Sea surface temperature data reconstruction method based on remote sensing image

Abstract

The invention discloses a sea surface temperature data reconstruction method based on a remote sensing image, and the method comprises the steps: obtaining a remote sensing image for preprocessing, and obtaining a sea surface temperature image of a target region; preprocessing sea surface temperature data in the image to obtain input variables of the sea surface temperature data, and dividing the input variables into a training set and a test set; an Inception module is added into the DINCAE model, an I-DINCAE model is constructed, the training set is input into the I-DINCAE model for training, and a sea surface data reconstruction model is obtained; inputting the test set into the sea surface data reconstruction model to obtain a sea surface temperature error value, and obtaining reconstructed sea surface temperature data; enabling the reconstructed sea surface temperature data to correspond to coordinates in the remote sensing image, and obtaining a sea surface temperature distribution image; according to the method, the characteristics of different scales of the input data can be captured at the same time through the Inception module, the parameter quantity and the calculation complexity can be reduced, and overfitting in the model training process is prevented, so that the accuracy of the reconstructed data is improved.

Description

Technical Field

[0001] The invention relates to the technical field of ocean remote sensing, and in particular to a sea surface temperature data reconstruction method based on remote sensing images. Background Art

[0002] Sea surface temperature (SST) is one of the important indicators for studying ocean dynamics, ocean-atmosphere interactions and climate change. The spatiotemporal characteristics of SST reflect the basic ecological information of the ocean area and are closely related to many marine environmental factors such as ocean currents, salinity, and nutrient distribution, which together affect the balance and evolution of the marine ecosystem. Although the traditional SST measurement method is accurate, it is limited by the number and coverage of sampling points, and it is difficult to meet the needs of comprehensive monitoring and accurate analysis of SST data in a wide range of areas and complex marine environments. With the rapid development of satellite remote sensing technology, a large amount of SST data can be obtained. These data not only cover the global ocean area, but also have a high update frequency, and have been widely used in marine research. However, during the collection process of satellite remote sensing data, SST data are often missing due to factors such as weather conditions, satellite scanning orbit range, and satellite sensor operation failures, which limits the use of data to a certain extent. Therefore, in order to address the problem of missing SST data in satellite remote sensing data, accurate data reconstruction is carried out to obtain high-quality and full-coverage sea surface temperature data sets, which is of great significance to marine research.

[0003] Traditional data reconstruction techniques mainly use geo-statistical interpolation techniques to fill in missing pixel data. These methods predict and fill in missing pixel values ​​by analyzing the spatial and temporal information of neighboring valid pixels. This method works well when the data is less missing and evenly distributed, but the accuracy of the results will be greatly affected when the data is severely missing or unevenly distributed. This is mainly because these methods rely on valid information around the missing pixels for prediction. When the surrounding information is insufficient or abnormal, the accuracy of the results will be reduced.

[0004] Data reconstruction based on deep learning has significant advantages. It can automatically learn the intrinsic laws and feature representations of data and extract complex and abstract features, which makes the data reconstruction process more accurate and efficient. In addition, these models are able to handle large-scale data sets and reconstruct high-quality information from damaged or incomplete data when necessary. The Data Interpolation Convolutional Autoencoder (DINCAE) proposed by Barth et al. adopts the OI-based convolutional autoencoder structure and achieves accurate filling of missing data by combining the convolutional autoencoder and data interpolation ideas. Unlike traditional OI methods based on linear assumptions, DINCAE gives OI nonlinear interpretability through a series of convolution operations. It solves the limitations of DINEOF in dealing with nonlinear relationships in time and space domains. By making full use of the power of neural networks, DINCAE can effectively handle complex nonlinear relationships.

[0005] However, increasing the size of the model is a direct and effective way to improve network performance, but increasing the depth or width of the network means that the model will have more layers and neurons, and thus be able to express more complex functions and patterns. This expansion of the network size is also accompanied by a sharp increase in the amount of computation. The deepening and widening of the network not only increases the number of learning parameters, but may also cause the model to overfit during training. Summary of the invention

[0006] The present invention provides a sea surface temperature data reconstruction method based on remote sensing images, so as to overcome the technical problems of insufficient multi-scale feature extraction of sea surface images and easy overfitting when using DINCAE to reconstruct sea surface temperature data.

