Method for determining typhoon rapid intensification trend, electronic device, and storage medium

Abstract

The application provides a typhoon rapid enhancement trend discrimination method, an electronic device and a storage medium, and belongs to the technical field of computer vision. The method comprises the following steps: preprocessing a remote sensing image sequence and a corresponding intensity sequence of an acquired typhoon to be measured to obtain a to-be-measured data sequence of the typhoon to be measured; calculating a life cycle identifier of the typhoon to be measured; inputting the to-be-measured data sequence and the life cycle identifier of the typhoon to be measured into a rapid enhancement trend discrimination model to obtain a confidence degree of rapid enhancement of the typhoon to be measured; comparing the confidence degree of rapid enhancement of the typhoon to be measured with a rapid enhancement discrimination threshold value; and when the confidence degree of rapid enhancement of the typhoon to be measured is greater than the rapid enhancement discrimination threshold value, determining that the typhoon to be measured has a rapid enhancement trend. The application aims to more accurately predict and discriminate the rapid enhancement trend of a typhoon.

Description

Technical Field

[0001] This application relates to the field of computer vision, and more specifically, to a method for determining the rapid intensification trend of a typhoon, an electronic device, and a storage medium. Background Technology

[0002] Tropical cyclones, or typhoons, are among the most destructive weather systems on Earth. Predicting typhoon trajectories and intensities is of great concern. Over the past half-century, significant improvements have been made in forecasting typhoon trajectories and intensities. However, predicting rapid intensification (RI) of typhoons remains challenging. Previous research on RI prediction has focused on using predictor variables as features, such as the SHIPS series of statistical typhoon intensity forecasts, or analyzing the climatic and spatiotemporal characteristics of tropical cyclones, such as season, month, latitude and longitude, pressure changes, and monthly average sea surface temperature fields. Meteorologists use statistics to identify patterns as important criteria for determining whether a tropical cyclone will rapidly intensify.

[0003] However, research relying on predictive variables lacks exploration of new technologies to better utilize raw data, and the patterns discovered can only serve as auxiliary means, making it difficult to quickly and accurately predict whether a typhoon will rapidly intensify. Summary of the Invention

[0004] This application provides a method, electronic device, and storage medium for determining the rapid intensification trend of typhoons, aiming to more accurately predict and determine the rapid intensification trend of typhoons.

[0005] In a first aspect, embodiments of this application provide a method for determining the rapid intensification trend of typhoons in remote sensing cloud image sequences, the method comprising:

[0006] The acquired remote sensing image sequence of the typhoon to be measured and the corresponding intensity sequence are preprocessed to obtain the data sequence of the typhoon to be measured.

[0007] Calculate the lifecycle identifier of the typhoon to be tested;

[0008] The preprocessed data sequence to be tested and the life cycle identifier of the typhoon to be tested are used as inputs. The rapid enhancement trend discrimination model is used for identification to obtain the confidence of the rapid enhancement of the typhoon to be tested. The rapid enhancement trend discrimination model is based on a sample training set composed of remote sensing image sequences, corresponding intensity sequences, rapid enhancement identifiers and life cycle identifiers of multiple sample typhoons, and is trained by a combination of resampling and reweighting.

[0009] The confidence level of the rapid intensification of the typhoon under test is compared with the rapid intensification discrimination threshold. When the confidence level of the rapid intensification of the typhoon under test is greater than the rapid intensification discrimination threshold, it is determined that the typhoon under test has a rapid intensification trend.

[0010] Optionally, the fast augmentation trend discrimination model includes multiple sequentially connected modules: a CNN model, an LSTM time series model, and a classifier; the training process of the fast augmentation trend discrimination model includes the following steps:

[0011] Based on the remote sensing image sequences of multiple sample typhoons and their corresponding intensity sequences, the CNN model is pre-trained to obtain the first feature extraction module after pre-training. The first feature extraction module after pre-training is used to extract the visual spatial features of a single remote sensing image of a typhoon.

[0012] Based on the remote sensing image sequences and corresponding intensity sequences of multiple sample typhoons, as well as the processing results of the pre-trained CNN module on the remote sensing image sequences and corresponding intensity sequences of multiple sample typhoons, the fast enhancement trend discrimination model is trained to obtain the trained second feature extraction module. The second feature module includes a CNN model and an LSTM temporal model. The trained second feature extraction module is used to extract the spatiotemporal correlation features of typhoon visual morphology changes and intensity changes.

