A multi-source meteorological field constrained time-series InSAR atmospheric phase intelligent reconstruction method and system

Abstract

The application discloses a kind of multi-source weather field constraint time sequence InSAR atmospheric phase intelligent reconstruction method and system, it is related to radar remote sensing atmospheric correction technical field.The method obtains target area time sequence InSAR data, terrain constraint data and multi-source weather field data, forms multi-source input data set after space-time matching, quality control and gridding processing;Based on multi-source weather field, terrain constraint and radar observation geometry, initial atmospheric delay field is constructed and is converted into initial atmospheric phase prior field;Refined residual phase field is output by U-Net+ConvLSTM depth residual reconstruction model;Final atmospheric phase estimation field is obtained by fusing initial atmospheric phase prior field and refined residual phase field, and is used for interference phase correction and deformation inversion.The scheme considers physical prior constraint and data-driven residual reconstruction, and can improve the time sequence InSAR atmospheric phase correction accuracy under complex terrain and complex weather conditions.

Description

Technical Field

[0001] This invention relates to the field of radar remote sensing and atmospheric correction technology, and in particular to a method and system for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields. Background Technology

[0002] Synthetic Aperture Radar Interferometry (InSAR) technology can acquire surface deformation information by utilizing the phase difference between multiple SAR images. It has been widely used in fields such as urban land subsidence, landslide hazard identification, mining subsidence monitoring, seismic coseismic deformation inversion, volcanic activity monitoring, and permafrost deformation analysis. With the continuous acquisition of data from satellites such as Sentinel-1, Gaofen-3, ALOS, and TerraSAR-X, time-series InSAR technology has become an important means of monitoring surface deformation at the millimeter to centimeter level.

[0003] However, InSAR interferometric phase analysis not only contains surface deformation information but is also affected by various factors such as topographic residuals, atmospheric delay, orbital errors, thermal noise, temporal decoherence, and spatial decoherence. Among these, atmospheric phase screens caused by atmospheric disturbances are one of the important error sources affecting the accuracy of time-series InSAR deformation inversion. Radar electromagnetic waves are affected by changes in atmospheric refractive index when passing through the troposphere and ionosphere, resulting in propagation path delays, which manifest as atmospheric phase errors in the interferometric phase analysis. Especially in mountainous areas, humid regions, and under severe convective weather conditions, the spatial distribution of water vapor is highly non-uniform, and the wet delay varies drastically, leading to significant spatial correlation, topographic correlation, and temporal randomness in the atmospheric phase, severely impacting the reliability of deformation monitoring results.

[0004] Existing InSAR atmospheric phase correction methods mainly include time averaging, elevation correlation regression, spatial filtering, external meteorological model correction, GNSS-assisted correction, and atmospheric delay product correction methods such as GACOS. Among them, time averaging relies on the assumption that atmospheric phase time is random and deformation time is continuous, which easily leads to deformation and atmospheric overlap in scenarios of rapid deformation or seasonal deformation; elevation correlation regression usually assumes a linear relationship between atmospheric phase and elevation, making it difficult to describe complex water vapor structures and local wet delay anomalies; numerical meteorological models or reanalysis data can provide large-scale temperature, humidity, and pressure structures, but their spatial and temporal resolutions are limited, making it difficult to characterize rapid changes in local water vapor near the surface; single GNSS or single sounding data have insufficient spatial coverage, making it difficult to directly form a continuous regional atmospheric phase field.

[0005] Existing technologies mainly fall into two categories: one is to directly utilize external meteorological products such as ERA5, WRF, GNSS, or GACOS to construct InSAR atmospheric correction fields; the other is to directly employ deep learning models such as CNN, U-Net, BiGRU, and GAN to predict atmospheric delays or remove atmospheric phase screens from interferograms or time-series results. The former is limited by the spatiotemporal resolution and local water vapor characterization capabilities of meteorological products, making it difficult to fully represent small-scale atmospheric disturbances under complex terrain and non-stationary boundary layer conditions; while the latter, although possessing strong nonlinear fitting capabilities, is prone to lacking clear atmospheric physical constraints and carries the risk of misjudging real deformations as atmospheric disturbances. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for intelligent reconstruction of temporal InSAR atmospheric phases constrained by multi-source meteorological fields.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A time-series InSAR atmospheric phase intelligent reconstruction method constrained by multi-source meteorological fields includes:

[0009] Acquire temporal InSAR data, terrain-constrained data, and multi-source meteorological field data of the target area, and process the temporal InSAR data, terrain-constrained data, and multi-source meteorological field data to form a multi-source input dataset;

[0010] Based on multi-source meteorological field data, terrain-constrained data and radar observation geometric parameters of multi-source input dataset, an initial atmospheric delay field is constructed and converted into an initial atmospheric phase prior field. The initial atmospheric phase prior field is used to characterize the large-scale background atmospheric phase components explained by physical mechanisms.

[0011] Based on the difference between the original interferometric phase and the initial atmospheric phase prior field of the multi-source input dataset, a candidate residual phase field is constructed. The candidate residual phase field is dominated by residual atmospheric components, but may also contain non-atmospheric components such as deformation residuals, orbital errors and noise that have not yet been separated. This field is not used as the final atmospheric correction result, but as the object to be refined in the subsequent deep residual reconstruction model.

[0012] A deep residual reconstruction model is constructed, taking the candidate residual phase field as the modeling object, and combining the multi-source input dataset and the initial atmospheric phase prior field as the condition input to reconstruct the candidate residual phase field and output the refined residual phase field.

[0013] The initial atmospheric phase prior field is fused with the refined residual phase field to obtain the final atmospheric phase estimation field;

[0014] The original interferometric phase is subtracted and corrected using the final atmospheric phase estimation field. The corrected interferometric phase is then subjected to phase unwrapping, residual topographic phase removal, orbital error correction, and noise suppression in sequence to obtain the surface deformation results of the target area.

[0015] Preferably, the multi-source meteorological field data includes at least three of the following: ERA5 reanalysis data, numerical weather prediction data, BeiDou radiosonde data, ground-based microwave radiometer data, GNSS / MET zenith delay data, wind profiler radar data, and automatic weather station observation data; the multi-source input dataset also includes ionospheric auxiliary constraint data, which includes ionospheric total electron content data and ionospheric delay characteristics derived therefrom.

