A local incidence angle weighted SAR image ice cover freeze-thaw detection method

Abstract

This invention discloses a method for detecting ice sheet freeze-thaw cycles using SAR imagery with local incident angle weighting, belonging to the field of remote sensing image processing technology. The method acquires SAR images and topographic data of the same polarization (HH) and cross-polarization (HV) during the freezing and thawing periods of the ice sheet region, and performs preprocessing. Based on the topographic data, the local incident angle is calculated to obtain the scattering difference characterization values ​​of the same polarization and cross-polarization relative to the winter reference scattering state. A piecewise nonlinear weighting function is constructed based on the local incident angle to generate a fused difference map. Then, a two-dimensional Otsu algorithm is used for threshold segmentation to obtain the ice sheet freeze-thaw classification result map. This invention can reduce the impact of topographic undulations and speckle noise on the classification results and improve the accuracy of dry snow and wet snow identification.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically, to a method for detecting ice sheet freeze-thaw cycles in locally incident angle-weighted SAR images. Background Technology

[0002] The freeze-thaw status of ice sheets is crucial information reflecting changes in the polar cryosphere, and is of great significance for climate change research, ice sheet stability assessment, and sea-level change analysis. Due to the harsh environment and limited observation conditions in polar regions, large-scale, continuous freeze-thaw monitoring is difficult to achieve relying solely on manual field measurements. Synthetic Aperture Radar (SAR) imagery, with its all-weather, day-and-night observation capabilities, can acquire surface scattering information of ice sheets under cloud cover, polar night, and complex weather conditions, and is therefore widely used for identifying the freeze-thaw status of ice sheets.

[0003] Existing methods for detecting ice sheet freeze-thaw cycles typically construct change maps based on backscatter variations between SAR images of the freezing and thawing periods. Wet snow regions are then extracted using fixed empirical thresholds, single-polarization threshold segmentation, dual-polarization change map analysis, or the maximum inter-class variance threshold algorithm. These methods leverage the differences in dielectric properties and radar scattering responses between dry and wet snow to achieve a certain degree of ice sheet freeze-thaw monitoring. Specifically, HH and HV polarizations exhibit different responses to dry and wet snow under different local incident angles, providing polarization information of different dimensions for freeze-thaw discrimination. Local incident angles and topographic features can also be used for backscatter correction, classification assistance, or topographic zoning.

[0004] However, in ice sheet regions with significant topographic relief, the local incident angles of different pixels vary, causing fluctuations in the backscattering response under the same freeze-thaw condition that are not caused by freeze-thaw factors. Simultaneously, the sensitivity of HH and HV polarizations to freeze-thaw conditions differs with changes in the local incident angle. Existing methods, even when considering local incident angles or topographic factors, typically use them as normalization correction parameters, classification aids, or topographic zoning criteria. They struggle to adjust the contribution ratio of the two polarization information in the fused difference map pixel-by-pixel based on the varying distinguishing abilities of co-polarization and cross-polarization for dry and wet snow under different local incident angles. Furthermore, SAR speckle noise affects the threshold determination of the difference map and the fused difference map, leading to an underestimation of wet snow extent, local misclassification, or blurred boundaries. Therefore, existing technologies still suffer from insufficient accuracy in ice sheet freeze-thaw detection under complex terrain and noise interference conditions. There is an urgent need for a SAR image-based ice sheet freeze-thaw detection method that can combine local incident angle variation patterns, HH / HV polarization sensitivity differences, and neighborhood spatial information. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for detecting ice sheet freeze-thaw in SAR images with local incident angle weighting, so as to solve the problem that existing SAR image-based ice sheet freeze-thaw detection methods are easily affected by local incident angle changes, terrain undulations and SAR speckle noise in complex terrain areas, resulting in insufficient classification accuracy of wet snow areas and dry snow areas.

[0006] To address the aforementioned technical problems, this invention provides a method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR images. This method acquires HH and HV polarimetric SAR images and topographic data of the ice sheet region during the freezing and thawing periods. The SAR images are then registered, radiometrically corrected, topographically corrected, and noise-suppressed. Based on the topographic data, the local incident angle θ is calculated. Furthermore, based on the pre-processed SAR images from the freezing and thawing periods, the scattering difference characteristics of co-polarimetric HH relative to the winter average scattering value and the scattering difference characteristics of cross-polarimetric HV relative to the winter average scattering value are calculated, and denoted as R0. HH and R HV ; Construct a piecewise nonlinear function model based on the local incident angle θ to determine the weights W; Use the weights W to apply to R HH and R HV Perform weighted fusion to generate a fusion difference map R. c Finally, the two-dimensional Otsu algorithm was used to fuse the difference map R. c The data is segmented to obtain the classification result image.

[0007] Specifically, the method includes the following steps:

[0008] S1. Acquire HH and HV polarimetric SAR images and topographic data of the ice sheet region during the freezing and thawing periods, and perform registration, radiometric correction, topographic correction and noise suppression on the SAR images.

[0009] S2. Calculate the local incident angle θ based on the terrain data, and calculate the scattering difference characterization values ​​of co-polarized HH relative to the winter average scattering value and cross-polarized HV relative to the winter average scattering value based on the preprocessed SAR images of the freezing and thawing periods, respectively, and denot them as R. HH and R HV ;

[0010] S3. Construct a piecewise nonlinear function model based on the local incident angle θ, and determine the weights W;

[0011] S4. Using weight W, the scattering difference characterization values ​​corresponding to the same polarization and cross-polarization are weighted and fused to generate a fused difference map R. c ;

[0012] S5. Using the two-dimensional Otsu algorithm to fuse the difference map R c The data is segmented to obtain the classification result image.

