CN114548606B - Construction method of cyclone strength prediction model and cyclone strength prediction method - Google Patents
Construction method of cyclone strength prediction model and cyclone strength prediction method Download PDFInfo
- Publication number
- CN114548606B CN114548606B CN202210438320.XA CN202210438320A CN114548606B CN 114548606 B CN114548606 B CN 114548606B CN 202210438320 A CN202210438320 A CN 202210438320A CN 114548606 B CN114548606 B CN 114548606B
- Authority
- CN
- China
- Prior art keywords
- latitude
- longitude
- convolution
- cyclone
- prediction model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention discloses a construction method of a cyclone strength prediction model and a cyclone strength prediction method, wherein the construction method comprises the following steps: obtaining training data, wherein each training data comprises a sequence of satellite images; sequentially inputting each satellite image in the satellite image sequence to an area curved surface convolution layer, and outputting a first spatial characteristic map by the area curved surface convolution layer, wherein in the area curved surface convolution layer, traversal convolution is carried out on the satellite image, in each convolution, the longitude and latitude of other sampling points of the current convolution are determined based on the longitude and latitude of a central sampling point of the current convolution, and the longitude and latitude of the central sampling point of the current convolution and the longitude and latitude of the other sampling points are used as sampling ranges of the current convolution; obtaining a prediction result of cyclone strength based on the first spatial feature map; and calculating a loss function value according to the cyclone strength prediction result and the real value, and training based on the loss function value to obtain a cyclone strength prediction model. The method can construct a model with higher accuracy of the vortex strength prediction.
Description
Technical Field
The invention relates to the technical field of deep learning and image processing, in particular to a construction method of a cyclone strength prediction model and a cyclone strength prediction method.
Background
Tropical cyclones frequently affect coastal regions and seriously threaten the life and property safety of coastal residents, so that the tropical cyclones can be accurately predicted with high value. With the development of geosynchronous satellite technology, meteorological observations are made over the entire earth region.
At present, satellite image data is used for tropical cyclone strength analysis, but the existing tropical cyclone strength analysis method based on the satellite image data has the problem of low accuracy.
Disclosure of Invention
The invention solves the problem that the existing tropical cyclone strength analysis method based on satellite image data is not high in accuracy.
The invention provides a construction method of a cyclone strength prediction model, which comprises the following steps: the method for constructing the cyclone strength prediction model comprises the following steps of:
obtaining training data from a preset training data set, wherein each of the training data comprises a sequence of satellite images;
sequentially inputting each satellite image in the satellite image sequence to the spatial feature extraction network, and outputting a first spatial feature map by the spatial feature extraction network, wherein in the area curved surface convolution layer of the spatial feature extraction network, the satellite image is subjected to traversal convolution, in each convolution, the longitude and latitude of other sampling points of the current convolution are determined based on the longitude and latitude of the central sampling point of the current convolution, and the longitude and latitude of the central sampling point of the current convolution and the longitude and latitude of the other sampling points are used as the sampling range of the current convolution;
obtaining a cyclone intensity prediction result based on the first spatial feature map;
and calculating a loss function value according to the cyclone strength prediction result and the real value, and training based on the loss function value to obtain the cyclone strength prediction model.
Optionally, the spatial feature extraction block further includes a longitude and latitude attention layer, and the longitude and latitude attention layer is disposed behind the area curved surface convolution layer; the obtaining a prediction of cyclone intensity based on the first spatial signature comprises:
inputting the first spatial feature map into the longitude and latitude attention layer, and calculating mean values of the first spatial feature map in the longitude direction and the latitude direction respectively in the longitude and latitude attention layer to obtain two feature blocks with the sizes of C1W and C H1 respectively, wherein C is channel dimensionality, H is longitude dimensionality, and W is latitude dimensionality;
exchanging the last two dimensions of any one of the two feature blocks, and fusing the last two dimensions with the other remaining feature block to obtain a fused feature block;
after the fused feature blocks are subjected to down sampling and up sampling in sequence, two feature blocks with the sizes of C x 1 x W and C x H x 1 are regenerated, and after the two regenerated feature blocks are subjected to convolution and expansion operations, two feature blocks with the sizes of C x H x W are generated;
fusing two feature blocks with the size of C H W with the first spatial feature map to obtain a second spatial feature map;
a cyclone intensity prediction result is obtained based on the second spatial signature.
Optionally, the spatial feature extraction block includes two of the area curved surface convolution layers and one of the longitude and latitude attention layers, and the longitude and latitude attention layer is disposed between the two of the area curved surface convolution layers.
Optionally, the spatial feature extraction block further includes a normalization-activation function layer and a max-pooling layer, the normalization-activation function layer is disposed behind the region curved surface convolution layer, and the max-pooling layer is disposed at the end of the spatial feature extraction block.