[0016] The terrain constraint data includes digital elevation model data and at least two of the following derived terrain elevation, slope, aspect, elevation gradient, local undulation, and terrain curvature features:

[0017] Preferably, using the SAR imaging time as a reference, the fusion weights of various neutral atmospheric meteorological data are determined based on data quality identifiers, the time difference between observation and imaging times, spatial resolution or representative parameters, and the height difference between the observation height and the target pixel terrain height. Temperature, humidity, air pressure, water vapor profiles, zenith delay, and surface meteorological elements from automatic weather stations are weighted and fused to form a multi-source meteorological constrained dataset with unified spatiotemporal resolution and vertical reference. The formula for calculating the fusion weight Wᵢ is:

[0018]

[0019] Wherein, Qᵢ is the data quality indicator (value 0~1), ΔTᵢ is the time difference between the observation time and the SAR imaging time (unit: hours), Rᵢ is the spatial resolution or spatial representativeness parameter (unit: km), and ΔHᵢ is the height difference between the observation height and the terrain height of the target pixel (unit: km). When zero or an abnormally small value appears in ΔTᵢ, Rᵢ, or ΔHᵢ, truncation protection is performed: ΔTᵢ < 0.5 hours is counted as 0.5 hours, Rᵢ < 0.1 km is counted as 0.1 km, and ΔHᵢ < 10 m is counted as 10 m. After truncation protection, the fusion weights of each meteorological data are normalized so that the sum of the weights of each data is 1. The total electron content data of the ionosphere is not included in the above-mentioned meteorological weighted fusion. Instead, it is processed separately by time interpolation and spatial gridding, and then directly converted into the ionospheric delay phase based on the radar carrier frequency and propagation geometry. This data is used as an independent input channel for the depth residual reconstruction model.

[0020] Preferably, the initial atmospheric delay field includes an initial tropospheric delay field and an initial ionospheric delay field; wherein, the initial tropospheric delay field is obtained by integrating along the radar line-of-sight path after calculating the atmospheric refractive index based on the temperature, pressure, and water vapor pressure parameters of the multi-source meteorological constraint dataset; the initial ionospheric delay field is estimated based on the total electron content data of the ionosphere, the radar carrier frequency, and the propagation geometry; the initial atmospheric phase prior field is obtained by superimposing the corresponding phase quantities of each delay field; the formula for calculating the atmospheric refractive index N is:

[0021] ;

[0022] Where N is the atmospheric refractive index, P is the pressure, T is the absolute temperature, e is the water vapor pressure, and k1, k2, and k3 are empirical coefficients of refractive index, which can be obtained using the Smith-Weintraub formula (k1 = 77.6 K / hPa, k2 = 71.6 K / hPa, k3 = 3.747 × 10⁻⁶). 5 K² / hPa);

[0023] Initial tropospheric delay field The calculation formula is:

[0024] ;

[0025] in, This represents the radar line-of-sight propagation path.

[0026] Initial tropospheric phase prior field The calculation formula is:

[0027] ;

[0028] in, The radar wavelength;

[0029] Initial ionospheric phase prior field It was estimated based on data on the total electron content of the ionosphere, radar carrier frequency, and propagation geometry.

[0030] Initial atmospheric phase prior field The calculation formula is:

[0031] .

[0032] Preferably, the candidate residual phase field The calculation formula is:

[0033]

[0034] in, For candidate residual phase fields, The original interference phase (differential interference phase after removing flat ground and terrain). This is the initial atmospheric phase prior field;

[0035] The candidate residual phase field is dominated by residual atmospheric components, but may also contain non-atmospheric components such as deformation residuals that have not yet been separated, orbital errors, and noise. This field is not used as the final atmospheric correction result, but as the object to be refined in the subsequent deep residual reconstruction model. The influence of its non-atmospheric components will be effectively suppressed under the physical consistency constraints and spatial and temporal smoothing constraints in the subsequent joint loss function. Finally, the model outputs the refined residual phase field.

[0036] Preferably, the conditional input includes at least two of the following: original interferometric phase, initial atmospheric phase prior field, candidate residual phase field, coherence coefficient, topographic elevation characteristics, topographic gradient characteristics, slope characteristics, water vapor or zenith delay characteristics, near-surface temperature, relative humidity, air pressure, precipitation, wind speed, wind direction characteristics, and time coding characteristics observed by automatic weather stations.

[0037] The time coding features include at least two of the following: radar imaging time, adjacent time phase time interval, and time-series position coding;

[0038] The deep residual reconstruction model is a U-Net+ConvLSTM architecture: the U-Net unit extracts multi-scale spatial features, terrain-related phase features and local water vapor disturbance features through encoder downsampling, and restores spatial resolution through decoder with skip connections upsampling; the ConvLSTM unit receives the spatial feature sequence output by U-Net, models the temporal evolution relationship of atmospheric phase of adjacent temporal residuals, and constrains the temporal continuity of the output results.

[0039] Preferably, the output of the deep residual reconstruction model is the refined residual phase field at time t. The calculation formula is:

[0040] ;

[0041] in, This represents the U-Net+ConvLSTM deep residual reconstruction model. These are the model parameters.

[0042] Preferably, the formula for calculating the joint loss function is:

[0043]

[0044] in, For residual reconstruction loss, For physical consistency loss, For spatial smoothing loss, For the loss of time continuity, , , , These are the loss weight coefficients for each corresponding term, and satisfy the following conditions: ;

[0045]

[0046] in, For the first The label residual phase field corresponding to each training sample (obtained in the manner described above). This is the refined residual phase field output by the model;

[0047]

[0048] in, It is a topographic elevation function. The empirical correlation coefficient is used to constrain the statistical correlation between the final atmospheric phase estimation field and the terrain elevation from abrupt changes, or to constrain the atmospheric phase variation between adjacent time phases from not exceeding the physical range defined by meteorological observations.

[0049]

[0050]

[0051] in, and These represent the refined residual phase fields corresponding to two adjacent imaging events.

