InSAR and GNSS fusion settlement monitoring method based on double-branch network architecture

Abstract

The invention provides an InSAR (Interferometric Synthetic Aperture Radar) and GNSS (Global Navigation Satellite System) fusion settlement monitoring method based on a double-branch neural network, which comprises the following steps of: converting low-time-resolution InSAR data into a high-resolution time sequence matched with GNSS through an interpolation algorithm, and constructing an input matrix with aligned dimensions; a heterogeneous dual-channel network is adopted to respectively extract InSAR spatial features and GNSS time sequence features, after key information is extracted through dimension reduction processing, fusion data is reconstructed, a data splitting strategy is utilized to train a model and optimize parameters, and finally accurate modeling of a multi-source data nonlinear relation is realized. According to the method, the limitation of traditional linear fusion is broken through, a heterogeneous neural network is designed, multi-source data fusion is realized by using the high spatial resolution feature of the InSAR image and the high temporal resolution feature of the GNSS, the problem of insufficient modeling of a non-linear relationship of heterogeneous data in a traditional method is solved, and the space-time precision and robustness of settlement monitoring are improved.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing technology, and specifically relates to a multi-source remote sensing data fusion method based on deep learning, in particular a settlement monitoring method based on the fusion of InSAR and GNSS using a dual-branch neural network architecture, which is used to improve the spatiotemporal accuracy and robustness of settlement monitoring. Background Art

[0002] Surface subsidence monitoring is a crucial technical tool for geological disaster prevention and infrastructure safety assessment. InSAR and GNSS are currently the primary technologies for surface deformation monitoring. InSAR can acquire large-scale, high-spatial-resolution surface deformation information, but its temporal resolution is limited by the satellite revisit period and is susceptible to atmospheric delays and loss of coherence. GNSS can provide high-precision, high-temporal-resolution single-point deformation data, but its spatial coverage is limited, preventing continuous monitoring over large areas.

[0003] Traditional data fusion methods, such as linear interpolation and weighted averaging, struggle to fully exploit the spatiotemporal complementarity of InSAR and GNSS data and are incapable of extracting nonlinear features from multi-source data, limiting fusion accuracy. Furthermore, monitoring data is susceptible to environmental noise, further impacting the accuracy of subsidence predictions.

[0004] In recent years, deep learning technology has demonstrated its advantages in the field of multi-source data fusion, but existing methods still have some problems. InSAR and GNSS data differ in spatiotemporal scales, making direct fusion prone to introducing errors. Traditional neural networks are weak in their ability to collaboratively learn the spatial characteristics of InSAR and the temporal characteristics of GNSS. In complex monitoring environments, existing models have limited anti-interference capabilities.

[0005] To address these issues, this paper builds on existing research by reducing and reconstructing the data to ensure that the data can be fused and calculated on the same time scale. A deep learning model with a dual-branch network architecture is constructed to extract features from InSAR and GNSS data, build subnetworks, and integrate them before training the model using a data splitting strategy. This data-driven training approach makes the model more rational, thereby improving the accuracy and robustness of the fusion prediction results. Summary of the Invention

[0006] The present invention proposes a subsidence monitoring method that integrates InSAR and GNSS based on a dual-branch network architecture. A heterogeneous neural network is designed to achieve multi-source data fusion by utilizing the high spatial resolution characteristics of InSAR images and the high temporal resolution characteristics of GNSS. This method solves the problem that traditional methods are insufficient in modeling the nonlinear relationship between heterogeneous data, and improves the spatiotemporal accuracy and robustness of subsidence monitoring.

[0007] 1. Interpolate InSAR data to GNSS resolution

[0008] Using the sampling time points of the original InSAR data and their corresponding observation values, with the help of interpolation algorithms, the estimated value of the interference point at the corresponding GNSS solution time is accurately calculated. After completing the calculation of all interference points, the estimated values ​​of each interference point are stored one by one in the corresponding column of the pre-created array in order, converting the low-time-resolution InSAR data to high-time-resolution data to obtain a more refined InSAR data time series. Assume that the InSAR observation data is {(t j , d j )}, where t j is the InSAR sampling time, d j is the corresponding observation value. The GNSS calculation time point is t k , Indicates InSAR at GNSS time t k The estimated value of f(t) is the interpolation function, which is obtained by the InSAR sampling point (t j , d j ) is constructed by interpolation algorithm. Interpolation calculation:

[0009]

[0010] All calculated Store them in an array in chronological order to form a new InSAR high temporal resolution data sequence:

[0011]

[0012] 2. Input structure

[0013] GNSS data contains N g The observation value G of each station is used as GNSS input, and the dimension is (D,N g ). InSAR data covers N i The interpolation data I of the interference points is used as the InSAR input, with the dimension (D, N i ). The fusion target dimension is kept consistent with the InSAR input dimension to ensure that the data are fused and calculated at the same time scale.