Optionally, the cyclone intensity prediction model further comprises a spatiotemporal feature extraction network, the spatiotemporal feature extraction network is arranged after the spatial feature extraction network, the spatiotemporal feature extraction network comprises an extended causal convolution network, and the extended causal convolution network comprises a hole convolution and a causal convolution.
Optionally, the cyclone strength prediction model further comprises a generalized linear model, and the generalized linear model is arranged after the spatiotemporal feature extraction network and outputs the cyclone strength prediction result.
Optionally, before obtaining the training data from the preset training data set, the method further includes:
acquiring an original satellite image set, and calculating the mean, variance and square mean of satellite images one by one, wherein the satellite image set comprises a training data set, a verification data set and a test data set;
generating an overall mean and an overall variance of the training data set;
and subtracting the integral mean value from all the satellite images in the satellite image set, and dividing the integral mean value by the integral variance to obtain a standardized training data set, a standardized verification data set and a standardized test data set, wherein the preset training data set is the standardized training data set.
Optionally, the determining the longitude and latitude of other sampling points of the current convolution based on the longitude and latitude of the center sampling point of the current convolution comprises: determining the longitude and latitude of other to-be-calculated sampling points of the current convolution by adopting the following formula:
wherein the content of the first and second substances,
the latitude is referred to as the number of years,
which refers to the longitude of the user's hand,
refers to the coordinates of the sampling points to be calculated on the tangent plane block,
，
respectively to the centre sample point latitude and longitude,
refers to the distance from the central sampling point of the sampling point to be calculated on the tangent plane block,
finger-shaped
The value of the arc tangent of (a) is,
。
the invention provides a cyclone strength prediction method, which comprises the following steps:
acquiring a satellite image sequence as a prediction basis;
and inputting the satellite image sequence into the cyclone strength prediction model constructed by the construction method of the cyclone strength prediction model, and outputting a predicted cyclone strength value.
The invention also proposes a computer device comprising a computer readable storage medium and a processor storing a computer program, which, when read and executed by the processor, implements the method for constructing a cyclone strength prediction model as described above or the method for predicting cyclone strength as described above.
According to the method, the regional curved surface convolution is introduced into the cyclone intensity prediction model, the satellite image is processed according to the longitude and latitude information, the distortion of the satellite image by the earth curved surface is overcome, and the accuracy of spatial characteristic representation is improved.
Drawings
FIG. 1 is a schematic flow chart illustrating a method for constructing a cyclone strength prediction model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a cyclone strength prediction model according to an embodiment of the present invention;
FIG. 3 is a schematic diagram illustrating a data processing flow of a latitudinal attention layer in a cyclonic intensity prediction model according to an embodiment of the present invention;
FIG. 4 is a schematic structural diagram of an extended causal convolution network in the cyclone strength prediction model according to an embodiment of the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below.
In an embodiment of the present invention, the cyclone strength prediction model includes: the spatial feature extraction network comprises at least one spatial feature extraction block, and the spatial feature extraction block comprises a region curved surface convolution layer. As shown in fig. 1, the construction method of the cyclone strength prediction model includes:
step S100, training data is obtained from a preset training data set, wherein each of the training data includes a satellite image sequence.
Acquiring geosynchronous orbit meteorological satellite data and tropical cyclone path data to generate a training data set. The satellite images in the geosynchronous orbit meteorological satellite data are cut, the cut satellite images take the circulation center of the tropical cyclone as the satellite image center, the satellite images taking the circulation center of the tropical cyclone as the center are taken as training data, therefore, the peripheral area of the position of the tropical cyclone is brought into the range of the satellite images for training as much as possible, effective features are extracted from the satellite images as much as possible by using a gas rotation strength prediction model, and a good model training effect is ensured.
The input of the cyclone strength prediction model comprises satellite images at a plurality of time instants, namely a satellite image sequence with a certain time sequence relation. For example, the satellite images at 8 time instants are selected and combined into one data sequence. The input of the cyclone intensity prediction model further includes longitude and latitude information, for example, including longitude and latitude information of each pixel point of each satellite image, which is used for calculating the convolution of the regional curved surface, or longitude and latitude information of a pixel point at a specific position of each satellite image, such as longitude and latitude information of a pixel point in the center of the satellite image, longitude and latitude information of a pixel point (upper left corner, upper right corner, lower left corner, lower right corner) of the satellite image, and the like. The output of the cyclone strength prediction model is a tropical cyclone strength value at a preset prediction time, and the preset prediction time is a future time corresponding to the model input data, for example, if the input of the cyclone strength prediction model is a satellite image from time T-7 to time T, the output of the model may be the tropical cyclone strength value at time T + 8. And (3) the satellite image sequence, the longitude and latitude information and the tropical cyclone strength value at the preset prediction moment are arranged into an input-output data pair to serve as training data.