[0007] In order to achieve the above object, the technical solution of the present invention is:

[0008] A method for reconstructing sea surface temperature data based on remote sensing images, comprising:

[0009] S1: Acquire a remote sensing image, preprocess the remote sensing image, and obtain a sea surface temperature image of a target area;

[0010] S2: preprocessing the sea surface temperature data in the sea surface temperature image of the target area, obtaining input variables of the sea surface temperature data, and dividing the input variables of the sea surface temperature data into a training set and a test set; the input variables include but are not limited to the abnormal value of the sea surface temperature data of a certain day, the inverse of the error variance of the sea surface temperature data of the same day and the two days before and after, the corresponding sea surface temperature abnormal value scaled by the inverse of the error variance, and the scaled value of the time data;

[0011] S3: introducing the DINCAE model, adding the Inception module to the DINCAE model, constructing the I-DINCAE model, inputting the training set into the I-DINCAE model for training, performing multi-scale feature extraction on the training set through the Inception module, and obtaining a sea surface data reconstruction model;

[0012] S4: inputting the test set into the sea surface data reconstruction model to obtain a sea surface temperature error value, and obtaining reconstructed sea surface temperature data according to the sea surface temperature error value;

[0013] S5: Matching the reconstructed sea surface temperature data with the coordinates in the remote sensing image to obtain a sea surface temperature distribution image.

[0014] Furthermore, the DINCAE model is introduced, and the Inception module is added to the DINCAE model to construct the I-DINCAE model, including:

[0015] The DINCAE model includes an input layer, a coding layer module, a fully connected layer module, a decoding layer module and an output layer connected in sequence;

[0016] The coding layer module comprises four coding layers connected in sequence, the fully connected layer module comprises two fully connected layers connected in sequence, and the decoding layer module comprises four decoding layers connected in sequence;

[0017] The first coding layer is skip-connected with the fourth decoding layer, the second coding layer is skip-connected with the third decoding layer, and the third coding layer is skip-connected with the second decoding layer;

[0018] An Inception module is added to each coding layer in the coding layer module to form an improved coding layer module, and the coding layer module is replaced to form an I-DINCAE model.

[0019] Furthermore, each encoding layer in the improved encoding layer module includes an Inception module, a convolutional layer, and a pooling layer connected in sequence;

[0020] The Inception module is used to extract multi-scale features from the feature tensor of the previous layer input;

[0021] The convolution layer is used to perform nonlinear combination on the multi-scale features output by the Inception module to obtain fused image features;

[0022] The pooling layer is used to compress the fused image features to obtain a compressed feature tensor.

[0023] Furthermore, the training set is input into the I-DINCAE model for training to obtain a sea surface data reconstruction model, including:

[0024] The input layer is used to randomly add Gaussian noise to the data of the input training set to obtain the first feature tensor of the input variable containing the sea surface temperature data;

[0025] The improved coding layer module is used to extract and compress the multi-scale features of the input image tensor step by step, that is, the first feature tensor is input to the first coding layer for processing, the multi-scale features of the first feature tensor are extracted, nonlinear combination is performed, and the feature map size of the multi-scale features is compressed, the output data is input to the second coding layer, after the second coding layer is processed, the output data is input to the third coding layer for processing, and the output data is input to the fourth coding layer to obtain a compressed second feature tensor;

[0026] The fully connected layer module is used to convert the second feature tensor output by the improved encoding layer module into a one-dimensional vector, that is, the second feature tensor is input into the fully connected layer module to obtain a third feature tensor after dimensionality reduction;

[0027] The decoding layer module and the first three layers of the improved encoding layer module transfer features through jump connections, which are used to splice the compressed multi-scale features output by the encoding layer with the feature map channels of the corresponding level of the decoding layer to restore detail information, that is, the third feature tensor after dimensionality reduction is input to the first decoding layer, and the result obtained is input to the second decoding layer, the second decoding layer processes the data output by the first decoding layer and the data output by the third encoding layer, and the result obtained is input to the third decoding layer, the third decoding layer processes the data output by the second decoding layer and the data output by the second encoding layer, and the result obtained is input to the fourth decoding layer, the fourth decoding layer processes the data output by the first encoding layer and the data output by the third decoding layer, and the spatial dimension is restored to obtain a fourth feature tensor;

[0028] The output layer introduces a residual connection, adds the fourth feature tensor data output by the first decoding layer, the second decoding layer, the third decoding layer and the fourth decoding layer, and outputs a two-dimensional feature tensor.

[0029] Further, the test set is input into the reconstruction model to obtain a sea surface temperature error value, and reconstructed sea surface temperature data is obtained according to the sea surface temperature error value, including:

[0030] S41, inputting the test set into the reconstruction model to obtain a two-dimensional feature tensor, namely, a sea surface temperature error value, wherein the sea surface temperature error value includes a reciprocal scaling of an expected error variance of the sea surface temperature and a logarithm of the reciprocal of the expected error variance;

[0031] S42, calculating the corresponding error variance and reconstructing the sea surface temperature value according to the inverse scaling of the expected error variance of the sea surface temperature and the logarithm of the inverse of the expected error variance, as shown in formulas (1) and (2),

[0032]

[0033] in, is the error variance, To reconstruct the sea surface temperature, Y ij1 is the sea surface temperature scaled by the inverse of the expected error variance, γ is the logarithm of the inverse of the expected error variance, and δ is a constant; i and j represent the longitude and latitude in the target sea area; max and min represent the functions of taking the maximum and minimum values ​​respectively, and m is the average value of the test set; Y ij2 is the logarithm of the inverse of the expected error variance.