[0013] The model parameters of the pre-trained LSTM time series model and CNN model are frozen. The remote sensing image sequences, corresponding intensity sequences, rapid enhancement labels and life cycle labels of each sample typhoon are repeatedly used. The classifier of the rapid enhancement trend discrimination model is retrained by combining resampling and reweighting. The trained classifier is used to predict the confidence level of rapid enhancement of typhoons.

[0014] Optionally, before pre-training the CNN model, the following steps are also included:

[0015] The remote sensing image sequences of the sample typhoons are processed to obtain image sequences in the target format;

[0016] The intensity sequences of the sample typhoons were processed to fill in the missing data in the intensity sequences;

[0017] Based on the complete life cycle of the sample typhoons, the maximum wind speed node of the sample typhoons is determined. Using time as the sequence, the life cycle of the typhoon sequence before the maximum wind speed node is determined as the rising phase, and the life cycle of the typhoon sequence after the maximum wind speed node is determined as the falling phase.

[0018] Optionally, when training the classifier, the sample training set is resampled according to a specified ratio to obtain positive and negative samples with a preset distribution to train the classifier. Data of typhoons with the first identifier are quickly enhanced and data of typhoons with the second identifier are quickly enhanced and are negative samples.

[0019] Optionally, when training the classifier, the loss of positive and negative samples is adjusted according to the weights corresponding to the positive and negative samples, so that the classifier is trained based on positive and negative samples from a preset distribution.

[0020] Optionally, the method further includes:

[0021] Based on statistical principles, the rapid enhancement discrimination threshold is determined according to the historical discrimination results of the rapid enhancement trend discrimination model.

[0022] Optionally, the lifecycle identifier of the typhoon under test is calculated, including:

[0023] The lifecycle identifier of the data sequence to be tested is determined based on the intensity change in the previous data sequence corresponding to the data sequence to be tested.

[0024] In a second aspect, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect of the embodiments.

[0025] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect of the embodiments.

[0026] Beneficial effects:

[0027] The acquired remote sensing image sequence and corresponding intensity sequence of the typhoon to be tested are preprocessed to obtain the test data sequence of the typhoon. Then, the life cycle identifier of the typhoon to be tested is determined. The preprocessed test data sequence and life cycle identifier are used as input and identified by a rapid enhancement trend discrimination model to obtain the confidence level of the rapid enhancement of the typhoon to be tested. If the confidence level of the rapid enhancement of the typhoon to be tested is greater than the rapid enhancement discrimination threshold, it can be determined that the typhoon to be tested has a rapid enhancement trend.

[0028] The rapid intensification trend discrimination model is based on a sample training set consisting of remote sensing image sequences of multiple sample typhoons, their corresponding intensity sequences, rapid intensification markers, and life cycle markers, and is trained using a combination of resampling and reweighting. It can predict the rapid intensification trend of the typhoon to be tested. This application uses the model for identification and adds the typhoon's life cycle marker during identification, thereby enabling a faster and more accurate judgment of the rapid intensification trend of the typhoon. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the steps of a method for determining the rapid intensification trend of a typhoon according to an embodiment of this application;

[0031] Figure 2 This is a flowchart of the steps of a training method for a rapid enhanced trend discrimination model proposed in an embodiment of this application;

[0032] Figure 3 This is a schematic diagram of the training process of a classifier according to an embodiment of this application;

[0033] Figure 4 This is a functional block diagram of a typhoon rapid intensification trend discrimination system proposed in an embodiment of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Tropical cyclones, or typhoons, are among the most destructive weather systems on Earth. While forecasts of typhoon tracks and intensity have improved significantly over the past half-century, predicting rapid intensification (RI) of typhoons remains challenging, impacting predictions of subsequent typhoon structure and precipitation.