Optionally, the cyclone intensity prediction model inputs included longitude and latitude information, which may be specifically acquired from path data, the center longitude and latitude information of the tropical cyclone center at the corresponding moment of each satellite image, and fixes the center of each satellite image at the center of the tropical cyclone at the corresponding moment through the previous data preprocessing, and the tropical cyclone center longitude and latitude information is longitude and latitude information of a center pixel point of the satellite image, and according to the tropical cyclone center longitude and latitude information, the longitude and latitude of each pixel point of the satellite image may be obtained for calculating the regional curved surface convolution. Therefore, each satellite image only needs to input longitude and latitude information of the center, the longitude and latitude of each pixel point of the satellite image does not need to be input into the cyclone intensity prediction model, the data volume input into the cyclone intensity prediction model can be reduced, meanwhile, the cyclone intensity prediction model can generate the longitude and latitude of each pixel point based on the longitude and latitude information of the center, calculation of area curved surface convolution is guaranteed to be smoothly achieved, and accurate extraction of spatial features is achieved.
Step S200, inputting each satellite image in the satellite image sequence to the spatial feature extraction network in sequence, outputting a first spatial feature map by the spatial feature extraction network, wherein traversing convolution is carried out on the satellite images in the area curved surface convolution layer of the spatial feature extraction network, in each convolution, the longitude and latitude of other sampling points of the current convolution are determined based on the longitude and latitude of the central sampling point of the current convolution, and the longitude and latitude of the central sampling point of the current convolution and the longitude and latitude of the other sampling points are used as the sampling range of the current convolution.
Fig. 2 shows a schematic diagram of satellite images sequentially input into the spatial feature extraction network. After the satellite image is input into the spatial feature extraction network, the convolution operation of the area curved surface is carried out in the area curved surface convolution layer. The convolution operation of the area curved surface specifically comprises the following steps: and traversing and convolving the satellite image, determining the longitude and latitude of other sampling points of the current convolution based on the longitude and latitude of the central sampling point of the current convolution in each convolution, and taking the longitude and latitude of the central sampling point of the current convolution and the longitude and latitude of other sampling points as the sampling range of the current convolution.
The tropical cyclone data adopts equidistant cylindrical projection, the projection mode is that an imaginary spherical surface and a cylindrical surface are tangent to the equator, the equator is a line without deformation, and the longitude and the latitude lines are projected into two groups of parallel straight lines which are vertical to each other. The map projected by equidistant cylinders has no length deformation along the warp direction, but the deformation lines such as angles and areas are parallel to the weft, and the deformation value is gradually increased from the equator to the high latitude, so that the problem of space shape distortion is caused. In the traditional convolutional neural network, the information is considered to have spatial locality, the parameter space can be reduced by using a shared weight mode of a sliding convolution kernel, the problem of spatial shape distortion is not completely applicable to the assumption, and if the regular convolution kernel sliding convolution with a fixed size is adopted according to the traditional mode, the sampling interval in a high-latitude area is insufficient, so that the representation accuracy of the extracted spatial features is low. Based on the method, the regional curved surface convolution layer is introduced into the cyclone intensity prediction model, the traditional convolution operation is expanded from a regular plane domain to a spherical domain, the regional curved surface convolution kernel is expanded to a block tangent to the spherical surface of the earth, each sampling point is projected to the spherical surface to obtain the longitude and latitude coordinates corresponding to each sampling point, and then the sampling range of each convolution operation on the satellite image is determined, so that the influence of the longitude and latitude on the sampling interval is eliminated, and the accuracy of spatial feature representation is improved.
The longitude and latitude of all pixel points of the satellite image are obtained, all the pixel points of the satellite image are traversed through a convolution kernel, for each convolution, the pixel point corresponding to the central sampling point of the convolution kernel is known, and the longitude and latitude of the pixel point corresponding to the central sampling point of the convolution kernel are also known, so that the longitude and latitude of other sampling points on the convolution kernel can be determined based on the position relation between the other sampling points on the convolution kernel and the central sampling point, and then the pixel points corresponding to the longitude and latitude on the satellite image are obtained, and the sampling range of the current convolution on the satellite image is determined. The central sampling point of the current convolution refers to the central point of a convolution kernel, and the longitude and latitude of the central sampling point of the current convolution refers to the longitude and latitude of a pixel point corresponding to the central point of the convolution kernel in the current convolution. In one embodiment, the cyclone intensity prediction model inputs the included longitude and latitude information as the longitude and latitude information of the center of the satellite image, and based on the longitude and latitude information, the longitude and latitude of all the pixel points of the satellite image can be calculated.