[0034] Furthermore, the loss function during model training is shown in formula (3):

[0035]

[0036] Among them, y ij is the original sea surface temperature value, To reconstruct the sea surface temperature value, is the standard deviation of the reconstructed SST, N is the total number of missing SST data points; is the standardized squared residual scaled by the error standard deviation, To reduce the variance of the error standard deviation, is a constant term used for normalization.

[0037] Furthermore, the sea surface temperature data in the sea surface temperature image of the target area is preprocessed to obtain input variables of the sea surface temperature data, including:

[0038] S21, subtracting the time average of the sea surface temperature from the sea surface temperature data in the sea surface temperature image of the target area to obtain an abnormal value of the sea surface temperature data;

[0039] S22, normalizing the sea surface temperature data in the sea surface temperature image of the target area, as shown in formula (4),

[0040]

[0041] Among them, x represents the original sea surface temperature data, x′ represents the normalized data, and x min and x max Respectively represent the minimum and maximum values ​​in the sea surface temperature data;

[0042] S23, randomly selecting a day, calculating the inverse of the error variance of the sea surface temperature data of the day and the two days before and after, and the sea surface temperature anomaly scaled by the inverse of the error variance;

[0043] S24, scaling the time data by sine and cosine transforms, as shown in formulas (5) and (6),

[0044]

[0045] time sin Indicates that time data is scaled by sine transform, time cos It indicates that the time data is scaled by cosine transformation; the time data is the collection date corresponding to the sea surface temperature value.

[0046] Beneficial effects: The present invention provides a sea surface temperature data reconstruction method based on remote sensing images. By adding the Inception module to the existing DINCAE model, the Inception module can simultaneously capture the features of different scales of the input data, reduce the number of parameters and computational complexity, and prevent overfitting during model training, so as to improve the accuracy of the reconstructed data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0048] Figure 1 A method flow chart of a method for reconstructing sea surface temperature data based on remote sensing images provided by the present invention;

[0049] Figure 2 A schematic diagram of the I-DINCAE model provided by the present invention;

[0050] Figure 3 The structure diagram of the improved coding layer of the present invention;

[0051] Figure 4 This is the structural diagram of the Inception module. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] This embodiment provides a method for reconstructing sea surface temperature data based on remote sensing images. Figure 1 As shown, including:

[0054] S1: Acquire a remote sensing image, preprocess the remote sensing image, and obtain a sea surface temperature image of a target area;

[0055] S2: normalizing the sea surface temperature data in the sea surface temperature image of the target area, obtaining input variables of the sea surface temperature data, and dividing the input variables of the sea surface temperature data into a training set and a test set; the input variables include but are not limited to abnormal values ​​of the sea surface temperature data of a certain day, and the inverse of the error variance of the sea surface temperature data of the same day and the two days before and after;

[0056] S3: introducing the DINCAE model, adding the Inception module to the DINCAE model, constructing the I-DINCAE model, inputting the training set into the I-DINCAE model for training, performing multi-scale feature extraction on the training set through the Inception module, and obtaining a sea surface data reconstruction model;

[0057] S4: inputting the test set into the sea surface data reconstruction model to obtain a sea surface temperature error value, and obtaining reconstructed sea surface temperature data according to the sea surface temperature error value;

[0058] S5: Matching the reconstructed sea surface temperature data with the coordinates in the remote sensing image to obtain a sea surface temperature distribution image.

[0059] Specifically, firstly, a remote sensing image is acquired, and the remote sensing image is preprocessed to obtain a sea surface temperature image of a target area. The image is preprocessed to identify and mark pixels with missing data in the image, that is, the location of the missing image, and an image of the target area, that is, the area where the sea surface data needs to be reconstructed, is acquired. A remote sensing image usually cannot directly represent a specific location and a corresponding temperature value, so it is necessary to preprocess the remote sensing image, correspond to the geographic coordinates, perform geographic registration, map the image pixels to a specific geographic location, and then convert the remote sensing data into a temperature value to reconstruct the data;

[0060] Secondly, the sea surface temperature data in the sea surface temperature image of the target area is preprocessed to obtain the input variables of the sea surface temperature data, and the input variables of the sea surface temperature data are divided into a training set and a test set. The data is preprocessed to simulate the correlation between the spatial data and the sea surface temperature data, the sea surface data is scaled, the abnormal values ​​of the sea surface data are obtained, and the scaling value of the acquisition time is obtained, and the time data is mapped to a periodic format to reflect the seasonal change characteristics. At this time, the input variable multi-channel spatiotemporal feature grid has a spatial dimension consistent with the original remote sensing image, and each pixel contains 10 normalized parameters, namely, input variables. The data structure is input into the model in the form of a tensor for subsequent processing;

