Sea ice roughness retrieval method based on ensemble learning method and two-dimensional wavelet transform

Abstract

The application provides a sea ice roughness inversion method based on an ensemble learning method and a two-dimensional wavelet transform, first extracts the spatial features of sea ice from a SAR image containing sea ice in a target period in a target region through two-dimensional continuous wavelet transform, then obtains the actual roughness of sea ice from all footprint point data in the target period in the target region, and matches the original information of the SAR image in the target period in the target region, the spatial features of sea ice and the corresponding actual roughness of sea ice in space to construct a total data set; finally, the original information of the SAR image and the spatial features of sea ice are taken as input, and the corresponding predicted roughness of sea ice is taken as output to train an ensemble learning model to obtain a sea ice roughness inversion model based on an ensemble learning method. The application can effectively solve the problems of large sea ice roughness inversion error and low resolution in the prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of marine remote sensing, and particularly relates to a sea ice roughness inversion method based on an integrated learning method and a two-dimensional wavelet transform. BACKGROUND

[0002] Sea ice roughness is an important factor affecting sea-air energy exchange and plays an important role in the Arctic climate system. Therefore, it is very important to invent an accurate high-resolution sea ice roughness inversion method. Although traditional field observation and airborne LiDAR observation can accurately obtain the elevation information of sea ice and thus calculate the relatively accurate sea ice roughness, the two methods have small spatial coverage and are scattered in distribution; they are restricted by the harsh weather in the polar region and cannot achieve all-time and all-weather observation; and both of them are very labor and material intensive. Satellite remote sensing (such as synthetic aperture radar) can obtain large-area and high-temporal and spatial resolution sea ice backscattering images. However, it is difficult to invert the sea ice roughness by a trivial linear regression model only from the backscattering coefficient. Therefore, the application combines SAR images (Sentinel-1 data) and airborne radar elevation data (OIB ATM L2 data) to invert the sea ice roughness.

[0003] There are mainly two ways to invert the sea ice roughness. The first way is to use the multi-angle reflection information obtained by different cameras in the multi-angle imaging spectrometer (MISR) [2] to calculate the normalized difference angle index (NDAI), and then use a certain empirical algorithm to establish the relationship between NDAI and sea ice roughness. Common empirical algorithms include K-nearest neighbor (KNN) regression algorithm, support vector machine (SVM) algorithm, etc. Although this method can obtain high-resolution (275m x 275m) sea ice roughness and is relatively simple to operate, the error of the inverted roughness is very large. Moreover, this method needs to use an entire ATM product to calculate the roughness of the entire product, and this calculation method often ignores the small-scale roughness changes within the ATM product.

[0004] The second way is to use the backscattering coefficient data of each pixel point of the SAR image to invert the sea ice roughness through the theoretical backscattering model of sea ice (such as IEM model) [3]. Compared with the first method, this method can obtain higher-resolution sea ice roughness (depending on the resolution of the SAR image pixel) and the backscattering coefficient calculated from the inverted roughness is closer to the backscattering coefficient of the SAR image itself. However, this method is very difficult to implement and has high program time complexity, and it is difficult to apply to large-area inversion.

[0005] Both methods are based on the statistical relationship between the roughness and the parameters directly measured by different sensors. However, the spatial distribution information of sea ice features is not considered in the construction of the model, and the spatial information of sea ice features can be obtained from high-resolution SAR images. Therefore, it has become a technical problem to be solved in the technical field to invent a simple and easy method for retrieving the roughness of Arctic sea ice to solve the problem of low accuracy and low resolution of the roughness retrieval in the prior art.

[0006] Document [1] A. Nolin and E. Mar, "Arctic sea ice surface roughness estimated from multi-angular reflectance satellite imagery," Remote Sensing, vol. 11, p. 50, 122018.

[0007] Document [2] E. Mosadegh and A. Nolin, "A new data processing system for generating sea ice surface roughness products from the multi-angle imaging spectroradiometer (MISR) imagery," Remote Sensing, vol. 14, p. 4979, 102022.

[0008] Document [3] X. Wen, C. Xue, and Q. Dong, "The Arctic sea ice surface roughness estimation and application," Proceedings of the International Offshore and Polar Engineering Conference, pp. 958-961, 012011. SUMMARY

[0009] The purpose of the present application is to invent a simple and easy method for retrieving the roughness of sea ice to solve the problem of large retrieval error and low resolution of sea ice roughness in the prior art.