[0014] 3. Feature Extraction

[0015] Using InSAR input as the initial value of the target, dimensionality reduction and data reconstruction are performed to extract important information. A subset related to GNSS observation points is extracted from the InSAR data and fused with the GNSS data to form a new data matrix:

[0016]

[0017] in, Represents data fusion operation, I is the InSAR data covering N i The interpolation data of interference points, G is the GNSS data containing N g Observations at each site.

[0018] 3.1 Dimensionality Reduction

[0019] Using dimensionality reduction methods Process X and extract the main features:

[0020]

[0021] Among them, P is the feature representation after dimensionality reduction, and C is the dimensionality reduction processing parameter.

[0022] 3.2 Data Reconstruction

[0023] Reconstructing the map through data Restore low-dimensional features to high-dimensional data space:

[0024]

[0025] Among them, Y is the reconstructed data matrix, which serves as the target output of the final model.

[0026] 4. Build a deep learning model

[0027] Based on the GNSS and InSAR data dimensions, two independent input channels are created as input layers:

[0028]

[0029] Through a series of transformations Extract features and build GNSS sub-network; by transforming Extract features and build an InSAR subnetwork. Finally, integrate features from the GNSS and InSAR subnetworks to form a fused representation:

[0030]

[0031] in Represents the fusion operation. Defines parameters such as loss function and optimizer to build the final trainable model.

[0032] 5. Training the Model

[0033] Using data splitting strategies Divide the data according to the ratio α:

[0034]

[0035] Among them, r is the random seed, As a training set for model learning, As a test set for evaluating model performance, I is the InSAR data covering N i The interpolation data of interference points, G is the GNSS data containing N g The observation values ​​of the stations, Y is the reconstructed data matrix. Perform training and record the training process:

[0036]

[0037] Among them, E is the number of training rounds, B is the batch size, and H is the training history, including indicators such as loss value and accuracy.

[0038] 6. Model Prediction

[0039] Use the trained model to infer new data, generate fusion prediction results, and use the model Make a prediction:

[0040]

[0041] in, Represents the fusion prediction result of the model output, X g 、X i denote GNSS and InSAR observations, respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Technical flow chart

[0043] Figure 2 Model training process

[0044] Figure 3 Subsidence time series at the interference point

[0045] Figure 4 Distribution of the root mean square error of InSAR interferometric point settlement before and after fusion

[0046] Figure 5 Robustness of fusion algorithm DETAILED DESCRIPTION

[0047] The present invention belongs to the field of remote sensing, and specifically relates to a multi-source remote sensing data fusion method based on deep learning, which is used to improve the spatiotemporal accuracy and robustness of subsidence monitoring. To make the implementation of the invention clearer, the specific implementation method is described in detail with reference to the accompanying drawings.

[0048] Figure 1 :Technical flow chart

[0049] This paper provides a deep learning-based InSAR and GNSS data fusion method to improve the spatiotemporal accuracy and robustness of subsidence monitoring. The specific implementation steps are as follows:

[0050] 1. Data preprocessing: Temporally interpolate the InSAR data to match the temporal resolution of the GNSS data and construct the GNSS and InSAR input matrices to ensure temporal alignment.

[0051] 2. Feature extraction: Reduce the dimensionality of InSAR data, extract key features, and then map the features back to high-dimensional space through data reconstruction

[0052] 3. Model construction and training: Establish a dual-branch neural network to process GNSS and InSAR data respectively, then integrate the dual-branch output through the feature fusion layer, and finally divide the training set and test set for model training

[0053] 4. Settlement prediction: Use the trained model to predict new data

[0054] Figure 2 : Model training process

[0055] The data is divided proportionally according to the data splitting strategy, and the model is trained using the experimental data. The root mean square error of the objective function during the training process is as follows: Figure 2 As shown, after several iterations, the model has reached a stable state.