[0052] Preferably, the final atmospheric phase estimation field The calculation formula is:

[0053] ;

[0054] in, This represents the initial atmospheric phase prior field. To refine the residual phase field;

[0055] Corrected interference phase The calculation formula is:

[0056] ;

[0057] in, The original interference phase, For the final atmospheric phase estimation field;

[0058] Final Deformation Phase The calculation formula is:

[0059] ;

[0060] in, This indicates the phase unwrapping operator. For residual topographic phase, For orbital error phase, This is the noise phase.

[0061] A time-series InSAR atmospheric phase intelligent reconstruction system constrained by multi-source meteorological fields includes:

[0062] Data acquisition and preprocessing module: Acquires time-series InSAR data, terrain-constrained data, and multi-source meteorological field data of the target area, and processes the time-series InSAR data, terrain-constrained data, and multi-source meteorological field data to form a multi-source input dataset;

[0063] Multi-source meteorological weighted fusion module: Based on multi-source meteorological field data, terrain constraint data and radar observation geometric parameters of multi-source input dataset, an initial atmospheric delay field is constructed and converted into an initial atmospheric phase prior field. The initial atmospheric phase prior field is used to characterize the large-scale background atmospheric phase components explained by physical mechanisms.

[0064] Candidate residual phase field construction module: Based on the difference between the original interferometric phase and the initial atmospheric phase prior field of the multi-source input dataset, a candidate residual phase field is constructed. The candidate residual phase field is used to characterize the local non-stationary, high-frequency residual atmospheric components not covered by the initial atmospheric phase prior field.

[0065] Deep residual reconstruction module: Constructs a deep residual reconstruction model, takes the candidate residual phase field as the modeling object, combines the multi-source input dataset and the initial atmospheric phase prior field as the condition input, reconstructs the candidate residual phase field, and outputs the refined residual phase field;

[0066] Joint constraint optimization module: It fuses the initial atmospheric phase prior field with the refined residual phase field to obtain the final atmospheric phase estimation field;

[0067] Deformation inversion output module: The original interferometric phase is subtracted and corrected using the final atmospheric phase estimation field. The corrected interferometric phase is then subjected to phase unwrapping, residual topographic phase removal, orbital error correction and noise suppression processing in sequence to obtain the surface deformation results of the target area.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] This invention constructs an initial atmospheric phase prior field based on meteorological field data, digital elevation model data, and propagation geometry. It constructs a candidate residual phase field based on the difference between the original interferometric phase and the initial atmospheric phase prior field. Then, it uses U-Net+ConvLSTM deep residual reconstruction to refine the candidate residual phase field. Finally, it fuses the refined residual phase field with the initial atmospheric phase prior field to form the final atmospheric phase estimation field.

[0070] Compared to methods that directly utilize meteorological models to construct correction fields, this invention can further compensate for local residual atmospheric structures that the initial atmospheric phase prior field fails to characterize. Compared to methods that directly use deep learning models to predict the total atmospheric phase, this invention uses physical prior fields as explicit constraints and the residual phase field, rather than the total atmospheric phase field, as the reconstruction object, thus balancing physical interpretability, local detail recovery capability, and model training stability. Therefore, this invention can improve the accuracy of atmospheric phase correction under complex terrain and meteorological conditions, reduce the impact of atmospheric residuals on surface deformation inversion results, and has significant engineering application value.

[0071] (1) The present invention first constructs an initial atmospheric phase prior field with clear physical meaning, and then performs deep residual reconstruction on the residual phase field that the initial atmospheric phase prior field fails to accurately express. This avoids the problems of insufficient physical constraints and weak interpretability of results caused by direct end-to-end prediction of the total atmospheric phase, thereby enhancing the physical interpretability and reliability of the atmospheric phase correction process.

[0072] (2) By integrating meteorological field information, topographic information and temporal features, this invention enables U-Net+ConvLSTM deep residual reconstruction to more finely characterize local atmospheric delay anomalies in complex mountainous areas, areas with strong water vapor disturbance and non-stationary boundary layer conditions under explicit prior constraints, thereby improving the accuracy of atmospheric phase estimation and correction.

[0073] (3) The present invention adopts a two-stage modeling method of constructing the initial atmospheric phase prior field and refining the residual phase field, which decomposes the complex total atmospheric phase estimation problem into two interconnected sub-problems: background atmospheric composition estimation and local residual composition compensation. This reduces the learning difficulty and output dynamic range of the deep learning model, and is conducive to improving the model training stability and convergence effect.

[0074] (4) This invention obtains the final atmospheric phase estimation field by fusing the refined residual phase field with the initial atmospheric phase prior field, and uses the final atmospheric phase estimation field to correct the original interferometric phase, thereby effectively reducing the interference of atmospheric residuals on subsequent phase unwrapping, deformation inversion and time series analysis results, and improving the accuracy and reliability of settlement monitoring, landslide monitoring, volcanic deformation monitoring and infrastructure deformation monitoring.

[0075] (5) This invention combines the interpretability advantage of physical modeling methods with the representation advantage of data-driven methods for complex nonlinear local details. It has strong adaptability, good engineering application prospects and promotion value, and is suitable for InSAR atmospheric phase correction tasks under various complex terrain and meteorological conditions. Attached Figure Description

[0076] Figure 1 This is a general flowchart of an embodiment of the present invention;

[0077] Figure 2 This is a block diagram of multi-source meteorological field fusion according to an embodiment of the present invention;

[0078] Figure 3 This is a schematic diagram illustrating the construction of a multi-source meteorological prior atmospheric phase field according to an embodiment of the present invention;

[0079] Figure 4 This is a diagram of the deep learning residual reconstruction network structure according to an embodiment of the present invention;

[0080] Figure 5 This is a schematic diagram of the atmospheric phase reconstruction and deformation output results according to an embodiment of the present invention;

[0081] Figure 6 This is a schematic diagram of multi-source meteorological data weighted fusion and quality control in an embodiment of the present invention;

[0082] Figure 7 This is a comparison diagram of the interference phase before and after correction in an embodiment of the present invention. Detailed Implementation

[0083] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0084] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0085] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0086] Reference Figures 1-7 As shown.

[0087] The embodiments further illustrate the intelligent reconstruction method and system for time-series InSAR atmospheric phase constrained by multi-source meteorological fields proposed in this invention.