[0036] Previous research on predicting rapid typhoon intensification has focused on using predictor variables as features, or analyzing the climatic and spatiotemporal characteristics of tropical cyclones, such as season, month, latitude and longitude, air pressure changes, and monthly average sea surface temperature fields. Meteorologists use statistical methods to identify patterns as important criteria for judging whether a tropical cyclone will intensify rapidly. However, research relying on predictor variables lacks exploration of new technologies to better utilize raw data, and the patterns discovered can only serve as auxiliary means, making it difficult to quickly and accurately predict whether a typhoon will intensify rapidly in real time.

[0037] With the development of information technology and the rapid advancement of machine learning and deep learning technologies in the field of artificial intelligence, these technologies have been widely applied in finance, agriculture, and the military, yielding numerous results. Therefore, with the continuous development of my country's meteorological satellites, utilizing artificial intelligence technologies to identify the rapid intensification trend of typhoons is a promising approach. Specifically, for identifying the rapid intensification of typhoons in the meteorological field, time-series prediction technology from the field of computer vision can be used to solve this problem, enabling real-time identification of the rapid intensification trend of typhoons based on big data meteorological data.

[0038] However, based on the statistical results of typhoons over the years, there are relatively few positive samples of typhoons rapidly intensifying. Therefore, when processing these samples, it is necessary to consider the impact of imbalanced sample distribution on the determination of the rapid intensification trend of typhoons. If the imbalanced sample distribution is not considered, or if the issues of cold start and missing data for typhoons are not taken into account, the determination of the rapid intensification trend of typhoons will be inaccurate.

[0039] Based on the above problems, this application proposes a method for judging the rapid intensification trend of typhoons based on remote sensing cloud image sequences, aiming to obtain a more accurate method for predicting or judging the rapid intensification trend of typhoons.

[0040] Reference Figure 1 The flowchart illustrates the steps of a method for determining the rapid intensification trend of typhoons based on remote sensing cloud image sequences, as described in an embodiment of the present invention. Figure 1 As shown, the method may specifically include the following steps:

[0041] S101. Preprocess the acquired remote sensing image sequence of the typhoon to be measured and the corresponding intensity sequence to obtain the data sequence of the typhoon to be measured.

[0042] The remote sensing image sequence of the typhoon to be measured is a remote sensing cloud image of the typhoon's optimal path obtained by meteorological satellites. The remote sensing cloud image is a high-resolution N grayscale image in time sequence. The N remote sensing cloud images can be a continuous time series or N remote sensing cloud images extracted at certain time intervals and arranged in chronological order. The value of N can be determined according to the input requirements of the model, and is not limited in this embodiment. The remote sensing image sequence has a corresponding typhoon intensity sequence, that is, each remote sensing cloud image has its corresponding intensity information.

[0043] After obtaining the remote sensing image sequence and corresponding intensity sequence of the typhoon to be measured, the first step is to preprocess the remote sensing image sequence and intensity sequence. The preprocessing of the remote sensing image sequence includes adjusting the data format of the remote sensing images. For example, in this embodiment, the resolution of each remote sensing image is adjusted to 400*400 and the size is adjusted to the specified model input size of 224*224.

[0044] For intensity sequences, since corresponding intensity information may be missing in N remote sensing cloud images, it is necessary to add the missing intensity information. Therefore, the preprocessing of intensity sequences can include assigning the intensity information of the adjacent previous or next time step to the remote sensing cloud images with missing intensity information. For example, if there are fewer than N intensity data points, or if there are missing N consecutive equally spaced data points, the most recent typhoon data can be used to complete the sequence. Specifically, if the initial time is t, then N-1 initial time data points are copied, i.e., [t, t, ..., t]; if the initial time is t and the current time is t+6, then N-2 time data points at time t are copied, i.e., [t, ..., t, t+6]; and so on.

[0045] The data sequence for the typhoon to be measured is composed of preprocessed remote sensing image sequences and corresponding intensity sequences.

[0046] S102. Calculate the life cycle identifier of the typhoon to be tested.

[0047] The life cycle identifier indicates whether a typhoon is in its ascending or descending phase within a certain period of time. The life cycle identifier for the ascending phase is 1, and the life cycle identifier for the descending phase is 0.

