A semi-automatic prediction method of subglacial melt-freeze distribution based on ice radar data

Abstract

The application provides a semi-automatic prediction method for ice thawing and freezing distribution under ice suitable for ice radar data, and belongs to the field of ice radar data processing. The method comprises the following steps: generating a base reflectivity change curve, correcting a thawing and freezing screening threshold, automatically detecting a thawing and freezing transition zone, and predicting the ice thawing and freezing distribution under ice. In view of the error problem of the echo power attenuation term, the ice under water in the region is used as a reference to correct the detection threshold of the thawing and freezing of the region. According to the characteristics of the thawing and freezing transition zone, the corresponding features are defined and extracted, the automatic detection of the thawing and freezing transition zone is realized, and more accurate thawing and freezing distribution in the thawing and freezing transition zone can be obtained. Finally, more accurate ice thawing and freezing distribution under ice can be generated, and the method is suitable for different regions and a large amount of ice radar data.

Description

Technical Field

[0001] This invention relates to the field of ice radar data processing, and specifically to a semi-automatic prediction method for the distribution of sub-ice melting and freezing in ice radar data. Background Technology

[0002] With the development of research on Antarctic ice sheet mass balance and subglacial hydrological systems, the distribution of melting and freezing beneath the Antarctic ice sheet, as information describing the subglacial environment, can indicate the location of melting and freezing basement within the region, possessing significant research and application value. Currently, this technology is mainly implemented through two methods: thermodynamic model solving and radio echo sounding (RES) data processing. Solving the thermodynamic model yields the subglacial temperature distribution and basement melting rate distribution, thus predicting the subglacial melting and freezing distribution, which can quantitatively explain current and even past subglacial melting and freezing. However, this method typically has regional limitations and relies on existing, potentially inaccurate, geothermal flux as boundary conditions, leading to significant uncertainty in the predicted subglacial melting and freezing distribution. RES data processing, which calculates the basement reflectivity based on the echo power of the basement or other layers to predict the subglacial melting and freezing distribution, can be easily extended and applied to different radar systems and regions. It does not require the use of potentially inaccurate geothermal flux as boundary conditions, and the process is simpler with less computational burden. However, potential errors in estimating the echo power attenuation term can reduce the accuracy of the results. Furthermore, insufficient attention and inaccurate estimations of melt-freezing in regions where transitional wet and dry conditions are significant for research purposes lead to a decrease in the accuracy of predicted subglacial melt-freezing distribution. With the development of Antarctic subglacial exploration and the increasing demand for high-resolution Antarctic subglacial topographic maps, a large amount of RES data will be generated in the future. Given the advantages of RES data processing methods in handling large amounts of data, predicting subglacial melt-freezing distribution based on RES data processing has become a research hotspot.

[0003] To achieve more accurate subglacial melting and freezing distribution using RES data processing methods, this method addresses two key issues. First, by correcting the screening thresholds for subglacial melting and freezing states, it largely resolves the errors in the estimation of echo power attenuation. Second, focusing on the special melting-freezing transition zone, this method analyzes its characteristics and automatically extracts this transition zone, resolving the issues of insufficient attention to this zone and inaccurate melting-freezing estimation. This allows for a more accurate prediction of the subglacial melting and freezing distribution in the region. Summary of the Invention

[0004] To overcome the problems of estimation errors in the echo power attenuation term and insufficient attention to the melt-freeze transition zone, which reduce the accuracy of existing methods for predicting subglacial melt-freeze based on RES data processing, this invention proposes a semi-automatic prediction method for subglacial melt-freeze distribution based on ice radar data. The advantages of this method are: First, it uses subglacial water bodies within the region as a reference to correct the regional melt and freeze detection thresholds, greatly solving the error problem in the echo power attenuation term; Second, based on the characteristics of the melt-freeze transition zone, features are defined and extracted, enabling automatic detection of the melt-freeze transition zone and obtaining a more accurate melt-freeze distribution within it. Based on the corrected thresholds and the detected melt-freeze transition zone, this method can efficiently and accurately obtain the regional subglacial melt-freeze distribution. The overall flowchart of the method is as follows... Figure 1 As shown, it is mainly divided into four modules: the substrate reflectance change curve generation module, the thaw and freeze screening threshold correction module, the thaw and freeze transition zone automatic detection module, and the ice-freeze distribution prediction module.

