CN-121980916-A - Ocean underwater water temperature inversion method and system based on convolution long-short-term memory network

CN121980916ACN 121980916 ACN121980916 ACN 121980916ACN-121980916-A

Abstract

The invention discloses a marine water temperature inversion method and a marine water temperature inversion system based on a convolution long-short-term memory network, wherein the method comprises the following specific steps of step 1, preprocessing satellite remote sensing data, wherein the satellite remote sensing data comprises sea surface data; step 2, clustering sea surface space in a K-means mode according to sea surface data in the step 1, step 3, extracting time features and space features of the sea data in different areas by adopting an attention convolution long-short-term memory network according to the clustering result in the step 2, and step 4, converting the time features and the space features extracted in the step 3 by adopting the attention convolution long-short-term memory network, and integrating the conversion result according to different clustering areas. The method not only remarkably improves the accuracy of water temperature inversion, but also has stronger advantages in the aspects of large-scale data processing and model training efficiency.

Inventors

WU JUNTONG
BI MEIHUA
ZHOU XUEFANG
HU MIAO
Geng Xiulin
XU MENGMENG
LI HAOZHEN
Mao Yuxiu
ZHANG XUN
Zeng Puming
WANG JINXIU

Assignees

杭州电子科技大学

Dates

Publication Date: 20260505
Application Date: 20251219

Claims (7)