[0048] During the testing phase, the lifecycle of the data sequence of the typhoon under test can be determined based on the intensity changes of the previous sequence. For example, if the wind speed of the data sequence under test continuously increases within 12 or 24 hours of the previous sequence, the lifecycle of the data sequence under test is determined to be the rising phase, and the lifecycle identifier is set to 1. If the wind speed of the data sequence under test decreases within 12 or 24 hours of the previous sequence, the lifecycle of the data sequence under test is determined to be the falling phase, and the lifecycle identifier is set to 0. If the wind speed of the data sequence under test remains unchanged within 12 or 24 hours of the previous sequence, the process continues to trace back to the previous sequence until the lifecycle can be determined. For the initial phase sequence of a typhoon, the lifecycle of the initial data sequence under test is defaulted to the rising phase, i.e., its lifecycle identifier is 1.

[0049] Historical typhoon data shows that typhoons are more likely to intensify rapidly during their growth phase. Therefore, when predicting the rapid intensification trend of a typhoon, considering its life cycle can lead to a more accurate prediction.

[0050] S103. The preprocessed data sequence to be tested and the life cycle identifier of the typhoon to be tested are used as inputs, and the rapid enhancement trend discrimination model is used for identification to obtain the confidence of the rapid enhancement of the typhoon to be tested. The rapid enhancement trend discrimination model is based on a sample training set composed of remote sensing image sequences of multiple sample typhoons, their corresponding intensity sequences, rapid enhancement identifiers and life cycle identifiers, and is trained using a combination of resampling and reweighting.

[0051] The rapid enhancement trend discrimination model is pre-trained on a sample training set and can calculate the confidence level of rapid enhancement of the typhoon under test based on the input remote sensing image sequence, intensity sequence, and life cycle identifier of the data sequence.

[0052] S104. Compare the confidence level of the rapid intensification of the typhoon under test with the rapid intensification discrimination threshold. When the confidence level of the rapid intensification of the typhoon under test is greater than the rapid intensification discrimination threshold, it is determined that the typhoon under test has a rapid intensification trend.

[0053] The rapid enhancement discrimination threshold can be set custom-defined. One way to determine the rapid enhancement discrimination threshold is based on statistical principles, using the historical discrimination results of the rapid enhancement trend discrimination model. In this embodiment, the rapid enhancement discrimination threshold is set to 0.55 based on the historical discrimination results. In other embodiments, the value of the rapid enhancement discrimination threshold can also be determined using other methods.

[0054] This application preprocesses the acquired remote sensing image sequence of the typhoon to be measured and the corresponding intensity sequence to obtain the data sequence of the typhoon to be measured. Then, it determines the life cycle identifier of the typhoon to be measured. The preprocessed data sequence and life cycle identifier are used as input and identified by a rapid enhancement trend discrimination model to obtain the confidence level of the rapid enhancement of the typhoon to be measured. If the confidence level of the rapid enhancement of the typhoon to be measured is greater than the rapid enhancement discrimination threshold, the rapid enhancement trend of the typhoon to be measured in the next 12 hours or 24 hours can be judged.

[0055] The rapid intensification trend discrimination model in this application can more accurately predict the rapid intensification trend of the typhoon under test. The rapid intensification trend discrimination model includes several sequentially connected modules: CNN model, LSTM time series model and classifier.

[0056] In one feasible implementation, this application also provides a method for training the rapidly enhanced trend discrimination model.

[0057] Reference Figure 2 This illustrates a training method for a fast enhanced trend discrimination model provided by an embodiment of the present invention, such as... Figure 2 As shown, the training method may include the following steps:

[0058] S201. Construct the training dataset and perform preprocessing.

[0059] The rapid enhancement trend discrimination model is based on a sample training set consisting of remote sensing image sequences of multiple sample typhoons, their corresponding intensity sequences, rapid enhancement labels, and life cycle labels, and is trained using a combination of resampling and reweighting. In the process of training the CNN model, LSTM time series model, and classifier respectively, the data format can be determined according to the actual implementation.

[0060] First, a training dataset is constructed based on multiple historical typhoon data sets. A sample typhoon sequence in the training dataset has the data format <{Image 1, Image 2, ..., Image N}, {Intensity 1, Intensity 2, ..., Intensity N}, Fast Enhancement Identifier>, which can consist of N consecutive typhoon data points at intervals of 3-6 hours. The Fast Enhancement Identifier is determined based on whether the typhoon experienced rapid intensification within 12 hours or 24 hours after the last moment in the typhoon sequence. The Fast Enhancement Identifier is used to label whether a sample is a positive sample with rapid intensification or a negative sample without rapid intensification.