Further, a conversion formula of the area curved surface convolution kernel in the spherical coordinate and the tangent plane space coordinate is as follows, and the longitude and latitude of other sampling points to be calculated in the current convolution are determined by adopting the formula:
wherein the content of the first and second substances,
the latitude is referred to as the latitude,
the number of the points is the longitude,
refers to the coordinates of the sample points to be calculated on the tangent plane block,
，
respectively the central sample point latitude and longitude,
refers to the distance of the sample point to be calculated from the center sample point on the block of the tangent plane, i.e.
，
Finger-shaped
The value of the arc tangent of (a) is,
。
optionally, a 3 × 3 convolution kernel is used in the region surface convolution layer.
And step S300, obtaining a prediction result of cyclone strength based on the first space characteristic map.
The result of predicting the cyclone strength may be the maximum average wind speed at the center of the cyclone, or may be the wind circle radius, where the wind circle radius of the tropical cyclone refers to the radius of a wind field area where the average wind speed is greater than a certain level, and the prediction result may be in the form of seven-level wind circle radius, ten-level wind circle radius, twelve-level wind circle radius, or the like.
And S400, calculating a loss function value according to the cyclone strength prediction result and a real value, and training based on the loss function value to obtain the cyclone strength prediction model.
The cyclone strength prediction model is trained through multiple iterations, and the trained cyclone strength prediction model is stored for subsequent use.
Optionally, new tropical cyclone data are sent to the cyclone strength prediction model periodically for incremental training, so that online learning of the cyclone strength prediction model is realized, and the performance of the cyclone strength prediction model is enhanced.
In the embodiment, the regional curved surface convolution is introduced into the cyclone intensity prediction model, the satellite image is processed according to the longitude and latitude information, the spatial distortion of data is corrected, the distortion of the satellite image by the earth curved surface is overcome, the information of the surrounding region is accurately sensed, and the accuracy of spatial feature representation is improved.
Optionally, as shown in fig. 2, the spatial feature extraction block further includes a longitude and latitude attention layer, and the longitude and latitude attention layer is disposed behind the area curved surface convolution layer. As shown in fig. 3, the obtaining the prediction result of cyclone intensity based on the first spatial feature map comprises:
and inputting the first spatial feature map into the longitude and latitude attention layer, and calculating an average value of the first spatial feature map in the longitude direction and the latitude direction by using Avg Pool in the longitude and latitude attention layer to obtain long-distance information in a single longitude direction or a single latitude direction, namely obtaining two feature blocks with the sizes of C1W and C H1 respectively, wherein C is a channel dimension, H is a longitude dimension, and W is a latitude dimension. And exchanging the last two dimensions of any one of the two feature blocks, and fusing the last two dimensions with the other remaining feature block to obtain a fused feature block, so that the longitude direction information and the latitude direction information are combined according to the original middle layer channel, and the long-distance dependence information of the single feature map in the longitude direction and the latitude direction is fused, in the example given in the figure 3, the last two dimensions of the C H1 feature block are exchanged to change the size of the C H1 feature block into C1H, and then the C1W feature block is fused to obtain the fused feature block with the size of C1 (H + W). And after sequentially performing down-sampling and up-sampling on the fused feature block, regenerating two feature blocks with the sizes of C & ltx 1 & gtW and C & ltx H & gt 1 respectively, and performing convolution and expansion operations on the two regenerated feature blocks respectively to generate two feature blocks with the sizes of C & ltx H & gtW, wherein the two regenerated feature blocks are subjected to 1 & ltx 1 & gtconvolution and expansion operations respectively to generate two feature blocks with the sizes of C & ltx & gtH & gtW, the down-sampling and the up-sampling realize the interaction of different feature layer information, and the final attention force value matched with the size of the input feature map is obtained through the convolution and expansion operations. And fusing two feature blocks with the size of C H W with the first spatial feature map to obtain a second spatial feature map, and specifically obtaining the second spatial feature map through Hadamard product fusion so as to realize the self-adaptive capture of the important region of the feature map. Obtaining a cyclone intensity prediction result based on the second spatial feature map.
The target value predicted by the cyclone strength prediction model is an index of the center position of the tropical cyclone, such as the maximum average wind speed or the wind circle radius of the center position, so that the information of the tropical cyclone area, particularly the information of the center area, in the satellite image is more important than other areas of the satellite image.
Optionally, the spatial feature extraction block includes two of the area curved surface convolution layers and one of the longitude and latitude attention layers, and the longitude and latitude attention layer is disposed between the two of the area curved surface convolution layers.
