CN117975298A - Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method - Google Patents

Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method Download PDF

Info

Publication number
CN117975298A
CN117975298A CN202410389490.2A CN202410389490A CN117975298A CN 117975298 A CN117975298 A CN 117975298A CN 202410389490 A CN202410389490 A CN 202410389490A CN 117975298 A CN117975298 A CN 117975298A
Authority
CN
China
Prior art keywords
mesoscale
sub
vortex
observation
typical
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.)
Granted
Application number
CN202410389490.2A
Other languages
Chinese (zh)
Other versions
CN117975298B (en
Inventor
年睿
袁明瀚
张正光
吴彤
李建辉
吴菲菲
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ocean University of China
Original Assignee
Ocean University of China
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Ocean University of China filed Critical Ocean University of China
Priority to CN202410389490.2A priority Critical patent/CN117975298B/en
Publication of CN117975298A publication Critical patent/CN117975298A/en
Application granted granted Critical
Publication of CN117975298B publication Critical patent/CN117975298B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/048—Activation functions
    • 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
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Biomedical Technology (AREA)
  • Databases & Information Systems (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention belongs to the technical field of image classification, and relates to an unsupervised classification method for a mesoscale vortex and a sub-mesoscale fine structure AI thereof, which comprises the following steps: extracting a mesoscale vortex observation sub-image dataset; labeling the mesoscale vortex observation sub-image data set, and randomly dividing the mesoscale vortex observation sub-image data set into a training sample set and a test sample set; constructing a self-adaptive neural structure searching and judging network; training a self-adaptive neural structure searching and distinguishing network, and outputting a mesoscale vortex and a sub-mesoscale typical structure distinguishing result thereof by using a test set; constructing an unsupervised learning model driven by a self-adaptive neural structure search discrimination network, and outputting a clustering result by utilizing a multiple clustering criterion; and respectively calculating the wave number spectrum and the slope of the typical and atypical structures induced by the mesoscale vortex and the distribution of the vortex-free background field according to the output clustering result of the typical fine structure of the sub-mesoscale, and carrying out auxiliary verification on the typical configuration according to the wave number spectrum and the slope difference. The invention can effectively capture the morphological characteristics of the sub-mesoscale fine structure.