[0061] Secondly, the training dataset is preprocessed, including processing the remote sensing image sequences of the sample typhoons to obtain image sequences in the target format. In this embodiment, the image sequence is a remote sensing cloud image sequence with a resolution of 400*400, which is cropped from the vortex center of the typhoon and extended by 10° (latitude and longitude) in all directions according to the optimal path information of the sample typhoon. The remote sensing cloud image is automatically scaled when input into the model to meet the input size requirements of the subsequent model. In this example, the image input size is 224*224.

[0062] It also includes handling missing data in the intensity sequences of the sample typhoons to fill in the missing data in the intensity sequences. For example, if there are fewer than N intensity data points, or if there are missing N consecutive equally spaced data points, the most recent typhoon data is used to fill in the missing data. Specifically, if the initial time is t, then N-1 initial time data points are copied, i.e., [t, t, ..., t]; if the initial time is t and the current time is t+6, then N-2 data points at time t are copied, i.e., [t, ..., t, t+6]; and so on.

[0063] Each sample typhoon sequence is also labeled with a corresponding lifecycle identifier. The preprocessing stage includes calculating the lifecycle of the sample typhoon sequence and setting the lifecycle identifier. Specifically, taking a typhoon with a complete cycle as an example, the moment when the typhoon reaches its maximum wind speed is obtained, and this moment is used as the dividing point. If the typhoon maintains maximum wind speed for a period of time, the first moment of that maximum wind speed period is used as the dividing point. If the typhoon maintains maximum wind speed for multiple periods, the first moment of the last maximum wind speed period is used as the dividing point. The lifecycle of the typhoon sequence before the dividing point is considered the rising phase, with a lifecycle identifier of 1. The lifecycle of the typhoon sequence after the dividing point is considered the falling phase, with a lifecycle identifier of 0.

[0064] S202. Based on the remote sensing image sequences of multiple sample typhoons and their corresponding intensity sequences, the CNN model is pre-trained to obtain the first feature extraction module after pre-training. The first feature extraction module after pre-training is used to extract the visual spatial features of a single remote sensing image of a typhoon.

[0065] In this embodiment, a ResNet model is used as a pre-trained CNN model. The input of ResNet is <{image 1, image 2, ..., image N} and {intensity 1, intensity 2, ..., intensity N}>. The model is then fine-tuned and pre-trained until the ResNet model training converges.

[0066] By training the ResNet model, a pre-trained first feature extraction module is obtained. This pre-trained first feature extraction module is used to extract the visual spatial features of a single remote sensing image of a typhoon.

[0067] S203. Based on the remote sensing image sequences and corresponding intensity sequences of multiple sample typhoons, and the processing results of the pre-trained CNN module on the remote sensing image sequences and corresponding intensity sequences of multiple sample typhoons, the fast enhancement trend discrimination model is trained to obtain the trained second feature extraction module. The second feature module includes a CNN model and an LSTM time series model. The trained second feature extraction module is used to extract the spatiotemporal correlation features of typhoon visual morphology changes and intensity changes.

[0068] In this step, the finely tuned CNN model and LSTM time series model are jointly trained. Since the CNN model used in this embodiment is the ResNet model, the ResNet model and LSTM time series model are jointly trained. The training uses the preprocessed sample typhoon sequences in the training dataset.

[0069] First, the remote sensing image sequence of the sample typhoon <image 1, image 2, ..., image N>, i.e. <p1,p2,...,p N The input is fed into the feature extraction network of a pre-trained ResNet model, which outputs N 512-dimensional visual space feature vectors from its first fully connected layer. <w1,w2,...,w N >

[0070] Secondly, the intensity sequence <intensity 1, intensity 2, ..., intensity N> corresponding to the remote sensing image sequence of the sample typhoon is... <tense1,tense2,...,tense N After embedding, it is transformed into N 512-dimensional intensity vectors. <t1,t2,...,t N The intensity vector and the visual space feature vector are then concatted. The `concat()` method is used to join two or more arrays. This method does not modify the existing arrays; it only returns a new array. Figure 2 In this context, the "+" sign represents the concat operation between data, resulting in a new sequence vector <[w1+t1], [w2+t2], ..., [w N +t N ]>.