The contents of the area curved surface convolution layer and the latitude and longitude attention layer are as described above, and are not described herein again. As an example shown in fig. 2, a first spatial feature map output by the first area curved surface convolution layer is input to the longitude and latitude attention layer, and after feature extraction is performed on the first spatial feature map by the longitude and latitude attention layer, a second spatial feature map is output to the second area curved surface convolution layer. The first area curved surface convolution layer corrects the spatial distortion of data by utilizing the curved surface convolution so as to accurately sense the information of the surrounding area and generate a preliminary characteristic map, the longitude and latitude attention layer captures long-distance dependence information of the preliminary characteristic map in the longitude and latitude directions so that the network can adaptively pay attention to the important area of the preliminary characteristic map, and the second area curved surface convolution layer is used for increasing the perception area of each pixel point of the characteristic map so as to improve the accuracy of spatial characteristic representation. Optionally, the number of output channels of the first area surface convolution layer is twice the number of input channels to capture more feature information from the previous layer, and the number of output channels of the second area surface convolution layer is consistent with the number of output channels to increase the sensing area of each pixel point of the feature map.
Optionally, the spatial feature extraction block further includes a normalization-activation function layer and a max-pooling layer, the normalization-activation function layer is disposed behind the region curved surface convolution layer, and the max-pooling layer is disposed at the end of the spatial feature extraction block.
The normalization-activation function layer is used for increasing the nonlinearity of the network and simultaneously can accelerate the training speed of the network. Because the size of the input satellite image is larger, the maximum pooling layer is used for removing redundancy of the characteristic diagram and increasing network nonlinearity, and meanwhile, the network parameter quantity is reduced and the training is accelerated.
In one embodiment, the spatial feature extraction block is internally provided with: the first area curved surface convolution layer, the normalization-activation function layer, the longitude and latitude attention layer, the second area curved surface convolution layer, the normalization-activation function layer and the maximum pooling layer. The number of output channels of the first area curved surface convolution layer is twice of the number of input channels, more feature information from the previous layer can be captured, and meanwhile, the spatial distortion of the data is corrected by utilizing the curved surface convolution, so that the information of the surrounding area is accurately sensed, and a preliminary feature map is generated; the normalization-activation function layer is used for increasing the nonlinearity of the network and simultaneously accelerating the training speed of the network; the longitude and latitude attention layer is used for capturing long-distance dependence information of the preliminary characteristic diagram in the longitude and latitude directions, so that the network can pay attention to the important area of the characteristic diagram in a self-adaptive manner; the number of output channels of the second area curved surface convolution layer is consistent with that of the output channels, and the second area curved surface convolution layer is used for increasing the sensing area of each pixel point of the characteristic diagram; the max-pooling layer is used to de-redundantly feature and increase the non-linearity of the network, while reducing the number of network parameters to speed up training.
In the example as given in fig. 2, the spatial feature extraction network is formed by stacking 5 spatial feature extraction blocks, where each spatial feature extraction block has the following internal parts: the system comprises a first area curved surface convolution layer, a first normalization-activation function layer, a longitude and latitude attention layer, a second area curved surface convolution layer, a second normalization-activation function layer and a maximum pooling layer. The number of the intermediate feature layers of each convolution layer in each spatial feature extraction block is consistent, because the final maximum pooling layer of each spatial feature extraction block can reduce the width and height of a feature map, in order to extract and retain more information, the number of output feature map channels of each spatial feature extraction block is sequentially increased by setting the number of convolution kernels of each intermediate feature layer, in each spatial feature extraction block, the number of output channels of a first region curved surface convolution layer is set to be twice the number of input channels, and the number of output channels of a second region curved surface convolution layer is set to be consistent with the number of output channels, in one embodiment, the number of intermediate feature layer channels of 5 spatial feature extraction blocks is respectively 64, 128, 256, 512 and 1024.
Optionally, the cyclone intensity prediction model further comprises a spatiotemporal feature extraction network, the spatiotemporal feature extraction network is arranged after the spatial feature extraction network, the spatiotemporal feature extraction network comprises an extended causal convolution network, and the extended causal convolution network comprises a hole convolution and a causal convolution.
FIG. 4 is a schematic diagram of an embodiment of an extended causal convolutional network. The extended causal convolutional network comprises an input layer, an implied layer 1, an implied layer 2 and an output layer, wherein the input of the implied layer 1 at the time T is the output of the previous layer (namely the input layer) at the time T-1 and the time T, the input of the implied layer 2 at the time T is the output of the previous layer (namely the implied layer 1) at the time T-2 and the time T, and the input of the output layer at the time T is the output of the previous layer (namely the implied layer 2) at the time T-4 and the time T.