[0005] Subsurface Reflectance Variation Curve Generation Module: Since most of the subsurface melting and freezing distributions in this method are obtained by filtering based on the substrate reflectance variation curve (representing the substrate reflectance) using melting and freezing thresholds, it is necessary to obtain the substrate reflectance variation curve from the echo amplitudes of the input air-ice interface and ice-subsurface interface. The flowchart of this module is as follows: Figure 2 As shown, the process includes three steps: calculating the echo attenuation term and the echo power at the air-ice interface and the ice-substrate interface; fitting the average ice absorption loss rate; and generating the substrate reflectivity change curve. In generating the substrate reflectivity change curve, transmission and reflection losses caused by multiple reflections from the intermediate layer are ignored. The substrate reflectivity change ΔR(dB) can be expressed as the echo power P from the air-ice interface. ai and echo power P from the ice substrate interface ib Represented as shown in Formula 1:

[0006] ΔR=P ib -P ai +L Ga -L Gb +R ai -2T ai -R ir -2L i (Formula 1)

[0007] The subscripts representing the interfaces are as follows: b represents the basement in an unknown state (melted or frozen), a represents air, i represents ice, and r represents bedrock. L Ga L represents the geometrical propagation loss of air. GbLet L represent the propagation loss at the substrate geometry, T represent the one-way transmission loss at different interfaces, R represent the reflection loss at different interfaces, and L represent the propagation loss at the substrate geometry. i The ice absorption loss term, ΔR, includes the change in basement reflectivity due to variations in bedrock surface water content. First, the echo attenuation term and the echo power at the air-ice interface and the ice-basement interface are calculated. The echo attenuation term in Equation 1 includes the geometric propagation loss L. Ga and L Gb The one-way transmission loss T at the air-ice interface ai Reflection loss R at the air-ice interface and ice-bedrock interface ai and R ir T ai R ai and R ir For the constant term whose theoretical value is known, L GQ (Q = a or b) can be expressed as (as shown in Formula 2):

[0008]

[0009] Among them G a Let λ be the antenna gain and λ be the radar wave wavelength. When Q = a, r Q =r a That is, the propagation distance of the radar wave to the air-ice interface. When Q = b, r Q =r b , where is the propagation distance of the radar wave to the substrate, and n is the refractive index of the air-ice interface. The echo power of the air-ice interface and the ice-substrate interface can be calculated using the following formula (as shown in Formula 3):

[0010] P = 10log(1000·A) 2 ) (Formula 3)

[0011] Where A represents the echo amplitude at the air-ice interface or ice-substrate interface. Thus, in Formula 1, except for L... i All terms other than ΔR are known, once L i Once a good estimate is obtained, ΔR can be calculated. Then, the average ice absorption loss rate is fitted, given L. i There is a linear model assumption for (dB) and depth z, i.e. (as shown in Equation 4):

[0012] in, It depends on the local ice state and temperature, representing the average ice absorption loss rate. According to Equation 1, L i (dB) can be expressed as (as shown in Equation 5):

[0013] L i +(1 / 2)ΔR=1 / 2(P ib -P ai +LGa -L Gb +R ai -2T ai -R ir ) (Formula 5)

[0014] By drawing L i The scatter plot uses a robust Random Consensus Linear Fitting (RANSAC) method to effectively mask the influence of outliers. The slope of the fitted linear curve is then obtained. Finally, the substrate reflectance change curve is generated, and L at any depth can be obtained according to Formula 4. i By substituting all known quantities into Formula 1, the base reflectance variation curves for all survey lines along the azimuth direction can be generated.

[0015] Melt-Freezing Screening Threshold Correction Module: In the module generating the substrate reflectivity change curve, some echo power attenuation terms—namely, the transmission attenuation T and reflection attenuation R at different interfaces—are known and are theoretically constant values ​​calculated by glaciologists. However, because the theoretical values ​​of each attenuation term may have regional errors in the actual ice sheet environment, and the losses during radar wave transmission may not be fully estimated, both factors will lead to inaccurate ΔR curves generated using these attenuation terms as input. Ideal melt-freezing and freezing thresholds may not accurately detect the subglacial environment. Since subglacial water is known to have relatively ideal reflectivity, and water near its pressure melting point has relatively stable physical properties, the melting-freezing and freezing screening thresholds within the region can be corrected using the subglacial water present in the area.