[0013] Furthermore, the HH polarization is homopolarization, and the HV polarization is cross-polarization. The R... HH The R is used to characterize the scattering difference of the same polarized HH relative to the winter average scattering value. HV This is used to characterize the scattering difference between cross-polarized HV and the winter average scattering value. R is calculated separately. HH and R HV This can preserve the different response characteristics of co-polarization and cross-polarization in ice sheet freeze-thaw detection.

[0014] Furthermore, the local incident angle θ is used to characterize the angle between the radar incident beam and the local surface normal vector. Since the topographic undulations in the ice sheet region cause changes in radar observation geometry at different locations, thus affecting the backscatter response, the introduction of the local incident angle θ can more accurately reflect the modulation effect of the topography on the SAR scattered echo.

[0015] Furthermore, the weight W represents the contribution ratio of cross-polarized HV in data fusion, and 1-W represents the contribution ratio of homopolarized HH in data fusion. The larger W is, the more emphasis is placed on the freeze-thaw sensitivity characteristics of cross-polarized HV during fusion; the smaller W is, the more emphasis is placed on the freeze-thaw sensitivity characteristics of homopolarized HH during fusion.

[0016] Furthermore, the piecewise nonlinear function model is as follows:

[0017]

[0018] Where a, b, and c are coefficients determined by the fitting process. In one specific implementation, a = 0.00079, b = 0.184, and c = 0.4.

[0019] Furthermore, the fusion difference map R c Characterized by weight W and cross-polarization scattering difference value R HV Characterization value R of the difference between the same polarization scattering HH Confirmed, expressed as:

[0020]

[0021] Furthermore, the two-dimensional Otsu algorithm is based on the fusion difference map R. c Two-dimensional feature pairs are constructed using pixel grayscale and neighborhood average grayscale, and the optimal segmentation threshold is determined based on the principle of maximizing inter-class dispersion to segment and fuse the difference map R. c The classification result image is obtained. By simultaneously considering pixel grayscale and neighborhood average grayscale, the impact of SAR speckle noise and local abnormal pixels on threshold determination can be reduced.

[0022] This invention uses the local incident angle as a control variable to determine the contribution ratio of co-polarized HH and cross-polarized HV in the fusion difference map. By analyzing 800 sets of measured snow and ice scattering samples, the variation in the ability of co-polarized HH and cross-polarized HV to distinguish between dry and wet snow within different local incident angle intervals was statistically determined. Based on this, three segment nodes at 20°, 40°, and 50° were established, enabling the fusion process to adjust the participation ratio of different polarization channels according to changes in the local incident angle.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects:

[0024] First, this invention uses the local incident angle as a control variable for the fusion contribution of co-polarized HH and cross-polarized HV, and determines three segment nodes of 20°, 40° and 50° based on 800 sets of measured ice and snow scattering samples. This allows the two polarization information under different local incident angle conditions to participate in the fusion according to their differences in distinguishing between dry and wet snow, thereby reducing the impact of topographic undulations and local incident angle differences on the ice sheet freeze-thaw detection results.

[0025] Second, this invention generates a fused difference map based on the scattering difference characterization values ​​of co-polarized HH and cross-polarized HV relative to the average scattering state in winter. Compared with single polarization channel, fixed weight fusion or uniform incident angle normalization, it can more fully express the scattering difference between dry snow and wet snow under different local incident angle conditions, and improve the fused difference map's ability to characterize freeze-thaw boundaries and complex terrain areas.

[0026] Third, this invention employs a piecewise nonlinear function model to determine the contribution weights, causing the contribution weights to exhibit quadratic decay in the 20° to 40° range, exponential convergence in the 40° to 50° range, and tending towards equal-weighted fusion when the local incident angle is greater than 50°. This model can simultaneously adapt to the two-stage variation law of rapid decay and slow convergence of cross-polarized HV sensitivity, and has a better fitting effect compared to linear functions, exponential functions, and single quadratic functions.

[0027] Fourth, this invention utilizes a two-dimensional Otsu algorithm to segment the fused difference map. The threshold determination process considers both pixel grayscale and neighborhood average grayscale, which reduces the impact of SAR speckle noise and isolated anomalous pixels on the classification results. Experimental results show that the overall accuracy of the method in this invention is 88.8% for wet snow and 84.3% for dry snow, which is higher than that of the single-polarization SAR fixed threshold method and the dual-polarization one-dimensional Otsu method, thus improving the accuracy and stability of ice sheet freeze-thaw classification results. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the method for detecting ice sheet freeze-thaw in local incident angle-weighted SAR images according to the present invention;

[0029] Figure 2 This is a schematic diagram showing the fitting curve between the local incident angle and the weight W in this invention, as well as a comparison of different function models;

[0030] Figure 3 This is a flowchart of the freeze-thaw classification of ice sheets based on the two-dimensional Otsu algorithm in this invention;

[0031] Figure 4 A comparison chart of ice sheet freeze-thaw classification results obtained by different methods. Detailed Implementation

[0032] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0033] This invention provides a method for detecting ice sheet freeze-thaw cycles using SAR imagery with local incident angle weighting. Based on the correlation between local incident angle and the HH / HV polarization scattering difference characterization value in snow melt detection, this method constructs a fused difference map with local incident angle as the weight, and uses a two-dimensional Otsu algorithm to classify the fused difference map, thereby obtaining a classification result map of ice sheet freeze-thaw cycles.