The spatial feature extraction network respectively extracts the spatial features of the tropical cyclone satellite images at a single moment, the embodiment captures time sequence information in space-time sequence meteorological data by introducing an extended causal convolution network, and guarantees that future information cannot be used for prediction of the current time step through causal convolution, because the output of the time step T can be obtained only according to T-1 and convolution operation on the previous time step, the sensing fields of all hidden layers can be enlarged through cavity convolution under the condition that a pooling layer is not increased, and each convolution output contains information in a large range.
Optionally, the cyclone strength prediction model further comprises a generalized linear model, which is disposed after the spatio-temporal feature extraction network, and is used for calculating a feature representation and outputting the cyclone strength prediction result. The generalized linear model can be selected as a maximum entropy model, Logistic regression, Softmax regression or the like.
Optionally, before obtaining the training data from the preset training data set, the method further includes:
acquiring an original satellite image set, and calculating the mean value, the variance and the mean of the squares of the satellite images one by one, wherein the original satellite image set is divided into a training data set, a verification data set and a test data set in model training, namely the satellite image set comprises the training data set, the verification data set and the test data set; generating an overall mean and an overall variance of the training data set; and subtracting the integral mean value from all the satellite images in the satellite image set, and dividing the integral mean value by the integral variance to obtain a standardized training data set, a standardized verification data set and a standardized test data set, wherein the preset training data set is the standardized training data set. When the mean value, the variance and the square mean of the satellite image are calculated, the mean value, the variance and the square mean of the satellite image after missing values are ignored are calculated, and the number of effective values is recorded as a weight value so as to accurately calculate the mean value and the variance of the satellite image.
Wherein, the number of the training set images is set as N. For the ith image
，
Represents the mean value of the valid data of the image,
represents the squared average of the image significant data,
representing the number of effective data of the image, and generating an integral mean value of a training data set by adopting the following formula
And overall variance
：
Through the standardization, the expected mean value of the satellite images in the obtained training data set, the obtained verification data set and the obtained test data set is 0, and the expected variance is 1, so that the pixel value of each pixel point in the satellite images is reduced, and the calculated amount is reduced.
The method comprises the steps of acquiring geosynchronous orbit meteorological satellite data and tropical cyclone path data, intercepting satellite images and tropical cyclone center longitude and latitude information at corresponding moments according to time sequence segments through a standardized mode with missing value processing, and generating a data set for model training and verification.
The invention provides a cyclone strength prediction method, which comprises the following steps:
acquiring a satellite image sequence as a prediction basis; and inputting the satellite image sequence into the cyclone strength prediction model constructed by the construction method of the cyclone strength prediction model, and outputting a predicted cyclone strength value.
The satellite image sequence used as a prediction basis is standardized according to the distribution condition of historical tropical cyclone data, and specifically comprises the steps of subtracting the integral mean value of a pre-stored training data set from the satellite image sequence, dividing the integral mean value by the integral variance to obtain a standardized satellite image sequence, inputting the standardized satellite image sequence into a trained cyclone intensity prediction model, and outputting a cyclone intensity prediction value, wherein the cyclone intensity prediction value can refer to the central maximum average wind speed of the tropical cyclone.
According to the embodiment of the invention, a deep neural network based on regional surface convolution and longitude and latitude attention technology is used for generating a spatial feature representation, then an extended causal convolution is used for generating a space-time feature representation, the feature representation is calculated through a generalized linear model, and a predicted value of tropical cyclone intensity is output. Therefore, aiming at the problems that the existing tropical cyclone strength prediction scheme is high in computing resource consumption and low in efficiency, the characteristic extraction process depends on artificial statistical experience and the like, a deep learning method is adopted, and tropical cyclone space-time characteristics are automatically extracted from satellite data and used for tropical cyclone strength prediction. Aiming at the problem that the distortion of an earth curved surface to a satellite image is not considered in the space of the existing deep learning scheme, the satellite image is effectively processed according to longitude and latitude information by using regional curved surface convolution. Aiming at the fact that the importance degrees of different image areas are different in the existing scheme, the method and the device for predicting the strength of the image in the image area analyze the importance levels of the different longitude and latitude areas on the strength prediction by using a longitude and latitude attention technology. Aiming at the problem that parallelization is insufficient and time-consuming in time sequence processing of the existing scheme adopting an LSTM model, the causal relationship of spatial characteristic representation of satellite data at different moments is effectively analyzed by adopting extended causal convolution, and space-time characteristic representation is generated.
In an embodiment of the invention, the computer device comprises a computer readable storage medium storing a computer program and a processor, the computer program being read and executed by the processor for implementing the cyclone strength prediction method as described above. The beneficial effects of the computer device of the present invention over the prior art are consistent with the cyclone strength prediction method described above, and are not described herein again.
The reader should understand that in the description of this specification, reference to the description of the terms "one embodiment," "some embodiments," "an example," "a specific example" or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.