1. The marine water temperature inversion method based on the convolution long-short-term memory network is characterized by comprising the following specific steps of: Step 1, preprocessing satellite remote sensing data, wherein the satellite remote sensing data comprises sea surface data; Step 2, clustering sea surface space by using a K-means mode according to the sea surface data in the step 1; step 3, extracting time features and space features of the ocean data in different areas by adopting an attention convolution long-short-term memory network according to the clustering result in the step 2; And 4, converting the time and space characteristics extracted in the step 3 by adopting an attention convolution long-short term memory network, and integrating conversion results according to different clustering areas.
2. The method for inverting the ocean water temperature based on the convolution long-short-term memory network according to claim 1, wherein in the step 1, the acquired satellite remote sensing data is subjected to standardized processing so that each characteristic is subjected to Gaussian distribution, x is set as an original characteristic value, Is the characteristic value which is subjected to the standardization process, Is the mean value of the features, Is the standard deviation of the features, and the normalization is achieved by the following formula: ; dividing the sequence into a plurality of small windows by adopting a sliding window algorithm for two dimensions of time and space; the space sliding window specifically comprises: ; Wherein, the Representing the i-th space in the longitudinal direction, Represents the j-th space in the dimension direction, The left upper corner coordinates of the space, i and j respectively represent row and column indexes of the sliding window, and s represents the sliding step length, namely the displacement interval of the adjacent window on the space coordinates; The time sliding window procedure is expressed as: ; Where T is the time sliding window size, s t is the time step, representing the displacement interval of adjacent time windows on the time axis, Is the starting point in time of the time series, N represents the total length of the original time series; The processed data dimension is expressed as: ; wherein, num is the number of samples; Timestep the length of the time sequence; Channels: number of characteristic Channels; height and Width, representing the grid size in two dimensions.
3. The ocean water temperature inversion method based on the convolution long-term memory network according to claim 2, wherein in the step 2, the K-means mode is specifically as follows: Initializing k clusters, and calculating a new centroid position by utilizing data points in each cluster group, specifically, the average value of all data point attributes in the clusters, wherein the centroid formula is as follows: ; wherein z represents a cluster Is a data point of (2); The Euclidean distance used in the calculation process of the distance has the following specific formula: ; Where d represents the dimension of the data, x i,j represents the value of data point x i in the j-th dimension, Representing cluster centers A value in the j-th dimension.
4. The ocean water temperature inversion method based on the convolution long-short-term memory network according to claim 3, wherein in the step 3, a self-attention mechanism is adopted by the attention convolution long-short-term memory network, and a ACLSTM unit is arranged; In the self-attention mechanism, input tensor data generates a query vector Q, a key vector K and a value vector V through linear transformation, and the calculation process is as follows: assuming that the input tensor is X, the linear transformation is: ; Wherein, the 、、 Respectively linear transformation matrixes, wherein the shapes of the matrixes are (C, dim), C represents the number of input channels, and dim is the characteristic dimension after mapping; calculating the weight parameter of the attention, calculating the attention score by using the dot product of Q and K, and normalizing by a softmax function to obtain the attention weight : ; Wherein S represents a matrix of feature similarity, which is normalized Determining a global relationship weight for each location; Combining the generated attention weight with the value vector V to obtain an output characteristic O of an attention mechanism: ; After the self-attention mechanism is introduced, the calculation of different gates in the convolution long-short-period memory network is changed, and the specific formula is as follows: ; ; ; ; ; ; Wherein, the Representing per-element products; the tan h is a hyperbolic tangent function; 、、 the weights of the forget gate, the input gate and the output gate are respectively; And W x and W h are the attention mapping matrices of x and h corresponding to different gates; 、、、 dynamically adjusted for bias, attention mechanism is hidden state for each time step The formula is as follows: ; 。
5. The method for inverting ocean water temperature based on the convolution long-short-term memory network according to claim 4, wherein in the step 3, the attention convolution long-short-term memory network fuses a converter and a convolution long-short-term memory network, the converter network is a residual network and is provided with a multi-head attention layer, the number of attention heads is selected to be 4, the mapping dimension dim is set to be 128, the dimension of each head is 32, the output of the multi-head attention is added with the input through residual connection and adopts a standardized layer, then the output enters a two-layer feedforward network, the first layer extracts nonlinear characteristics by using RELU activation functions, the hidden dimension of the first layer is set to be 256, the second layer also extracts nonlinear characteristics by using RELU activation functions, the output dimension of a final model is set to be 128, and the output vector serves as the input of the two-dimensional convolution network to realize inversion target tasks.
6. The ocean water temperature inversion method based on the convolution long-term memory network according to claim 5, wherein, In step 4, the output hidden state of the convolution long-short-term memory network contains space-time characteristics, the decoding stage adopts two-dimensional convolution operation to process the hidden state and map it into output space, and the hidden state at the last moment is set as The formula of the two-dimensional convolution is: ; Wherein: Is a two-dimensional convolution kernel; Is biased; for the tensor of the output, C y represents the number of output channels, Representing the spatial height of the output feature map, A spatial width representing the output feature map; measuring predicted values using a loss function And the true value y, the loss function is defined as the mean square error MSE: ; wherein N is the total number of samples; Is the true value of the ith sample; Is a predicted value of the model; by calculating the gradient of the loss function relative to the model parameters, the convolution kernel weight W and the bias b of the model are updated as follows: Let W be the convolution kernel weight, For learning rate, the update formula of gradient descent is: ; Wherein: is the gradient to the loss function; In the gradient calculation process, the convolution operation is unfolded through a chain rule: ; Wherein: Gradient to hidden state for loss; gradient of hidden state to weight W; The hidden state of the output of the attention convolution long-term and short-term memory network is changed into a matrix to be used as the input of a two-dimensional convolution network, and then the two-dimensional convolution neural network is utilized, wherein for an H multiplied by W image, the convolution calculation formula is specifically as follows: ; Wherein: Is to output the characteristic diagram on the channel k, the position Is a value of (2); is to input the characteristic diagram on the channel m, position Is a value of (2); Is the convolution kernel on channels m through k, positions Weights of (2); Kernel is the convolution kernel size, C in is the number of characteristic channels input; the convolution kernel parameter learning is realized by the following formula: ; wherein W r is the parameter matrix of the model at the r-th iteration, For the learning rate, r is the number of learning iterations.
7. Ocean water temperature inversion system based on a convolution long and short term memory network for performing the method according to any one of claims 1-6, comprising the following modules: the preprocessing module is used for preprocessing satellite remote sensing data, wherein the satellite remote sensing data comprises sea surface data; the clustering module is used for clustering the sea surface space in a K-means mode according to sea surface data; the characteristic extraction module is used for extracting time characteristics and space characteristics of the ocean data in different areas by adopting an attention convolution long-short-term memory network according to the clustering result; and the conversion and integration module is used for converting the extracted time and space characteristics by adopting an attention convolution long-short term memory network and integrating the conversion results according to different clustering areas.