[0034] In this invention, SAR stands for Synthetic Aperture Radar. HH polarization represents horizontal transmission and horizontal reception polarization, which is co-polarization; HV polarization represents horizontal transmission and vertical reception polarization, which is cross-polarization. Due to the differences in dielectric properties, liquid water content, and scattering mechanisms between dry and wet snow, their backscattering responses under HH and HV polarizations differ. Furthermore, this difference is influenced by local incident angles and terrain undulations. Therefore, this invention uses the local incident angle as a crucial basis for determining the weight W, enabling co-polarization and cross-polarization to participate in fusion according to their freeze-thaw sensitivities under different local incident angle conditions.

[0035] Example 1

[0036] This embodiment provides a method for detecting ice sheet freeze-thaw cycles using locally incident angle-weighted SAR imagery. (Refer to...) Figure 1 The method includes data acquisition and preprocessing, calculation of local incident angle, calculation of the characterization values ​​of the difference between homopolarized and crosspolarized scattering, construction of piecewise nonlinear function model, and fusion of difference map R. c The process includes steps such as generation and two-dimensional Otsu classification.

[0037] Step S1: Acquire HH and HV polarimetric SAR images and topographic data of the ice sheet region during the freezing and thawing periods, and preprocess the SAR images.

[0038] Specifically, the ice sheet region can be the Antarctic ice sheet, Antarctic ice shelf, Greenland ice sheet, mountain glaciers, or other ice-covered areas experiencing seasonal freeze-thaw changes. SAR images during the freezing period are used to characterize the reference scattering state when there is no significant melting on the ice sheet surface, while SAR images during the ablation period are used to characterize the scattering state to be classified when freeze-thaw changes occur on the ice sheet surface.

[0039] The SAR imagery includes at least HH polarization SAR imagery and HV polarization SAR imagery. HH polarization is used to characterize the scattering response of the ice sheet surface under the same polarization channel, while HV polarization is used to characterize the scattering response of the ice sheet surface under the cross-polarization channel. The terrain data can be digital elevation model data used to calculate the local incident angle.

[0040] Preprocessing of SAR images includes registration, radiometric correction, topographic correction, and noise suppression. Registration is used to ensure that frozen and ablated SAR images correspond to each other under the same spatial reference; radiometric correction is used to convert SAR images into comparable backscattered representations; topographic correction is used to reduce geometric distortion caused by topographic relief; and noise suppression is used to reduce the impact of thermal noise and SAR speckle noise on subsequent processing.

[0041] In one specific implementation, the reference scattering value for the freezing period is obtained by averaging multiple SAR images from the freezing period. Specifically, a Sentinel-1 co-orbit HH / HV dual-polarized GRD image from the stable freezing period during the Antarctic winter polar night can be selected as the reference image for the freezing period. The stable freezing period can be from June to August each year. The selection criteria for the reference image for the freezing period include: the study area is in a state without significant ablation, the image covers the entire study area, and images with abnormal stripes or significant interference are excluded. In this embodiment, two SAR images from the stable freezing period in winter are selected, and their backscattering values ​​are averaged over time to reduce the impact of single-image noise and local topographic distortion on subsequent calculations of scattering differences.

[0042] Step S2: Calculate the local incident angle θ based on terrain data, and calculate the same polarization scattering difference characterization value R based on the preprocessed SAR images during the freezing and ablation periods. HH Characterization value R of cross-polarization scattering difference HV .

[0043] Specifically, the surface slope, aspect, or local normal vector can be determined based on digital elevation model data, and the local incident angle θ can be determined based on the angle between the radar incident direction and the local surface normal vector. The local incident angle θ is used to reflect the influence of terrain undulation on radar observation geometry and the modulation of backscattered echoes.

[0044] Compared with directly using the ellipsoidal incident angle, the local incident angle can more accurately characterize the influence of terrain slope and aspect on SAR scattering signals, and is therefore more suitable for freeze-thaw detection in complex terrain ice cover areas.

[0045] After obtaining the preprocessed SAR images during the frozen period and the SAR images during the ablation period, the scattering difference characterization value R between the same polarization HH and the winter average scattering value was calculated. HH And the scattering difference characterization value R between cross-polarized HV and the winter average scattering value HV R HH R is used to characterize the scattering difference between the freezing and ablation periods under the same polarization channel. HV Used to characterize the scattering difference between the freezing and ablation periods under cross-polarization channels.

[0046] In one specific implementation, the scattering difference characterization values ​​of co-polarization and cross-polarization relative to the winter average scattering value can be determined based on the backscattering value and the winter average scattering value. When using a linear-scaled backscattering value, it can be expressed as:

[0047]

[0048] When using dB-scale backscattering values, it can be equivalently expressed as:

[0049]

[0050] Among them, S XX S represents the backscattering value of the period to be classified corresponding to XX polarization. avg,XX This represents the average backscattering value in winter corresponding to polarization XX. This represents the dB-scale backscattering value corresponding to the period to be classified for polarization XX. This represents the average dB-scale backscattering value corresponding to the XX polarization during winter. When, the scattering difference characterization value corresponding to the same polarization HH is obtained; when At that time, the scattering difference characterization value corresponding to the cross-polarization HV is obtained. In this embodiment, the dB scale is used for scattering difference calculation.

