the requirements of both thresholds as the regions of mesoscale convective systems (MCSs). Among them, the regions of mesoscale convective systems (MCSs) are marked as 1, and the non-mesoscale convective systems (MCSs) are marked as 0. After the above processing, the labels of mesoscale convective systems (MCSs) are obtained, and the label size is 5143×1715;

[0104] S203. Fill the null values in the cloud top brightness temperature data using the interpolation method. In addition, process the outliers in the data. Specifically, those with a value greater than 345 in the cloud top brightness temperature data are replaced with 345, and those with a value less than 170 are replaced with 170. The obtained cloud top brightness temperature data size is 5143×1715;

[0105] S204. Perform edge padding on the mesoscale convective system (MCSs) labels obtained in S202 and the cloud top brightness temperature data obtained in S203. Specifically, fill the cloud top brightness temperature data with a value of 300, and fill the label data with a value of 0. The sizes of both after padding are 5632×2048.

[0106] S205. Divide the padded mesoscale convective system (MCSs) labels and cloud top brightness temperature data obtained in S204 into a training set and a test set. Among them, both the training set and the test set include cloud top brightness temperature data and mesoscale convective system (MCSs) labels, and the storage format is numpy.

[0107] Through the above design, the present invention can construct the data set required for training the described mesoscale convective system recognition model.

[0108] In this embodiment, the specific implementation steps of the mesoscale convective system recognition model (MCSeg) proposed by the present invention are as follows:

[0109] A1. Identify mesoscale convective systems (MCSs) from the original cloud top brightness temperature data. In this step, the threshold method is used to identify and label the mesoscale convective systems (MCSs) in the cloud top brightness temperature data. First, set the temperature threshold for mesoscale convective systems (MCSs) to 233K, and preliminarily label the regions with a temperature less than 233K.

[0110] A2. Set the regional coverage threshold for mesoscale convective systems (MCSs) to 5000 km 2 , and further screen the regions marked in A1. Select the regions with a coverage area larger than 5000 km 2Regions, and mark these regions as mesoscale convective systems MCSs. Among them, the regions of mesoscale convective systems MCSs are marked as 1, and the regions that are not mesoscale convective systems MCSs are marked as 0.

[0111] A3. Convert the original cloud top brightness temperature data and the annotation data obtained in A2 into the numpy format. The size of a single image is 5143×1715, and there are a total of 4392 groups of data.

[0112] A4. For the cloud top brightness temperature data whose median is greater than 345, use 345 for replacement, and for the cloud top brightness temperature data whose median is less than 170, use 170 for replacement. The size of the obtained cloud top brightness temperature data is 5143×1715.

[0113] A5. Divide the training set and the test set. The data volumes of the finally obtained training set and test set are 2920 and 1472 respectively.

[0114] S3. Build a mesoscale convective system identification model, and its implementation method is as follows:

[0115] S301. Process the cloud top brightness temperature data obtained in S204 using max - min normalization and convert it into the tensor format, where the maximum value is taken as 345 and the minimum value is taken as 170. The MCS label data obtained in S204 remains unchanged;

[0116] S302. Build a mesoscale convective system identification model according to the normalization processing result of S301.

[0117] The present invention improves the stability in the model training process and accelerates the model convergence by pre - processing the constructed data set.

[0118] In this embodiment, the mesoscale convective system identification model includes:

[0119] An encoder, which is used to extract hierarchical pyramid features after the normalization processing of S301. Among them, the encoder includes four semi - residual multi - scale feature extractors SRFE; the semi - residual multi - scale feature extractor SRFE is used to extract multi - scale features to prevent the loss of small - scale features during the down - sampling process. The semi - residual design can prevent the forward propagation of redundant features;

[0120] A decoder, which is used to decode and restore pyramid features at different levels. Among them, the decoder includes a convolution and up - sampling module. The structural schematic diagram of the decoder is as Figure 5 shown.