[0010] To achieve the purpose of the present application, the present application proposes a sea ice roughness retrieval method based on an integrated learning method and two-dimensional wavelet transform, mainly including the following steps:

[0011] Step 1: Obtain SAR images containing sea ice in the target period in the target area, and perform data preprocessing on each SAR image, wherein the preprocessing includes the following steps in sequence: orbit correction, thermal noise removal, radiation calibration, speckle filtering, decibel, and two-dimensional linear interpolation. is the coordinate vector of a pixel in the image;

[0012] Step 2: Obtain all sea ice footprint point data in the target period in the target area, and calculate the elevation standard deviation of the nearest preset number of footprint points for each footprint point to obtain the actual roughness of each footprint point.

[0013] Step 3: For each two-dimensional spatial SAR image , perform a complex two-dimensional continuous wavelet transform to obtain the complex wavelet coefficient of each pixel in each two-dimensional spatial SAR image . Wherein φ is the wavelet rotation angle of each pixel, a is the scale parameter of the complex wavelet transform in each pixel, is the translation parameter of the complex wavelet transform in each pixel.

[0014] In the two-dimensional complex continuous wavelet transform, the calculation formula of the complex wavelet coefficient is as shown in the following equation:

[0015]

[0016] And And r -φ The formula is as follows:

[0017]

[0018]

[0019]

[0020] Wherein, c ψ is a constant to make the equation satisfy the normalization condition, is the spatial frequency, is the conjugate of the two-dimensional Fourier transform function of the wavelet function ψ, is the two-dimensional Fourier transform of the spatial image ; The scale parameter a is taken as 1-32, and the rotation angle φ is taken as 0-2π interval .

[0021] Wherein, the wavelet mother function is a two-dimensional Cauchy wavelet function, and the result of the two-dimensional Fourier transform thereof is:

[0022]

[0023] where ω x is the frequency in x direction, ω y is the frequency in y direction, A is the scaling parameter of wavelet function, and α is the half-open angle of the convex cone in which the domain of two-dimensional Cauchy wavelet function is located.

[0024] Step 4: By traversing all scale parameters and rotation angle parameters for each pixel point in each SAR image, the scale parameter a m and the rotation angle parameter φ m corresponding to the maximum modulus of complex wavelet coefficients are obtained, and a m and φ m constitute the spatial information of sea ice, wherein |·| represents modulus;

[0025] Step 5: The SAR image original information and the sea ice spatial feature information in the target region and the target period, and the corresponding actual sea ice roughness are matched in space to construct a total data set containing the above information;

[0026] Step 6: The SAR image original information and the sea ice spatial feature information are taken as input, and the corresponding predicted sea ice roughness is taken as model output. The Adaboost regression learner is trained by using the training set to obtain the trained sea ice roughness inversion model based on Adaboost regression;

[0027] Step 7: The regression model obtained in step 6 is used to predict the test set to obtain the predicted sea ice roughness of the test set and to perform preliminary evaluation, authenticity test and in-depth evaluation of the model.

[0028] wherein the preliminary evaluation of the model is to evaluate the model by using the actual roughness and the predicted roughness of the test set to calculate model evaluation indexes, and the calculation formulas of the indexes are as follows:

[0029]

[0030]

[0031]

[0032]

[0033] wherein R 2 is the determination coefficient, MAE is the mean absolute error, RMSE is the root mean square error, MAPE is the mean absolute percentage error, y i is the true value, is the predicted value, is the average value of the true value, N is the number of samples.

[0034] Beneficial effects: compared with the prior art, the technical scheme of the present application has the following beneficial technical effects: the present application proposes a simple and easy method of using SAR images to invert sea ice roughness by combining Adaboost regression and two-dimensional wavelet transform, solving the problems of large inversion error, low resolution, long program calculation time and difficulty in implementation in the prior art. It provides a data source for the inversion of sea ice thickness and density and can also provide a certain reference for the prediction of target area climate evolution. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The flowchart of the sea ice roughness inversion method based on the integrated learning method and two-dimensional wavelet transform of the present application;

[0036] Figure 2 The sea ice density inversion result graph during the ice melting period based on the Adaboost regression model;

[0037] Figure 3 The verification result graph based on the Adaboost regression model. DETAILED DESCRIPTION

[0038] The present application will be further illustrated below in combination with the drawings and specific embodiments, and it should be understood that these examples are only used to illustrate the present application and not to limit the scope of the present application, and after reading the present application, various equivalent modifications of the present application by those skilled in the art all fall within the scope defined by the claims attached hereto.

[0039] The sea ice roughness inversion method based on the integrated learning method and two-dimensional wavelet transform described in the present application, as shown in Figure 1 Step 1: Obtain the SAR image containing sea ice in the target period in the target area, and perform data preprocessing on each SAR image, wherein the preprocessing includes the following steps in turn: orbit correction, thermal noise removal, radiation calibration, speckle filtering, decibel, two-dimensional linear interpolation. Obtain the uniform decibel backscattering coefficient image in space, which is the two-dimensional space SAR image The coordinate vector of the pixel point in the image;

[0040] Step 2: Obtain all sea ice footprint point data in the target period in the target area, calculate the elevation standard deviation of the nearest preset number of footprint points for each footprint point, and obtain the actual sea ice roughness of each footprint point;

[0041] Step 3: For each two-dimensional space SAR image Perform complex two-dimensional continuous wavelet transform to obtain each two-dimensional space SAR image Complex wavelet coefficient of each pixel where φ is the wavelet rotation angle of each pixel, a is the scale parameter of the complex wavelet transform in each pixel, is the translation parameter of the complex wavelet transform in each pixel.

[0042] In the two-dimensional complex continuous wavelet transform, the calculation formula of the complex wavelet coefficient is shown in the following equation.

[0043]

[0044] and and r -φ The formula is as follows:

[0045]

[0046]

[0047]

[0048] where c ψ is a constant to make the equation satisfy the normalization condition, is the spatial frequency, is the conjugate of the two-dimensional Fourier transform function of the wavelet function ψ, is the two-dimensional Fourier transform of the spatial image The scale parameter a is taken as 1-32, and the rotation angle φ is taken as 0-2π interval value.

[0049] where the wavelet mother function is a two-dimensional Cauchy wavelet function, and the result of its two-dimensional Fourier transform is:

[0050]

[0051] where ω x is the frequency in the x direction, ω y is the frequency in the y direction, A is the scaling parameter of the wavelet function, and α is the half-open angle of the convex cone in which the two-dimensional Cauchy wavelet function is defined.

[0052] Step 4: By traversing all scale parameters and rotation angle parameters for each pixel point in each SAR image, the scale parameter a m and the rotation angle parameter φ m corresponding to the maximum modulus of the complex wavelet coefficient are obtained, and a m and φ m constitute the spatial information of sea ice, wherein |·| represents the modulus.

[0053] Step 5: The SAR image original information and the sea ice spatial feature information in the target area and the corresponding sea ice actual roughness in the target period are matched in space to construct a total data set containing the above information;

[0054] Step 6: The SAR image original information and the sea ice spatial feature information are taken as inputs, and the corresponding sea ice predicted roughness is taken as the model output. The Adaboost regression learner is trained by using the training set to obtain the trained sea ice roughness inversion model based on Adaboost regression;

[0055] Step 7: The regression model obtained in step 6 is used to predict the test set to obtain the sea ice predicted roughness of the test set and perform preliminary evaluation, authenticity test and in-depth evaluation of the model.

[0056] The preliminary evaluation of the model is to evaluate the model by using the actual roughness and the predicted roughness of the test set to calculate the model evaluation indexes. The calculation formulas of the indexes are as follows:

[0057]

[0058]

[0059]

[0060]

[0061] wherein R 2 is the determination coefficient, MAE is the mean absolute error, RMSE is the root mean square error, MAPE is the mean absolute percentage error, y i is the true value, is the predicted value, is the average value of the true value, and N is the sample number.

[0062] Figure 2 is the sea ice roughness inversion result map of the ice melting period based on the Adaboost regression model, including the comparison chart of the SIR true value spatial distribution and the SIR predicted value spatial distribution of the Beaufort Sea mixed area of summer annual ice and multi-year ice (range: 74-78 ° N, 158-162 ° W, hereinafter referred to as the study area) on July 13 and 14. According to Figure 2 It can be seen that there is only a difference between the SIR true value and the SIR predicted value at individual sample points, so it can be preliminarily considered that the model is effective;

[0063] Figure 3 is the verification result based on the Adaboost regression model, which shows the performance of the test set before and after the spatial information obtained by two-dimensional wavelet analysis is added. According to Figure 3It can be seen that the effect of inversion after adding spatial information is significantly better than the result before adding spatial information, and the error indicators (MAE, RMSE and MAPE) are significantly reduced after adding spatial information, and the performance indicator R 2 is significantly improved, reaching 0.91 after adding spatial information. This is sufficient to show that the inversion method proposed in the present application has very high precision.