[0051] The radar backscattering characteristics of snow and ice surfaces are influenced by a variety of factors, including the dielectric properties of the snow itself, the radar operating wavelength, the surface roughness, the incident angle, the polarization combination, and the regional topographic conditions. The difference in scattering intensity between dry and wet snow exhibits different patterns with varying radar incident angles, and shows significant differences under conditions of same polarization and cross-polarization.

[0052] In the same polarization observation mode, the increase in surface scattering contribution will improve the overall backscattering coefficient; in the cross polarization channel, the difference in scattering signals between wet snow and dry snow is more prominent under low local incident angle conditions, which can more clearly reflect the changes in the water content of the snow layer and provide effective features for freeze-thaw discrimination.

[0053] When the local incident angle is below 20°, the backscattered signals of dry and wet snow regions differ significantly under the cross-polarization HV method. Using the cross-polarization scattering difference characterization value as the discrimination criterion can effectively distinguish between dry and wet snow. When the local incident angle is between 20° and 50°, the ability of co-polarization HH and cross-polarization HV to distinguish between dry and wet snow changes. The sensitivity of cross-polarization HV decreases as the local incident angle increases, while the contribution of co-polarization HH gradually increases. When the local incident angle is greater than 50°, the reference weights of co-polarization HH and cross-polarization HV in freeze-thaw discrimination tend to be close.

[0054] In one specific implementation, to further illustrate the ability of co-polarization and cross-polarization to distinguish between dry and wet snow under different local incident angles, the differences in backscattering between dry and wet snow within different local incident angle intervals can be statistically analyzed. The statistical results are shown in Table 1.

[0055] Table 1. Differences in backscattering of dry and wet snow under different local incident angle ranges.

[0056]

[0057] Table 1 shows that within the range of local incident angles less than 20°, the difference in cross-polarization scattering is greater than that in co-polarization scattering, indicating that cross-polarization has a stronger ability to distinguish between dry and wet snow. Within the range of 20° to 40°, the difference in co-polarization scattering increases significantly with increasing local incident angle, while the difference in cross-polarization scattering, although varying, remains lower than that in co-polarization. Within the range of 40° to 50°, the scattering differences of both co-polarization and cross-polarization begin to decrease. After the local incident angle exceeds 50°, the scattering differences corresponding to the two polarization modes tend to converge. These statistical results provide data support for constructing a piecewise nonlinear function model using 20°, 40°, and 50° as piecewise nodes.

[0058] Furthermore, the three segment nodes of 20°, 40°, and 50° were defined based on the statistical results of 800 sets of measured snow and ice scattering samples. Specifically, the sample statistical results show that when the local incident angle is less than 20°, the difference between dry and wet snow scattering corresponding to cross-polarized HV is significantly higher than that of homopolarized HH, making it suitable for cross-polarized HV to dominate the fusion; in the 20° to 40° range, the distinguishing advantage of cross-polarized HV decreases rapidly, while the difference between dry and wet snow scattering of homopolarized HH gradually increases, thus making it suitable for the contribution weight corresponding to cross-polarized HV to decay rapidly; in the 40° to 50° range, the decay rate of cross-polarized HV slows down, and the distinguishing ability of homopolarized HH and cross-polarized HV gradually approaches that of cross-polarized HV, thus making it suitable for the contribution weight to converge slowly; when the local incident angle is greater than 50°, the distinguishing ability of the two polarization modes for dry and wet snow tends to be similar, thus making it suitable for the two to participate in the fusion in an equal weight manner.

[0059] Based on the above principles, the relationship between co-polarization and cross-polarization and the local incident angle is crucial for enhancing ice sheet freeze-thaw inversion. Therefore, this invention utilizes the local incident angle θ and the characterization value R of the co-polarization scattering difference. HH Cross-polarization scattering difference characterization value R HV The relationships between them are used to construct a fusion difference diagram.

[0060] Step S3: Construct a piecewise nonlinear function model based on the local incident angle θ, and determine the weights W.

[0061] In this embodiment, the weight W represents the contribution ratio of cross-polarized HV in data fusion, and 1-W represents the contribution ratio of homopolarized HH in data fusion. The larger W is, the more emphasis is placed on the freeze-thaw sensitivity characteristics of cross-polarized HV during fusion; the smaller W is, the more emphasis is placed on the freeze-thaw sensitivity characteristics of homopolarized HH during fusion.

[0062] The core of constructing the piecewise nonlinear function model is establishing the correspondence between the local incident angle θ and the weight W. Since the difference in backscattering between dry and wet snow does not change strictly linearly with the local incident angle, the cross-polarization HV sensitivity exhibits a process of rapid decay followed by slow convergence within the 20° to 50° range. A single linear function, a single exponential function, or a single quadratic function is insufficient to fully capture this variation. Therefore, this invention employs a piecewise nonlinear function model to determine the weight W.

[0063] In this embodiment, the weight W is determined according to the following piecewise nonlinear function model:

[0064]

[0065] Where a, b, and c are coefficients. In a preferred embodiment, a = 0.00079, b = 0.184, and c = 0.4. These coefficients are the final fixed parameters obtained based on 800 sets of measured ice and snow scattering samples and combined with double-boundary constraint fitting.

[0066] The technical implications of the above piecewise nonlinear function model are as follows:

[0067] Within the 20° to 40° range, as the local incident angle increases, the SAR microwave incident path gradually tilts, and the sensitivity of cross-polarized HV to changes in the liquid water content inside the snow rapidly decreases, manifested as a rapid attenuation of the difference in backscattering between dry and wet snow. Therefore, a quadratic attenuation method is used to describe the change in contribution weights in this range. Within the 40° to 50° range, the attenuation rate of cross-polarized HV sensitivity gradually slows down, and the ability of co-polarized HH and cross-polarized HV to distinguish between freeze-thaw states gradually becomes similar. Therefore, an exponential convergence method is used to describe the process of contribution weights approaching an equally weighted fusion state.