[0088] (I) Data Acquisition and Preprocessing

[0089] Acquire temporal InSAR data, terrain-constrained data, multi-source meteorological field data, and ionospheric auxiliary constraint data for the target area.

[0090] The time-series InSAR data includes SAR imagery, interferometric phase, unwrapping phase, coherence coefficient, incident angle, azimuth angle, and imaging time information.

[0091] The multi-source meteorological field data includes at least three of the following: ERA5 reanalysis data, numerical weather prediction data, BeiDou radiosonde data, ground-based microwave radiometer data, GNSS / MET zenith delay data, wind profiler radar data, and automatic weather station observation data; the ionospheric auxiliary constraint data includes total ionospheric electron content data and ionospheric delay characteristics derived therefrom.

[0092] The terrain constraint data includes digital elevation model data and at least two of the following derived terrain elevation, slope, aspect, elevation gradient, local undulation, and terrain curvature features:

[0093] The temporal InSAR data undergoes registration, flatland phase removal, topographic phase removal, filtering, phase unwrapping, and low-coherence area masking. Multi-source meteorological data is processed for outlier removal, missing data handling, profile smoothing, and quality control. For automatic weather station observations, the process also includes station quality identification checks, anomaly jump identification, missing time period marking, minute-level or hourly observation time merging, and station representativeness assessment to ensure consistency with radar imaging timestamps and the InSAR pixel grid.

[0094] Using SAR imaging time as a reference, meteorological data are time-matched; InSAR data, DEM data and meteorological data are unified to the same coordinate system, spatial resolution and vertical reference to form a multi-source input dataset.

[0095] (II) Multi-source meteorological weighted fusion and construction of the initial atmospheric phase prior field

[0096] Using SAR imaging time as a reference, the fusion weights of various neutral atmospheric meteorological data are determined based on data quality identifiers, the time difference between observation and imaging times, spatial resolution or representative parameters, and the height difference between observation altitude and target pixel terrain altitude. Temperature, humidity, air pressure, water vapor profiles, zenith delay, and surface meteorological elements from automatic weather stations are weighted and fused to form a multi-source meteorological constrained dataset with unified spatiotemporal resolution and vertical reference. The formula for calculating the fusion weight Wᵢ is:

[0097]

[0098] Wherein, Qᵢ is the data quality indicator (value 0~1), ΔTᵢ is the time difference between the observation time and the SAR imaging time (unit: hours), Rᵢ is the spatial resolution or spatial representativeness parameter (unit: km), and ΔHᵢ is the height difference between the observation height and the terrain height of the target pixel (unit: km). When zero or an abnormally small value appears in ΔTᵢ, Rᵢ, or ΔHᵢ, truncation protection is performed: ΔTᵢ < 0.5 hours is counted as 0.5 hours, Rᵢ < 0.1 km is counted as 0.1 km, and ΔHᵢ < 10 m is counted as 10 m. After truncation protection, the fusion weights of each meteorological data are normalized so that the sum of the weights of each data is 1.

[0099] The total electron content data of the ionosphere is not included in the above-mentioned meteorological weighted fusion. Instead, it is processed separately by time interpolation and spatial gridding, and then directly converted into the ionospheric delay phase based on the radar carrier frequency and propagation geometry. This data is used as an independent input channel for the depth residual reconstruction model.

[0100] Based on the fusion weights, the information on temperature, humidity, air pressure, water vapor profile, zenith delay, and boundary layer structure is weighted and fused to obtain a multi-source meteorological constrained dataset with unified spatiotemporal resolution and unified vertical reference.

[0101] Based on multi-source meteorological field data, terrain-constrained data, and radar observation geometric parameters from a multi-source input dataset, an initial atmospheric delay field is constructed and converted into an initial atmospheric phase prior field. The initial atmospheric delay field includes an initial tropospheric delay field and an initial ionospheric delay field. The initial tropospheric delay field is obtained by integrating along the radar line-of-sight after calculating the atmospheric refractive index using meteorological parameters such as temperature, pressure, and water vapor pressure. The initial ionospheric delay field is independently estimated based on the total electron content data of the ionosphere (not involved in meteorological weighted fusion), radar carrier frequency, and propagation geometry.

[0102] The formula for calculating the atmospheric refractive index N is:

[0103]

[0104] Where N is the atmospheric refractive index, P is the pressure (in hPa), T is the absolute temperature (in K), e is the water vapor pressure (in hPa), and k1, k2, and k3 are empirical coefficients of refractive index. In a preferred embodiment, the Smith-Weintraub formula coefficients are used: k1 = 77.6 K / hPa, k2 = 71.6 K / hPa, k3 = 3.747 × 10⁻⁶. 5K² / hPa. This set of coefficients has an accuracy of approximately 0.5% in frequencies below 30 GHz and within the range of conventional atmospheric parameters, and is applicable to SAR satellite data from the L-band (approximately 1.2 GHz) to the C-band (approximately 5.3 GHz). In another alternative implementation, Thayer's formula coefficients (k1 = 77.604 K / hPa, k2 = 64.79 K / hPa, k3 = 3.776 × 10⁻⁶) can also be used. 5 The specific value (K² / hPa) is selected based on the climate characteristics of the target area and the radar band.

[0105] The formula for calculating the initial tropospheric delay field L_trop is:

[0106]

[0107] Where s is the radar line-of-sight propagation path (integrated along the slant range direction).

[0108] Initial tropospheric phase prior field φ_trop⁽ 0 The formula for calculating ⁾ is:

[0109]

[0110] Where λ is the radar wavelength.

[0111] Initial ionospheric phase prior field φ_iono⁽ 0 ⁾Estimated based on total ionospheric electron content data, radar carrier frequency, and propagation geometry.

[0112] Initial atmospheric phase prior field φ_atm⁽ 0 The formula for calculating ⁾ is:

[0113]

[0114] This prior field is used to provide large-scale background constraints and physical prior references required for subsequent residual modeling.

[0115] (III) Construction of candidate residual phase field

[0116] Candidate residual phase fields are constructed based on the difference between the original interference phase and the initial atmospheric phase prior field. The formula for calculating the candidate residual phase field φ_res⁽ᶜᵃⁿᵈ⁾ is:

[0117]

[0118] Where φ_res⁽ᶜᵃⁿᵈ⁾ represents the candidate residual phase field, φ_int represents the original interferometric phase (the differential interferometric phase after de-flattening and de-terrainting processing), and φ_atm⁽ 0 ⁾ represents the initial atmospheric phase prior field.