Claims (8)
1. A construction method of a cyclone strength prediction model is characterized in that the cyclone strength prediction model comprises the following steps: the system comprises a space characteristic extraction network, a cyclone strength prediction model and a dynamic characteristic prediction model, wherein the space characteristic extraction network comprises at least one space characteristic extraction block, the space characteristic extraction block comprises two area curved surface convolution layers and a longitude and latitude attention layer, the longitude and latitude attention layer is arranged between the two area curved surface convolution layers, the cyclone strength prediction model further comprises a space-time characteristic extraction network, the space-time characteristic extraction network is arranged behind the space characteristic extraction network, and the space-time characteristic extraction network comprises an extended causal convolution network; the cyclone intensity prediction model also comprises a generalized linear model, the generalized linear model is arranged after the space-time feature extraction network and outputs a cyclone intensity prediction result, and the construction method of the cyclone intensity prediction model comprises the following steps:
obtaining training data from a preset training data set, wherein each of the training data comprises a sequence of satellite images;
sequentially inputting each satellite image in the satellite image sequence to the spatial feature extraction network, and outputting a first spatial feature map by the spatial feature extraction network, wherein in the area curved surface convolution layer of the spatial feature extraction network, the satellite images are subjected to traversal convolution, in each convolution, the longitude and latitude of other sampling points of the current convolution are determined based on the longitude and latitude of the central sampling point of the current convolution, and the longitude and latitude of the central sampling point of the current convolution and the longitude and latitude of the other sampling points are used as the sampling range of the current convolution;
obtaining a cyclone intensity prediction result based on the first spatial feature map;
and calculating a loss function value according to the cyclone strength prediction result and the real value, and training based on the loss function value to obtain the cyclone strength prediction model.
2. The method of constructing a model for predicting cyclone strength of claim 1, wherein the spatial feature extraction block further comprises a latitude and longitude attention layer disposed behind the region curved surface convolution layer; the obtaining a prediction of cyclone intensity based on the first spatial signature comprises:
inputting the first spatial feature map into the longitude and latitude attention layer, and calculating mean values of the first spatial feature map in the longitude direction and the latitude direction respectively in the longitude and latitude attention layer to obtain two feature blocks with the sizes of C1W and C H1 respectively, wherein C is channel dimensionality, H is longitude dimensionality, and W is latitude dimensionality;
exchanging the last two dimensions of any one of the two feature blocks, and fusing the last two dimensions with the other remaining feature block to obtain a fused feature block;
after the fused feature blocks are subjected to down sampling and up sampling in sequence, two feature blocks with the sizes of C x 1 x W and C x H x 1 are regenerated, and after the two regenerated feature blocks are subjected to convolution and expansion operations, two feature blocks with the sizes of C x H x W are generated;
fusing two feature blocks with the size of C H W with the first spatial feature map to obtain a second spatial feature map;
a cyclone intensity prediction result is obtained based on the second spatial signature.
3. The method of constructing a cyclonic strength prediction model as claimed in claim 2, wherein the spatial feature extraction block further comprises a normalized-activation function layer and a max-pooling layer, the normalized-activation function layer being disposed after the region surface convolution layer, the max-pooling layer being disposed at the end of the spatial feature extraction block.
4. The method of constructing a cyclonic strength prediction model of claim 1, wherein the extended causal convolution network comprises a hole convolution and a causal convolution.
5. The method of constructing a cyclone strength prediction model as claimed in claim 1, wherein before the obtaining training data from the preset training data set, further comprising:
acquiring an original satellite image set, and calculating the mean, variance and square mean of satellite images one by one, wherein the satellite image set comprises a training data set, a verification data set and a test data set;
generating an overall mean and an overall variance of the training data set;
and subtracting the integral mean value from all the satellite images in the satellite image set, and dividing the integral mean value by the integral variance to obtain a standardized training data set, a standardized verification data set and a standardized test data set, wherein the preset training data set is the standardized training data set.
6. The method of constructing a cyclone intensity prediction model of claim 1, wherein the determining the latitude and longitude of the other sample points of the current convolution based on the latitude and longitude of the center sample point of the current convolution comprises: determining the longitude and latitude of other sampling points to be calculated in the current convolution by adopting the following formula:
wherein the content of the first and second substances,
the latitude is referred to as the latitude,
the number of the points is the longitude,
refers to the coordinates of the sampling points to be calculated on the tangent plane block,
，
respectively the central sample point latitude and longitude,
refers to the distance from the central sampling point of the sampling point to be calculated on the tangent plane block,
finger-shaped
The value of the arc tangent of (a) is,
。
7. a method of predicting cyclone intensity, comprising:
acquiring a satellite image sequence as a prediction basis;
inputting the satellite image sequence into a cyclone strength prediction model constructed by the construction method of the cyclone strength prediction model according to any one of claims 1 to 6, and outputting a predicted cyclone strength value.