[0016] The method for correcting the melting screening threshold is to use the subglacial water within the region as a reference, calculate the average value of the ΔR curve within the subglacial water range, and denote the corrected melting screening threshold ΔR for the j-th subglacial water body. wet-c (j)(dB)(as shown in Formula 6):

[0017]

[0018] Where n1 and n2 determine the range of the subglacial water body, and ΔR(i) represents the ΔR value of the j-th subglacial water body at position i between n1 and n2.

[0019] For the case of multiple subglacial water bodies in a region, the final corrected melting screening threshold ΔR wet-c (dB) is expressed as (as shown in Equation 7):

[0020]

[0021] Where m represents the number of subglacial water bodies in the region. Each subglacial water body is assigned the same weight to obtain a more universal ΔR. wet-cAt the same time, the freeze screening threshold was adjusted to in It represents the reflectivity of a known frozen ice-bedrock interface.

[0022] Automatic Detection Module for Melt-Freeze Transition Zones: Based on the generated ΔR curve and the corrected threshold, we achieve automatic detection of all regions of this type according to the characteristics of the melt-freeze transition zone. The flowchart of this module is as follows. Figure 3 As shown. Within a short distance with similar topography and an approximately monotonic trend, we define the region where the basement reflectance is correlated with bedrock depth and exhibits melting and freezing transitions as the melting-freezing transition zone. Based on the assumption that the geographical structure and physical properties of the bedrock are stable over a short distance (i.e., the reflectance of the ice-bedrock interface is stable) and the understanding that the spatial variation in ice absorption attenuation rate will not have a major impact, within a short distance region where the basement topography satisfies an approximately monotonic trend, the basement reflectance in the melting-freezing transition zone will correspondingly show significant variations between the ice-bedrock interface reflectance and the ice-water interface reflectance, exhibiting a trend opposite to that of the basement topography.

[0023] Based on the characteristics of the thaw-freeze transition zone described above, we define five features to achieve automatic detection of the thaw-freeze transition zone. To ensure the capture of effective features corresponding to each azimuth sample point on the substrate and to guarantee the feature recognition capability, we first determine the feature calculation window. We use statistical methods to plot along the azimuth direction so that the reflectance corresponding to each azimuth sample generates a value of ΔR. D The minimum number of azimuth samples required to determine the change is found by analyzing the number of azimuth samples N corresponding to the first peak in the statistical distribution plot. W , which serves as the size of the feature calculation window W. Then, for any azimuth sample point x0: we define feature D to characterize whether a transition between thawing and freezing states occurs within the feature calculation window, and D is expressed as follows (as shown in Equation 8):

[0024] D=(ΔR max -ΔR min )-ΔR D (Formula 8)

[0025] Where, ΔR max and ΔR min These represent the maximum and minimum values ​​of ΔR within the calculation window, respectively. We define feature A to characterize whether the reflectance within the feature calculation window has a significant trend of change, and A is expressed as follows (as shown in Equation 9):

[0026]

[0027] in, It is the slope of ΔR within the window fitted by the least squares method. We define a feature S to characterize whether the terrain within the feature calculation window has an opposite trend to the reflectivity. S is expressed as follows (as shown in Equation 10):

[0028]

[0029] in N is the slope of the terrain within the window fitted using the least squares method. S This refers to the number of range-oriented sampling points corresponding to a range of one kilometer in the radar image. We define the feature. To characterize whether the reflectance within the feature calculation window has a large span. It is represented as follows (as shown in Equation 11):

[0030]

[0031] in This represents the ΔR value at the x-axis sample point. X0 represents the average value of ΔR within the calculation window corresponding to x0, where X0 = [x0 - n] x ,x0+n x ],n x =(N W -1) / 2. We define the feature. This is used to characterize whether the terrain trend within the feature calculation window has changed. It is represented as follows (as shown in Equation 12):

[0032]

[0033] This represents the elevation value of the ice base interface at the x-axis sample point. This represents the elevation value of the azimuth sample point x of the fitted straight line within the calculation window corresponding to x0, representing the terrain. The final feature vector is expressed as follows (as shown in Equation 13):

[0034]

[0035] After feature extraction, the final dataset, as defined by the thaw-freeze transition region, is an imbalanced dataset with an uneven number of positive and negative class samples. To reduce overfitting, we employ synthetic minority oversampling and the edit nearest neighbor algorithm (SMOTE+ENN) to balance the dataset. This is combined with a classification model to adapt to the imbalanced dataset and significantly reduce overfitting. After balancing, the dataset is only standardized to ensure broad applicability. We use a radial basis function (RBF) support vector machine (SVM) as the classification model to suit the dataset's characteristics. This is because the classifier must be adaptable to the imbalanced dataset. Furthermore, since we define features based on the most prominent features of the thaw-freeze transition region, ensuring feature effectiveness, a simple and fast classifier is ideal; therefore, we chose the SVM classifier. Due to the non-linear nature of the extracted features, an RBF kernel is used. This classification model ultimately enables automatic detection of the thaw-freeze transition region.