[0119] The candidate residual phase field is dominated by residual atmospheric components, but may also contain non-atmospheric components such as deformation residuals that have not yet been separated, orbital errors, and noise. This field is not used as the final atmospheric correction result, but as the object to be refined in the subsequent deep residual reconstruction model. The influence of its non-atmospheric components will be effectively suppressed under the physical consistency constraints and spatial and temporal smoothing constraints in the subsequent joint loss function. Finally, the model outputs the refined residual phase field.

[0120] When constructing the initial atmospheric phase prior field, this invention does not simply use multi-source meteorological data as input to the deep learning model. Instead, it first performs quality control, time matching, spatial interpolation, vertical benchmark unification, and weight fusion on meteorological data from different sources to form a multi-source meteorological constraint field consistent with the SAR imaging time, radar line-of-sight direction, and InSAR pixel grid.

[0121] Specifically, ERA5 reanalysis data or numerical weather prediction data are used to provide large-scale background fields of temperature, humidity, pressure, and geopotential height for the study area; BeiDou radiosonde data are used to correct the vertical temperature, humidity, and pressure profiles of the meteorological field; ground-based microwave radiometer data are used to supplement the continuous water vapor profile and integrated water vapor information in the near-surface layer; GNSS / MET data are used to constrain the integral of the total zenith delay or the wet zenith delay; automatic weather station observation data are used to supplement high temporal resolution surface meteorological elements such as near-surface temperature, relative humidity, pressure, precipitation, wind speed, and wind direction; wind profiler radar data are used to characterize boundary layer height, low-level wind field, and local dynamic disturbances; and digital elevation model data are used to provide topographic constraint information such as topographic elevation, slope, aspect, and elevation gradient.

[0122] (iv) Construction and training of deep residual reconstruction model

[0123] A deep residual reconstruction model is constructed, which takes the candidate residual phase field as the modeling object and combines the multi-source input dataset and the initial atmospheric phase prior field as the condition input to reconstruct the candidate residual phase field and output the refined residual phase field.

[0124] The input conditions include multiple combinations of the original interferometric phase, the initial atmospheric phase prior field, the candidate residual phase field, the coherence coefficient, the terrain elevation characteristics, the terrain gradient characteristics, the slope characteristics, the water vapor or zenith delay characteristics, and the near-surface temperature, relative humidity, air pressure, precipitation, wind speed, wind direction characteristics, and time coding characteristics observed by automatic weather stations.

[0125] The time coding features include at least two of the following: radar imaging time, time interval between adjacent time phases, and time-series position coding.

[0126] The deep residual reconstruction model is a U-Net+ConvLSTM architecture: the U-Net unit extracts multi-scale spatial features, terrain-related phase features and local water vapor disturbance features through encoder downsampling, and restores spatial resolution through decoder with skip connections upsampling; the ConvLSTM unit receives the spatial feature sequence output by U-Net, models the temporal evolution relationship of atmospheric phase of adjacent temporal residuals, and constrains the temporal continuity of the output results.

[0127] The multi-source features corresponding to the candidate residual phase field are constructed as conditional inputs c⁽ᵗ⁾, which include: the original interferometric phase φ_int and the initial atmospheric phase prior field φ_atm⁽ 0 ⁾, Topographic elevation feature h, topographic gradient feature ∇h, time coding feature τ, slope feature, water vapor or zenith delay feature m, coherence coefficient γ, the conditional input can be expressed as:

[0128]

[0129] Wherein, Concat indicates stitching by channel dimension, and t represents the imaging phase or interference phase.

[0130] In a preferred embodiment, the water vapor or zenith delay feature m includes a rasterized feature formed by time matching, spatial interpolation, and unified gridding of surface meteorological elements such as near-surface temperature, relative humidity, air pressure, precipitation, wind speed, and wind direction obtained from automatic weather station observations.

[0131] The conditional input 'c' is fed into the U-Net spatial feature extraction unit. The U-Net encoder extracts multi-scale spatial features from the candidate residual phase field through multi-layer convolution and downsampling, including local water vapor disturbance features, terrain-related phase change features, and small-scale residual atmospheric structure features.

[0132] The U-Net decoder restores spatial resolution through upsampling and skip connections, ensuring the output matches the original interferometric phase grid. These skip connections preserve local detail information and reduce the loss of residual phase boundaries and small-scale structures during feature extraction.

[0133] The spatial feature sequence output from the U-Net encoder or intermediate layer is input into the ConvLSTM temporal modeling unit. ConvLSTM is used to learn the evolution relationship of residual atmospheric phase between adjacent temporal phases, extracting the continuity and variation trend of atmospheric disturbances in the time dimension.

[0134] For consecutive time phases t−1, t, t+1, ConvLSTM updates the hidden state based on the spatial characteristics of adjacent time phases, which is used to constrain the reconstruction results of the atmospheric phase of the current time phase residual, so that the model output remains continuous and stable in the time dimension.

[0135] The output of the deep residual reconstruction model is the refined residual phase field φ_reŝ⁽ᵗ⁾ at time t, calculated as follows:

[0136]

[0137] Where ℱ_Θ represents the U-Net+ConvLSTM deep residual reconstruction model, and Θ is the model parameter.

[0138] The refined residual phase field is used to characterize the local non-stationary atmospheric residual structure that the initial atmospheric phase prior field fails to fully express.

[0139] The model training employs a joint loss function, and the label values ​​of the training samples are constructed using one of the following methods: (a) selecting interferometric pairs within the target area with no obvious deformation or known deformation (such as stable foundation areas or bedrock exposed areas), and using the candidate residual phase field as the supervision signal; (b) generating simulated interferograms with known atmospheric phase true values ​​using an atmospheric simulation model, and using the simulated residual phase field as the supervision signal; (c) introducing independent external validation data (such as atmospheric zenith delay inverted from a synchronous GNSS continuously operating reference station) for indirect supervision during the training process to ensure that the residuals learned by the model are dominated by atmospheric components.