[0121] In this embodiment, build and initialize the mesoscale convective system identification model, and send the training set pre - processed by S2 into the small - target recognition network respectively, specifically including:

[0122] B1. Convert the cloud top brightness temperature data into the tensor format tensor, and normalize all pixel values in the cloud top brightness temperature data to 0 to 1 using min-max normalization;

[0123] B3. Mesoscale convective system identification model;

[0124] In this embodiment, the mesoscale convective system identification model includes an encoder with a multi-scale feature extractor SRFE embedded with semi-residuals and a decoder, and the overall model is an encoder-decoder architecture. The encoder of the mesoscale convective system identification model consists of four multi-scale feature extractors SRFE with semi-residuals, which are used to extract multi-scale hierarchical pyramid features of the input data. The decoder is stacked by convolutional and upsampling modules and an object recognition head, which is used to decode and recover the pyramid features at different levels step by step, and the recognition head is used for the final recognition.

[0125] In this embodiment, the input of the mesoscale convective system identification model is the cloud top brightness temperature data obtained by S205, denoted as Input. First, Input is split to obtain 11×4 data with a size of 512×512. Every 11 groups of data are divided into a batch, and a total of 4 batch data are obtained, denoted as B i . Among them, the first batch B1 and the fourth batch B4 are cloud top brightness temperature data in the mid-latitude region, and the second B2 and the third batch B3 are data in the low-latitude region. Pass B i into the encoder to extract multi-scale feature information, and obtain the feature maps of each layer, which are respectively denoted as feature map Ei, where i represents the i-th layer of the encoder. Subsequently, the feature map Ei is passed into the decoder. The decoder takes the output E4 of the last layer of the encoder as the input, and obtains the output D1 of the first layer of the decoder. The input of the second layer of the decoder comes from the output D1 of the first layer of the decoder and the output E3 of the third layer of the encoder. The input of the third layer of the decoder comes from the output D2 of the second layer of the decoder and the output E2 of the second layer of the encoder. The input of the fourth layer of the decoder comes from the output D3 of the third layer of the decoder and the output E1 of the first layer of the encoder. Finally, the output of the fourth layer of the decoder is input into the recognition head to obtain the final mesoscale convective system MCSs recognition map Output. Finally, the dice coefficient loss function, binary cross-entropy loss, and semantic consistency regularization loss are used to constrain the difference degree between the prediction map Output and the original label mask, and the calculated loss value is passed into the network for backpropagation to automatically adjust the parameters of the network.

[0126] Figure 6 It is a schematic structural diagram of the calculation method of the semantic consistency regularization proposed by the present invention.

[0127] In this embodiment, when the input data volume reaches 1 / 5 of the total data volume, the test set is passed into the mesoscale convective system recognition model for testing. It is determined whether the current Dice metric of the network test is the highest. If so, the current model parameters are saved; otherwise, training continues. Dice represents the Dice coefficient.

[0128] Step S4: Train the mesoscale convective system recognition model using the training set, and its implementation method is as follows:

[0129] S401: For the cloud top brightness temperature data after normalization in S301, the size is 5632×2048. First, the data is split to obtain 11×4 data with a size of 512×512. Every 11 groups of data are divided into a batch, and a total of 4 batches of data are obtained, denoted as B i . Among them, the first batch B1 and the fourth batch B4 are cloud top brightness temperature data in the mid-latitude region, and the second batch B2 and the third batch B3 are data in the low-latitude region;

[0130] S402: Input the first batch B1 into the semi-residual multi-scale feature extractor SRFE to extract hierarchical multi-scale pyramid features, and output the feature maps of each layer of the encoder , and its implementation method is as follows:

[0131] S4021: Input the first batch B1 into the first semi-residual multi-scale feature extractor SRFE of the encoder. First, copy B1 4 times in the channel dimension. The first copy is subjected to feature extraction using a 1×1 convolution to obtain the first branch feature map F1, the second copy is subjected to feature extraction using a 3×3 convolution to obtain the second branch feature map F2, the third copy is subjected to feature extraction using a 5×5 convolution to obtain the third branch feature map F3, and the fourth copy is subjected to feature extraction using a 7×7 convolution to obtain the fourth branch feature map F4;