[0071] Let the sequence vectors <[w1+t1], [w2+t2], ..., [w N +t NAfter passing through a two-layer LSTM network, a rapid enhancement feature vector r of the typhoon is obtained. The obtained rapid enhancement feature vector r and the life cycle identifier corresponding to the sample typhoon sequence are used to train the classifier. By jointly training the ResNet model and the LSTM time series model, the rapid enhancement trend discrimination model can learn the spatiotemporal correlation features of the visual morphological changes and intensity changes of the typhoon, so as to obtain the rapid enhancement features of the typhoon.

[0072] The training process, from inputting remote sensing image sequences into ResNet to outputting RI confidence scores, is a full model training process, during which the parameters of ResNet, LSTM, and the classifier network are updated. Standard cross-entropy loss is used to calculate the loss during training.

[0073] in

[0074] In the formula, p i This represents the probability that sample i is predicted to be a positive (mutant) sample by the output of the fast-amplified trend discrimination model. After softmax, we obtain 1-p. i y represents the probability that sample i is predicted as a negative (non-mutant) sample by the output of the fast-enhanced trend discriminant model; i The mutation label represents sample i (1 for positive samples and 0 for negative samples); C represents the total number of sample categories, which is C=2 in this example. Here, positive (mutated) refers to rapidly enhanced samples and negative (non-mutated) refers to non-rapidly enhanced samples.

[0075] At this point, the model structure of the classifier network has been trained and can be regarded as a feature extractor. When training the classifier, the parameters of the trained CNN model and LSTM time series model need to be frozen.

[0076] S204. Freeze the model parameters of the trained LSTM time series model and CNN model, and repeatedly use the remote sensing image sequences, corresponding intensity sequences, rapid enhancement labels and life cycle labels of multiple sample typhoons. Retrain the classifier of the rapid enhancement trend discrimination model by combining resampling and reweighting. The trained classifier is used to predict the confidence level of rapid enhancement of typhoons.

[0077] Reference Figure 3 The flowchart illustrating the training process of the classifier in this application is shown below. Figure 3As shown, when training a classifier, the parameters of the trained CNN model and LSTM time series model remain unchanged. In order to train the model to achieve better results, it is necessary to use re-sampler and re-weighting methods to retrain the classifier.

[0078] First, for the pre-processed training dataset, samples are randomly selected and resampled according to a specified ratio to obtain positive and negative samples with a preset distribution to train the classifier. Data of typhoons with the first identifier are quickly enhanced and data of typhoons with the second identifier are quickly enhanced and are negative samples. In this embodiment, the weights of positive and negative samples are set to 15:1. The problem of uneven distribution of positive and negative samples in the training dataset is changed by the re-sampler method.

[0079] The resampled typhoon sequences are input into the trained CNN and LSTM time-series models to obtain a fast-enhanced feature vector r. Then, the lifecycle identifier of the typhoon sequence is one-hot encoded to obtain l, and l and the feature vector r are concatenated to obtain...<l,r> The data is fed into a classifier network, consisting of two fully connected layers and a softmax layer, to output the confidence level of a rapid increase in typhoon intensity within the next 12 or 24 hours.

[0080] In the process of calculating the loss of the classifier, it is necessary to assign weights to the standard cross-entropy loss and apply different degrees of penalty to positive and negative samples. According to the weights corresponding to the positive and negative samples, the loss of positive and negative samples is adjusted so that the classifier is trained based on positive and negative samples with a preset distribution. That is, the loss function Loss of positive and negative samples during training is adjusted by re-weighting to rebalance the sample distribution and make the model achieve a better effect. In this example, the weights of positive and negative samples are set to [0.54:0.46].

[0081] in

[0082] In the formula, ω1 represents the Loss weight assigned to samples that are actually positive (mutated); p i This represents the probability that sample i is predicted to be a positive sample by the output of the fast-amplified trend discrimination model. Obtained through softmax; ω0 represents the loss weight assigned to samples that are actually negative (non-mutated); 1-p i y represents the probability that sample i is predicted as a negative sample by the output of the fast-augmented trend discriminant model; iThe mutation label represents sample i (1 for positive samples and 0 for negative samples); C represents the total number of sample categories, which is C=2 in this example.