Description

Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method
Technical Field
The invention belongs to the technical field of image classification, and particularly relates to an unsupervised classification method for mesoscale vortexes and sub-mesoscale fine structures AI thereof.
Background
Ocean mesoscale vortex is a phenomenon commonly found in the global ocean and is characterized by a rotating vortex with horizontal scales from tens to hundreds of kilometers and time scales from weeks to months. The inherent spatio-temporal dimensions of mesoscale vortices determine their unique kinetic characteristics, a very important ring in whole ocean energy cascade chains. Mesoscale vortices are created from the background flow, transferring energy from large to mesoscale, but the dissipation process of energy occurs on small scale, and if it is desired to cascade energy from mesoscale to smaller scale, it is necessary to destructively turn the balance, creating non-geodetic motion, which requires a sub-mesoscale process. Current research of sub-mesoscale processes is kinetically focused on unstable processes that produce sub-mesoscale processes, but inherent features and types of sub-mesoscale processes themselves still lack systematic research and classification, largely due to observation difficulties and challenges presented by the short spatio-temporal scales of sub-mesoscales. One way to overcome this obstacle is to study the sub-mesoscale fine structure of mesoscale vortices using satellite remote sensing image data with high resolution and broad coverage.
The image classification (Image Classification) tasks can be divided into supervised classification (Supervised Classification) and unsupervised (Unsupervised Classification) classification. The image classification task at the present stage is realized largely by means of supervised learning, that is, each sample has a label corresponding to the sample, and classification is finally realized by continuously learning the characteristics corresponding to each label, such as a support vector machine (Support Vector Machine, SVM), random Forest (RF), extreme learning machine (Extreme LEARNING MACHINES, ELM) and the like. The supervision classification using the marking information can obtain higher precision, but when the fine-grained classification and multi-label classification tasks are involved, the marking cost can exponentially increase along with the number of targets and the identifiable difficulty, the training cost is too high, and the application in production practice is not facilitated, so that a lot of researches on the non-supervision image classification are generated. The unsupervised method classifies images based on the characteristics of the images, and solves the classification problem of automatic labeling to a certain extent.
At present, unsupervised image classification research based on deep learning is still in a development stage, and because of the large problem difficulty, the research results are only tested on certain simple data sets, and great challenges still exist for identifying and classifying complex and diverse images.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides an unsupervised classification method for a mesoscale vortex and a sub-mesoscale fine structure AI thereof, which comprises the following steps:
s1: collecting a mesoscale vortex high-resolution remote sensing observation data set, and extracting a mesoscale vortex observation sub-image data set according to a known mesoscale vortex center position reference standard from the data set;
S2: labeling the mesoscale vortex observation sub-image data set as whether the mesoscale vortex observation sub-image data set has a candidate typical structure or not, and randomly dividing the mesoscale vortex observation sub-image data set into a training sample set and a testing sample set; the training sample set consists of samples with candidate typical structures and samples without candidate typical structures; constructing a self-adaptive neural structure searching and judging network;
s3: constructing an adaptive neural structure searching and judging network, training the adaptive neural structure searching and judging network in the step S2 by a training sample set, using a trained network model for testing a test sample set, and outputting a mesoscale vortex and a sub-mesoscale typical structure judging result thereof;
S4: constructing an unsupervised learning model driven by a self-adaptive neural structure search discrimination network, further dividing a sub-mesoscale typical fine structure by utilizing a constructed multiple clustering criterion according to the mesoscale vortex and the sub-mesoscale typical structure discrimination result outputted in the step S3, and outputting a clustering result;
S5: and (3) respectively calculating the wave number spectrum and the slope of the typical and atypical structures induced by the mesoscale vortex and the distribution of the vortex-free background field according to the clustering result of the typical fine structure of the sub-mesoscale output in the step S4, and assisting in verifying the typical configuration according to the wave number spectrum and the slope difference.
Preferably, the step S1 specifically includes:
Introducing a normalized vortex center coordinate to extract a mesoscale vortex observation sub-image: for the track of a given one of the mesoscale vortices, is arranged in At the moment, its vortex center position isRadius isLet the original coordinates of the mesoscale vortex observation sub-image beProjecting the mesoscale vortex observation sub-image into a coordinate system taking the vortex center as an origin, namely, the horizontal coordinate of the mesoscale vortex observation sub-image isWherein，；
Normalizing horizontal coordinates to vortex radius，，；
By normalizing the vortex center coordinates, inExtracting a mesoscale vortex observation sub-image data set in the vortex radius region; and extracting the mesoscale vortex observed sub-image of the vortex-free region as a vortex-free background field distribution data set.
Preferably, in the step S2, the specific process of labeling the mesoscale vortex observation sub-image dataset is:
(1) Firstly, taking a residual neural network as an encoder, and performing depth feature mapping on a mesoscale vortex observation sub-image data set to form an encoding feature set;
(2) Using the depth mapping feature set as input of a K-means clustering method, and calculating clustering errors of all mapping features;
(3) And judging whether the mesoscale vortex observation sub-image has a candidate typical structure or not according to the clustering error.