[0132] S4022: Concatenate the four feature maps F1, F2, F3, and F4 obtained in S4021 in the channel dimension and then use a 1×1 convolution for multi-scale feature fusion to obtain the preliminary fusion feature map M1

[0133] S4023: Split the preliminary fusion feature map M1 into 4 parts in the channel dimension. The first part is subjected to feature extraction using a 1×1 convolution to obtain the first branch feature map F 11 , the second part is subjected to feature extraction using a 3×3 convolution to obtain the second branch feature map F 22 , the third part is subjected to feature extraction using a 5×5 convolution to obtain the third branch feature map F 33 , and the fourth part is subjected to feature extraction using a 7×7 convolution to obtain the fourth branch feature map F44 ;

[0134] S4024. Concatenate the four feature maps F 11 , F 22 , F 33 , F 44 on the channel dimension and then use a 1×1 convolution for multi-scale feature fusion to obtain the final fused feature map M 11 . Add M 11 to M1 to obtain the final output E1 of the multi-scale feature extractor SRFE of the first semi-residual.

[0135] S4025. Input the final output E1 of the multi-scale feature extractor SRFE of the first semi-residual into the multi-scale feature extractor SRFE of the next semi-residual, and perform the operations of S4021 - S4025 to obtain the final output E2 of the multi-scale feature extractor SRFE of the second semi-residual.

[0136] S4026. Repeat S4025 twice to obtain the final outputs E3 and E4 of the multi-scale feature extractors SRFE of the third and fourth semi-residuals;

[0137] S403. In the decoder part, upsample the feature map Di output by the decoder, concatenate it with the feature map Ei output by the encoder on the channel dimension, and then input it into a convolution and perform upsampling to obtain Di, where Di represents the feature restored by the i-th layer of the decoder;

[0138] S404. Input the output of the last layer of the decoder into the convective recognition head SegHead to obtain the recognition result of the mesoscale convective system.

[0139] S405. When optimizing the model through backpropagation, we use semantic consistency regularization to optimize the model. The implementation method is as follows:

[0140] S4051. For the output Ei of each layer of the encoder, calculate the L2 norm of each channel, reorder the channels according to the value size, and represent the reordered result as Es_i;

[0141] S4052. Divide the reordered result Es_i into two parts on the channel. The first half is the part with large feature differences, and the second half is the part with small feature differences. By calculating the losses of these two parts, the model is constrained to extract more diverse features;

[0142]

[0143] Among them, represents the intra-class semantic consistency regularization loss, It represents the first half of the result after reordering the features extracted by the i-th layer decoder. It represents the second half of the result after reordering the features extracted by the i-th layer decoder, and N represents the total number of layers of the decoder.

[0144] S4053. Downsample the last layer of features D4 restored by the S403 decoder so that the scale of D4 is the same as the scale of the features Ei extracted by each layer of the encoder. The four downsampled features are respectively denoted as d1, d2, d3, d4.

[0145] S4054. For the features Ei extracted by each layer of the encoder, perform random channel selection so that the number of channels of Ei is the same as that of D4. The features after random channel selection are denoted as d1, d2, d3, d4.

[0146] S4055. After S4053 and S4054, d1, d2, d3, d4 and e1, e2, e3, e4 with the same spatial dimension and channel dimension will be obtained. Calculate the mean square error of the corresponding positions of the two to constrain the feature symmetry between the features extracted by the model and the features in the decoding process.

[0147]

[0148] Among them, It represents the regular loss of the semantic consistency between encoding and decoding. It represents the result after random channel selection of the output of each layer of the encoder. It represents the results of D4 being downsampled four times respectively, and N represents the total number of layers of the encoder.

[0149] S406. Repeat S402 - S405 until 4 batches of data B i are all input into the model.

[0150] The present invention prevents the loss of small-scale features in the feature extraction process by designing a semi-residual multi-scale feature extractor SRFE. The semi-residual method can enable the model to more effectively transmit the information flow and prevent the transmission of redundant features.