[0068] When θ < 20°, the cross-polarized HV is more sensitive to the backscattering difference of snow freeze-thaw and has better anti-interference ability. Therefore, W = 1 is fixed to make the cross-polarized HV dominate in the fusion process.

[0069] When 20°≤θ≤40°, as the local incident angle increases, the sensitivity of cross-polarized HV to freeze-thaw conditions decreases rapidly, while the contribution of co-polarized HH gradually increases. Therefore, a quadratic decay function is used to describe the rapid decrease of weight W in this interval.

[0070] When 40°<θ≤50°, the sensitivity decay rate of cross-polarized HV begins to slow down, and the contributions of co-polarized HH and cross-polarized HV gradually approach each other. Therefore, an exponential convergence function is used to make the weight W gradually approach 0.5.

[0071] When θ > 50°, the contributions of co-polarized HH and cross-polarized HV in freeze-thaw discrimination tend to be close. Therefore, W = 0.5 is fixed to make the two equally weighted.

[0072] To determine the function model, this embodiment selects dry snow pixels and wet snow pixels as sample points, and statistically analyzes the ability of co-polarization (HH) and cross-polarization (HV) to distinguish freeze-thaw states under different local incident angles. Linear functions, exponential functions, quadratic functions, and piecewise nonlinear functions are used to fit the relationship between the local incident angle θ and the weight W. By comparing the coefficient of determination R², mean square error (MSE), and mean absolute error (MAE), the piecewise nonlinear function model is determined as the weight function.

[0073] In one specific implementation, 800 real dry and wet snow sample points are selected for fitting analysis. These 800 real dry and wet snow sample points are mixed samples of dry and wet snow, with a generally balanced sample size, but not limited to 400 dry snow samples and 400 wet snow samples. These sample points are based on NDWI generated from Sentinel-2 L1C imagery. ice Binary reference image selection, where NDWI ice The binary reference map is generated by combining ice and snow bands corresponding to the ice surface normalized water index, and is divided into dry snow and wet snow regions based on an adaptive threshold of the whole scene histogram. During sample selection, the NDWI... ice Uniform random sampling was performed within the dry and wet snow areas marked on the binary reference map, and invalid pixels corresponding to shadows, ice cliffs, seawater, and exposed bedrock were manually removed. The sample points covered a local incident angle range of 10° to 70°, with sample points allocated within each 10° local incident angle segment to avoid fitting bias caused by missing local incident angle intervals. The sample values ​​for contribution weights were determined based on the relative magnitude of the backscattering differences between co-polarized HH and cross-polarized HV for dry and wet snow. Specifically, within the same local incident angle interval, when the backscattering difference between dry and wet snow corresponding to cross-polarized HV was greater than that corresponding to co-polarized HH, the contribution weight corresponding to cross-polarized HV was increased; when the backscattering differences between dry and wet snow corresponding to co-polarized HH and cross-polarized HV gradually approached each other, the contribution weight corresponding to cross-polarized HV gradually approached the remaining proportion corresponding to co-polarized HH. The resulting contribution weight sample values ​​are directly derived from the statistical differences in scattering of dry and wet snow by different polarization modes, rather than from a reverse calculation based on classification accuracy or statistical distance. Comparison of fitting curves for different function models is shown below. Figure 2 As shown in Table 2, the goodness-of-fit indices for different function models are presented.

[0074] Table 2 Comparison of goodness-of-fit indices for four function models

[0075]

[0076] Table 2 shows that the piecewise nonlinear function model has the highest coefficient of determination R², and the lowest mean square error (MSE) and mean absolute error (MAE), indicating that this model can more accurately characterize the correspondence between the local incident angle θ and the weight W. Therefore, the piecewise nonlinear function model does not arbitrarily fit the relationship between the local incident angle and the weight, but rather corresponds to the two-stage variation law of rapid decay of cross-polarization HV sensitivity in the 20° to 40° interval and slow convergence of cross-polarization HV sensitivity in the 40° to 50° interval. A single linear function, exponential function, or quadratic function cannot simultaneously adapt to the above two-stage variation characteristics, while the piecewise nonlinear function model can adopt different variation forms in different angle intervals, making the contribution weights match the actual scattering statistics.

[0077] Specifically, while the linear function model is simple in form, its weight change trend is too simplistic, failing to reflect the two-stage variation of cross-polarization HV sensitivity: rapid decay in the 20° to 40° range and slow convergence in the 40° to 50° range. The exponential function model can describe the decay trend, but it struggles to simultaneously account for the rapid decline in the low-angle range and the relatively stable change in the high-angle range. The quadratic function model provides a better fit than the linear and exponential functions, but it still cannot simultaneously adapt to the different decay mechanisms in different local incident angle ranges. The piecewise nonlinear function model, by setting piecewise nodes at 20°, 40°, and 50°, allows the weights W to vary in different angle ranges, thus better reflecting the actual variation of cross-polarization HV sensitivity with local incident angles.

[0078] Step S4: Weight W is used to weight and fuse the characteristic values ​​of the difference between co-polarized and cross-polarized scattering to generate a fused difference map R. c .