Preferably, in the step S3, the adaptive neural structure searching and distinguishing network uses a neural structure searching technology and a compound scaling technology, and is composed of a fusion mobile overturn bottleneck convolution, a mesoscale vortex observation sub-image key feature weight guiding unit and a mesoscale vortex observation sub-image key feature significance feedback unit; the training process comprises the following steps:
(1) Taking arbitrary labeled mesoscale vortex observation sub-image data as input, and extracting a mesoscale vortex observation sub-image convolution feature map through a self-adaptive neural structure searching and distinguishing network;
(2) The method comprises the steps that a mesoscale vortex observation sub-image key feature weight guiding unit is utilized, a larger weight is given to a remarkable feature in a channel dimension, and a mesoscale vortex observation sub-image feature map with a higher degree of distinction is obtained;
(3) And extracting the salient features of the space dimension by using a key feature salient feedback unit of the mesoscale vortex observation sub-image to obtain a discrimination result of the mesoscale vortex typical structure.
Preferably, in the step S4, the unsupervised learning model driven by the adaptive neural structure search and discrimination network includes the adaptive neural structure search and discrimination network and a transducer coding module; the transducer coding module consists of a multi-head self-attention mechanism, a residual error connecting layer, a normalization layer and a feedforward neural network.
Preferably, in the step S4, the multiple clustering criterion is a multiple clustering loss function constructed by using inter-sample similarity, label inferred similarity, mutual information correlation, and cooperative contrast classification loss, and specifically includes:
；
Wherein, 、、Respectively representing the weight duty ratio of each loss function; /(I)A similarity loss function between samples; /(I)Deducing a similarity loss function for the tag; /(I)Is a mutual information correlation loss function; /(I)Classifying the loss function for cooperative comparison; representing the transform depth map function parameters.
Preferably, the sample-to-sample similarity loss function is:
；
Wherein, 、Mesoscale vortex observation samples/>, respectively, with typical structure、Is a coding feature of (a); /(I)Is a typical structure sample setIs a similarity loss pseudo-graph.
Preferably, the tag extrapolates a similarity loss function as:
；
Wherein, Represents the/>, in the typical structure discrimination resultA set of potential class samples; /(I)A pseudo tag self-monitoring matrix representing the potential class; /(I)Representing the total number of potential categories; /(I)Representing cross entropy loss; forMesoscale vortex observation samples/>, with typical structures, anywhere within the individual potential categoriesLetFor the characteristics of the code thereof,Tags for their potential categories.
Preferably, the inter-sample mutual information correlation loss function is:
；
Wherein, Is a deep feature of each sample in a sample set of typical structures,Is a shallow feature of each sample in a typical structure sample set; /(I)Is to calculate deep features/>, by using Jensen-shannon divergenceAnd shallow featuresIs a piece of mutual information of the mobile terminal.
Preferably, for a mesoscale vortex observation sample with a typical structurePositive samples/>, with the same class of sub-mesoscale fine structuresNegative samples with maximized inter-similarity and different class sub-mesoscale fine structureThe cooperative contrast class loss function with minimized inter-similarity is:
；
Wherein, Representing the total number of mesoscale vortex observed samples with typical structure; /(I)Deep feature extraction for representing a mesoscale vortex observation sample; /(I)Is a positive sampleTotal number of/(I)Is a negative sampleIs the sum of (3); representing the inner product of positive samples of the same class of fine structures in a high-dimensional space,/> Representing the inner product of negative samples of different types of fine structures in a high-dimensional space as a sub-mesoscale fine structure similarity measure; /(I)And the parameters are regulated and used for controlling the smoothness degree of the distribution of the same-class sub-mesoscale fine structures.
Compared with the prior art, the invention has the beneficial effects that: aiming at the characteristic that a mesoscale vortex high-resolution remote sensing observation image is influenced by the rotation direction, rotation speed, vortex size, vortex shape and other characteristics of the mesoscale vortex, the invention presents complex and changeable morphological characteristics, utilizes the neural network searching and compound scaling technology to construct a self-adaptive neural structure searching and distinguishing network to pay more attention to a typical structure formed by the mesoscale vortex motion, and can effectively capture morphological characteristics of a sub-mesoscale fine structure. And the classification of the sub-mesoscale fine structure is realized in an unsupervised mode by utilizing the self-adaptive neural structure searching and distinguishing network as a drive, and finally the classification of the sub-mesoscale fine structure is verified and assisted through wave number spectrum and slope difference.
Drawings
FIG. 1 is a flow chart of an unsupervised classification method for a mesoscale vortex and its sub-mesoscale fine structure AI in an embodiment of the invention;
FIG. 2 is a diagram of a model structure of a typical structure adaptive neural structure search discrimination network constructed in an embodiment of the present invention.
Detailed Description
The following describes in further detail the embodiments of the present invention with reference to the drawings and examples. The following examples are illustrative of the invention and are not intended to limit the scope of the invention.
In this embodiment, sea surface chlorophyll high-resolution remote sensing observation data is taken as an example, and the mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method provided by the invention is described in detail. The flow of the method is shown in fig. 1, and specifically comprises the following steps:
Step one, collecting a mesoscale vortex high-resolution remote sensing observation data set, and extracting a mesoscale vortex observation sub-image data set from the data set according to a known mesoscale vortex center position reference standard.
Collecting sea surface chlorophyll high-resolution remote sensing observation data setWhereinRepresenting the total number of images in the dataset, stSheet imageIs，。