[0151] Constrains the redundant feature extraction and the feature asymmetry in the feature recovery process through the intra-class semantic consistency regular loss and the encoding-decoding semantic consistency regular loss.

[0152] In this embodiment, calculate the loss between the final predicted segmentation map Output in step S5 and the label data mask in step S302. The expression of the loss function of the mesoscale convective system recognition model is as follows:

[0153]

[0154] Among them, represents the total loss, represents the regular loss of codec semantic consistency, represents the regular loss of intra-class semantic consistency, represents the weighted binary cross-entropy loss function, represents the dice coefficient loss function, Output represents the recognition result of the mesoscale convective image, and mask represents the label data.

[0155] Through the above loss function, the present invention can optimize the mesoscale convective system recognition model and accelerate the model convergence.

[0156] In this embodiment, the mesoscale convective system recognition model with the best performance during the test process is taken out, and then the test set is sequentially fed into this network. Seven quantitative indicators are used to evaluate the performance of this model: average dice coefficient mDice, average intersection over union mIoU, weighted F-index , structural similarity , enhanced alignment index E, mean absolute error , and maximum F-index , and save the MCSs recognition map Output. It is used to evaluate the average value of the difference between the predicted value and the true value. The smaller the difference, the closer the predicted result is to the true result. represents the best performance of the segmentation result. is the weighted average of precision and recall, and evaluates the overall performance of the recognition result by balancing precision and recall. is used to evaluate the structural similarity between the recognition result and the true result. E measures the average structural similarity between the recognition result and the true result. mIoU is used to compare the similarity and diversity of the sample set. mDice is used to calculate the similarity of two samples. Except for M, all the above indicators are closer to 1, indicating a better recognition result. The specific calculation formulas of the above indicators are as follows:

[0157]

[0158]

[0159]

[0160]

[0161]

[0162]

[0163]

[0164] Among them, represents the set of network prediction results of the mesoscale convective system identification model MCSeg, represents the set of corresponding true labels, K represents the number of categories, P represents precision, and R represents recall. and represent the regional structural similarity and the structural similarity of objects respectively. represents the enhanced alignment matrix, and w and h represent the width and height of the output result respectively.

[0165] As Figure 7 shown, a fast and efficient mesoscale convective system identification system in the mid - low latitude region provided by an embodiment of the present invention specifically includes:

[0166] An input module for obtaining cloud - top brightness temperature data in the mid - low latitude region;

[0167] A pre - processing module for constructing a mesoscale convective system identification data set and pre - processing it;

[0168] A model construction module for constructing a mesoscale convective system identification model;

[0169] A training module for training the mesoscale convective system identification model;

[0170] An output module for inputting a test set into the mesoscale convective system identification model that meets the index requirements to obtain the final mesoscale convective system identification result.

[0171] The present invention belongs to the field of image processing, and discloses a fast and efficient mesoscale convective system identification method and system in the mid - low latitude region, including obtaining cloud - top brightness temperature data; constructing a mesoscale convective system identification data set based on the cloud - top brightness temperature data; constructing a mesoscale convective system identification model; training the model using the constructed training set; after reaching the specified number of training rounds, testing the model with a test set to obtain the final mesoscale convective system identification result.

[0172] To verify the effectiveness of the method of the present invention, the proposed method is compared with other advanced methods. For a fair comparison, the official released codes of other methods are used, and the experimental settings are followed, where all methods are implemented in the same computing environment, and quantitative and qualitative analyses are carried out simultaneously. Table 1 gives the quantitative comparison results of seven indicators of the mesoscale convective systems MCSs data set on 11 different network structures.