[0061] Again, the DINCAE model is introduced, and the Inception module is added to the DINCAE model to construct an I-DINCAE model. The training set is input into the I-DINCAE model for training. The multi-scale feature extraction of the training set is performed through the Inception module to obtain a sea surface data reconstruction model. The Inception module is added to the DINCAE model, and the features of different scales of the input data can be captured through the Inception module at the same time, and the number of parameters and the computational complexity can be reduced to prevent overfitting during model training, so as to improve the accuracy of the reconstructed data; the test set is input into the sea surface data reconstruction model to obtain a sea surface temperature error value, and the reconstructed sea surface temperature data is obtained according to the sea surface temperature error value. The inverse scaling of the expected error variance of the sea surface temperature and the logarithm of the inverse of the expected error variance are obtained through the model, and the actual reconstructed sea surface temperature value is calculated through these two data;

[0062] Finally, the reconstructed sea surface temperature data is matched with the coordinates in the remote sensing image to obtain a sea surface temperature distribution image. The reconstructed data is a two-dimensional SST numerical matrix. Each pixel value is a temperature value (unit: ℃). The spatial arrangement corresponds to the geographic coordinates of the original remote sensing image. The temperature matrix is ​​converted into a visual image through color scale mapping (such as pseudo color) to obtain a sea surface temperature distribution image.

[0063] In a specific embodiment, a scheme for acquiring a remote sensing image and preprocessing the remote sensing image to obtain a sea surface temperature image of a target area is:

[0064] In this embodiment, Python is used to perform projection conversion, image cropping and other remote sensing image data preprocessing on the downloaded remote sensing images, identify and mark missing pixels, obtain remote sensing images of the target area, and construct a complete time series data set:

[0065] S11, projection conversion: remap the pixels on the original image to the pixels on the target coordinate system. In this embodiment, the data is set to WGS 84 (EPSG:4326) projection. WGS 84 is a geographic coordinate system that uses latitude and longitude to describe the location of data.

[0066] S12, Image cropping: obtain a smaller remote sensing image of the study area from the original large-scale full-scene image;

[0067] S13, numerical conversion: convert the original numerical values ​​of the image data points into temperature values, and mark the missing data as NAN.

[0068] In this embodiment, the image is reprojected, an image is converted from one geographic coordinate system to another geographic coordinate system, and the pixel points of the image are described by longitude and latitude to facilitate subsequent data processing;

[0069] By cropping the image, disk storage space can be saved, data processing time can be reduced, and a remote sensing image of a smaller area of ​​the study area can be obtained from the original large-scale full-scene image;

[0070] By performing numerical conversion on the image, the raw numerical values ​​are converted into temperature values. This conversion is usually because remote sensing data are often stored in raw digital form (such as satellite sensor readings), and these numbers themselves have no direct physical meaning until they are converted into actual physical quantities, such as temperature, to participate in subsequent data processing.

[0071] In a specific embodiment, the sea surface temperature data in the sea surface temperature image of the target area is normalized to obtain the input variables of the sea surface temperature data, and the scheme of dividing the input variables of the sea surface temperature data into a training set and a test set is:

[0072] S21, using the sea surface temperature data in the sea surface temperature image of the target area minus the time average value of the sea surface temperature (the average value of the sea surface temperature in a specific time period), to obtain an abnormal value of the sea surface temperature data;

[0073] S22. In order to better simulate the correlation between spatial data and SST data, the sea surface temperature data in the sea surface temperature image of the target area is normalized, that is, the longitude and latitude are linearly scaled between -1 and 1, as shown in formula (7),

[0074]

[0075] Among them, x represents the original sea surface temperature data (expressed in longitude and latitude), x′ represents the normalized data, and x min and x maxRespectively represent the minimum and maximum values ​​in the sea surface temperature data;

[0076] S23, randomly select a day, calculate the inverse of the error variance of the sea surface temperature data of that day and the two days before and after (if the data is missing, it is set to zero) and the sea surface temperature anomaly scaled by the inverse of the error variance;

[0077] S24. Model the correlation between time data and sea surface temperature data, that is, the time data is scaled by sine and cosine transformations, as shown in formulas (8) and (9),

[0078]

[0079] time sin Indicates that time data is scaled by sine transform, time cos Indicates that the time data is scaled by cosine transformation; the time data is the acquisition date corresponding to the sea surface temperature value;

[0080] The input variables of sea surface temperature data are shown in Table 1.

[0081] Table 1 Input variables of I-DINCAE model

[0082]

[0083] The dimension of the final input data tensor is 10, which contains the above 10 types of data.