[0079] In obtaining the characterization value R of the difference in isopolarization scattering HH Cross-polarization scattering difference characterization value R HV And after weight W, use weight W to apply to R HH and R HV Perform weighted fusion to generate a fusion difference map R. c In one specific implementation, the difference map R is fused. c Represented as:

[0080]

[0081] Among them, R HV R represents the scattering difference characterization value of cross-polarized HV relative to the winter average scattering value. HH The scattering difference value represents the difference between the same polarization HH and the average winter scattering value, W represents the contribution ratio of the cross-polarization HV in the fusion, and 1-W represents the contribution ratio of the same polarization HH in the fusion. In this embodiment, the fusion difference map is directly generated by linear weighted fusion. After generating the fusion difference map, no additional normalization, smoothing filtering, threshold constraint, outlier removal, or masking processing is performed. Instead, the fusion difference map is directly input into the two-dimensional Otsu algorithm for segmentation.

[0082] The above fusion method allows for the organic combination of different polarization information with terrain angular features. Compared to using homopolarized or cross-polarized data alone, the fused difference map R... cIt can simultaneously utilize the different response characteristics of the same polarization (HH) and cross-polarization (HV) to the freeze-thaw state of the ice sheet; compared with fixed-weight fusion, this invention dynamically determines the weight W through the local incident angle θ, so that the fused difference map R c It can better adapt to changes in polarization sensitivity under different terrain and local incident angle conditions.

[0083] In regions with low local incident angles, the fusion difference map R c More emphasis is placed on cross-polarization HV information; in the medium local incident angle region, as the sensitivity of cross-polarization HV decreases, the fusion difference map R... c The contribution of co-polarized HH gradually increases; in regions with higher local incident angles, co-polarized HH and cross-polarized HV tend to participate in fusion with equal weight. Therefore, the fusion difference diagram R... c It can more accurately depict the scattering differences between dry and wet snow.

[0084] Step S5: Use the two-dimensional Otsu algorithm to segment and fuse the difference map R. c The classification result image is obtained. The specific classification process is as follows: Figure 3 As shown.

[0085] In obtaining the fusion difference map R c Then, the two-dimensional Otsu algorithm was used to fuse the difference map R. c Classification is performed. Due to speckle noise in SAR images, segmentation using only a one-dimensional grayscale threshold is easily affected by isolated noise points and local anomalous scattering values, leading to unstable threshold determination. The two-dimensional Otsu algorithm considers both pixel grayscale and neighborhood average grayscale, which can utilize neighborhood spatial information to improve the noise resistance of threshold segmentation.

[0086] Specifically, let the fusion difference graph R be... c The grayscale level is divided into L levels, and the average grayscale level of the pixel's neighborhood is also divided into L levels. For the fused difference map R... c For each pixel in the dataset, calculate the average gray level of its neighborhood, and construct a two-dimensional feature pair using the pixel's gray level and the average gray level of its neighborhood. In this embodiment, the average gray level of the neighborhood is calculated using a fixed 3×3 neighborhood window. Let the two-dimensional feature pair... The frequency of occurrence is Then the corresponding joint probability density for:

[0087]

[0088] Where N is the fusion difference map R c The total number of pixels in the image.

[0089] Let the candidate threshold be Where s is the pixel grayscale threshold and t is the neighborhood average grayscale threshold. Based on the candidate threshold... The two-dimensional grayscale space is divided into background class B and target class T. The class probabilities of background class B and target class T are as follows:

[0090]

[0091]

[0092] The mean vector of background class B can be represented as:

[0093]

[0094] The mean vector of the target class T can be represented as:

[0095]

[0096] Let the population mean vector of the two-dimensional grayscale space be:

[0097]

[0098] Candidate threshold The corresponding inter-class dispersion can be expressed as:

[0099]

[0100] The optimal segmentation threshold is:

[0101]

[0102] Based on the optimal segmentation threshold, the fusion difference map R is analyzed. c Segmentation is performed, and the difference map R is fused. c The backscattering difference direction is used to divide the image region into wet snow region and dry snow region, resulting in a classification result image.

[0103] Through the aforementioned two-dimensional Otsu algorithm, the classification process not only utilizes the fused difference map R... c The grayscale information is used, and the average grayscale of the neighborhood is used to reflect the local spatial consistency, thereby reducing the impact of SAR speckle noise on threshold determination and improving the stability of the classification result image.

[0104] Example 2

[0105] This embodiment provides an application of the method of the present invention in the Larsen C ice shelf region of Antarctica.

[0106] In one specific application, Sentinel-1 SAR images of the Larsen C ice shelf region during both the freezing and thawing periods are acquired, along with digital elevation model data covering the area. In one specific implementation, the SAR images can be HH and HV polarimetric SAR images acquired by the Sentinel-1 satellite, using the IW mode. The terrain data uses GETASSE30 30 arcsecond global terrain data, combined with radar local incidence angle data acquired by the platform for subsequent weight determination. Sentinel-1 images have the capability to acquire HH, HV, and HH+HV polarizations, and can characterize the difference in scattering across the same polarization, Ri. HH Characterization value R of cross-polarization scattering difference HV The calculations provide the data foundation. After registration, radiometric correction, topographic correction, and noise suppression of the SAR images, frozen-period SAR images and ablation-period SAR images with comparable scattering expressions are obtained.

[0107] Subsequently, the local incident angle θ within the study area was calculated based on digital elevation model data, and the characteristic value R of the same polarization scattering difference was calculated respectively. HH Characterization value R of cross-polarization scattering difference HV A piecewise nonlinear function model is constructed based on the local incident angle θ, and the weights W are determined; the weights W are then used to adjust R. HH and R HV Perform weighted fusion to generate a fusion difference map R. c Finally, the two-dimensional Otsu algorithm was used to fuse the difference map R. c The data is segmented to obtain the classification result image.