And then introducing a normalized vortex center coordinate according to the known reference standard of the position of the vortex center of the mesoscale vortex, namely the polarity, longitude, latitude, amplitude, radius and rotation speed of the mesoscale vortex, so as to extract the sea surface chlorophyll observation sub-image.
For the track of a given one of the mesoscale vortices, is arranged inAt the moment, its vortex center position isRadius isLet the original coordinates of the sea surface chlorophyll observation sub-image beThe sub-images of the sea surface chlorophyll observation can be projected into a coordinate system with the vortex center as the origin, namely the horizontal coordinate of the sub-images of the sea surface chlorophyll observation around the mesoscale vortex isWherein，. Further normalizing horizontal coordinates to vortex radiusWherein，。
By normalizing the vortex center coordinates, in（Generally take 2.5 but different remote sensing data values are different, and the data set is determined according to specific conditions) in the area with double vortex radius to extract the sub-image data set of sea surface chlorophyll observationWhereinComprises(Is a natural number representing the number of sub-images) sea surface chlorophyll observation sub-images extracted according to a mesoscale vortex core position reference standard, and in addition, the sea surface chlorophyll observation sub-images of the vortex-free region are extracted as a vortex-free background field distribution dataset。
Step two, the sea surface chlorophyll observation sub-image data setLabeling according to whether there are candidate canonical structures; dividing a training sample set and a testing sample set; and constructing a self-adaptive neural structure searching and distinguishing network.
1. Sea surface chlorophyll observation sub-image data setTagging, data set/>, according to whether there is a candidate canonical chlorophyll structurePreliminary partitioning into two sets: /(I)Representing that the set of seasurface chlorophyll observation sub-images has a candidate representative chlorophyll structure; /(I)Indicating that the set of seasurface chlorophyll observation sub-images does not have candidate representative chlorophyll structures. The labeling process comprises the following specific steps:
(1) Firstly, taking a residual neural network as a base of an encoder, and observing a sub-image dataset of sea surface chlorophyll Performing depth feature mapping to form a coded feature set；
(2) Using the above depth map feature set as input to the K-means clustering method, the clusters will be determined synthetically by the spatial distance between the encoded feature sets in a similarity measure in a lower dimension:
（1）；
Wherein, Representing a set of coding features with candidate representative chlorophyll structures and a set of coding features without candidate representative chlorophyll structures,Represents theMean vector of individuals,Encoding features for a single depth map,Cluster errors for all mapping features;
(3) Finally according to the clustering error Judging sea surface chlorophyll observation sub-imageWhether or not there is a candidate typical chlorophyll structure.
2. Sub-image dataset from labeled seasurface chlorophyll observationsThe training samples need only consist of a small number of balanced samples with candidate representative chlorophyll structures and samples without candidate representative chlorophyll structures.
3. The method comprises the steps of constructing a self-adaptive neural structure searching and distinguishing network, and specifically comprises the following steps: by utilizing the neural structure search technology and the compound scaling technology, an efficient self-adaptive neural structure search discrimination network is constructed based on the mobile overturn bottleneck convolution and the fusion mobile overturn bottleneck convolution module by comprehensively considering the depth, the width and the resolution of the network. And merging the mobile overturn bottleneck convolutions in the shallow network stack multilayer, and merging the mobile overturn bottleneck convolutions in the deep network stack multilayer. The mobile flip bottleneck convolution consists of a standard convolution, a depth separable convolution and a Squeeze and Excitation (SE) module, and the fusion mobile flip bottleneck convolution consists of a standard convolution and an SE module. The model structure is shown in fig. 2, and the self-adaptive neural structure searching and distinguishing network is composed of a fusion mobile overturn bottleneck convolution, a mesoscale vortex observation sub-image key feature weight guiding unit and a mesoscale vortex observation sub-image key feature significance feedback unit.
And thirdly, training the self-adaptive neural structure searching and distinguishing network by using a training set, using the trained network model for testing a testing set, and outputting a mesoscale vortex and a sub-mesoscale typical structure distinguishing result thereof.
Training process: 1. observing sub-images with arbitrarily labeled sea surface chlorophyllAs input, extracting the convolutional feature map/>, of the sea surface chlorophyll observation sub-image via an adaptive neural structure search discrimination network，Channel number representing convolution feature map,Representing the height of the convolution feature map,Representing the width of the convolution signature.
2. Key feature weight guiding unit for observing sub-image by using mesoscale vortexThe salient features are given greater weight in the channel dimension. The method specifically comprises the following steps:
(1) First to Each channel is subjected to global average pooling to acquire information of global feature contexts;
(2) Calculation of Global maximum pooling of each channel is carried out, and sea surface chlorophyll observation sub-image characteristics with more distinguishing degree are obtained; respectively obtain average pooling characteristicsAnd maximize pooling feature；
(3) Utilizing average pooling featuresAnd calculating the similarity of each channel and all other channels, and weighting the characteristics of each channel by the similarity. The method comprises the following steps: calculate a similarity matrixWhereinRepresents theIndividual channels andThe similarity of the pooling features of the individual channel averages;
weighting with similarity matrix AndRespectively obtaining update vectors:
（2）；
Wherein, Representing global average pooling,Representing global maximum pooling; is/> Update vector ofIsIs used to update the vector.
Will beAndAdding to obtain a final sea surface chlorophyll observation sub-image key feature weight guide matrix，Representing the transpose.