[0084] In this embodiment, during the data preprocessing process, the time average value of SST is subtracted to obtain the mean normalization result, thereby ensuring the consistency of the data. For the missing values ​​in the data set, they are set to zero to maintain the integrity of the data structure. The time data (the day of the year) is scaled by sine and cosine transformations. This transformation is to map the time data to a periodic format and generate periodic coding features to reflect the seasonal variation characteristics. As a model input parameter, this encoding enables the model to explicitly capture the annual cycle variation law of SST and enhance the correlation modeling capability of the time dimension.

[0085] In a specific embodiment, the DINCAE model is introduced, the Inception module is added to the DINCAE model, the I-DINCAE model is constructed, the training set is input into the I-DINCAE model for training, and the scheme for obtaining the sea surface data reconstruction model is:

[0086] like Figure 2 As shown, the DINCAE model includes an input layer, a coding layer module, a fully connected layer module, a decoding layer module and an output layer connected in sequence;

[0087] The encoding layer module includes four encoding layers, the fully connected layer module includes two fully connected layers, and the decoding layer module includes four decoding layers;

[0088] The first coding layer is skip-connected with the fourth decoding layer, the second coding layer is skip-connected with the third decoding layer, and the third coding layer is skip-connected with the second decoding layer;

[0089] Adding an Inception module to each coding layer in the coding layer module to form an improved coding layer module, and replacing the coding layer module to form an I-DINCAE model;

[0090] like Figure 3 As shown, each encoding layer in the improved encoding layer module includes an Inception module, a convolutional layer, and a pooling layer connected in sequence;

[0091] The training set is input into the I-DINCAE model for training. The structure and function of each layer are as follows:

[0092] The input layer is used to randomly add Gaussian noise to the data of the input training set to obtain the first feature tensor of the input variable containing the sea surface temperature data;

[0093] The training set is represented by a multidimensional tensor (shape: batch_size×300×300×10), which contains information such as SST anomalies, error variance, and spatiotemporal parameters:

[0094] The input layer receives preprocessed sea surface temperature (SST) data with a dimension of 10, including SST anomalies scaled by the inverse of the error variance, the inverse of the error variance, the scaled SST anomalies and the inverse of the error variance of the day before and after, normalized longitude and latitude (range [-1,1]), and sine / cosine transforms of time information (cos(day / 365.25) and sin(day / 365.25)). The input shape is (batch_size,height,width,10). The input layer randomly adds Gaussian noise (mean 0, standard deviation 0.05) to the data to enhance robustness, and splits the data into small batches (batch size = 50) with 50 images in each small batch;

[0095] The improved coding layer module is used to extract and compress the multi-scale features of the input image tensor step by step, the first feature tensor is input to the first coding layer for processing, the multi-scale features of the first feature tensor are extracted, nonlinear combination is performed, and the feature map size of the multi-scale features is compressed, the output data is input to the second coding layer, after the second coding layer is processed, the output data is input to the third coding layer for processing, and the output data is input to the fourth coding layer to obtain the compressed second feature tensor:

[0096] The Inception module is used to extract multi-scale features from the feature tensor of the previous layer input. The structure is as follows Figure 4 As shown, the module contains four parallel branches:

[0097] Branch 1: 1×1 convolution, output 12 channels;

[0098] Branch 2: 3×3 convolution (padding=1), output 12 channels;

[0099] Branch 3: 5×5 convolution (padding=2), output 12 channels;

[0100] Branch 4: 3×3 max pooling (step size = 1, padding = 1) → 1×1 convolution, output 12 channels;

[0101] The four-branch output channels are concatenated and reduced to 24 channels by 1×1 convolution, resulting in a feature tensor of size (300, 300, 24);

[0102] The convolution layer is used to perform nonlinear combination on the multi-scale features output by the Inception module to obtain fused image features. In this embodiment, a 3×3 convolution kernel is used to extract spatial features, and the activation function is a Leaky ReLU function (α=0.2). The output shape keeps the spatial size unchanged, which is a feature tensor of (300, 300, 24), and the number of channels increases.

[0103] The pooling layer is used to compress the fused image features to obtain a compressed feature tensor. In this embodiment, the pooling layer uses 2×2 maximum pooling with a step size of 2 to compress the feature map size. The output shape is a half of the spatial size (such as 300×300→150×150), and a feature tensor of size (150, 150, 24) is obtained. The maximum pooling layer takes the maximum value within the window and compresses the local area of ​​the input feature map into a single value, thereby reducing the spatial size of the feature map.

[0104] The feature tensor of (150, 150, 24) is input into the second encoding layer to obtain a feature tensor of size (75, 75, 48). The feature tensor of (75, 75, 48) is input into the third convolutional layer to obtain a feature tensor of size (37, 37, 96). The feature tensor of (37, 37, 96) is input into the fourth encoding layer to obtain a feature tensor of size (18, 18, 192).