[0083] When the entire model training is complete, that is, when the classification network training is complete.

[0084] This application has at least the following beneficial effects:

[0085] 1. The trained model is used to identify the rapid intensification trend of the typhoon under test. The life cycle identifier of the typhoon is added during the identification process, which enables the rapid and accurate judgment of the rapid intensification trend of the typhoon.

[0086] 2. Jointly train the ResNet model and the LSTM time series model so that the rapid enhancement trend discrimination model can learn the spatiotemporal correlation features of the visual morphological changes and intensity changes of typhoons, so as to obtain the features of rapid typhoon enhancement.

[0087] 3. When training the classifier, a resampler is used to perform imbalanced sampling of samples. By repeatedly sampling positive samples, the deficiency of positive samples can be appropriately alleviated, so that the model can be trained based on a balanced distribution of positive and negative samples, thereby enhancing the training effect of the classifier.

[0088] 4. When training the classifier, a re-weighting method is used to assign different weights to the loss of positive and negative samples, thereby further alleviating the problem of imbalanced distribution of positive and negative samples.

[0089] Reference Figure 4 This application illustrates a discrimination system for the rapid intensification trend of typhoons in remote sensing cloud image sequences, as provided in an embodiment of this application. Figure 4 As shown, the system includes:

[0090] The preprocessing module 100 is used to preprocess the acquired remote sensing image sequence of the typhoon to be tested and the corresponding intensity sequence to obtain the test data sequence of the typhoon to be tested.

[0091] The first determining module 200 is used to calculate the life cycle identifier of the typhoon to be tested;

[0092] The identification module 300 is used to take the preprocessed data sequence to be tested and the life cycle identifier of the typhoon to be tested as input, and identify it through a fast enhancement trend discrimination model to obtain the confidence of the rapid enhancement of the typhoon to be tested. The fast enhancement trend discrimination model is based on a sample training set composed of remote sensing image sequences of multiple sample typhoons, their corresponding intensity sequences, fast enhancement identifiers and life cycle identifiers, and is trained using a combination of resampling and reweighting.

[0093] The second determining module 400 is used to compare the confidence level of the rapid intensification of the typhoon under test with the rapid intensification discrimination threshold, and when the confidence level of the rapid intensification of the typhoon under test is greater than the rapid intensification discrimination threshold, determine that the typhoon under test has a rapid intensification trend.

[0094] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in the embodiments.

[0095] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described in the embodiments.

[0096] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0097] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0101] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0102] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0103] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for determining the rapid intensification trend of typhoons in remote sensing cloud image sequences, characterized in that, The method includes: The acquired remote sensing image sequence of the typhoon to be measured and the corresponding intensity sequence are preprocessed to obtain the data sequence of the typhoon to be measured. Calculate the lifecycle identifier of the typhoon to be tested; The preprocessed data sequence to be tested and the life cycle identifier of the typhoon to be tested are used as inputs. The rapid enhancement trend discrimination model is used for identification to obtain the confidence of the rapid enhancement of the typhoon to be tested. The rapid enhancement trend discrimination model is based on a sample training set composed of remote sensing image sequences, corresponding intensity sequences, rapid enhancement identifiers and life cycle identifiers of multiple sample typhoons, and is trained by a combination of resampling and reweighting. The confidence level of the rapid intensification of the typhoon under test is compared with the rapid intensification discrimination threshold. When the confidence level of the rapid intensification of the typhoon under test is greater than the rapid intensification discrimination threshold, it is determined that the typhoon under test has a rapid intensification trend.