Convolving sea surface chlorophyll observation sub-image with characteristic diagramAnd key feature weight steering matrixMultiplying corresponding elements of the sea surface chlorophyll observation sub-image feature images with more discrimination：
（3）；
Wherein,As a sigmoid function,Representing element level multiplication.
3. Key feature significance feedback unit for observing sub-image by using mesoscale vortexSignificant features of the spatial dimension are extracted. Sea surface chlorophyll observation sub-image feature map/>, more differentiated in the channel dimensionRespectively carrying out large-scale and small-scale convolution operation, capturing spatial information in a larger range and finer local spatial information, and obtaining a large-scale spatial convolution characteristic diagramAnd small-scale spatial convolution feature map：
（4）。
Respectively pairs in channel dimensionAndCarrying out average pooling and maximum pooling, and reserving and feeding back pixel space characteristic information of sea surface chlorophyll observation sub-images to obtain large-scale cross-channel average pooling characteristicsSmall-scale cross-channel average pooling featureLarge scale max pooling featuresSmall-scale max pooling feature. Two kinds of pooling features of a large scale and a small scale are spliced on the channel dimension respectively through the large scale convolution layer and the small scale convolution layer:
（5）；
Wherein, 、Respectively representing large-scale and small-scale convolution operations; /(I)、And the saliency feedback weight matrix is respectively used for the characteristic diagram saliency feedback weight matrix of the output large-scale and small-scale sea surface chlorophyll observation sub-images.
Corresponding the significance feedback weight matrix to a large-scale space convolution characteristic mapAnd small-scale spatial convolution feature mapIn the channel dimension splicing, mesoscale convolution/>, is utilizedFusing the large-scale and small-scale saliency feedback characteristic information to obtain a finally output saliency feedback sea surface chlorophyll observation sub-image fusion characteristic diagram：
（6）。
Fusing saliency feedback sea surface chlorophyll observation sub-images into characteristic imagesSerializing to obtainObtaining the actual discrimination output of the mesoscale vortex typical structure:
（7）；
Wherein, Representing any sea surface chlorophyll observation sub-imageConfidence in whether or not to have a typical chlorophyll structure, according toAnd judging whether the structure has a mesoscale and sub-mesoscale typical structure or not by taking the value.
And (3) testing a network model and outputting results: the trained network model is used for testing a test set, and a discrimination result of a mesoscale vortex and a sub-mesoscale typical structure is output to form a typical chlorophyll structure sample setAnd atypical chlorophyll Structure sample setWherein、Respectively, the total number of typical chlorophyll structure and atypical chlorophyll structure observation samples output by the adaptive neural structure searching and distinguishing network,，。
Step four, constructing an unsupervised learning model driven by a self-adaptive neural structure searching and distinguishing network, and according to the output mesoscale vortex and a sub-mesoscale typical structure sample set (namely) Establishing a multiple clustering criterion by utilizing the similarity among samples, the label inferred similarity and the mutual information correlation and the cooperation contrast classification loss, further dividing a sub-mesoscale typical fine structure, and outputting a clustering result;
1. and constructing an unsupervised learning model comprising the self-adaptive neural structure searching and distinguishing network, wherein the unsupervised learning model consists of a fusion mobile overturn bottleneck convolution, a mesoscale vortex observation sub-image key feature weight guiding unit, a mesoscale vortex observation sub-image key feature significance feedback unit and a transform coding module. The transducer coding module consists of a multi-head self-attention mechanism, a feedforward neural network, and a residual connection layer and a normalization layer are arranged between the two parts.
2. The sea surface chlorophyll observation sample with the typical structure is preparedAs input, shallow feature/>, captured with fusion mobile inversion bottleneck convolution of a multi-layer stackCapturing deep features by multi-layer mobile rollover bottleneck convolution is then continuedThe key feature weight guiding unit/>, of the mesoscale vortex observation sub-image constructed by the method, is further used for guiding the key feature weight of the mesoscale vortex observation sub-imageAnd mesoscale vortex observation sub-image key feature significance feedback unitCapturing fine-grained local features, combining normalization and ReLU functions, and outputting sea surface chlorophyll fine structure clustering featuresFinally, the output set is serialized by utilizing point-by-point convolutionWhereinRepresenting the sequence dimension:
（8）；
Wherein, 、Respectively representing a multi-layer stacked fusion mobile overturning bottleneck rolling module and a mobile overturning bottleneck convolution module; /(I)Representing ReLU activation function,Representing a point-by-point convolution.
Serializing output sets for cluster featuresPosition coding/>, is calculated for each cluster feature sequenceEmbedding the position codes into the clustering feature sequences to update and obtain：
（9）。
Embedding position codes into clustered feature sequencesInput to a transducer coding module, and learn a transducer-based depth mapping function/>, through a multi-head self-attention mechanism, residual connection and layer normalization operationA transducer code representation of the distinguishing characteristics of the typical structure of each sea surface chlorophyll is obtained:
， representing the transform depth map function parameters.
3. Multiple clustering criteria
(1) And constructing a similarity loss function between samples.
Transformer coding module based on the self-adaptive neural structure searching and distinguishing network-driven unsupervised learning model, and for typical chlorophyll structure sample setSea surface chlorophyll observation samples/>, with typical structures, are calculated by Gaussian kernel functions in pairsAndSimilarity between coding featuresObtaining a similarity matrix：
（10）；
Wherein,、Respectively represents any two sea surface chlorophyll observation samples with typical structuresAndEncoding features ofTo calculate the Euclidean distance operation,Is its variance.
Calculating a sample set of typical chlorophyll structuresSimilarity loss pseudo-graph：