[0105] The fully connected layer module is used to convert the second feature tensor output by the improved encoding layer module into a one-dimensional vector, that is, input the second feature tensor into the fully connected layer module to obtain a third feature tensor after dimensionality reduction; the fully connected layer module includes a fully connected layer and a Dropout layer connected in sequence;

[0106] The fully connected layer is used to combine all the extracted multi-scale feature tensors in a nonlinear manner, nonlinearly combine the flattened multi-scale features, and output a one-dimensional vector (low-dimensional representation) with a size of (50, N); in this embodiment, the number of neurons in the first fully connected layer is N / 5 (N is the number of neurons in the last pooling layer of the encoder), the number of neurons in the second layer is N, and the activation function is Leaky ReLU (α=0.2);

[0107] The Dropout layer is used to train the one-dimensional vector (dropout rate = 0.3) to prevent the model from overfitting;

[0108] The decoding layer module and the first three layers of the encoding layer module transfer features through jump connections, which are used to splice the compressed multi-scale features output by the encoding layer with the feature map channels of the corresponding level of the decoding layer to restore detail information, that is, the third feature tensor after dimensionality reduction is input to the first decoding layer, and the result obtained is input to the second decoding layer. The second decoding layer processes the data output by the first decoding layer and the data output by the third encoding layer, and the result obtained is input to the third decoding layer. The third decoding layer processes the data output by the second decoding layer and the data output by the second encoding layer, and the result obtained is input to the fourth decoding layer. The fourth decoding layer processes the data output by the first encoding layer and the data output by the third decoding layer, restores the spatial dimension, and obtains the fourth feature tensor; the decoding layer includes an upsampling layer and a convolution layer connected in sequence;

[0109] In this embodiment, the upsampling layer uses the nearest neighbor interpolation method to restore the spatial dimension, and the spatial size of the output shape is doubled (e.g., 150×150→300×300), obtaining a feature tensor of size (150, 150, 54); the convolution layer uses a 3×3 convolution kernel to refine the features, and the activation function is Leaky ReLU;

[0110] The jump connection concatenates the output of the pooling layer corresponding to the encoding layer with the upsampled output of the decoding layer in the channel dimension to restore the detail information (such as edges and textures) lost during the encoding process.

[0111] The output layer introduces a residual connection, adds the output of each decoding layer in the decoding module to the output of the network, outputs a two-dimensional feature tensor with a size of 300×300×2, and finally outputs the inverse scaling of the error variance of the sea surface temperature and the logarithm of the inverse of the error variance:

[0112] The output layer adds the fourth feature tensor data output by the first decoding layer, the second decoding layer, the third decoding layer and the fourth decoding layer, and outputs a two-dimensional feature tensor.

[0113] In this embodiment, the Inception module is added to the encoding layer of the original DINCAE. By using convolution kernels and pooling layers of different sizes in parallel in the network, the Inception architecture successfully captures feature information of different scales in the image, further enriching the hierarchy of feature representation. Specifically, the module integrates three convolution kernels of different scales, including 1×1, 3×3, and 5×5, so that image feature information of different levels can be extracted. This design enables the feature map to cover receptive fields of different sizes, effectively balancing the width and depth of the network, thereby improving the overall performance of the network. In addition, in order to reduce the computational complexity and adjust the size of the feature map, the Inception model also introduces a 1×1 convolution kernel for dimensionality reduction. This strategy not only reduces the amount of calculation, but also helps to optimize the dimension of the feature map, making it more suitable for subsequent network processing.

[0114] The preprocessed data and images are divided in the form of a two-dimensional network. The longitude and latitude in the target sea area represent the position of each unit in the data grid. The data grid is a regular two-dimensional spatial structure. Each unit (i, j) corresponds to a fixed geographical location (longitude, latitude). The target sea area (such as a 300×300 km area) is divided into regularly arranged two-dimensional grid units. Each unit (i, j) corresponds to a fixed geographical coordinate (longitude, latitude); the valid sea surface temperature (SST) pixels in the grid that are not marked as missing (NaN) are used as data points. The total number N determines the normalized benchmark of the loss function, ensuring that the model focuses on areas with physical significance.