8. Computer equipment, characterized in that it comprises a computer readable storage medium storing a computer program and a processor, the computer program being read and executed by the processor to implement the method of constructing a cyclone strength prediction model according to any one of claims 1-6 or the method of cyclone strength prediction according to claim 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210438320.XA CN114548606B (en) | 2022-04-25 | 2022-04-25 | Construction method of cyclone strength prediction model and cyclone strength prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210438320.XA CN114548606B (en) | 2022-04-25 | 2022-04-25 | Construction method of cyclone strength prediction model and cyclone strength prediction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114548606A CN114548606A (en) | 2022-05-27 |
| CN114548606B true CN114548606B (en) | 2022-08-23 |
Family
ID=81667583
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210438320.XA Active CN114548606B (en) | 2022-04-25 | 2022-04-25 | Construction method of cyclone strength prediction model and cyclone strength prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114548606B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115082439B (en) * | 2022-07-22 | 2022-11-29 | 浙江大学 | Tropical cyclone strength determining method, medium and equipment fused with satellite cloud picture space-time information |
| CN116010812B (en) * | 2022-12-13 | 2023-11-21 | 南京信息工程大学 | North cyclone identification method, storage medium and device based on traditional method and deep learning |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110851957A (en) * | 2019-10-15 | 2020-02-28 | 南京航空航天大学 | Atmospheric data sensing system resolving method based on deep learning |
| CN111722306A (en) * | 2020-07-27 | 2020-09-29 | 郑州航空工业管理学院 | Typhoon landing intensity prediction method and system based on TCN network model |
| CN112070753A (en) * | 2020-09-10 | 2020-12-11 | 浙江科技学院 | Multi-scale information enhanced binocular convolutional neural network saliency image detection method |
| CN113111835B (en) * | 2021-04-23 | 2022-08-02 | 长沙理工大学 | Semantic segmentation method and device for satellite remote sensing image, electronic equipment and storage medium |
-
2022
- 2022-04-25 CN CN202210438320.XA patent/CN114548606B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN114548606A (en) | 2022-05-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110458939B (en) | Indoor scene modeling method based on visual angle generation | |
| CN108961235B (en) | Defective insulator identification method based on YOLOv3 network and particle filter algorithm | |
| CN114548606B (en) | Construction method of cyclone strength prediction model and cyclone strength prediction method | |
| US20210027532A1 (en) | Primitive-based 3d building modeling, sensor simulation, and estimation | |
| CN105654492A (en) | Robust real-time three-dimensional (3D) reconstruction method based on consumer camera | |
| CN110941999B (en) | Method for adaptively calculating size of Gaussian kernel in crowd counting system | |
| CN114332385A (en) | Monocular camera target detection and spatial positioning method based on three-dimensional virtual geographic scene | |
| CN112419202A (en) | Wild animal image automatic identification system based on big data and deep learning | |
| CN115457492A (en) | Target detection method and device, computer equipment and storage medium | |
| CN114863062B (en) | Industrial scene 3D point cloud model construction method based on point and voxel feature representation | |
| CN111079826B (en) | Construction progress real-time identification method integrating SLAM and image processing | |
| CN116310219A (en) | Three-dimensional foot shape generation method based on conditional diffusion model | |
| CN116612468A (en) | Three-dimensional target detection method based on multi-mode fusion and depth attention mechanism | |
| CN113449612A (en) | Three-dimensional target point cloud identification method based on sub-flow sparse convolution | |
| Laupheimer et al. | The importance of radiometric feature quality for semantic mesh segmentation | |
| US7486815B2 (en) | Method and apparatus for scene learning and three-dimensional tracking using stereo video cameras | |
| CN114663880A (en) | Three-dimensional target detection method based on multi-level cross-modal self-attention mechanism | |
| CN114170438A (en) | Neural network training method, electronic device and computer storage medium | |
| CN111881919B (en) | Line element intelligent simplification method and device based on tracking type grid subdivision | |
| CN116758219A (en) | Region-aware multi-view stereo matching three-dimensional reconstruction method based on neural network | |
| CN113591685B (en) | Geographic object spatial relationship identification method and system based on multi-scale pooling | |
| CN114881850A (en) | Point cloud super-resolution method and device, electronic equipment and storage medium | |
| CN113496159B (en) | Multi-scale convolution and dynamic weight cost function smoke target segmentation method | |
| Cao et al. | Understanding 3D point cloud deep neural networks by visualization techniques | |
| CN112287740B (en) | Target detection method and device for power transmission line based on YOLOv3-tiny, and unmanned aerial vehicle |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |