Automatic identification method of Antarctic ice bottom feature unit based on ice radar data

Abstract

The invention discloses an automatic identification method of a south pole ice bottom feature unit based on ice radar data, and belongs to the field of ice radar data processing. Comprising an image feature recognition ice bottom unit, signal feature weight calculation, a signal feature recognition ice bottom unit, two-dimensional entropy-based module weight calculation and automatic recognition of a South Pole ice bottom feature unit. According to the method, an ice radar profile map and signal features are used as input, and the signal features and image features of ice radar data are used, so that the inaccuracy caused by using a single feature is reduced; weights are given to the ice bottom units with the large distance backward scattering change rate, interference of the large radar echo intensity change rate on signal features is reduced, and the capacity of recognizing the ice bottom units is improved. According to the pixel distribution rule of the ice radar profile map, the proportion of different modules in the comprehensive identification result is adjusted, and the reduction of the identification accuracy caused by the pixel distribution difference is reduced. Finally, a more accurate Antarctic ice bottom unit identification result can be obtained, and the method is suitable for different areas and a large amount of ice radar data.

Description

Technical Field

[0001] The present invention relates to the field of ice radar data processing, and in particular to an automatic recognition method for Antarctic ice bottom feature units applicable to ice radar data. Background Art

[0002] With the advancement of research on the mass balance and motion of the Antarctic ice sheet, the subglacial units beneath the Antarctic ice sheet have become of great research and application value, as they can enhance deformation and sliding of the underlying ice and alter the ice sheet's dynamic behavior. The presence of subglacial units not only reflects the characteristics of the ice substrate environment and its historical evolution, but also enhances deformation and sliding of the underlying ice, thereby altering the ice sheet's dynamic behavior. Therefore, accurately identifying and extracting subglacial units is of great scientific significance for understanding the subglacial environment and processes, and for studying ice sheet dynamics and evolution. Currently, this technology is primarily implemented through three methods: traditional extraction methods, machine learning model extraction, and neural network image recognition. Traditional methods use phase shift response functions for identification. Based on the relationship between signal-to-noise ratio and phase shift, the phase shift response function of a specific layer is plotted. The standard deviation of the Gaussian fit results of the phase shift response functions of different layers is compared with the standard deviation of the intermediate layer and subglacial units to obtain the subglacial unit classification results. However, this method is computationally intensive and time-consuming, and requires secondary classification based on the already classified subglacial units. Methods based on feature selection and support vector machines select location range, KL divergence, edge detectors, energy attenuation, and specularity as features for identifying ice bottom feature units. The selected feature vectors are then fed into a classifier, which uses a supervised classifier to learn features for different categories and assign labels to each pixel to obtain classification results. However, this method suffers from inappropriate feature selection. The location range feature requires a specific distance from the bed or ice bottom unit to the ice surface, making it difficult to accurately identify ice bottom units near the ice surface and unable to identify ice bottom feature units with large depth differences. Methods using neural networks for image recognition label radar profiles and input them into a ResNet network trained using a training set and parameter optimization to complete model training. Test data is then fed into the optimized model for recognition. However, this method, relying solely on image features, has low recognition accuracy when ice bottom feature units are present in the presence of high concentrations of high-dielectric particles, making it unsuitable for high-accuracy ice bottom unit identification. With the advancement of Antarctic subglacial exploration and the increasing demand for further understanding of ice sheet dynamics, a large amount of Antarctic ice radar data will be generated in the future. Automatic ice bottom unit identification from this large amount of ice radar data has become a research hotspot. Summary of the Invention

[0003] In order to overcome the problems of low recognition efficiency and recognition accuracy being easily affected by the echo intensity change rate in the existing ice bottom unit distribution recognition methods based on Antarctic ice radar data processing, the present invention proposes an automatic recognition method for Antarctic ice bottom units based on ice radar data. The advantages of this method are: 1) the signal characteristics and image characteristics of ice radar data are used simultaneously, reducing the inaccuracy caused by the use of a single feature; 2) weights are assigned to "ice bottom units" with large distance backscattering change rates, reducing the interference of large radar echo intensity change rates on signal characteristics, and improving the ability to identify "ice bottom units"; 3) according to the pixel distribution law of the ice radar profile, the proportion of different modules in the comprehensive recognition results is adjusted to reduce the decline in recognition accuracy caused by pixel distribution differences. In this way, the distribution of ice bottom units in the Antarctic region can be more accurately identified. This method combines image recognition of residual neural networks and signal feature recognition coupled with weights, and can efficiently and accurately obtain the distribution results of regional ice bottom units. The overall flow chart of the method is shown as follows. Figure 1 As shown, it is mainly divided into four modules, namely signal feature calculation module, image feature recognition module, signal feature recognition module and judgment module.

[0004] Signal feature calculation module: To extract candidate areas of ice bottom cells from ice radar data, we first calculate the steepness of the range signal to obtain the changing trend of the ice radar signal. The steepness calculation formula is (as shown in Formula 1):

[0005]

[0006] Among them, x i is the number of distance points, N is the width of the steepness calculation window, Z(x i ) is the echo power in the range direction, SL N (x i ) is the upward steepness of the ice radar signal over distance. Because the ice bottom cell is located in a reflection void, there are regions of low echo power on both sides of the ice bottom cell echo signal. Based on this characteristic, we can exclude the ice bottom region and identify the candidate ice bottom cell area.

[0007] After obtaining the candidate regions for the ice base unit, we used reflectivity as a signal feature to correctly identify whether the candidate regions contained the ice base unit. At the same time, to address the reflectivity loss caused by the glacier decay effect, we used a method where the glacier decay rate varies linearly with depth and used sliding Gaussian filters of different lengths to calculate the mean and azimuthal derivative of the glacier decay rate to calculate reflectivity. The formula for calculating relative reflectivity (as shown in Formula 2) is:

[0008]

[0009] Among them, Rr is the relative reflectivity, S is the radar system parameter, B is the birefringence loss, P is the ice radar echo intensity, G is the geometric transmission loss, N a is the glacier decay rate, is the rate of change of glacier decay rate, is the azimuthally averaged glacier decay rate, The overall calculation process of glacier decay, x is the number of azimuth points, is the average position in azimuth, d is the depth, and l is the Gaussian filter length. When the ice-bottom cell has a large echo intensity change rate, the geometric transmission loss decreases as a function of the negative power of one half. As shown in Equation 2, when backscattering increases, consistent with high particle concentrations, the reflectivity also decreases in an inverse square relationship.

[0010] For roughness, we chose a multi-scale, dual-parameter roughness calculation formula that can quantify the adaptive weights and parameters of terrain extraction within a moving window at different scales to obtain roughness quantification results. The multi-scale, dual-parameter method describes the roughness formula (Formulas 3 and 4) as follows:

[0011]

[0012]

[0013] Among them, ξ MS (x i ) is the vertical deviation parameter of the ice bed, η MS (x i ) is the slope frequency parameter, N min and N max is the value range of window length N. N (x i ) is the vertical deviation factor, η N (x i ) is the slope frequency factor, W N (x i ) is the weight of the two parameters, and the calculation formula (Formula 5) is as follows:

[0014]

[0015] in, as well as are the front and rear steepness of the bed elevation, SK N (x i ) is the deflection of the bed elevation, CV N (x i ) is the coefficient of variation of bed elevation. sgn(·) is the sign function, is the element-by-element addition symbol, α(x i) is the weight distribution parameter, which can be calculated from the standard deviation of skewness and divergence, N valid for . Equations 3, 4, and 5 show that the weights of the two roughness parameters, the ice bed vertical deviation coefficient, and the slope frequency coefficient of the deviation are entirely determined by the layer elevation and are independent of the echo power at that layer. Therefore, the roughness characteristics are not affected by changes in the high-dielectric particle concentration.

[0016] Image feature recognition module: The residual structure of ResNet helps reduce the blurring of image texture and edge features in Res images as the depth of the network layer increases. We replace the spatial pyramid pooling (ASPP) and image pooling layers originally used to enhance semantic segmentation with fully connected layers and softmax layers to convert the extracted image features into ice-bottom unit classification probabilities.

[0017] The RES image is first input into the bottleneck, which consists of three convolutional layers: a dimensionality reduction layer, a feature extraction layer, and a dimension expansion layer. The dimensionality reduction layer uses 1x1 convolution to reduce the dimensionality of the RES contour image, thereby reducing the computational load and the number of model parameters, which significantly reduces computational cost. The feature extraction layer applies a convolution operation to the input data using a sliding convolution kernel to extract the grayscale, texture, and edge features of the RES image. The dimension expansion layer restores the data to its original dimension to prevent information loss. Subsequently, a fully connected layer flattens the extracted image feature data and performs a linear transformation (matrix multiplication plus bias) to generate a vector in which each element corresponds to the score (logits) of the ice bottom unit class. Finally, a softmax layer converts the scores output by the fully connected layer into a probability distribution, converting the logits into a probability value between 0 and 1, representing the probability of identifying the ice bottom unit.

[0018] Signal Feature Recognition Module: This module uses roughness and reflectivity to identify candidate ice-base cell regions. However, since reflectivity can be affected by the concentration of high dielectric particles within the region, and the reflectivity and roughness differ in magnitude, a convolutional layer is required to map the layer features to the candidate regions, a normalization layer to normalize features of different magnitudes, and a contrast loss operation to analyze the concentration distribution within the candidate regions.

[0019] First, the reflectivity and roughness features are fed into the convolutional layer, mapping the roughness and reflectivity information at the layer to the entire ice-base unit candidate region. Next, the roughness and reflectivity features are normalized, unifying the two features of different magnitudes and units to between 0 and 1. A contrast loss calculation is then performed to reflect the difference between the two features after normalization. When the candidate medium is low-dielectric, the bedrock region has greater roughness and relatively lower reflectivity, while the ice-base unit region has less roughness and relatively higher reflectivity than the bedrock region. When the concentration of high-dielectric particles in the ice-base unit region is high, the reflectivity of the ice-base unit decreases while the roughness remains unchanged. Therefore, the contrast loss calculation can be used to determine the concentration distribution of high-dielectric dielectrics in the ice-base unit region, and thus the rate of change of radar echo intensity. Based on this high-dielectric particle concentration distribution, the reflectivity weighting (Equation 6) and the roughness weighting (Equation 7) are assigned. According to Formula 6, Formula 7 is given a weight distribution that decreases by a power of two as the concentration increases and is complementary, so that Formula 6 and Formula 7 produce a stable weighted sum, accurately reflecting the distribution of high dielectric particle concentration in the candidate area.

[0020] Since the reflectivity shows a trend of decreasing as the negative power of one-half as the concentration of high dielectric particles increases. In order to ensure that the reflectivity feature can correctly describe the ice bottom unit, we compensate for the reflectivity feature by multiplying the reflectivity weight that increases with the power of one-half of the concentration in Formula 6 with the reflectivity that decays as the negative power of one-half. At the same time, since the roughness feature of the ice radar data is a geological morphological feature, it does not change with the different properties of the ice bottom material. However, since the reflectivity weight increases with the concentration of high dielectric particles in the candidate area of the ice bottom unit, the overall feature weighted sum value will not be able to identify the concentration distribution as it increases, which will interfere with the identification of the ice bottom unit. We use Formula 7 to give the roughness a weight that decreases as the power of one-half of the concentration, so that the feature weighted sum remains stable, can correctly reflect the high dielectric particle concentration distribution in the candidate area, and accurately identify the ice bottom unit in the candidate area. And since roughness is a two-parameter signal feature, we divide the roughness weight by 2, and the weight distribution formula (as shown in Formulas 6 and 7) is as follows:

[0021]

[0022] Among them, Q F is the reflectivity weight, Q C is the roughness weight, f F (x i ) is the reflectivity input of the contrast loss function, f C (x i ) is the roughness input of the contrast loss function, is the result of the contrast loss function calculation, and N is the size of the contrast loss function operation. as well as is the reflectivity and roughness weight within the network, α is the learning rate, and the factor is the back propagation error of the l+1th layer, W C0 is the starting weight of the roughness in the input layer, W F0 is the starting weight of reflectivity in the input layer, rot180(·) is the transposition operation, ReLU(·) is the activation function, is the error of the p-th fully connected layer.

[0023] After assigning weights to reflectivity and roughness, we need to convert the weighted sum of these two features into an ice-base cell probability based on their weighted inputs. To do this, we use convolutional layers to extract feature information, as well as fully connected and softmax layers to convert the weighted sum of the features into a probability distribution. We first input the weighted reflectivity and roughness values into the convolutional layers to obtain the weighted sums of reflectivity and roughness at different locations. This is then reduced in dimensionality through a max-pooling layer. Next, through fully connected and softmax layers, the weighted sums are converted into class predictions, outputting a classification result indicating whether the ice radar data contains an ice base cell. Finally, at the network's output layer, we use the cross-entropy loss function, which is sensitive to target relevance and performs well in classification tasks, as a metric to assess training completion. The cross-entropy loss function incorporates a logarithmic transformation, which more severely penalizes incorrect predictions, thereby better guiding the classification model to learn the correct ice base cell classification decisions.

[0024] During the training phase, to ensure the network generates accurate classification results, we assign appropriate weights to the convolution kernels in the network through backpropagation. By taking the partial derivative of Equation 8, we perform error calculations from the softmax layer to the input layer, completing a new round of convolution kernel weight updates. Model training is complete when the cross-entropy loss approaches zero. After inputting the statistical characteristics, reflectivity, and roughness of the ice radar signal, we output the probability that the data contains an ice bottom unit. The cross-entropy loss function formula (shown in Equation 8) is:

[0025]

[0026] Among them, J(y,a) is the cross entropy loss, a i is the predicted distribution, y i is the true distribution.

[0027] Decision Module: The rate of change in backscattering consistent with high particle concentrations will affect the accuracy of ice base unit image feature recognition. Therefore, simply averaging the probability outputs of the signal feature recognition module and the image feature recognition module may lead to inaccurate recognition results. To account for variations in dielectric particle concentration within ice base units, it is necessary to assign module weights based on the image's two-dimensional entropy. When a distinct and bright ice base unit is present in the RES image, or when the unit is absent, the image's grayscale distribution is more regular and its two-dimensional entropy is lower. Conversely, when a high concentration of dielectric particles is present within the ice base unit, the grayscale values in certain areas of the image decrease, forming a more random distribution and increasing its two-dimensional entropy. Therefore, we weight the modules based on the calculated image entropy. Because the ice radar echo power decreases exponentially with increasing particle concentration, the grayscale values of the RES profile decrease. We address this issue by multiplying the exponentially increasing image entropy weight (as shown in Equation 9) with the output probability of the signal feature recognition module. Finally, by weighted summing the ice base unit detection probabilities from the image and signal feature recognition modules (as shown in Equation 9), we obtain a comprehensive recognition probability for regions with varying particle concentrations. The final comprehensive recognition result is determined by comparing the probability of detecting an ice-bottom unit with the probability of detecting a non-ice-bottom unit. If the probability of an ice-bottom unit is greater than the probability of a non-ice-bottom unit, the area is considered to contain an ice-bottom unit structure. The formula for calculating the comprehensive recognition probability (as shown in Equation 9) is as follows:

[0028]

[0029] Among them, L is the number of segmented regions used for recognition, p(i,j) is the element value of the position (i,j) in the middle layer of the neural network, P is the overall recognition probability, P i and P s are the recognition probabilities of the image and signal feature recognition modules respectively. is the two-dimensional entropy calculation result of the cross-section diagram, is the weight of the image feature recognition module, and 1 is set as the weight of the signal feature recognition module.

[0030] Beneficial effects

[0031] This paper proposes an automatic identification method for Antarctic ice base units from radar data. This method can obtain a more accurate distribution of regional Antarctic ice base units, adapts to regional characteristics, and can be applied to different regions and large amounts of RES data. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is an overall flow chart for implementing the present invention.

[0033] Figure 2 This is a flow chart of the signal feature calculation module of the present invention.

[0034] Figure 3 Flowchart of the image feature recognition module of the present invention.

[0035] Figure 4 This is a flow chart of the signal feature recognition module of the present invention.

[0036] Figure 5 Automatic identification of network structures for Antarctic ice-bottom units.

[0037] Figure 6 This is a cross-sectional view of an example survey line.

[0038] Figure 7 This is the steepness curve of the ice base unit for the example survey line.

[0039] Figure 8 This is the reflectivity curve of the ice bottom unit of the example survey line.

[0040] Figure 9 The two-parameter roughness curve of the ice bottom unit of the example survey line.

[0041] Figure 10 is the contrast loss function curve of the example survey line.

[0042] Figure 11 It is the accuracy and loss curve of the network training process.

[0043] Figure 12 The confusion matrix result.

[0044] Figure 13 For accuracy, precision, recall and F1 score

[0045] Figure 14 Results for ice-base unit identification. DETAILED DESCRIPTION

[0046] The present invention will be described in detail below with reference to the specific embodiments shown in the accompanying drawings.

[0047] Figure 1 This is the overall flow chart of the method for automatically identifying Antarctic ice bottom units based on ice radar data proposed in the present invention. It includes a signal feature calculation module, an image feature recognition module, a signal feature recognition module, and a judgment module. It takes ice radar data as input and outputs the recognition results of the ice bottom units in this data segment.

[0048] Figure 2To generate a flowchart for the signal feature calculation module, the main process includes steepness calculation, reflectivity calculation, and roughness calculation. First, we obtain high-gain focused data for the selected survey line, extract the layers of the profile generated by this survey line, and obtain the layer information of the ice bottom unit interface and the ice base interface to obtain the depth and echo amplitude to calculate the echo power. At the same time, the signal steepness is calculated in the distance direction. The calculation window width is set, and the sliding window moves along the distance direction. According to Formula 1, the change trend of the ice radar signal echo intensity is obtained. In the steepness calculation process, due to different data sets, the intervals between adjacent distance sampling points are also different. We resampled the data of Mount Gambutsev (AGAP) in Antarctica by interpolation to eliminate the influence of different data sets on the characteristic value calculation.

[0049] Based on the layer information, depth, echo power, etc. of the ice bottom unit interface and the ice base interface, the reflectivity can be calculated according to Formula 2. The radar system parameter S and the birefringence loss B can be regarded as constants, and the geometric transmission loss G can be obtained according to Formula 10:

[0050]

[0051] Where d is the ice depth, h is the aircraft altitude, and ε is the dielectric constant of ice. The relative reflectivity is then calculated using Equation 2, based on fixed radar system parameters, birefringence loss, geometric transmission loss, and glacier attenuation. The radar system used is the HiCARS radar system, and the sum of the radar system parameter S and the birefringence loss B is set to 9.2 dB.

[0052] We use an adaptive two-parameter method to characterize roughness. This method requires the elevation change of a layer as input. The elevation data can be obtained from the layer extraction results of the survey line data, and the two-parameter roughness can be calculated according to Equations 3, 4, and 5.

[0053] Figure 3This is the flowchart for the image feature recognition module. It takes Antarctic ice radar profiles as input and uses a trained ResNet101 model to identify the images and present classification results. First, a dataset of image annotations corresponding to the ice radar signal data is constructed, containing a large number of ice radar profiles and annotated to indicate the presence of ice-bottom units. Data preprocessing includes image normalization, grayscale conversion to three-channel, adaptation to ResNet input, and a 7:2:1 split into training, validation, and test sets corresponding to the ice radar signal data. During training, ResNet101 utilizes pretrained ImageNet weights for transfer learning, extracting features layer by layer through convolutional layers: shallow layers detect edges and textures, mid-layers identify internal reflective structures within the ice layer, and deep layers capture global anomalies of ice-bottom units (such as strongly reflective interfaces or continuous undulations). The residual structure of ResNet101 ensures efficient gradient backpropagation, preventing degradation in the deep layers. The Adam optimizer adjusts hyperparameters by minimizing the cross-entropy loss and monitoring metrics (accuracy and F1 score) on the validation set. The validation phase used an independent dataset to evaluate model performance, analyzing misclassified samples (such as ice clutter misclassified as ice bottom cells) through a confusion matrix. Ultimately, the network mapped the high-dimensional features of the ice radar image to a classification space, with the fully connected layer outputting probabilities.

[0054] Figure 4 This is a flowchart of the signal feature recognition module, which includes candidate region extraction, contrast loss calculation, normalized weighted summation, and the signal feature recognition network. At the boundary between the reflection void and the ice bed, the signal steepness exhibits a positive-negative-positive trend, indicating the presence of a transition zone from the reflection void to the ice layer. This characteristic can be used to extract candidate regions. After extracting the candidate regions, roughness and reflectivity features are used to identify the presence of ice bottom units. Because the two features, roughness and reflectivity, have different units, they can interfere with the subsequent estimation of particle concentration consistent with backscattering. A normalization layer is used to normalize the roughness and reflectivity characteristics to values between 0 and 1. Contrast loss (Equation 11) is then applied to reflect the differences between the normalized features.

[0055]

[0056] Among them, f F (x i ) is the reflectivity input, f C (x i) is the roughness input, and N is the size of the contrast loss function. When a base cell region has a high particle concentration consistent with backscatter, the reflectivity of the base cell decreases, but the roughness remains unchanged, resulting in a larger contrast loss function result. Therefore, contrast loss can help determine the particle concentration distribution consistent with backscatter from the base cell region. After calculating the contrast loss function, signal feature weights are assigned and weighted summed as shown in Equations 6 and 6. The weighted sum is then input into the signal feature recognition network. The signal feature recognition network is based on a convolutional neural network and can be used to identify "base cell" structures in Antarctic ice radar signal signatures. The dataset structure mirrors the profile dataset, also splitting the training, validation, and test sets corresponding to the ice radar signal data in a 7:2:1 ratio. The network consists of multiple convolutional and pooling layers. Through training, it automatically learns to identify key characteristic patterns in the weighted signal signatures, such as unusually strong signal signatures at specific locations. During training, the cross-entropy loss function and the Adam optimizer are used for parameter optimization, and model performance is monitored on the validation set to prevent overfitting. The trained network can finally scan the input weighted signal features, extract features layer by layer, and analyze the change pattern of signal features to determine whether the signal contains the geological structure feature of the ice bottom unit.

[0057] Figure 5 To automatically identify Antarctic under-ice units, a network for under-ice unit classification was designed based on the qualitative and semi-quantitative characteristics of under-ice units in ice radar data. The network primarily consists of a convolutional neural network and a ResNet101 network. The batch size was set to 16, and training was performed for 100 epochs. The initial learning rate was set to 0.1, and optimization was performed using the Adam optimizer. The cross-entropy loss function, as shown in Equation 8, was used as the loss function. The input image was 512×512 pixels, segmented using an image segmentation program, and the intermediate layers were manually removed. The output dimension was 2, indicating whether an under-ice unit was present or not. Since this is a binary classification task, a softmax function was used to convert the raw neural network output into a probability distribution and output the classification result. The hardware environment for network development was an Intel(R) Core(TM) i9-9880H @ 2.30GHz CPU with 64GB of memory. The network was programmed in Python. The network first identified candidate under-ice units by analyzing the statistical characteristics of the ice radar data. The concentration of particles consistent with backscatter in these candidate regions is then estimated by comparing normalized roughness and reflectivity. Finally, weighted roughness and reflectivity are used to determine the likelihood of an ice base unit. The image feature recognition module uses the ResNet101 architecture. It takes the ice radar profile image as input and outputs the probability of the presence of an ice base unit. Finally, the recognition results of the two modules are combined, weighted according to the two-dimensional entropy of the image (Equation 12), and the final recognition result is output.

[0058]

[0059] Among them, L is the number of segmented regions used for recognition, p(i,j) is the element value of the position (i,j) in the middle layer of the neural network, P is the overall recognition probability, P i and P s are the recognition probabilities of the image and signal feature recognition modules, respectively. When there is a distinct, bright ice base cell in the ice radar profile, or when there is no ice base cell, the image's grayscale distribution is more regular, and the image's two-dimensional entropy is lower. Conversely, when a high concentration of dielectric particles is present in the ice base cell, the grayscale levels in certain areas of the image decrease, resulting in a more random distribution and increasing the image's two-dimensional entropy. Therefore, we assign weights to the modules based on the calculated two-dimensional entropy of the image. Since the radio echo power decreases exponentially with increasing particle concentration, the grayscale level of the RES profile decreases. We address this issue by multiplying the weights by the output probability of the signal feature recognition module, as shown in (Equation 9), where the weights increase exponentially with the image entropy. Finally, by weighted summing the ice base cell detection probabilities from the image and signal feature recognition modules (Equation 9), we obtain the combined recognition probability for regions with different particle concentrations. The final comprehensive recognition result is determined by comparing the probability of detecting an ice base cell with the probability of detecting a non-ice base cell. If the probability of an ice base cell is greater than the probability of a non-ice base cell, the region is considered to contain an ice base cell structure.

[0060] Figure 6 This is a cross-section of survey line F10150 from the AGAP project, which includes ice-bottom units. The red dashed line indicates the location where steepness data was extracted, and the yellow dashed line indicates the window width for calculating reflectivity and roughness characteristics. Figure 8 、 9 The dotted lines in , 10 have the same meaning as here.

[0061] Figure 7 The steepness data for the survey line shows that at the boundary between the reflection void and the ice bed, the range signal steepness changes from 10.7 dB / m to -5.6 dB / m and then to 8.7 dB / m. The overall steepness shows a positive-negative-positive trend, indicating the presence of a transition zone from the reflection void to the ice bed. Figure 8 、 9 They are the reflectivity and roughness data of the survey line, respectively. We set the feature calculation window length M to 17.4 kilometers because it can include various glacial landforms appearing in this interval. Figure 8 、 9The reflectivity and roughness characteristics of the survey line after convolution and normalization are shown. At the test point, the reflectivity of the ice-base unit region is close to 1, and the roughness values are close to 0.17 and 0.15. This shows that ice-base units with high particle concentrations have higher reflectivity and unchanged roughness compared to normal ice-base units. Figure 10 The calculated contrast loss of the ice bottom unit is shown to be approximately 0.47, which is smaller than that of the general ice bottom unit area.

[0062] Figure 11 To plot the accuracy and loss curves during network training, we initialized the weights of each convolutional layer using a zero-mean Gaussian distribution with a standard deviation of zero. The batch size was set to 8. Both curves demonstrate that our neural network model can be trained effectively. The loss gradually decreases with increasing training iterations, reaching a value of 0.102 and remaining largely stable after epoch 70, indicating that the model is gradually improving during learning. The accuracy also reaches 90.39% and remains stable at epoch 70, demonstrating that accuracy improves as training progresses.

[0063] Figure 12 To create a confusion matrix, the experiment used 2,146 ice radar data sets in the test set, of which 1,112 contained ice bottom cells and 1,034 did not. Of the ice radar data sets that contained ice bottom cells, 1,021 were correctly identified by the network model as containing them, while 91 were incorrectly identified as not containing them. Conversely, 114 were incorrectly identified as containing them, while 920 were correctly identified as not containing them.

[0064] Figure 13 The accuracy, precision, recall, and F1 score of the method are shown in Table 1. Accuracy is the ratio of correctly predicted samples to the total number of samples. Precision is the accuracy of the positive example prediction relative to the predicted result of the positive example. Recall can assess the coverage of all actual positive examples. The F1 score is the harmonic mean of precision and recall, which comprehensively considers the accuracy and coverage of the model and provides a robust evaluation in different scenarios. Our method has an accuracy of 90.39%, which shows that this method for identifying ice bottom units is feasible and highly accurate. The accuracy is 91.77%, the recall is 89.9%, and the calculated F1 score is 90.83%, indicating that the proposed method performs well and can accurately identify ice bottom units.

[0065] Figure 14This is the result of identifying an ice-bottom unit for AGAP survey line F10150. The ice-bottom unit is circled in red. In this identification, the image recognition module gave a probability of 0.8961 for the ice-bottom unit, and the signal recognition module gave a probability of 0.9325 for the ice-bottom unit. The weighted overall identification result is 0.9110, with an overall probability greater than 0.5. Therefore, this data segment is determined to contain an ice-bottom unit.

Claims

1. A method for automatically identifying Antarctic ice bottom units based on ice radar data, characterized in that include: Signal feature calculation module: To extract candidate areas of ice bottom units from ice radar data, the steepness calculation is first performed on the range signal to obtain the change trend of the ice radar signal; the steepness calculation formula is shown in Formula 1: Among them, x i is the number of distance points, N is the width of the steepness calculation window, Z(x i ) is the echo power in the range direction, SL N (x i ) is the upward steepness of the ice radar signal range; After obtaining the candidate area of the ice bottom unit, the reflectivity signal feature is used to correctly identify whether the candidate area contains the ice bottom unit. At the same time, in order to solve the reflectivity loss problem caused by the glacier decay effect, the glacier decay rate is linearly changed with depth, and the mean and azimuthal derivative of the glacier decay rate are calculated using sliding Gaussian filters of different lengths to calculate the reflectivity. The relative reflectivity calculation formula is: Among them, R r is the relative reflectivity, S is the radar system parameter, B is the birefringence loss, P is the ice radar echo intensity, G is the geometric transmission loss, N a is the glacier decay rate, is the rate of change of glacier decay rate, is the azimuthally averaged glacier decay rate, The overall calculation process of glacier decay, x is the number of azimuth points, is the average position in azimuth, d is the depth, and l is the Gaussian filter length; The roughness calculation formula of multi-scale dual-parameter is selected to quantify the adaptive weights and parameters of terrain extraction within the moving window at different scales to obtain the roughness quantification result; the formula describing the roughness of the multi-scale dual-parameter method is as follows: Among them, ξ MS (x i ) is the vertical deviation parameter of the ice bed, η MS (x i ) is the slope frequency parameter, N min and N max is the value range of the window length N; N (x i ) is the vertical deviation factor, η N (x i ) is the slope frequency factor, W N (x i ) is the weight of the two parameters, and the calculation formula (Formula 5) is as follows: in, as well as are the front and rear steepness of the bed elevation, SK N (x i ) is the deflection of the bed elevation, CV N (x i ) is the coefficient of variation of bed elevation; sgn(·) is the sign function, is the element-by-element addition symbol, α(x i ) is the weight distribution parameter, which is calculated by the standard deviation of skewness and divergence, N valid for The value of Image feature recognition module: The RES image is first input into the bottleneck, which consists of three convolutional layers: a dimensionality reduction layer, a feature extraction layer, and a dimension expansion layer. The dimensionality reduction layer uses 1x1 convolution to reduce the dimension of the RES contour image. The feature extraction layer applies a convolution operation to the input data using a sliding convolution kernel to extract the grayscale, texture, and edge features of the RES image. The dimension expansion layer restores the data to its original dimension. Subsequently, the fully connected layer flattens the extracted image feature data and performs a linear transformation to generate a vector in which each element corresponds to the score (logits) of the ice bottom unit category. Finally, the softmax layer converts the scores output by the fully connected layer into a probability distribution and converts logits into a probability value between 0 and 1, indicating the recognition probability of the ice bottom unit. Signal feature recognition module: This module uses roughness and reflectivity to identify whether a candidate region of an ice bottom unit contains an ice bottom unit. It also includes a convolutional layer that maps layer features to the candidate region, a normalization layer that normalizes features of different orders of magnitude, and a contrast loss operation that analyzes the concentration distribution within the candidate region. First, the reflectivity and roughness features are fed into the convolutional layer, mapping the roughness and reflectivity information at the layer to the entire ice base unit candidate area. Next, the roughness and reflectivity features are normalized to unify the two features of different magnitudes and units to between 0 and 1. Then, a contrast loss operation is performed to reflect the difference between the two features after normalization. By multiplying the reflectivity weight that increases with the power of one-half of the concentration in Formula 6 by the reflectivity that decays with the power of negative one-half, the reflectivity characteristic is compensated. The weight of the roughness that decreases with the power of one-half of the concentration is given using Formula 7. Since roughness is a two-parameter signal characteristic, the roughness weight is divided by 2. The weight distribution formula is as follows: Among them, Q F is the reflectivity weight, Q C is the roughness weight, f F (x i ) is the reflectivity input of the contrast loss function, f C (x i ) is the roughness input of the contrast loss function, is the result of the contrast loss function calculation, N is the size of the contrast loss function operation; factor as well as is the reflectivity and roughness weight within the network, α is the learning rate, and the factor is the back propagation error of the l+1th layer, W C0 is the starting weight of the roughness in the input layer, W F0 is the starting weight of reflectivity in the input layer, rot180(·) is the transposition operation, ReLU(·) is the activation function, is the error of the p-th fully connected layer; First, the reflectivity and roughness with corresponding weights are input into the convolution layer to obtain the weighted sum of reflectivity and roughness in different areas, and the data dimension is reduced through the maximum pooling layer. Next, the extracted weighted sum is converted into a category prediction through the fully connected layer and the softmax layer, and the classification result of whether the ice radar data contains ice bottom units is output. Finally, at the output layer of the network, the cross entropy loss function, which is sensitive to target correlation and performs well in classification tasks, is used as an indicator to evaluate whether the training is complete. During the training phase, by taking the partial derivative of Formula 8, the error is calculated layer by layer from the softmax layer to the input layer from back to front to complete the update of the new round of convolution kernel weights. When the cross entropy loss tends to zero, the model training is completed. After inputting the statistical characteristics, reflectivity, and roughness of the ice radar signal, the probability of the data containing ice bottom units can be output. The cross entropy loss function formula (as shown in Formula 8) is: Among them, J(y,a) is the cross entropy loss, a i is the predicted distribution, y i is the true distribution; Decision module: By taking a weighted sum of the ice-bottom unit detection probabilities from the image and signal feature recognition modules, a comprehensive recognition probability for regions with different particle concentrations is obtained. The final comprehensive recognition result is determined by comparing the probability of detecting an ice-bottom unit with the probability of detecting a non-ice-bottom unit. If the probability of an ice-bottom unit is greater than the probability of a non-ice-bottom unit, the region is considered to contain an ice-bottom unit structure. The formula for calculating the comprehensive recognition probability is as follows: Among them, L is the number of segmented regions used for recognition, p(i,j) is the element value of the position (i,j) in the middle layer of the neural network, P is the overall recognition probability, P i and P s are the recognition probabilities of the image and signal feature recognition modules respectively; factors is the two-dimensional entropy calculation result of the cross-section diagram, is the weight of the image feature recognition module, and 1 is set as the weight of the signal feature recognition module.

2. The method for automatically identifying Antarctic ice bottom units based on ice radar data according to claim 1, characterized in that: The signal steepness is calculated in the range direction. The calculation window width is set, and the sliding window moves along the range direction. The changing trend of the ice radar signal echo intensity is obtained according to Formula 1. The signal steepness at the boundary between the reflection blank area and the ice bed will show a positive-negative-positive trend, and candidate areas are extracted according to this trend.

3. The method for automatically identifying Antarctic ice bottom units based on ice radar data according to claim 1, characterized in that: In the signal feature recognition module, after using steepness to extract candidate areas for ice bottom units, reflectivity and roughness are used to identify ice bottom units. The normalized values of the two features are weighted and summed using a parameter related to the rate of change of distance-directed signal intensity. The weighted sum is then used as input to identify ice bottom units in one-dimensional signals, and the recognition probability is used as the output of the signal feature recognition module.

4. The method for automatically identifying Antarctic ice bottom units based on ice radar data according to claim 1, characterized in that: The contrast loss operation is used to reflect the difference between the reflectivity and roughness features after normalization. When the medium in the candidate area is a low dielectric medium, the roughness of the bedrock area is large and the reflectivity is relatively low, and the ice bottom unit area is smaller in roughness and relatively larger in reflectivity than the bedrock area. When the concentration of high dielectric particles in the ice bottom unit area is large, the reflectivity of the ice bottom unit decreases while the roughness remains unchanged. Therefore, through the contrast loss operation, the concentration distribution of the high dielectric medium in the ice bottom unit area can be obtained, and then the radar echo intensity change rate can be obtained. According to the obtained high dielectric particle concentration distribution, the reflectivity weight distribution of formula 6 and the roughness weight distribution shown in formula 7 are performed.

5. The method for automatically identifying Antarctic ice bottom units based on ice radar data according to claim 1, characterized in that: The signal characteristics and image characteristics of ice radar data are used simultaneously, and parameters related to the two-dimensional entropy of the ice radar profile are used to weight the signal feature recognition module and the image feature recognition module respectively, and the comprehensive recognition result is used as the final ice bottom unit recognition result.

Images