[0064] The above merely provides a specific implementation of the present application, which aims to enable those skilled in the art to understand the present application and implement it, and cannot limit the protection scope of the present application. Any equivalent transformation or replacement according to the essence of the present application should be covered within the protection scope of the present application.

Claims

1. A sea ice roughness retrieval method based on ensemble learning method and two-dimensional wavelet transform, characterized in that, Includes the following steps: Step 1: Obtain SAR images containing sea ice in the target region and target period, and pre-process each SAR image to obtain a spatially uniform decibel backscattering coefficient image, i.e., a two-dimensional spatial SAR image is a coordinate vector of a pixel point in the image. Step 2: Obtain all sea ice footprint data within the target area and during the target time period. Calculate the elevation standard deviation of each footprint point from the nearest preset number of footprint points to obtain the actual sea ice roughness of each footprint point. Step 3: For each two-dimensional spatial SAR image a complex two-dimensional continuous wavelet transform is performed to obtain a complex wavelet coefficient of each pixel point in each two-dimensional spatial SAR image a complex wavelet coefficient of each pixel point in each two-dimensional spatial SAR image wherein φ is a wavelet rotation angle of each pixel point, a is a scale parameter of the complex wavelet transform in each pixel point, is a translation parameter of the complex wavelet transform in each pixel point; Step 4: Obtain the scale parameter a corresponding to the maximum modulus of the complex wavelet coefficient by traversing all scale parameters and rotation angle parameters for each pixel in each SAR image m and rotation angle parameters m a m and φ m constitute the spatial information of sea ice, wherein | · | represents the modulus Step 5: Spatially match the original SAR image information, sea ice spatial feature information, and corresponding actual sea ice roughness within the target area and time period to construct a total dataset containing the above information; Step 6: Take the original SAR image information and sea ice spatial feature information as input, and the corresponding sea ice predicted roughness as the model output. Use the training set to train the Adaboost regression learner to obtain the trained sea ice roughness inversion model based on Adaboost regression. Step 7: Use the regression model obtained in Step 6 to predict the sea ice roughness of the test set, and conduct preliminary evaluation, authenticity test and in-depth assessment of the model.

2. The sea ice roughness retrieval method based on ensemble learning method and two-dimensional wavelet transform according to claim 1, characterized in that, In the complex two-dimensional continuous wavelet transform described in step 3, the formula for calculating the complex wavelet coefficients is shown in the following equation; and and r -φ The formula is as follows: where c ψ is a constant for normalizing the equation, is the spatial frequency, is the conjugate of the two-dimensional Fourier transform of the wavelet function ψ, is the two-dimensional Fourier transform of the spatial image s(x); the scale parameter a is taken to be 1-32 and the rotation angle φ is taken to be in the interval 0-2π. is taken.

3. The sea ice roughness inversion method based on ensemble learning and two-dimensional wavelet transform as described in claim 2, characterized in that, In the complex two-dimensional continuous wavelet transform described in step 3, the wavelet mother function is a two-dimensional Cauchy wavelet function, and the result of its two-dimensional Fourier transform is as follows: where ω x is the frequency in the x direction, ω y is the frequency in the y direction, A is a scaling parameter of the wavelet function, and α is a half-open angle of a convex cone in which the domain of the two-dimensional Cauchy wavelet function is located.

4. The sea ice roughness inversion method based on ensemble learning and two-dimensional wavelet transform as described in claim 1, characterized in that, The data preprocessing of the SAR image described in step 1 includes the following steps in sequence: orbit correction, thermal noise removal, radiometric calibration, speckle filtering, decibel conversion, and two-dimensional linear interpolation.

5. The sea ice roughness inversion method based on ensemble learning and two-dimensional wavelet transform as described in claim 1, characterized in that, Step 5 involves randomly shuffling all samples from the obtained total dataset and dividing them into a training set and a test set in a 7:3 ratio.

6. The sea ice roughness inversion method based on ensemble learning and two-dimensional wavelet transform as described in claim 1, characterized in that, Step 7 describes the preliminary model evaluation, which involves calculating model evaluation indices using the actual roughness and predicted roughness of the test set. The calculation formulas for each index are shown below: Among them, R 2 The coefficient of determination is y, MAE is mean absolute error, RMSE is root mean square error, and MAPE is mean absolute percentage error. i It is the true value. It is a predicted value. It is the average of the true values, and N is the number of samples.

Images