Description

Ocean underwater water temperature inversion method and system based on convolution long-short-term memory network Technical Field The invention belongs to the technical crossing field of ocean observation and deep learning, and particularly relates to an underwater water temperature inversion method and system based on a convolution long-short-term memory network, which are used for processing fusion analysis and inversion calculation of satellite observation data and ocean site data. Background With the development of science and technology and the deep research of ocean, the accurate inversion of the underwater temperature has important significance for global climate research, ocean ecological protection and ocean resource development. However, conventional statistical methods and dynamic modeling methods exhibit significant shortcomings in dealing with complex nonlinear relationships and large-scale data. These methods rely on a great deal of prior knowledge and manual adjustment, and it is difficult to accurately capture the dynamic change characteristics of the ocean temperature in time and space. In recent years, inversion methods based on deep learning have received a great deal of attention. Compared with the traditional method, the deep learning method has the remarkable advantages that 1) the nonlinear modeling capability is strong, and the deep learning method can effectively process complex nonlinear relations hidden in data, and particularly shows prominence in the multivariable coupling problem. 2) The deep learning method can automatically extract the space and time characteristics of high-dimension data, and reduce the dependence on manual intervention and expert domain knowledge. 3) The neural network framework established by deep learning can process large-scale and multidimensional observation data and support rapid parallel calculation. For example, chinese patent publication No. CN 120369151A discloses a method for inverting the internal temperature of sea by fusing PIES with sea surface multisource parameters, and the technical scheme uses deep learning and machine learning methods to obtain corresponding technical effects. But it does not consider both temporal and spatial features, which presents an inaccuracy problem in inversion. Disclosure of Invention In the prior art, the ocean water temperature inversion task is independently divided into different time tasks, and the space correlation and the time correlation are not concerned at the same time, so that the problem of low inversion accuracy is caused. In the invention, when the convolution long-period memory network architecture carries out space convolution, the time characteristics of sea surface remote sensing observation are focused at the same time, so that the time characteristics can be transmitted and extracted through the long-period memory network. The task of water temperature inversion is completed, a powerful feature extraction network is needed, in the traditional water temperature inversion network design, a commonly used Convolutional Neural Network (CNN) is used for realizing feature extraction of images through a convolutional kernel, and the features of time and space are simultaneously used as input features, so that the capability of extracting important information of the images is seldom brought into the attention of the field. Meanwhile, in order to better extract the space information, the convolution process in the convolution long-short-term memory network is improved, the convolution process is combined with a converter, and the attention is paid to the convolution long-short-term memory network (ACLSTM). The invention combines a convolution long-short-term memory network and an attention mechanism technology, and realizes the inversion of the water temperature under the sea by utilizing a spatial clustering mode. The invention aims at realizing the following technical scheme: The marine water temperature inversion method based on the convolution long-short-term memory network uses the sea surface remote sensing observation data through training to invert the network structure of the water temperature distribution under the sea surface, and comprises the following specific steps: Step 1, preprocessing satellite remote sensing data, wherein the satellite remote sensing data comprises sea surface data; Step 2, clustering sea surface space by using a K-means mode according to the sea surface data in the step 1; Step 3, extracting time features and space features of ocean data in different areas by ACLSTM according to the clustering result in the step 2; and 4, converting the time and space characteristics extracted in the step 3 by ACLSTM, and integrating conversion results according to different clustering areas. Preferably, in step 1, sea surface data observed by the satellite is collected, including parameters such as sea surface height, temperature, salinity, wind speed and the like. The marine da