[0036] Subglacial melting and freezing distribution prediction module: The flowchart of this module is as follows Figure 4 As shown, based on the detected thaw-freezing transition zone and the thaw-freezing screening threshold, we generate the final subglacial thaw-freezing distribution for this region: First, as... Figure 7 The example shown is a magnified view of the reflectance change curve. The lowest point of the ΔR curve in the feature calculation window (black box) corresponding to the melt-freeze transition zone is taken as the reference freezing position; the difference between ΔR and the reference freezing position is greater than or equal to ΔR. D The location is determined as the reference melting position, and the midpoint of ΔR between the reference freezing position and the reference melting position is used as the boundary to determine the melting-freezing distribution within the melting-freezing transition zone. Areas above the boundary are considered likely to be melting, and areas below the boundary are considered likely to be frozen. Then, for other areas (areas outside the black box), freezing thresholds (such as...) are applied... Figure 7 ΔR wet-c (as shown by the corresponding horizontal line) and melting threshold (as shown by the horizontal line) Figure 7 ΔR dry-c (As shown by the corresponding horizontal line) Detect the frozen and thawed positions, i.e., ΔR ≤ ΔR dry-c Indicates the frozen position, ΔR ≥ ΔR wet-c This indicates the melting location, thus generating a complete distribution of the sub-ice melting and freezing state. The prediction effect of the sub-ice melting and freezing distribution is as follows: Figure 8 As shown.

[0037] Beneficial effects

[0038] This invention proposes a semi-automatic prediction method for the distribution of subglacial melt-freezing in ice radar data. This method can obtain more accurate regional subglacial melt-freezing distributions and, while adapting to regional characteristics, can be applied to different regions and large amounts of RES data. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.

[0040] Figure 2 This is a flowchart of the module for generating the substrate reflectivity change curve of the present invention.

[0041] Figure 3 This is a flowchart of the automatic detection module for the melting-freezing transition zone of the present invention.

[0042] Figure 4 This is a flowchart of the ice melting and freezing distribution prediction module of the present invention.

[0043] Figure 5 The curve showing the change in substrate reflectance is shown for an example survey line.

[0044] Figure 6 The results of the melting-freezing transition zone detection are shown for example survey lines.

[0045] Figure 7 This is a magnified view of the base reflectance change curve for an example survey line.

[0046] Figure 8 The results show the thaw and freeze distribution of an example survey line. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings.

[0048] Figure 1 This is a flowchart of the overall process of a semi-automatic prediction method for the distribution of ice melting and freezing under ice in ice radar data proposed in this invention. It includes a module for generating a curve of changes in substrate reflectivity, a module for correcting melting and freezing screening thresholds, a module for automatically detecting the melting and freezing transition zone, and a module for predicting the distribution of ice melting and freezing under ice.

[0049] Figure 2This is a flowchart for generating the base reflectivity change curve. The main steps include calculating the echo attenuation term and echo power, fitting the average ice absorption loss rate, and generating the base reflectivity change curve. First, high-gain focusing data for the selected survey line is obtained. Stratigraphic extraction is performed on the cross-sectional map generated for this survey line to obtain the stratigraphic information of the gas-ice interface and the ice-base interface, thus obtaining the echo amplitude and calculating the echo power. The geometric propagation loss of the radar wave is calculated based on the transmission depth, and the theoretical reflection attenuation and theoretical transmission attenuation of the radar wave at different interfaces are known. Then, using the echo power and various attenuation terms at the gas-ice interface and the ice-base interface, the ice absorption loss term L is plotted according to Formula 5. i Scatter plots of depth z and ice absorption loss were used to obtain the average ice absorption loss rate by fitting a linear slope using RANSAC to avoid interference from outliers and other outliers, based on a linear model of ice absorption loss. The ice absorption loss term is calculated using the linear model shown in Equation 4. Finally, the change curve of substrate reflectance is generated along the azimuth of the survey line using Equation 1. Figure 5 The result shows the basement reflectance variation curve for an example survey line 20100117_01_004. This curve represents the change in basement reflectance relative to the preset frozen bedrock reflectance for each azimuth sampling point. It differs from the actual basement reflectance by a constant offset and is representative of the basement reflectance. This basement reflectance variation curve forms the basis for subsequent thaw-freezing screening threshold correction and the automatic detection module for thaw-freezing transition zones.