[0108] To verify the classification effect of the method of the present invention, this embodiment utilizes Sentinel-2 images and NDWI. ice A freeze-thaw reference image of the ice sheet was generated and used as the basis for evaluating classification accuracy. The Sentinel-2 imagery used L1C level data and NDWI. ice The reference image is generated based on the combination of ice and snow bands corresponding to the normalized water index of the ice surface, and dry snow and wet snow regions are divided by adaptive thresholding using the whole-scene histogram. After generating the binary reference image, minor manual corrections are made to the pixels confused with ice cliffs and seawater, and the corrected reference image is used as the true value to verify the classification accuracy. Based on the same study area, classification is performed using the single-polarization fixed threshold method, the dual-polarization one-dimensional Otsu method, and the method of this invention, respectively, to obtain the classification results of different methods. The ice sheet freeze-thaw classification results obtained by different methods are compared as follows: Figure 4 As shown.

[0109] In one specific implementation, the classification accuracy of different methods is compared in the table below.

[0110] Table 3 Comparison of classification accuracy of different methods on the reference image

[0111]

[0112] Among them, the overall accuracy of wet snow and the overall accuracy of dry snow are the classification accuracy of the corresponding categories. Table 3 shows that the overall accuracy of the method of this invention is 88.8% for wet snow and 84.3% for dry snow, both higher than the single-polarization SAR fixed threshold method and the dual-polarization one-dimensional Otsu method. The above results demonstrate that this invention, by constructing a piecewise nonlinear weighting function based on the local incident angle and utilizing the weight W on R... HH and R HV Perform weighted fusion to generate a fusion difference map R c This allows for a better representation of the scattering differences between dry and wet snow; simultaneously, the difference map R is segmented and fused using the two-dimensional Otsu algorithm. c This can reduce the impact of noise on threshold determination, thereby improving the accuracy and stability of the classification result image.

[0113] This invention does not rely on polarization decomposition parameters or supervised classifiers. Instead, it is based on the scattering differences between the co-polarized HH and cross-polarized HV during the freezing and ablation periods relative to the winter reference scattering state. The fusion contribution of the two polarization information is determined by the local incident angle, and threshold segmentation is completed by the two-dimensional Otsu algorithm.

[0114] Compared to the single-polarization fixed threshold method, the method of this invention does not rely on a single polarization channel and a fixed empirical threshold. Instead, it adjusts the contribution ratio of co-polarization HH and cross-polarization HV in the fusion based on the local incident angle θ. Therefore, it can reduce misclassification caused by local incident angle differences in complex terrain regions. Compared to the dual-polarization one-dimensional Otsu method, the method of this invention not only constructs a fusion difference map R based on the local incident angle, but also... c Furthermore, the two-dimensional Otsu algorithm is used in the classification stage to introduce neighborhood average gray information, which can further reduce the impact of SAR speckle noise on the classification results.

[0115] Example 3

[0116] This embodiment provides a computer implementation method.

[0117] The method of this invention can be executed by a remote sensing image processing platform, server, workstation, or other computer device with image processing capabilities. The computer device may include a processor and a memory, the memory storing a computer program. When the processor executes the computer program, it implements the aforementioned method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR images.

[0118] In one specific implementation, a computer device receives HH and HV polarimetric SAR images and topographic data during the frozen and ablation periods; performs registration, radiometric correction, topographic correction, and noise suppression on the SAR images; calculates the local incident angle θ based on the topographic data; and calculates the same polarimetric scattering difference characterization value R based on the preprocessed frozen and ablation period SAR images. HH Characterization value R of cross-polarization scattering difference HV ; Determine the weights W based on the piecewise nonlinear function model; Apply the weights W to R HH and R HV Perform weighted fusion to generate a fusion difference map R. c The difference map R was segmented and fused using the two-dimensional Otsu algorithm. c The classification result image is obtained.

[0119] In another specific implementation, the computer device can output the classification result map as a raster image, a thematic map, or vectorized boundary data. The classification result map can be used to display the distribution of wet snow areas and dry snow areas.

[0120] In summary, this invention characterizes the local incident angle θ and the difference in same polarization scattering R... HH Cross-polarization scattering difference characterization value R HV Piecewise nonlinear function model, fusion difference graph R c The combination of this method and the two-dimensional Otsu algorithm enables the ice sheet freeze-thaw detection process to simultaneously consider topographic relief, differences in polarization sensitivity, and the influence of SAR image noise, thereby improving the accuracy and stability of dry and wet snow classification results.

[0121] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, equivalent substitutions or adaptive adjustments can be made to the data sources, imaging modes, polarization methods, local incident angle calculation methods, weight function parameters, neighborhood window sizes, threshold segmentation methods, and classification result output formats in the above embodiments without departing from the concept of the present invention. All equivalent transformations made based on the content of this specification and drawings should fall within the scope of protection of the present invention.