[0140] The formula for calculating the joint loss function is:

[0141]

[0142] Among them, L res For residual reconstruction loss, L phys For physical consistency loss, L smooth For spatial smoothing loss, L time For time continuity loss, λ res , λ phys , λ smooth , λ time These are the loss weight coefficients for the corresponding terms, and they satisfy λ. res +λ phys +λ smooth +λ time =1. In a preferred embodiment, the weights are taken as follows: λ res = 0.35-0.50, λ phys =0.20-0.30, λ smooth =0.10-0.20, λ time =0.10-0.20. Wherein, the residual reconstruction weight λ res The physical consistency weight λ is the largest of the four to ensure the model's dominance in training the core task; phys Secondly, it is used to constrain the model output from deviating from atmospheric physical laws; spatial smoothing weight λsmooth and time continuity weight λ time The weights are relatively small to avoid excessive smoothing that could lead to the loss of local details. In an alternative implementation, the aforementioned weight coefficients can be adaptively adjusted during training using an uncertainty-weighted method.

[0143] The loss terms are defined as follows:

[0144]

[0145] Among them, φ_res⁽ 1 ⁾ represents the label residual phase field corresponding to the i-th training sample (obtained as described above), φ_reŝ⁽ 1 ⁾ represents the refined residual phase field output by the model.

[0146]

[0147] Where f(h) is the topographic elevation function, and α is the empirical correlation coefficient. In a preferred embodiment, the value of α is determined as follows: (a) Select a stable area without significant deformation (such as a bedrock exposed area or a stable foundation area) within the target area, perform linear regression on the initial atmospheric phase prior field and the topographic elevation, and use the regression slope as the initial estimate of α, with a typical value range of -0.005 to -0.02 rad / m; (b) In another optional embodiment, α is used as a trainable parameter and automatically optimized through backpropagation during model training; (c) For target areas with large topographic differences, α values ​​can be set separately according to topographic zones. The physical consistency loss is used to constrain the statistical correlation between the final atmospheric phase estimation field and the topographic elevation from abrupt changes, or to constrain the atmospheric phase variation amplitude between adjacent time phases from not exceeding the physical range defined by meteorological observations.

[0148]

[0149]

[0150] Where φ_res⁽ᵗ⁾ and φ_res⁽ᵗ⁺¹⁾ represent the refined residual phase fields corresponding to two adjacent imaging events, respectively.

[0151] The parameters for the deep residual reconstruction model architecture are set as follows:

[0152] The U-Net component consists of four encoder downsampling layers, each downsampled by a factor of 2. The first convolutional layer has 64 channels, and the number of channels increases exponentially with each downsampling layer (64→128→256→512). The kernel size is 3×3. ReLU is used as the activation function. Batch normalization is applied after each convolutional layer to accelerate convergence. Dropout is introduced into the network with a dropout rate of 0.1 to prevent overfitting.

[0153] ConvLSTM section: The ConvLSTM has 3 layers; the kernel size is 3×3, the number of hidden state channels is 64; the dropout rate is 0.2.

[0154] The training hyperparameters were set as follows: the optimizer was Adam, with an initial learning rate of 1×10⁻³, using cosine annealing scheduling or a stepwise decay strategy of 0.5% every 50 epochs; the batch size was set to 8; the maximum number of training epochs was set to 150, and the optimal number of epochs was determined using early stopping; the loss function was L1 loss. The input data consisted of image patches with the same crop size as the InSAR image, 512×512 pixels; the number of input channels was the sum of the number of conditional input features, and the number of output channels was 1, corresponding to the refined residual phase field.

[0155] After the model training is completed, the conditional input corresponding to the target region to be corrected is fed into the trained U-Net+ConvLSTM deep residual reconstruction model to obtain the refined residual phase field of the corresponding time phase.

[0156] A U-Net+ConvLSTM deep residual reconstruction model based on meteorological prior constraints is constructed, and the constructed candidate residual phase field is refined and reconstructed to obtain the refined residual phase field of the target area.

[0157] Instead of directly estimating the total atmospheric phase field or replacing the constructed initial atmospheric phase prior field, modeling is performed only on the remaining phase components that still exist after subtracting the initial atmospheric phase prior field. The refined residual phase field serves as a supplementary estimate for the unexplained portion of the initial atmospheric phase prior field and is used for subsequent fusion with the initial atmospheric phase prior field to form the final atmospheric phase estimation field.

[0158] The U-Net+ConvLSTM deep residual reconstruction model takes the candidate residual phase field as the learning object and the original interferometric phase, the initial atmospheric phase prior field, terrain features, coherence coefficient and temporal features as conditional inputs. Through spatial feature extraction and temporal feature modeling, it realizes the refined recovery of the local non-stationary, small-scale, high-frequency residual atmospheric structure and outputs a refined residual phase field.

[0159] (v) Final atmospheric phase estimation and deformation inversion

[0160] The initial atmospheric phase prior field is fused with the refined residual phase field to obtain the final atmospheric phase estimation field.

[0161] The formula for calculating the final atmospheric phase estimation field φ_atm̂ is:

[0162]

[0163] Where, φ_atm⁽ 0 ⁾ represents the initial atmospheric phase prior field, and φ_reŝ represents the refined residual phase field.

[0164] The original interferometric phase is subtracted and corrected using the final atmospheric phase estimation field to obtain the corrected interferometric phase.

[0165] The formula for calculating the corrected interference phase φ_cor is:

[0166]

[0167] Where φ_int is the original interferometric phase and φ_atm̂ is the final atmospheric phase estimation field.

[0168] The corrected interferometric phase is sequentially subjected to phase unwrapping, residual topographic phase removal, orbital error correction, and noise suppression to obtain the surface deformation results of the target area.

[0169] Final deformation phase φ def The calculation formula is:

[0170]

[0171] Where 𝒰 represents the phase unwrapping operator, φ topo For residual terrain phase, φ orb For the orbital error phase, φ noise This is the noise phase.