（11）；
Wherein,A given inter-sample similarity loss threshold.
Constructing a similarity loss function between samples by minimizing cosine similarity between the same classes and guiding the similarity loss pseudo-graph：
（12）；
Wherein,、Sea surface chlorophyll observation samples/>, respectively, having typical structuresAndIs a coded feature of (a).
(2) Constructing a label to infer a similarity loss function.
Constructing a similarity graph based on the established Gaussian kernel function, and calculating a corresponding symmetrical adjacency matrix，Representative chlorophyll Structure sample setAny two verticesAndWeights of the edges in between.
Cutting the similarity graph into segmentsEach sub-graph is respectivelyAnd for the weight between any two sub-graphsThe method comprises the following steps:
（13）；
Wherein, AndRespectivelySubgraph,，RespectivelyFirst/>, among a set of sub-graph verticesSumSamples.
For the followingA collection of individual subgraphs, a cut graph is defined as：
（14）；
Wherein,ForIs to minimizeThe cut map can be divided into optimalIndividual subgraphs, and thus a sample set of typical chlorophyll structuresClassified asPotential classes of，。
If the sea surface chlorophyll observation sample has a typical structureAtCategoryIn (1), willSet toAnd constructing the potential class label into a one-hot coded form:
（15）。
at the same time, for any sea surface chlorophyll observation sample with typical structure in the potential category The coding features output by itIts pseudo tag can be calculated, expressed asThe corresponding predicted pseudo tag probability isBy setting a thresholdTo select high confidence pseudo tag self-supervised learning:
（16）。
Thereby obtaining the first All pseudo-tag self-supervised learning matrices of category. OnlyWhen (1)SamplePseudo tagTraining is performed and the tag deduces the similarity loss functionThe definition is as follows:
（17）；
Wherein, Representing cross entropy loss.
(3) And constructing a mutual information correlation loss function.
Calculation of deep features with Jensen-Shannon divergence (JSD)And shallow featuresMutual information of (a). WhereinIs a typical chlorophyll structure sample set/>, which is obtained by a multilayer stacked mobile flip bottleneck convolution deep networkDeep features of each sample/(Is a typical chlorophyll structure sample set/>, which is obtained by a multilayer stacked fusion mobile flip bottleneck convolution shallow networkShallow features of each sample:
（18）；
Wherein, Representing a joint distribution between deep features and shallow features,Representing deep features and shallow feature edge distributions,As a discriminator,As a softplus function. If/>, between samplesAndFrom the same category, then the joint distribution is followed, otherwise the edge distribution is followed. By selecting a sample pair with high similarity between deep features and shallow features as a positive sample pair and a sample pair with low similarity as a negative sample pair, a mutual information correlation loss function：
（19）。
(4) And constructing a cooperative comparison classification loss function.
For a given sea surface chlorophyll observation sample with a typical structureDesign positive samples/>, of similar sub-mesoscale fine structuresMaximizing inter-similarity and negative-sample/>, with different classes of sub-mesoscale fine structuresCollaborative contrast class loss function with minimized inter-similarity:
（20）；
Wherein, Deep feature extraction of sea surface chlorophyll observation sample,AndSea surface chlorophyll observation samples/>, respectively, having typical structuresThe total number of positive samples of the same category and negative samples of different categories,Representing the inner product of a positive sample of the same class of fine structures in high dimensional space,Representing the inner product of negative samples of different types of fine structures in a high-dimensional space as a sea surface chlorophyll fine structure similarity measure,Is an adjusting parameter used for controlling the smoothness degree of the fine structure distribution of the sea surface chlorophyll of the same class.
(5) The constructed sample-to-sample similarity loss functionTag inferred similarity loss functionMutual information correlation loss functionCollaborative comparison Classification loss functionCombining to construct a sea surface chlorophyll fine structure clustering total loss function:
（21）；
Wherein, 、AndRepresenting the weight duty ratio/>, of each loss function。
Sample set of typical chlorophyll structuresInputting the result into an unsupervised learning model driven by a built self-adaptive neural structure search discrimination network, and further establishing a total loss function based on multiple clustering criteriaDivided into having aClass of sea surface chlorophyll typical fine structure sample set，。
And fifthly, respectively calculating wave number spectrums and slopes of typical and atypical structures of the sea surface chlorophyll induced by the mesoscale vortex and the distribution of the vortex-free background field. According to the sea surface chlorophyll fine structure clustering result, introducing 2D fast Fourier transform, and respectively calculating sea surface chlorophyll typical fine structure sample set induced by mesoscale vortexSimultaneously calculating a mesoscale vortex atypical chlorophyll structure sample setVortex-free background field distribution setWavenumber spectrum and slope of (a). After the spectrum analysis of different clustering results, the wave number spectrums and the spectrum slopes of the clustering results have obvious differences, and the typical structures can be classified according to the assistance.
(1) Calculating wave number spectrum by radial summation of 2D fast Fourier transformation for all sea surface chlorophyll observation sample data：
（22）；
Wherein,For 2D fast Fourier transform,Represents a given sea surface chlorophyll observation sampleAt the central positionDistribution on the upper surface; /(I)Data set representing sea surface chlorophyll observation sub-imageAnd vortex free background field distribution datasetIs a collection of (3); /(I)、Wave numbers in the horizontal and vertical directions,/>, respectivelyThe total wave number.
(2) For wave number spectrumSum wave numberTaking logarithm, and fitting wave number spectrum slope/>, by using least square method：
（23）；
（24）；
Wherein,To fit the offset,Is the number of wavenumbers in the mesoscale band range,Represents theWave number,，Confidence interval for wave number spectrum slope,To haveSignificance level andDegree of freedomQuantiles of distribution,Representing the base logarithm of 10,Representing the average value. /(I)