2. The method according to claim 1, characterized in that, The fast augmented trend discrimination model comprises several sequentially connected modules: a CNN model, an LSTM time series model, and a classifier; the training process of the fast augmented trend discrimination model includes the following steps: Based on the remote sensing image sequences of multiple sample typhoons and their corresponding intensity sequences, the CNN model is pre-trained to obtain the first feature extraction module after pre-training. The first feature extraction module after pre-training is used to extract the visual spatial features of a single remote sensing image of a typhoon. The ResNet model is used as the pre-trained CNN model. The input of ResNet is <{image 1, image 2, ..., image N} and {intensity 1, intensity 2, ..., intensity N}>. The model is then fine-tuned and pre-trained until the ResNet model training converges. By training the ResNet model, a pre-trained first feature extraction module is obtained. The pre-trained first feature extraction module is used to extract the visual spatial features of a single remote sensing image of a typhoon. Based on the remote sensing image sequences and corresponding intensity sequences of multiple sample typhoons, as well as the processing results of the pre-trained CNN module on the remote sensing image sequences and corresponding intensity sequences of multiple sample typhoons, the fast enhancement trend discrimination model is trained to obtain the trained second feature extraction module. The second feature extraction module includes a CNN model and an LSTM time series model. The trained second feature extraction module is used to extract the spatiotemporal correlation features of typhoon visual morphology changes and intensity changes. In this step, the fine-tuned CNN model and LSTM time series model are jointly trained. The CNN model used is the ResNet model. The ResNet model and LSTM time series model are jointly trained using preprocessed sample typhoon sequences from the training dataset. First, the remote sensing image sequence of the sample typhoon <image 1, image 2, ..., image N>, i.e. <p1,p2,…,p N The input is fed into the feature extraction network of a pre-trained ResNet model, which outputs N 512-dimensional visual space feature vectors from its first fully connected layer. <w1,w2,…,w N >; Secondly, the intensity sequence <intensity 1, intensity 2, ..., intensity N> corresponding to the remote sensing image sequence of the sample typhoon is... <tense1,tense2,…,tense N After embedding, it is transformed into N 512-dimensional intensity vectors. <t1,t2,…,t N The intensity vector and the visual space feature vector are concatenated. The concat() method is used to join two or more arrays to obtain a new sequence vector <[w1+t1], [w2+t2], ..., [w N +t N ]>; Let the sequence vectors <[w1+t1], [w2+t2], ..., [w N +t N After passing through a two-layer LSTM network, the typhoon rapid enhancement feature vector r is obtained. The obtained rapid enhancement feature vector r and the life cycle identifier corresponding to the sample typhoon sequence are used to train the classifier. By jointly training the ResNet model and the LSTM time series model, the rapid enhancement trend discrimination model can learn the spatiotemporal correlation features of the visual morphological changes and intensity changes of the typhoon, so as to obtain the rapid enhancement features of the typhoon. The model parameters of the pre-trained LSTM time series model and CNN model are frozen. The remote sensing image sequences, corresponding intensity sequences, rapid enhancement labels and life cycle labels of each sample typhoon are repeatedly used. The classifier of the rapid enhancement trend discrimination model is retrained by combining resampling and reweighting. The trained classifier is used to predict the confidence level of rapid enhancement of typhoons.

3. The method according to claim 2, characterized in that, Before pre-training the CNN model, the following steps are also included: The remote sensing image sequences of the sample typhoons are processed to obtain image sequences in the target format; The intensity sequences of the sample typhoons were processed to fill in the missing data in the intensity sequences; Based on the complete life cycle of the sample typhoons, the maximum wind speed node of the sample typhoons is determined. Using time as the sequence, the life cycle of the typhoon sequence before the maximum wind speed node is determined as the rising phase, and the life cycle of the typhoon sequence after the maximum wind speed node is determined as the falling phase.

4. The method according to claim 3, characterized in that, When training the classifier, the sample training set is resampled according to a specified ratio to obtain positive and negative samples with a preset distribution to train the classifier. Data of typhoons with the first identifier are quickly enhanced and data of typhoons with the second identifier are quickly enhanced and are negative samples.

5. The method according to claim 4, characterized in that, When training the classifier, the loss of positive and negative samples is adjusted according to the weights corresponding to the positive and negative samples, so that the classifier is trained based on positive and negative samples from a preset distribution.

6. The method according to claim 1, characterized in that, The method further includes: Based on statistical principles, the rapid enhancement discrimination threshold is determined according to the historical discrimination results of the rapid enhancement trend discrimination model.

7. The method according to claim 1, characterized in that, Calculating the lifecycle identifier of the typhoon under test includes: The lifecycle identifier of the data sequence to be tested is determined based on the intensity change in the previous data sequence corresponding to the data sequence to be tested.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the method of any one of claims 1 to 7.

Images