[0050] For the melt-freeze screening threshold correction module, taking all subglacial water bodies within the region as references, the average reflectance change curve of each subglacial water body location is calculated. All subglacial water bodies are treated as having equal weight within the region, and the final average reflectance change curve of all subglacial water bodies is denoted as ΔR. wet-c This serves as a melting screening threshold. It is based on the known average reflectance of the frozen ice-bedrock interface. To define the reflectivity in the frozen state, the corresponding freeze screening threshold is adjusted to...

[0051] Figure 3 This is a flowchart of the automatic detection module for the thaw-freeze transition zone. The main process includes five steps: feature calculation window determination, effective segment selection, feature extraction, dataset equalization and standardization, and automatic detection of the thaw-freeze transition zone using a classification model. First, the feature calculation window is determined. Since the feature calculation window used for feature extraction has a significant impact on the feature recognition ability, we use statistical methods to draw a window along the azimuth direction, so that the reflectance corresponding to each azimuth sample generates a value of ΔR. DThe minimum number of azimuth samples required to determine the change in the value is determined by analyzing the statistical distribution plot to find the number of azimuth samples N corresponding to the first peak. W The size was set to N. W A feature calculation window of 21 was used. Ten survey lines were randomly selected within the Dome A region of Antarctica to cover a larger area and more diverse terrain. Next, effective segments were selected. Considering the resulting imbalanced dataset, segments containing thaw-freeze transition zones, other areas, and some special terrain features were chosen from these survey lines to balance the positive-to-negative sample ratio, training set size, and the completeness of the inclusion. Then, feature extraction was performed. The feature calculation window slid along the orientation towards each sample, calculating all defined features to complete feature extraction. This resulted in 5546 labeled samples with a positive-to-negative sample ratio of approximately 1:24, including 225 positive samples and 5321 negative samples (positive samples represent thaw-freeze transition zones, and negative samples represent other areas). Finally, the dataset was balanced and standardized. For a highly imbalanced dataset, the SMOTE+ENN algorithm was used to balance the dataset to minimize overfitting. The processed dataset had a positive-to-negative sample ratio of 1:5, which was used as input to the classification model. For the processed dataset, we only performed standardization to make the model universal. Then, we split the training and test sets in a 7:3 ratio. The test set was obtained by randomly sampling 30% of the dataset from the original dataset with the same positive and negative sample ratio as the original dataset. The remaining portion of the dataset was used as the training set, resulting in completely non-overlapping training and test set samples. Finally, the classification model automatically detected the thaw-freeze transition region. The classification model is based on an SVM classifier, using an RBF kernel, and the positive and negative sample weights are set to 5:1 to achieve final adaptation to the imbalanced dataset. We optimized the hyperparameters of the classification model using hierarchical 5-fold cross-validation to find the optimal classifier. The performance of the method was evaluated using five metrics based on the confusion matrix: accuracy, precision, recall, F1 score (2 * precision * recall / (precision + recall)), and AUC (area under the ROC curve). Our optimal classifier achieved the following results on the test set (all expressed as percentages): accuracy 99.83%, precision 99.15%, recall 99.49%, F1 score 99.49%, and AUC 99.99%, demonstrating excellent classification performance. Finally, the classification model was applied to the target region to automatically detect the melt-freeze transition zone in all test lines within the target region. Figure 6 The results of the thaw-freeze transition zone detected by applying the method to example survey line 20100117_01_004 are shown, where the positions shown as 1 all correspond to the thaw-freeze transition zone.

[0052] Figure 4This is a flowchart of the subglacial melt-freeze distribution prediction module. First, based on the detected melt-freeze transition zone, it generates the melt-freeze distribution within that zone. For example... Figure 7 As shown, we find the lowest point of the reflectance change curve within the thawing-freezing transition zone as the reference freezing position, which has a reflectance ΔR higher than the freezing position. D The first location is used as a reference melting location. Locations with reflectance greater than the reference melting location are considered to be in a melting state. The midpoint of reflectance between the reference melting and reference freezing locations is considered the boundary between melting and freezing. Using this boundary as a reference, the melting-freezing distribution between the reference freezing and reference melting locations is determined, thus generating the melting-freezing distribution within the melting-freezing transition zone. Then, for other regions, locations where the reflectance change curve is below the freezing threshold are identified as freezing locations, and locations where the reflectance change curve is above the melting threshold are identified as melting locations, generating the melting-freezing distribution for other regions. Finally, the two results are combined to obtain the final regional subglacial melting-freezing distribution. Figure 8 This shows the thawing and freezing distribution map for survey line 20100117_01_004. "o" represents the frozen base, and "-" represents the thawing base. For one of the subglacial water bodies, [the map is shown]. Figure 5 It can be seen that a very high reflectivity is observed in the water body beneath the ice, due to Figure 6 It was found that thawing-freezing transition zones were detected at both ends of the water body beneath the ice. Figure 8 The final predicted melting-freezing distribution detected a continuous melting state within the subglacial water body, clearly demonstrating the existence of the subglacial water body; at the same time, the boundary of the predicted melting state corresponds to the boundary of the subglacial water body, showing accurate prediction results.

[0053] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in various embodiments can also be appropriately combined and implemented according to the understanding of those skilled in the art.

[0054] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the inventive technique should be included within the scope of protection of the present invention.

Claims

1. A semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data, characterized in that... Includes the following steps: The generation stage of the substrate reflectivity change curve: (1) Calculate the echo power of the air-ice interface based on the echo amplitude of the input air-ice interface and ice substrate interface. Echo power at the ice substrate interface And calculate the geometric propagation loss term at the air-ice interface. Geometric propagation loss term to the ice substrate interface Obtain the one-way transmission loss at the air-ice interface for other known echo power attenuation terms. Reflection loss at the air-ice interface Reflection loss at the ice-bedrock interface Theoretical value; (2) Based on the obtained , , , , , and , generating ice absorption loss term and depth The scatter plot was used to perform a linear fit on the RANSAC method, with the slope of the fit being the average ice absorption loss rate. ; (3) Based on the obtained Substitute and A linear model is used to calculate the ice-base interface corresponding to samples in all orientations. And based on the change in substrate reflectivity The calculation formula generates along the azimuth direction. curve; Melting / Freezing Screening Threshold Correction Stage: (4) Using the subglacial water in the area where the method is to be applied as a reference, the threshold used to screen for melting and freezing states in the substrate reflectance change curve is corrected. and ; Automatic detection phase of the thaw-freeze transition zone: (5) According to The curve, along the azimuth direction, seeks the sample whose reflectance corresponds to a magnitude of [value missing]. To determine the minimum number of azimuth samples required to measure the change, a statistical distribution plot is drawn. Analysis of this plot reveals the minimum number of azimuth samples corresponding to the first peak. , as a feature calculation window The size, where This is the difference between the melting and freezing thresholds; (6) Based on the characteristics of the thawing-freezing transition zone, five features are defined to characterize: 1) whether a transition between thawing and freezing states occurs within the feature calculation window; 2) whether the reflectance within the feature calculation window has a significant trend; 3) whether the terrain within the feature calculation window has an opposite trend to the reflectance; 4) whether the reflectance within the feature calculation window has a large span; 5) whether the terrain trend within the feature calculation window has changed; finally, the feature calculation window is slid along the azimuth direction of all obtained survey line segments to calculate all defined features to complete feature extraction and obtain the original dataset; (7) The original dataset is balanced by using synthetic minority oversampling and edit nearest neighbor algorithm SMOTE+ENN. The balanced dataset is then standardized to obtain the input dataset for the classification model. (8) Input the dataset into a support vector machine (SVM) binary classification model with radial basis function (RBF) kernel to complete the binary classification of the thaw-freeze transition zone and other regions, thereby achieving automatic detection of the thaw-freeze transition zone; Prediction phase of subglacial melting and freezing distribution: (9) For the melting-freezing distribution in the melting-freezing transition zone, the melting-freezing distribution in the melting-freezing transition zone is generated using the reference freezing position and reference melting position inside the detected melting-freezing transition zone; for the melting-freezing distribution in other areas, the melting-freezing distribution in other areas is generated based on the corrected melting and freezing screening thresholds; the two results are merged to generate the sub-ice melting-freezing distribution of the region.

2. The semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data as described in claim 1, characterized in that: In the and Linear fitting was performed on the scatter plot to obtain In this study, a robust sample random consistency fitting method, RANSAC, is used to effectively shield the influence of outliers and obtain a more accurate linear fitting slope.

3. The semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data as described in claim 1, characterized in that: Using the subglacial water bodies included in the survey line within the region as a reference, the thresholds used to detect melting and freezing are corrected, and the calculation is performed. Within the subglacial water body The average value of the curve is denoted as the corrected melting screening threshold for the subglacial water body. (dB): (Official 1) in and The extent of the subglacial water body has been determined; Indicates the first Location of subglacial water place value; The final revised version regarding the situation of multiple subglacial water bodies in the region. (dB) is represented as: (Official 2) in This represents the number of subglacial water bodies present in the application area, with each body assigned the same weight for greater universality. ; The freeze screening threshold has been corrected to ,in This represents the reflectivity of the known frozen bedrock interface. This represents the reflectivity of an ideal ice-water interface.

4. The semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data as described in claim 1, characterized in that: Based on the characteristics of the thawing-freezing transition zone, five features are defined to enable automatic detection of the thawing-freezing transition zone in the survey line; For any azimuth sample point : Define features This characterizes whether a transition between thawed and frozen states occurs within the feature calculation window. It is represented as follows: (Official 3) in, and windows respectively Inside The maximum and minimum values; Define features To characterize whether the reflectance within the feature calculation window has a significant trend of change. It is represented as follows: (Official 4) in, It is within the window fitted by the least squares method. The slope; Define features To characterize whether the terrain within the feature calculation window and reflectivity exhibit opposite trends. It is represented as follows: (Official 5) in It is the slope of the terrain within the window fitted using the least squares method. It is the number of range samples corresponding to one kilometer in the range direction in the radar image; Define features To characterize whether the reflectance within the feature calculation window has a large span. It is represented as follows: (Official 6) in Represents the orientation of the sample point of value, represent exist The average value within the corresponding calculation window, , ; Define features This is used to characterize whether the terrain trend within the feature calculation window has changed. It is represented as follows: (Official 7) Represents the orientation of the sample point The elevation value of the ice base interface, Representing terrain in The azimuth sample points of the fitted line within the corresponding calculation window Elevation value; The final feature vector is represented as follows: (Official 8).

5. The semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data as described in claim 1, characterized in that: Since the input data for the classification model that realizes the automatic detection of the thaw-freeze transition zone is an imbalanced dataset, in order to take into account the ratio of positive and negative samples, dataset size and completeness of inclusion, a segment containing the thaw-freeze transition zone and other areas outside the thaw-freeze transition zone is selected in the test line for subsequent feature extraction.

6. The semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data as described in claim 1, characterized in that: For the datasets used to train and test the classification model, synthetic minority oversampling and the edit nearest neighbor algorithm SMOTE+ENN are used to balance the data. This, combined with the penalty parameter of the SVM classification model, helps to adapt to the imbalanced dataset. Only the processed dataset is standardized to make the classification model universal.

7. The semi-automatic prediction method for the distribution of sub-ice melting and freezing based on ice radar data as described in claim 1, characterized in that: The specific stages of the prediction for the distribution of subglacial melting and freezing are as follows: When generating the final subglacial thaw and freeze distribution for the region based on the detected thaw-freeze transition zone and the thaw-freeze screening threshold: First, for the reflectance change curve, the window corresponding to the thaw-freeze transition zone is... The lowest point of the curve is used as the reference freeze point; the curve will be compared with the reference freeze point. The difference is greater than The location is determined as the reference melting location, and the reference freezing location and the reference melting location are... The midpoint is used as the boundary to determine the thawing-freezing distribution within the thawing-freezing transition zone. Areas above the boundary are considered likely to be thawing, and areas below the boundary are considered likely to be frozen. Then, for other areas, freezing and thawing locations are detected using freezing and thawing thresholds respectively. Indicates the freeze position. This indicates the melting location, thus generating a complete distribution of the thawing and freezing states beneath the ice.

Images