Claims (10)

1. The unsupervised classifying method for the mesoscale vortex and the sub-mesoscale fine structure AI is characterized by comprising the following steps:
s1: collecting a mesoscale vortex high-resolution remote sensing observation data set, and extracting a mesoscale vortex observation sub-image data set according to a known mesoscale vortex center position reference standard from the data set;
s2: labeling the mesoscale vortex observation sub-image data set as whether the mesoscale vortex observation sub-image data set has a candidate typical structure or not, and randomly dividing the mesoscale vortex observation sub-image data set into a training sample set and a testing sample set; the training sample set consists of samples with candidate typical structures and samples without candidate typical structures;
S3: constructing a self-adaptive neural structure searching and judging network; training the self-adaptive neural structure searching and judging network in the step S2 by a training sample set, using the trained network model for testing the test sample set, and outputting a mesoscale vortex and a sub-mesoscale typical structure judging result thereof;
S4: constructing an unsupervised learning model driven by a self-adaptive neural structure search discrimination network, further dividing a sub-mesoscale typical fine structure by utilizing a constructed multiple clustering criterion according to the mesoscale vortex and the sub-mesoscale typical structure discrimination result outputted in the step S3, and outputting a clustering result;
S5: and (3) respectively calculating the wave number spectrum and the slope of the typical and atypical structures induced by the mesoscale vortex and the distribution of the vortex-free background field according to the clustering result of the typical fine structure of the sub-mesoscale output in the step S4, and assisting in verifying the typical configuration according to the wave number spectrum and the slope difference.
2. The method for unsupervised classification of mesoscale vortices and sub-mesoscale fine structures AI thereof according to claim 1, wherein said step S1 comprises:
Introducing a normalized vortex center coordinate to extract a mesoscale vortex observation sub-image: for the track of a given one of the mesoscale vortices, is arranged in At the moment, its vortex center position isRadius isLet the original coordinates of the mesoscale vortex observation sub-image beProjecting the mesoscale vortex observation sub-image into a coordinate system taking the vortex center as an origin, namely, the horizontal coordinate of the mesoscale vortex observation sub-image isWherein，; Normalizing horizontal coordinates to/> according to vortex radius，，；
By normalizing the vortex center coordinates, inExtracting a mesoscale vortex observation sub-image data set in the vortex radius region; and extracting the mesoscale vortex observed sub-image of the vortex-free region as a vortex-free background field distribution data set.
3. The method for unsupervised classification of mesoscale vortices and sub-mesoscale fine structures AI according to claim 1, wherein in step S2, the specific process of labeling the mesoscale vortex observation sub-image dataset is as follows:
(1) Firstly, taking a residual neural network as an encoder, and performing depth feature mapping on a mesoscale vortex observation sub-image data set to form an encoding feature set;
(2) Using the depth mapping feature set as input of a K-means clustering method, and calculating clustering errors of all mapping features;
(3) And judging whether the mesoscale vortex observation sub-image has a candidate typical structure or not according to the clustering error.
4. The method for unsupervised classification of mesoscale vortices and sub-mesoscale fine structures AI thereof according to claim 1, wherein in step S3, the adaptive neural structure search discrimination network is composed of a fusion mobile flip bottleneck convolution, a mesoscale vortex observation sub-image key feature weight guiding unit and a mesoscale vortex observation sub-image key feature significance feedback unit by using a neural structure search technique and a compound scaling technique; the training process comprises the following steps:
(1) Taking arbitrary labeled mesoscale vortex observation sub-image data as input, and extracting a mesoscale vortex observation sub-image convolution feature map through a self-adaptive neural structure searching and distinguishing network;
(2) The method comprises the steps that a mesoscale vortex observation sub-image key feature weight guiding unit is utilized, a larger weight is given to a remarkable feature in a channel dimension, and a mesoscale vortex observation sub-image feature map with a higher degree of distinction is obtained;
(3) And extracting the salient features of the space dimension by using a key feature salient feedback unit of the mesoscale vortex observation sub-image to obtain a discrimination result of the mesoscale vortex typical structure.
5. The method according to claim 1, wherein in the step S4, the unsupervised learning model driven by the adaptive neural structure search discrimination network includes an adaptive neural structure search discrimination network and a transducer coding module; the transducer coding module consists of a multi-head self-attention mechanism, a residual error connecting layer, a normalization layer and a feedforward neural network.
6. The method according to claim 5, wherein in the step S4, the multiple clustering criterion is a multiple clustering loss function constructed by using inter-sample similarity, label inferred similarity, mutual information correlation, and cooperative contrast classification loss, and the multiple clustering loss function is specifically:
；
Wherein, 、、Respectively represent the weight duty ratio/>, of each loss function；A similarity loss function between samples; /(I)Deducing a similarity loss function for the tag; /(I)Is a mutual information correlation loss function; /(I)Classifying the loss function for cooperative comparison; /(I)Representing the transform depth map function parameters.
7. The method for unsupervised classification of mesoscale vortices and sub-mesoscale fine structures AI according to claim 6, wherein said sample-to-sample similarity loss function is:
；
Wherein, 、Mesoscale vortex observation samples/>, respectively, with typical structure、Encoding features ofIs a typical structure sample setIs a similarity loss pseudo-graph.
8. The method of unsupervised classification of mesoscale vortices and sub-mesoscale fine structures AI thereof according to claim 6, wherein said tag extrapolates a similarity loss function as:
；
Wherein, Represents the/>, in the typical structure discrimination resultSample set of potential classes,Pseudo tag self-monitoring matrix representing the potential class,Representing the total number of potential categories,Representing cross entropy loss; forMesoscale vortex observation samples/>, with typical structures, anywhere within the individual potential categories，For its coding features,Tags for their potential categories.
9. The method for unsupervised classification of mesoscale vortices and sub-mesoscale fine structures AI according to claim 6, wherein said inter-sample mutual information correlation loss function is:
；
Wherein, Is a deep feature of each sample in a typical structure sample set; /(I)Is a shallow feature of each sample in a typical structure sample set; /(I)Is to calculate deep features/>, by using Jensen-shannon divergenceAnd shallow featuresIs a piece of mutual information of the mobile terminal.
10. The method of unsupervised classification of mesoscale vortices and their sub-mesoscale fine structures AI according to claim 6, characterized in that for mesoscale vortex observation samples with typical structurePositive samples/>, with the same class of sub-mesoscale fine structuresNegative-sample/>, with maximized inter-similarity, and with different class sub-mesoscale fine structuresThe cooperative contrast class loss function with minimized inter-similarity is:
；
Wherein, Representing the total number of mesoscale vortex observed samples with typical structure; /(I)Deep feature extraction for representing a mesoscale vortex observation sample; /(I)Is a positive sampleIs the sum of (3); /(I)Is a negative sampleIs the sum of (3); /(I)Representing the inner product of positive samples of the same class of fine structures in a high-dimensional space,Representing the inner product of negative samples of different types of fine structures in a high-dimensional space as a sub-mesoscale fine structure similarity measure; /(I)Is a tuning parameter for controlling the smoothness of the distribution of the same-class sub-mesoscale fine structures.
CN202410389490.2A 2024-04-02 2024-04-02 Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method Active CN117975298B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202410389490.2A CN117975298B (en) 2024-04-02 2024-04-02 Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202410389490.2A CN117975298B (en) 2024-04-02 2024-04-02 Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method