[0115] In a specific embodiment, the test set is input into the reconstruction model to obtain a sea surface temperature error value, and a scheme for obtaining reconstructed sea surface temperature data according to the sea surface temperature error value is:

[0116] S41, inputting the test set into the reconstruction model to obtain a sea surface temperature error value; the sea surface temperature error value includes a reciprocal scaling of an expected error variance of the sea surface temperature and a logarithm of the reciprocal of the expected error variance;

[0117] S42, calculating the corresponding error variance and reconstructing the sea surface temperature value according to the inverse scaling of the expected error variance of the sea surface temperature and the logarithm of the inverse of the expected error variance, as shown in formulas (10) and (11),

[0118]

[0119] in, is the error variance, To reconstruct the sea surface temperature value, Y ij1 is the sea surface temperature scaled by the inverse of the expected error variance, γ is the logarithm of the inverse of the expected error variance; γ = 10, δ = 10 -3 ℃ -2 , is a constant to prevent the denominator from approaching zero; i and j represent the longitude and latitude of the target sea area, max and min represent the functions of taking the maximum and minimum values ​​respectively, and m is the average value of the original data set; Y ij2 is the logarithm of the inverse of the expected error variance;

[0120] In order to achieve the best reconstruction effect, the model uses the loss function as an optimization indicator during the training process. Through repeated iterative training, the model continuously learns and fine-tunes its internal parameters to gradually reduce the value of the loss function until it is minimized. The loss function is shown in formula (12):

[0121]

[0122] Among them, y ij is the original sea surface temperature value, To reconstruct the sea surface temperature value, is the standard deviation of the reconstructed SST, N is the total number of missing SST data points;

[0123] is the standardized squared residual scaled by the error standard deviation, To reduce the variance of the error standard deviation, is a constant term used for normalization. Therefore, the parameters (i.e., weights and biases) are determined to maximize the likelihood (i.e., in the form of a Gaussian probability distribution) between the estimates and the observed values.

[0124] In a specific embodiment, the scheme for obtaining a sea surface temperature distribution image by matching the reconstructed sea surface temperature data with the coordinates in the remote sensing image is:

[0125] The reconstructed data generates a complete temperature field by replacing the missing pixels (marked as NaN) in the original image. The output is a two-dimensional floating-point matrix, which can be directly used for spatial analysis or visualization. Therefore, the reconstructed data is a two-dimensional SST numerical matrix (300×300 grid), each pixel value is a temperature value (unit: ℃), the spatial arrangement strictly corresponds to the geographic coordinates of the original remote sensing image, and the geographic coordinate system is consistent with the input remote sensing image (300×300 pixels). The temperature matrix is ​​converted into a visual image through color scale mapping (such as pseudo color):

[0126] Select color scale: Select the appropriate color scale for different temperatures according to your needs, and correspond different temperature values ​​to different colors to intuitively display the temperature distribution;

[0127] Plotting: Visualize the temperature matrix using a plotting library such as Matplotlib.

[0128] Add color bar: display the mapping relationship between temperature value and color;

[0129] Adjust the display: set the title, axis labels, etc. to make the image clearer.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing sea surface temperature data based on remote sensing images, characterized in that: include: S1: Acquire a remote sensing image, preprocess the remote sensing image, and obtain a sea surface temperature image of a target area; S2: preprocessing the sea surface temperature data in the sea surface temperature image of the target area, obtaining input variables of the sea surface temperature data, and dividing the input variables of the sea surface temperature data into a training set and a test set; the input variables include but are not limited to abnormal values ​​of the sea surface temperature data of a certain day, and the inverse of the error variance of the sea surface temperature data of the same day and the two days before and after; S3: introducing the DINCAE model, adding the Inception module to the DINCAE model, constructing the I-DINCAE model, inputting the training set into the I-DINCAE model for training, performing multi-scale feature extraction on the training set through the Inception module, and obtaining a sea surface data reconstruction model; S4: inputting the test set into the sea surface data reconstruction model to obtain a sea surface temperature error value, and obtaining reconstructed sea surface temperature data according to the sea surface temperature error value; S5: Matching the reconstructed sea surface temperature data with the coordinates in the remote sensing image to obtain a sea surface temperature distribution image.

2. The method for reconstructing sea surface temperature data based on remote sensing images according to claim 1, characterized in that: The DINCAE model is introduced, and the Inception module is added to the DINCAE model to construct the I-DINCAE model, including: The DINCAE model includes an input layer, a coding layer module, a fully connected layer module, a decoding layer module and an output layer connected in sequence; The coding layer module comprises four coding layers connected in sequence, the fully connected layer module comprises two fully connected layers connected in sequence, and the decoding layer module comprises four decoding layers connected in sequence; The first coding layer is skip-connected with the fourth decoding layer, the second coding layer is skip-connected with the third decoding layer, and the third coding layer is skip-connected with the second decoding layer; An Inception module is added to each coding layer in the coding layer module to form an improved coding layer module, and the coding layer module is replaced to form an I-DINCAE model.

3. The method for reconstructing sea surface temperature data based on remote sensing images according to claim 2, characterized in that: Each encoding layer in the improved encoding layer module includes an Inception module, a convolutional layer, and a pooling layer connected in sequence; The Inception module is used to extract multi-scale features from the feature tensor of the previous layer input; The convolution layer is used to perform nonlinear combination on the multi-scale features output by the Inception module to obtain fused image features; The pooling layer is used to compress the fused image features to obtain a compressed feature tensor.