[0173] Table 1

[0174]

[0175] As shown in Table 1, this table illustrates the performance of our model in MCSs recognition. It can be seen that the model (MCSeg) proposed in the present invention outperforms other models in all metrics. Specifically, on the MCSs recognition dataset, MCSeg performs excellently in each of the seven metrics: it is 0.0016 in and 0.9692 in , 0.9866 in E, and 0.9809 in , 0.9624 in , 0.9119 in IoU, and 0.9539 in Dice. In addition, in this embodiment, a precision-recall curve and an F-measure curve are also provided to further evaluate the performance of all methods on the MCSs recognition dataset, as shown in Figure 8 . It can be seen that the red solid line represents MCSeg, which outperforms other models at most thresholds. The excellent performance of this model is attributed to the carefully designed semi-residual encoder and semantic consistency regularity constraint. The semi-residual encoder enhances the multi-scale feature extraction ability of the model without increasing the number of parameters. At the same time, the semantic consistency regularity further improves the ability of the model to extract diverse features.

[0176] In this embodiment, in order to verify the generalization ability of our model, we performed MCSs recognition on the brightness temperature data from 2011 to 2014 for a total of 4 years, and the results are shown in Tables 2, 3, 4, and 5. All 7 metrics are similar to the results obtained from the test set, which indicates that our model has good generalization ability. Although the model was only trained with the data in 2022, it still performs well in the convective recognition results from 2011 to 2014, which is a large time span. It can be concluded that our model accurately learns the feature distribution of MCSs, and the performance of the model does not decline over the years.

[0177] Table 2

[0178]

[0179] Table 3

[0180]

[0181] Table 4

[0182]

[0183] Table 5

[0184]

[0185] Table 6

[0186]

[0187] In this embodiment, to verify the effectiveness of the module design, learning strategy, and loss function, we conducted ablation experiments, as shown in Table 6. In the table, M represents the multi-scale residual semantic encoder (MRSE), S represents the slice learning strategy, and L represents the semantic consistency regularization. Taking the metric as an example, after adding the multi-scale residual semantic encoder (MRSE), the performance of the baseline model improved from 0.0027 to 0.0019. After adding the slice training strategy, the metric further improved to 0.0020, and the addition of semantic consistency regularization increased it to 0.0024. By gradually integrating these components into the baseline model, the overall performance of the model was significantly improved.

[0188] In this embodiment, Figure 9 shows the recognition results of MCSs in the mid-latitude and low-latitude regions at 00:00 on March 1, 2023. Overall, there is little difference between the results of the MCSs recognized by the model proposed in the present invention and the threshold-based method. Based on the slice training strategy we designed, the model can well learn the features of MCSs in different latitudes. The cross-learning method enables the model to balance the low-latitude region and the mid-latitude region and avoid the model forgetting the features learned previously. To more clearly observe the differences, we plotted a difference map ( Figure 9 ). In the difference map, the green area represents the over-recognized area (non-MCS is recognized as MCS), and the red area represents the under-recognized area (MCS is not accurately recognized). We selected four regions (including the mid-latitude and low-latitude regions) for magnification, and some over-recognized and under-recognized situations can be observed. In Region 1, most of the misrecognized areas are located at the edge of the main body, which is the result of too large a difference in the convection pattern. In Region 2, the MCSs are mostly isolated from each other. In Regions 3 and 4, the misrecognized areas are independent of the main body, which can be attributed to the result of too large a scale difference in the MCSs. Since a single image is too large, we did not compare the recognition results of the mid-latitude and low-latitude regions of other models here.

[0189] In this embodiment, Figure 10 and Figure 11Shows the spatio-temporal recognition performance of all models (recognition in consecutive time periods). The selected time period is from 00:00 to 21:00 on March 1, 2023, with a total of eight frames of convective data. From the graph, the recognition results of our model are closer to the true labels. Some relatively good models, such as SwinU and CPD, have better edge recognition ability for MCSs, but the recognition errors for small-scale MCSs are relatively serious. The remaining models (C2FNet, HitNet, Poly-PVT, PraNet, SINet) can only recognize the main part of MCSs and are not detailed enough in the recognition of the edge region. In the edge recognition of MCSs, the problems of over-recognition and mutual adhesion of adjacent MCSs are more serious. Compared with these methods, our model can accurately locate and recognize the content and morphological structure of MCSs in complex scenarios.

[0190] Regardless of the quantitative or qualitative results, our model performs excellently. In terms of the recognition results, the MCSs recognized by the model are very similar to those obtained by the threshold-based method. The use of deep learning technology makes our method much faster in recognition speed than the threshold-based method. It takes 2 hours and 42 minutes to recognize MCSs with a time span of one month using the threshold method, while it only takes 1 minute and 40 seconds using our method, and the recognition speed is increased by 200 times.

[0191] In summary, based on satellite data, traditional threshold methods can accurately capture the evolution of MCSs in short-term and local regions, including their germination, maturity, and extinction. However, these methods are all based on thresholds, that is, by calculating the scale size and temperature of a single target to determine whether it is an MCS. When applied globally, especially in tropical and mid-latitude regions, the time required for threshold methods to determine whether a target is an MCS is quite long. The huge time cost will lead to slow progress in downstream tasks (tracking and prediction), affecting the analysis of the climatological characteristics of MCSs. The present invention constructs a mesoscale convective system identification model (MCSeg) to identify mesoscale convective systems MCSs in the mid-latitudes and low-latitudes. In addition, the present invention also constructs a mid-latitudes and low-latitudes MCSs identification dataset for the training and testing of the mesoscale convective system identification model. The present invention proposes a piecewise learning strategy, a semi-residual multi-scale feature extractor SRFE, and semantic consistency regularization. The piecewise learning strategy slices the MCSs in the mid-latitudes and low-latitudes, and inputs the slices into the model in sequence for learning, so that the model can learn the characteristics of MCSs in the mid-latitudes (low-latitudes) without being affected by the characteristics of MCSs in the low-latitudes (mid-latitudes). In this way, the model can stably learn the feature expressions of MCSs in different regions, improve the stability during the model training process, and accelerate the model convergence. The semi-residual multi-scale feature extraction SRFE extracts hierarchical multi-scale features, improves the model's perception ability of MCSs of different scales with fewer parameters, prevents small-scale features from being lost during downsampling, and provides sufficient guidance for accurately identifying MCSs of different scales. The semantic consistency regularization is used to constrain the model to extract diverse features during feature extraction, and to constrain the model to extract features symmetric to those obtained during the decoding process. Finally, it is proved on the constructed dataset that the present invention can quickly and accurately identify mesoscale convective systems MCSs in the mid-latitudes and low-latitudes. Compared with other deep learning models, the mesoscale convective system identification model proposed by the present invention has better performance, solving the problems of large computational resource consumption, slow identification speed, and small identification area in the identification of mesoscale convective systems MCSs by traditional methods.

[0192] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated designed hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, such as provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software such as firmware.

[0193] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. A method for identifying a mesoscale convective system quickly and efficiently in the mid - low latitude region, characterized in that, The method specifically includes: S1. Obtain the cloud top brightness temperature data of the mid - low latitude region; S2. Construct a mesoscale convective system identification dataset for the cloud top brightness temperature data, where the dataset includes a training set and a test set; S3. Construct a mesoscale convective system identification model; S4. Use the training set to train the mesoscale convective system identification model; S5. Based on the trained mesoscale convective system identification model, obtain the mesoscale convective system identification result; S6. After S4 and S5 reach the specified number of training times in sequence, input the test set into the trained mesoscale convective system identification model to determine whether the current mesoscale convective system identification model meets the index requirements. If so, save the parameters of the current model and enter S7; otherwise, return to S4; S7. Input the test set into the mesoscale convective system identification model that meets the index requirements to obtain the final mesoscale convective system identification result; The mesoscale convective system identification model includes: An encoder, which is used to extract the hierarchical pyramid features after being normalized by S301. Among them, the encoder includes four semi - residual multi - scale feature extractors SRFE; the semi - residual multi - scale feature extractor SRFE is used to extract multi - scale features; A decoder, which is used to decode and restore the pyramid features of different levels. Among them, the decoder includes a convolution and up - sampling module; S4 includes the following steps: S401. For the cloud top brightness temperature data after normalization in S301, with a size of 5632×2048, first split the data to obtain 11×4 data with a size of 512×512. Divide every 11 groups of data into a batch, and a total of 4 batches of data are obtained, denoted as B i , where the first batch B1 and the fourth batch B4 are cloud top brightness temperature data in the mid-latitude region, and the second B2 and the third batch B3 are data in the low-latitude region; S402. Input the data in the first batch B1 into the semi - residual multi - scale feature extractor SRFE to extract hierarchical multi - scale pyramid features, and output the feature map Ei of each layer of the encoder; S403. In the decoder part, splice the feature map Ei output by the encoder and the feature map Di output by the decoder in the channel dimension and then input them into the convolution and perform up - sampling to obtain Di, where Di represents the feature restored by the i - th layer of the decoder; S404. Input the output of the last layer of the decoder into the convective recognition head SegHead to obtain the mesoscale convective system identification result; S405. When optimizing the model through backpropagation, use semantic consistency regularization to optimize the model; S406. Repeat S401 - S405 until 4 batches of data B i are all input into the model.