Publications (2)

Publication Number Publication Date
CN117975298A true CN117975298A (en) 2024-05-03
CN117975298B CN117975298B (en) 2024-07-05

Family

ID=90858285

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202410389490.2A Active CN117975298B (en) 2024-04-02 2024-04-02 Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method

Country Status (1)

Country Link
CN (1) CN117975298B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113392961A (en) * 2021-08-17 2021-09-14 中国海洋大学 Method for extracting mesoscale eddy track stable sequence and predicting cyclic neural network
WO2022001489A1 (en) * 2020-06-28 2022-01-06 北京交通大学 Unsupervised domain adaptation target re-identification method
CN115311502A (en) * 2022-09-16 2022-11-08 西安电子科技大学 Remote sensing image small sample scene classification method based on multi-scale double-flow architecture

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022001489A1 (en) * 2020-06-28 2022-01-06 北京交通大学 Unsupervised domain adaptation target re-identification method
CN113392961A (en) * 2021-08-17 2021-09-14 中国海洋大学 Method for extracting mesoscale eddy track stable sequence and predicting cyclic neural network
CN115311502A (en) * 2022-09-16 2022-11-08 西安电子科技大学 Remote sensing image small sample scene classification method based on multi-scale double-flow architecture

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
ZHANG, ZG等: "Surface Chlorophyll Enhancement in Mesoscale Eddies by Submesoscale Spiral Bands", 《GEOPHYSICAL RESEARCH LETTERS》, 31 July 2020 (2020-07-31) *
王改华;李涛;吕朦;袁国亮;: "采用无监督学习算法与卷积的图像分类模型", 华侨大学学报(自然科学版), no. 01, 20 January 2018 (2018-01-20) *

Also Published As

Publication number Publication date
CN117975298B (en) 2024-07-05

Similar Documents

Publication Publication Date Title
Bouti et al. A robust system for road sign detection and classification using LeNet architecture based on convolutional neural network
EP2808827B1 (en) System and method for OCR output verification
Bevandić et al. Discriminative out-of-distribution detection for semantic segmentation
Chen et al. Vehicle detection in high-resolution aerial images via sparse representation and superpixels
Cheriyadat Unsupervised feature learning for aerial scene classification
Sun et al. On-road vehicle detection using evolutionary Gabor filter optimization
Attari et al. Nazr-CNN: Fine-grained classification of UAV imagery for damage assessment
Torralba et al. Sharing visual features for multiclass and multiview object detection
Paclík et al. Building road-sign classifiers using a trainable similarity measure
Yu et al. Vehicle detection from high-resolution remote sensing imagery using convolutional capsule networks
EP2443612B1 (en) Image recognition method and image recognition apparatus
Zhang et al. Road recognition from remote sensing imagery using incremental learning
CN110210534B (en) Multi-packet fusion-based high-resolution remote sensing image scene multi-label classification method
CN106326938B (en) SAR image target discrimination method based on Weakly supervised study
Chang et al. SpoxelNet: Spherical voxel-based deep place recognition for 3D point clouds of crowded indoor spaces
CN113808166B (en) Single-target tracking method based on clustering difference and depth twin convolutional neural network
Yarlagadda et al. Voting by grouping dependent parts
CN114329031B (en) Fine-granularity bird image retrieval method based on graph neural network and deep hash
Lai et al. Improving classification with semi-supervised and fine-grained learning
Prystavka et al. Automated object recognition system based on convolutional autoencoder
Barodi et al. An enhanced artificial intelligence-based approach applied to vehicular traffic signs detection and road safety enhancement
Nguyen Fast traffic sign detection approach based on lightweight network and multilayer proposal network
CN112329798B (en) Image scene classification method based on optimized visual word bag model
CN117975298B (en) Mesoscale vortex and sub-mesoscale fine structure AI unsupervised classification method
Ellis et al. Autonomous navigation and sign detector learning

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