Claims

1. A method for detecting ice sheet freeze-thaw cycles using locally incident angle-weighted SAR imagery, characterized in that, include: S1. Acquire SAR images and topographic data of co-polarized HH and cross-polarized HV during the freezing and melting periods of the ice sheet region, and preprocess the SAR images. S2. Calculate the local incident angle based on the terrain data, and calculate the scattering difference characterization values ​​of co-polarized HH and cross-polarized HV relative to the winter average scattering value based on the preprocessed SAR images of the freezing and thawing periods. S3. Based on the local incident angle corresponding to each pixel, and with 20°, 40° and 50° as segment nodes, construct a piecewise nonlinear function model to determine the contribution weights of cross-polarization HV and homopolarization HH in data fusion. S4. Based on the contribution weights corresponding to each pixel, the scattering difference characterization values ​​corresponding to the same polarization HH and the cross polarization HV are weighted and fused to generate a fused difference map. S5. Use the two-dimensional Otsu algorithm to segment the fusion difference map to obtain the classification result map.

2. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, The SAR images of co-polarized HH and cross-polarized HV include SAR images of co-polarized HH during the freezing period, SAR images of cross-polarized HV during the freezing period, SAR images of co-polarized HH during the ablation period, and SAR images of cross-polarized HV during the ablation period. The SAR images during the freezing period are used to characterize the reference scattering state when there is no significant melting in the ice sheet region, and the SAR images during the ablation period are used to characterize the scattering state to be classified when the ice sheet region undergoes freeze-thaw changes.

3. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, In step S1, the preprocessing includes registration, radiometric correction, topographic correction, and noise suppression; wherein, the registration is used to make the frozen SAR image correspond to the ablation SAR image under the same spatial reference, the radiometric correction is used to obtain comparable backscattering representation, the topographic correction is used to reduce geometric distortion caused by topographic undulation, and the noise suppression is used to reduce thermal noise and speckle noise in the SAR image.

4. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, In step S2, the terrain data includes digital elevation model data. The surface slope, aspect, or local normal vector is determined based on the digital elevation model data, and the local incident angle θ is determined based on the angle between the radar incident direction and the local surface normal vector.

5. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, In step S2, the scattering difference characterization value of the same polarization HH relative to the winter average scattering value is used to characterize the scattering difference between the freezing period and the ablation period under the same polarization channel, and the scattering difference characterization value of the cross polarization HV relative to the winter average scattering value is used to characterize the scattering difference between the freezing period and the ablation period under the cross polarization channel.

6. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, In step S3, the contribution weight is used to characterize the contribution ratio of cross-polarized HV in data fusion, and the remaining ratio other than the contribution weight is used to characterize the contribution ratio of homopolarized HH in data fusion. The piecewise nonlinear function model determines the contribution weight according to the interval of the local incident angle. Specifically, when the local incident angle is less than 20°, cross-polarized HV dominates the fusion process; when the local incident angle is not less than 20° and not greater than 40°, the contribution weight corresponding to cross-polarized HV gradually decreases according to a quadratic decay method; when the local incident angle is greater than 40° and not greater than 50°, the contribution weight corresponding to cross-polarized HV gradually approaches 0.5 according to an exponential convergence method, so that the fusion contribution of cross-polarized HV and homopolarized HH gradually approaches each other; when the local incident angle is greater than 50°, cross-polarized HV and homopolarized HH participate in fusion in an equal weight manner.

7. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 6, characterized in that, The parameters of the piecewise nonlinear function model are determined through sample fitting. The sample fitting includes: selecting dry snow pixels and wet snow pixels as sample points based on the ice sheet freeze-thaw reference image; dividing the sample points into bins according to the local incident angle interval; statistically analyzing the backscattering differences of co-polarized HH and cross-polarized HV on dry and wet snow within each bin; and determining the contribution weight sample value based on the distinguishing advantage of cross-polarized HV relative to co-polarized HH; fitting the relationship between the local incident angle and the contribution weight sample value using linear functions, exponential functions, quadratic functions, and piecewise nonlinear functions respectively; and determining the piecewise nonlinear function as the weight function based on the comparison results of the coefficient of determination, mean square error, and mean absolute error.

8. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, In step S4, generating the fusion difference map includes: determining the participation ratio of the cross-polarization scattering difference characterization value in the fusion based on the contribution weight corresponding to the cross-polarization HV; determining the participation ratio of the same-polarization scattering difference characterization value in the fusion based on the remaining proportion other than the contribution weight; and performing weighted fusion of the cross-polarization scattering difference characterization value and the same-polarization scattering difference characterization value to obtain a fusion difference map used to characterize the scattering difference between dry snow and wet snow.

9. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, In step S5, the two-dimensional Otsu algorithm includes: constructing two-dimensional feature pairs based on the pixel grayscale and neighborhood average grayscale of the fused difference map, statistically analyzing the joint probability distribution of the two-dimensional feature pairs, calculating the class probability, mean vector and inter-class dispersion corresponding to the candidate threshold, and determining the candidate threshold corresponding to the maximum inter-class dispersion as the optimal segmentation threshold.

10. The method for detecting ice sheet freeze-thaw cycles using local incident angle-weighted SAR imagery according to claim 1, characterized in that, The segmentation nodes at 20°, 40°, and 50° were determined based on the changes in polarization discrimination ability of measured snow and ice scattering samples within different local incident angle ranges. Specifically, when the local incident angle is less than 20°, the cross-polarized HV has a higher discrimination ability between dry and wet snow than the same-polarized HH. When the local incident angle is between 20° and 40°, the discrimination advantage of the cross-polarized HV decreases while the discrimination ability of the same-polarized HH increases. When the local incident angle is between 40° and 50°, the discrimination abilities of the same-polarized HH and the cross-polarized HV gradually become similar. When the local incident angle is greater than 50°, the same-polarized HH and the cross-polarized HV tend to participate in fusion with equal weight.

Images