2. The method for identifying a mesoscale convective system quickly and efficiently in the mid - low latitude region as described in claim 1, wherein S2 includes the following steps: S201. Set the temperature threshold of the mesoscale convective system MCSs, and mark the area smaller than the temperature threshold from the original cloud top brightness temperature data; S202. Set the regional coverage threshold of the mesoscale convective system MCSs. For the area marked in S201, further screen it using the regional coverage threshold, select the area in S201 with an area larger than the regional coverage threshold, and mark the places that meet both threshold requirements as the mesoscale convective system MCSs area. Among them, the mesoscale convective system MCSs area is marked as 1, and the non - mesoscale convective system MCSs is marked as 0. After the mesoscale convective system MCSs area marking process, obtain the mesoscale convective system MCSs label, and the label size is 5143×1715; S203. Fill the null values in the cloud top brightness temperature data using the interpolation method, and process the outliers in the data. Replace the values greater than 345 in the cloud top brightness temperature data with 345, and replace the values less than 170 with 170. The size of the obtained cloud top brightness temperature data is 5143×1715; S204. Perform edge padding on the mesoscale convective system MCSs labels obtained in S202 and the cloud top brightness temperature data obtained in S203. Fill the cloud top brightness temperature data with a value of 300, and fill the label data with a value of 0. The size of both after padding is 5632×2048; S205. According to the padded mesoscale convective system MCSs labels and cloud top brightness temperature data obtained in S204, divide them into a training set and a test set. Both the training set and the test set include cloud top brightness temperature data and mesoscale convective system MCSs labels, and the data is saved in the npy format. npy is a data format.

3. The method for identifying a mesoscale convective system in the mid - low latitude region quickly and efficiently as described in claim 1, wherein, The said S3 includes the following steps: S301. Process the cloud top brightness temperature data obtained in S204 using maximum-minimum normalization and convert it to the tensor format, where the maximum value is taken as 345 and the minimum value is taken as 170. The MCS label data obtained in S204 remains unchanged; S302. Construct a mesoscale convective system recognition model according to the normalization result of S301.