4. The method for reconstructing sea surface temperature data based on remote sensing images according to claim 2, characterized in that: The training set is input into the I-DINCAE model for training to obtain a sea surface data reconstruction model, including: The input layer is used to randomly add Gaussian noise to the data of the input training set to obtain the first feature tensor of the input variable containing the sea surface temperature data; The improved coding layer module is used to extract and compress the multi-scale features of the input image tensor step by step, that is, the first feature tensor is input to the first coding layer for processing, the multi-scale features of the first feature tensor are extracted, nonlinear combination is performed, and the feature map size of the multi-scale features is compressed, the output data is input to the second coding layer, after the second coding layer is processed, the output data is input to the third coding layer for processing, and the output data is input to the fourth coding layer to obtain a compressed second feature tensor; The fully connected layer module is used to convert the second feature tensor output by the improved encoding layer module into a one-dimensional vector, that is, the second feature tensor is input into the fully connected layer module to obtain a third feature tensor after dimensionality reduction; The decoding layer module and the first three layers of the improved encoding layer module transfer features through jump connections, which are used to splice the compressed multi-scale features output by the encoding layer with the feature map channels of the corresponding level of the decoding layer to restore detail information, that is, the third feature tensor after dimensionality reduction is input to the first decoding layer, and the result obtained is input to the second decoding layer, the second decoding layer processes the data output by the first decoding layer and the data output by the third encoding layer, and the result obtained is input to the third decoding layer, the third decoding layer processes the data output by the second decoding layer and the data output by the second encoding layer, and the result obtained is input to the fourth decoding layer, the fourth decoding layer processes the data output by the first encoding layer and the data output by the third decoding layer, and the spatial dimension is restored to obtain a fourth feature tensor; The output layer introduces a residual connection, adds the fourth feature tensor data output by the first decoding layer, the second decoding layer, the third decoding layer and the fourth decoding layer, and outputs a two-dimensional feature tensor.

5. The method for reconstructing sea surface temperature data based on remote sensing images according to claim 1, characterized in that: Inputting the test set into the reconstruction model to obtain a sea surface temperature error value, and obtaining reconstructed sea surface temperature data according to the sea surface temperature error value, including: S41, inputting the test set into the reconstruction model to obtain a two-dimensional feature tensor, namely, a sea surface temperature error value, wherein the sea surface temperature error value includes a reciprocal scaling of an expected error variance of the sea surface temperature and a logarithm of the reciprocal of the expected error variance; S42, calculating the corresponding error variance and reconstructing the sea surface temperature value according to the inverse scaling of the expected error variance of the sea surface temperature and the logarithm of the inverse of the expected error variance, as shown in formulas (1) and (2), in, is the error variance, To reconstruct the sea surface temperature value, Y ij1 is the sea surface temperature scaled by the inverse of the expected error variance, γ is the logarithm of the inverse of the expected error variance, and δ is a constant; i and j represent the longitude and latitude in the target sea area; max and min represent the functions of taking the maximum and minimum values ​​respectively, and m is the average value of the test set; Y ij2 is the logarithm of the inverse of the expected error variance.

6. The method for reconstructing sea surface temperature data based on remote sensing images according to claim 5, characterized in that: The loss function during model training is shown in formula (3): Among them, y ij is the original sea surface temperature value, To reconstruct the sea surface temperature value, is the standard deviation of the reconstructed SST, N is the total number of missing SST data points; is the standardized squared residual scaled by the error standard deviation, To reduce the variance of the error standard deviation, is a constant term used for normalization.

7. The method for reconstructing sea surface temperature data based on remote sensing images according to claim 1, characterized in that: Preprocess the sea surface temperature data in the sea surface temperature image of the target area to obtain the input variables of the sea surface temperature data, including: S21, subtracting the time average of the sea surface temperature from the sea surface temperature data in the sea surface temperature image of the target area to obtain an abnormal value of the sea surface temperature data; S22, normalizing the sea surface temperature data in the sea surface temperature image of the target area, as shown in formula (4), Among them, x represents the original sea surface temperature data, x′ represents the normalized data, and x min and x max Respectively represent the minimum and maximum values ​​in the sea surface temperature data; S23, randomly selecting a day, calculating the inverse of the error variance of the sea surface temperature data of the day and the two days before and after, and the sea surface temperature anomaly scaled by the inverse of the error variance; S24, scaling the time data by sine and cosine transforms, as shown in formulas (5) and (6), time sin Indicates that time data is scaled by sine transform, time cos It indicates that the time data is scaled by cosine transformation; the time data is the collection date corresponding to the sea surface temperature value.

Images