[0056] Figure 3 Subsidence time series at the interference point

[0057] According to the trained model, the InSAR and GNSS observations are fused. An example of the settlement of a randomly selected InSAR interferometer point is shown in the figure below. Figure 3 As shown in Figure 3, the fused sedimentation results are smoother than the InSAR interpolation results and are closer to the actual displacement.

[0058] Figure 4 Distribution of the root mean square error of InSAR interferometric point settlement before and after fusion

[0059] For all interference points, the root mean square error of each InSAR interference point before and after fusion is calculated. The root mean square error after fusion is generally smaller, such as Figure 4 As shown in Figure 3, the statistical values ​​of the root mean square error of the settlement of all interference points are significantly reduced after fusion.

[0060] Figure 5 Robustness of fusion algorithm

[0061] Adding different levels of noise to the InSAR observations, as the noise increases, the root mean square error of the original InSAR observations increases significantly, while it increases slowly after fusion. The results are as follows: Figure 5 shown.

Claims

1. A surface deformation monitoring method based on GNSS and InSAR data fusion, characterized in that: The following steps are involved: Step 1: Data preprocessing: Interpolate the InSAR data according to the GNSS solution time point to convert the low-time-resolution InSAR data into a high-time-resolution sequence that matches the GNSS time resolution. Step 2: Feature extraction: Extract the subset related to the GNSS observation points from the interpolated InSAR data, and generate a fusion feature matrix through dimensionality reduction and data reconstruction. Step 3: Model construction and training: Build a dual-channel deep learning model consisting of a GNSS subnetwork and an InSAR subnetwork. Use the feature fusion layer to integrate the two types of data features. Then, train the model based on the divided training and test sets. Step 4: Fusion prediction: Use the trained model to fuse the GNSS and InSAR observations and output the surface subsidence prediction results.

2. The method according to claim 1, characterized in that The interpolation calculation is used in step 1: Where f(t) is the interpolation function, which is obtained by the InSAR sampling point (t j , d j ) is constructed by interpolation algorithm; t j is the InSAR sampling time, d j is the corresponding observation value. The GNSS calculation time point is t k , Indicates InSAR at GNSS time t k estimated value.

3. The method according to claim 1, characterized in that The calculated value in step 1 is Store them in an array in chronological order to form a new InSAR data sequence: This enables the conversion of InSAR low temporal resolution to high temporal resolution, resulting in a more refined InSAR time series.

4. The method according to claim 1, wherein The step 1 involves the construction of GNSS and InSAR data, and the fusion target dimension is kept consistent with the InSAR input dimension.

5. The method according to claim 1, wherein In step 2, a subset related to the GNSS observation points is extracted from the InSAR data and fused with the GNSS data to form a new data matrix: in, Represents data fusion operation, I is the InSAR data covering N i The interpolation data of interference points, G is the GNSS data containing N g Observations at each site.

6. The method according to claim 1, characterized in that In step 2, the dimensionality reduction method is used Process X and extract the main features: Among them, P is the feature representation after dimensionality reduction, and C is the dimensionality reduction processing parameter.

7. The method according to claim 1, characterized in that In step 2, the data is reconstructed and mapped Restore low-dimensional features to high-dimensional data space: Where Y is the reconstructed data matrix.

8. The method according to claim 1, characterized in that In step 3, the nonlinear transformation Extract features to build GNSS sub-network, and transform Extract features and construct InSAR sub-network.

9. The method according to claim 1, characterized in that In step 3, the features from the GNSS and InSAR sub-networks are integrated to form a fused representation: in Represents fusion operations such as concatenation and weighted combination. Defines parameters such as loss function and optimizer to build the final trainable model.

10. The method according to claim 1, characterized in that In step 3, use the data splitting strategy Divide the data by the ratio ɑ: Among them, r is the random seed, As a training set for model learning, As a test set for evaluating model performance, I is the InSAR data covering N i The interpolation data of interference points, G is the GNSS data containing N g The observation values ​​of the stations, Y is the reconstructed data matrix. At the same time, the model Perform training and record the training process: Among them, E is the number of training rounds, B is the batch size, and H is the training history, including indicators such as loss value and accuracy.

11. The method according to claim 1, wherein In step 3, the model training uses the root mean square error as the loss function.

12. The method according to claim 1, characterized in that In step 4, use the model Make a prediction: in, Represents the fusion prediction result of the model output, X g 、X i denote GNSS and InSAR observations, respectively.

Images