4. The method for identifying a mesoscale convective system in the mid - low latitude region quickly and efficiently as described in claim 1, wherein, The said S402 includes the following steps: S4021. Input the first batch B1 into the first semi-residual multi-scale feature extractor SRFE of the encoder. First, duplicate B1 four times in the channel dimension. Use a 1×1 convolution to extract features from the first copy to obtain the first branch feature map F1, use a 3×3 convolution to extract features from the second copy to obtain the second branch feature map F2, use a 5×5 convolution to extract features from the third copy to obtain the third branch feature map F3, and use a 7×7 convolution to extract features from the fourth copy to obtain the fourth branch feature map F4; S4022. Concatenate the four feature maps F1, F2, F3, and F4 obtained in S4012 in the channel dimension and then use a 1×1 convolution for multi-scale feature fusion to obtain the preliminary fusion feature map M1; S4023. Split the preliminary fusion feature map M1 into four parts along the channel dimension. Use a 1×1 convolution to extract features from the first part to obtain the first branch feature map F 11 , use a 3×3 convolution to extract features from the second part to obtain the second branch feature map F 22 , use a 5×5 convolution to extract features from the third part to obtain the third branch feature map F 33 , use a 7×7 convolution to extract features from the fourth part to obtain the fourth branch feature map F 44 ; S4024. Concatenate the four feature maps F 11 , F 22 , F 33 , F 44 along the channel dimension, and then use 1×1 convolution for multi-scale feature fusion to obtain the final fused feature map M 11 . Add M 11 to M1 to get the final output E1 of the first half-residual multi-scale feature extractor SRFE; S4025. Input the final output E1 of the first semi-residual multi-scale feature extractor SRFE into the next semi-residual multi-scale feature extractor SRFE, and perform the operations of S4021 - S4025 to obtain the final output E2 of the second semi-residual multi-scale feature extractor SRFE; S4026. Repeat S4025 twice to obtain the final outputs E3 and E4 of the third and fourth semi-residual multi-scale feature extractors SRFE.

5. The method for identifying a mesoscale convective system quickly and efficiently in the mid - low latitude region according to claim 1, wherein, The said S405 includes the following steps: S4051. For the output Ei of each layer of the encoder, calculate the L2 norm of each channel, reorder Ei by channel according to the value size, and represent the reordered result as Es_i; S4052. Divide the re - ordered result Es_i into two parts on the channel. The first half is the part with large feature differences, and the second half is the part with small feature differences. Constrain the model to extract more diverse features by calculating the losses of these two parts. The formula is as follows; Among them, represents the intra-class semantic consistency regularization loss, represents the result of the first half of the features extracted by the i-th layer decoder after reordering, represents the result of the second half of the features extracted by the i-th layer decoder after reordering, and N represents the total number of layers of the decoder; S4053. Downsample the last - layer feature D4 restored by the S403 decoder so that the scale of D4 is the same as the scale of the features Ei extracted by each layer of the encoder. The four downsampled features are respectively denoted as d1, d2, d3, d4; S4054. For the feature Ei extracted by each layer of the encoder, perform random channel selection so that the number of channels of Ei is the same as that of D4. The feature after random channel selection is denoted as e1, e2, e3, e4; After S4053 and S4054, d1, d2, d3, d4 and e1, e2, e3, e4 with the same spatial dimension and channel dimension will be obtained. Constrain the symmetry of the features extracted by the model and the features in the decoding process by calculating the losses at the corresponding positions. The formula is as follows; Among them, represents the regular loss of codec semantic consistency, represents the result after random channel selection of the output of each layer of the encoder, represents the results of D4 after four downsamplings respectively, and N represents the total number of layers of the encoder.

6. The method for identifying a mesoscale convective system quickly and efficiently in the mid - low latitude region as described in claim 1, wherein The expression of the loss function of the mesoscale convective system recognition model is as follows: Among them, represents the total loss, represents the regular loss of codec semantic consistency, represents the regular loss of intra-class semantic consistency, represents the weighted binary cross-entropy loss function, represents the dice coefficient loss function, Output represents the recognition result of the mesoscale convective image, and mask represents the label data.

7. A rapid and efficient mesoscale convective system identification system for the mid - low latitude region, which realizes the rapid and efficient mesoscale convective system identification method for the mid - low latitude region described in any one of claims 1 - 6, is characterized in that The fast and efficient mesoscale convective system recognition system in the mid - low latitude region specifically includes: An input module for obtaining the cloud - top brightness temperature data in the mid - low latitude region; A pre - processing module for constructing and pre - processing the mesoscale convective system recognition data set; A model construction module for constructing the mesoscale convective system recognition model; A training module for training the mesoscale convective system recognition model; An output module for inputting the test set into the mesoscale convective system recognition model that meets the index requirements to obtain the final mesoscale convective system recognition result.

Images