[0172] (vi) System Implementation Examples

[0173] A time-series InSAR atmospheric phase intelligent reconstruction system constrained by multi-source meteorological fields includes:

[0174] Data acquisition and preprocessing module: Acquires temporal InSAR data, terrain-constrained data, and multi-source meteorological field data of the target area, as well as ionospheric auxiliary constraint data; performs registration, flatland phase removal, terrain phase removal, filtering, phase unwrapping, and low-coherence area masking on the temporal InSAR data, terrain-constrained data, and multi-source meteorological field data, and performs spatiotemporal matching, quality control, and gridding to form a multi-source input dataset;

[0175] Multi-source meteorological weighted fusion module: Based on neutral atmospheric meteorological data, topographic constraint data, and radar observation geometric parameters from the multi-source input dataset, an initial tropospheric delay field is constructed; the ionospheric delay field is independently estimated based on the total electron content data of the ionosphere (which does not participate in meteorological weighted fusion); the tropospheric and ionospheric delay fields are converted into phase quantities and then superimposed to obtain the initial atmospheric phase prior field, which is used to characterize the large-scale background atmospheric phase components explained by physical mechanisms;

[0176] Candidate residual phase field construction module: Based on the difference between the original interferometric phase and the initial atmospheric phase prior field of the multi-source input dataset, a candidate residual phase field is constructed. The candidate residual phase field is dominated by the residual atmospheric component, but may also contain non-atmospheric components such as deformation residuals, orbital errors and noise that have not yet been separated. This field is not used as the final atmospheric correction result, but as the object to be refined in the subsequent deep residual reconstruction model.

[0177] Deep residual reconstruction module: Constructs a deep residual reconstruction model, takes the candidate residual phase field as the modeling object, combines the multi-source input dataset and the initial atmospheric phase prior field as the condition input, reconstructs the candidate residual phase field, and outputs the refined residual phase field;

[0178] Joint constraint optimization module: It fuses the initial atmospheric phase prior field with the refined residual phase field to obtain the final atmospheric phase estimation field;

[0179] Deformation inversion output module: The original interferometric phase is subtracted and corrected using the final atmospheric phase estimation field. The corrected interferometric phase is then subjected to phase unwrapping, residual topographic phase removal, orbital error correction and noise suppression processing in sequence to obtain the surface deformation results of the target area.

[0180] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0181] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A time-series InSAR atmospheric phase intelligent reconstruction method constrained by multi-source meteorological fields, characterized in that, include: Acquire temporal InSAR data, terrain-constrained data, and multi-source meteorological field data of the target area, and process the temporal InSAR data, terrain-constrained data, and multi-source meteorological field data to form a multi-source input dataset; Based on multi-source meteorological field data, terrain-constrained data and radar observation geometric parameters of multi-source input dataset, an initial atmospheric delay field is constructed and converted into an initial atmospheric phase prior field. The initial atmospheric phase prior field is used to characterize the large-scale background atmospheric phase components explained by physical mechanisms. Based on the difference between the original interferometric phase and the initial atmospheric phase prior field of the multi-source input dataset, a candidate residual phase field is constructed. The candidate residual phase field is dominated by the residual atmospheric component and includes unseparated deformation residuals, orbital errors, and noise non-atmospheric components. This field is not used as the final atmospheric correction result, but as the object to be refined in the deep residual reconstruction model. The deep residual reconstruction model is constructed, using the candidate residual phase field as the modeling object, and combining the multi-source input dataset and the initial atmospheric phase prior field as conditional inputs to reconstruct the candidate residual phase field and output a refined residual phase field. The initial atmospheric phase prior field is fused with the refined residual phase field to obtain the final atmospheric phase estimation field; The original interferometric phase is subtracted and corrected using the final atmospheric phase estimation field. The corrected interferometric phase is then subjected to phase unwrapping, residual topographic phase removal, orbital error correction, and noise suppression in sequence to obtain the surface deformation results of the target area.

2. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 1, characterized in that, The multi-source meteorological field data includes at least three of the following: ERA5 reanalysis data, numerical weather prediction data, BeiDou radiosonde data, ground-based microwave radiometer data, GNSS / MET zenith delay data, wind profiler radar data, and automatic weather station observation data; the multi-source input dataset also includes ionospheric auxiliary constraint data, which includes total ionospheric electron content data and ionospheric delay characteristics derived therefrom. The terrain constraint data includes digital elevation model data and at least two of the following derived terrain elevation, slope, aspect, elevation gradient, local undulation, and terrain curvature features:

3. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 2, characterized in that, Using SAR imaging time as a reference, the fusion weights of various neutral atmospheric meteorological data are determined based on data quality identifiers, the time difference between observation and imaging times, spatial resolution or representative parameters, and the height difference between observation height and target pixel terrain height. Temperature, humidity, air pressure, water vapor profiles, zenith delay, and surface meteorological elements from automatic weather stations are weighted and fused to form a multi-source meteorological constrained dataset with unified spatiotemporal resolution and vertical reference. The formula for calculating the fusion weight Wᵢ is as follows: Wherein, Qᵢ is the data quality indicator, ranging from 0 to 1; ΔTᵢ is the time difference between the observation time and the SAR imaging time; Rᵢ is the spatial resolution or spatial representativeness parameter; and ΔHᵢ is the height difference between the observation height and the terrain height of the target pixel. When zero or an anomalous minimum value appears in ΔTᵢ, Rᵢ, or ΔHᵢ, truncation protection is performed: ΔTᵢ < 0.5 hours is counted as 0.5 hours, Rᵢ < 0.1 km is counted as 0.1 km, and ΔHᵢ < 10 m is counted as 10 m. After truncation protection, the fusion weights of each meteorological data are normalized so that the sum of the weights of each data is 1. The total electron content data of the ionosphere does not participate in the above meteorological weighted fusion, but is separately processed by temporal interpolation and spatial gridding, and is directly converted into the ionospheric delay phase according to the radar carrier frequency and propagation geometry, and used as an independent input channel for the depth residual reconstruction model.

4. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 1, characterized in that, The initial atmospheric delay field includes an initial tropospheric delay field and an initial ionospheric delay field. The initial tropospheric delay field is obtained by integrating the atmospheric refractive index along the radar line-of-sight path after calculating the temperature, pressure, and water vapor pressure parameters based on a multi-source meteorological constraint dataset. The ionospheric delay field is independently estimated based on the total electron content data of the ionosphere, radar carrier frequency, and propagation geometry. The initial atmospheric phase prior field is obtained by superimposing the corresponding phase quantities of each delay field. The formula for calculating the atmospheric refractive index N is: Where N is the atmospheric refractive index, P is the pressure, T is the absolute temperature, e is the water vapor pressure, and k1, k2, and k3 are empirical coefficients of refractive index, which can be obtained using the Smith-Weintraub formula. Initial tropospheric delay field The calculation formula is: in, This represents the radar line-of-sight propagation path. Initial tropospheric phase prior field The calculation formula is: in, The radar wavelength; Initial ionospheric phase prior field It was estimated based on data on the total electron content of the ionosphere, radar carrier frequency, and propagation geometry. Initial atmospheric phase prior field The calculation formula is: 。 5. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 1, characterized in that, Candidate residual phase field The calculation formula is: in, For candidate residual phase fields, The original interferometric phase, and the differential interferometric phase after removing flat areas and terrain features. This is the initial atmospheric phase prior field; The candidate residual phase field is dominated by residual atmospheric components, but may also contain non-atmospheric components such as deformation residuals, orbital errors and noise that have not yet been separated. This field is not used as the final atmospheric correction result, but as the object to be refined in the deep residual reconstruction model. The influence of its non-atmospheric components will be effectively suppressed under the physical consistency constraints and spatial and temporal smoothing constraints in the joint loss function. Finally, the model outputs the refined residual phase field.

6. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 1, characterized in that, The conditional inputs include at least two of the following: original interferometric phase, initial atmospheric phase prior field, candidate residual phase field, coherence coefficient, topographic elevation characteristics, topographic gradient characteristics, slope characteristics, water vapor or zenith delay characteristics, and near-surface temperature, relative humidity, air pressure, precipitation, wind speed, wind direction characteristics, and time coding characteristics observed by automatic weather stations. The time coding features include at least two of the following: radar imaging time, adjacent time phase time interval, and time-series position coding; The deep residual reconstruction model is a U-Net+ConvLSTM architecture: the U-Net unit extracts multi-scale spatial features, terrain-related phase features and local water vapor disturbance features through encoder downsampling, and restores spatial resolution through decoder with skip connections upsampling; the ConvLSTM unit receives the spatial feature sequence output by U-Net, models the temporal evolution relationship of atmospheric phase of adjacent temporal residuals, and constrains the temporal continuity of the output results.

7. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 6, characterized in that, The output of the deep residual reconstruction model is the refined residual phase field at time t. The calculation formula is: ； in, This represents the U-Net+ConvLSTM deep residual reconstruction model. These are the model parameters.

8. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 7, characterized in that, The model training employs a joint loss function, and the label values ​​of the training samples are constructed using one of the following methods: (a) selecting interferometric pairs with no obvious deformation or known deformation within the target region, using the candidate residual phase field as the supervision signal; (b) generating simulated interferograms with known atmospheric phase ground truth using an atmospheric simulation model, using the simulated residual phase field as the supervision signal; (c) introducing independent external validation data for indirect supervision during training to ensure that the residuals learned by the model are dominated by atmospheric components. The formula for calculating the joint loss function is: in, For residual reconstruction loss, For physical consistency loss, For spatial smoothing loss, For the loss of time continuity, , , , These are the loss weight coefficients for each corresponding term, and satisfy the following conditions: ; in, For the first The label residual phase field corresponding to each training sample This is the refined residual phase field output by the model; in, It is a topographic elevation function. The empirical correlation coefficient is used to constrain the statistical correlation between the final atmospheric phase estimation field and the terrain elevation from abrupt changes, or to constrain the atmospheric phase variation between adjacent time phases from not exceeding the physical range defined by meteorological observations. in, and These represent the refined residual phase fields corresponding to two adjacent imaging events.

9. The method for intelligent reconstruction of temporal InSAR atmospheric phase constrained by multi-source meteorological fields according to claim 1, characterized in that, Final atmospheric phase estimation field The calculation formula is: ； in, This represents the initial atmospheric phase prior field. To refine the residual phase field; Corrected interference phase The calculation formula is: ; in, The original interference phase, For the final atmospheric phase estimation field; Final Deformation Phase The calculation formula is: ; in, This indicates the phase untangling operator. For residual topographic phase, For orbital error phase, This is the noise phase.

10. A multi-source meteorological field-constrained temporal InSAR atmospheric phase intelligent reconstruction system, applied to the multi-source meteorological field-constrained temporal InSAR atmospheric phase intelligent reconstruction method according to any one of claims 1 to 9, characterized in that, include: Data acquisition and preprocessing module: Acquires time-series InSAR data, terrain-constrained data, and multi-source meteorological field data of the target area, and processes the time-series InSAR data, terrain-constrained data, and multi-source meteorological field data to form a multi-source input dataset; Multi-source meteorological weighted fusion module: Based on neutral atmospheric meteorological data, topographic constraint data, and radar observation geometric parameters from the multi-source input dataset, an initial tropospheric delay field is constructed; the ionospheric delay field is independently estimated based on the total electron content data of the ionosphere; the tropospheric and ionospheric delay fields are converted into phase quantities and then superimposed to obtain the initial atmospheric phase prior field, which is used to characterize the large-scale background atmospheric phase components explained by physical mechanisms; Candidate residual phase field construction module: Based on the difference between the original interferometric phase and the initial atmospheric phase prior field of the multi-source input dataset, a candidate residual phase field is constructed. The candidate residual phase field is dominated by the residual atmospheric component and includes unseparated deformation residuals, orbital errors, and noise non-atmospheric components. This field is not used as the final atmospheric correction result, but as the object to be refined in the deep residual reconstruction model. Deep residual reconstruction module: Constructs a deep residual reconstruction model, using the candidate residual phase field as the modeling object, and combining the multi-source input dataset and the initial atmospheric phase prior field as conditional inputs to reconstruct the candidate residual phase field and output a refined residual phase field. Joint constraint optimization module: It fuses the initial atmospheric phase prior field with the refined residual phase field to obtain the final atmospheric phase estimation field; Deformation inversion output module: The original interferometric phase is subtracted and corrected using the final atmospheric phase estimation field. The corrected interferometric phase is then subjected to phase unwrapping, residual topographic phase removal, orbital error correction and noise suppression processing in sequence to obtain the surface deformation results of the target area.

Images