Time-Series InSAR with Deep-Learning-Based Topography-Dependent Atmospheric Delay Correction for Potential Landslide Detection

Abstract

Synthetic aperture radar interferometry (InSAR) has emerged as an effective technique for monitoring potentially unstable landslides and has found widespread application. Nevertheless, in mountainous reservoir regions, the precision of time-series InSAR outcomes is often constrained by topography-dependent atmospheric delay (TDAD) effects. To address this limitation, we propose a novel InSAR time-series method that integrates TDAD correction. This approach employs advanced deep learning algorithms to individually model and mitigate TDAD for each interferogram, thereby enhancing the accuracy of small baseline subset InSAR (SBAS-InSAR) and stacking InSAR time-series analyses. Utilizing Sentinel-1 data, we apply this method to identify potential landslides in the Baihetan reservoir area, located in southwestern China, where we successfully identified 26 potential landslide sites. Comparative experimental results demonstrate a significant reduction (averaging 70% and reaching up to 90%) in phase standard deviation (StdDev) in the corrected interferograms, indicating a marked decrease in phase–topography correlation. Furthermore, the corrected time-series InSAR results effectively remove TDAD signals, leading to clearer displacement boundaries and a remarkable reduction in other spurious displacement signals. Overall, this method efficiently addresses TDAD in time-series InSAR, enabling precise identification of potentially unstable landslides influenced by TDAD, and providing essential technical support for early landslide hazard detection using time-series InSAR.

1. Introduction

Landslide disasters represent a significant threat in high mountainous and canyon regions across the globe, resulting in thousands of injuries, fatalities, and substantial damage to critical infrastructure [1]. The southwestern region of China is particularly susceptible to frequent landslides due to its intricate terrain and diverse geological structures, which pose significant societal and economic risks. Early identification of landslide hazards becomes imperative to mitigate casualties and avert losses resulting from such disasters.

In recent decades, synthetic aperture radar (SAR) interferometry (InSAR) technology has been widely applied in landslide monitoring due to its all-weather, all-day coverage capabilities, among other advantages [1,2,3,4,5,6,7,8,9,10,11,12,13]. Researchers have extensively utilized time-series InSAR techniques to identify potential landslide hazards [5,6,14,15,16,17,18,19,20]. Time-series InSAR techniques are still limited by atmospheric delays [21,22,23,24,25,26,27,28,29], although many methods have been proposed to correct different atmospheric delays to improve the accuracy of displacement monitoring (e.g., [27,28,30,31,32]). Among these atmospheric delays, the influence of topography-dependent atmospheric delay (TDAD) is particularly severe in many cases. This component is caused by variations in atmospheric refractivity in the vertical profile and it is highly correlated with the topography and primarily affects areas with significant terrain variations [33]. Numerous methods have been proposed to correct TDAD in InSAR, which can be broadly categorized into three types. The first type is the mathematical modeling method, where TDAD is estimated based on the phase data themselves. The most widely used method assumes a linear relationship between the interferometric phase and topography to estimate the TDAD [4,34]. The second type uses external auxiliary data to estimate TDAD, including meteorological models [34,35], GNSS data [36,37,38,39,40,41,42], spectral measurements [43,44], or a fusion of these data sources [1,45]. The third type is based on deep neural network approaches, such as [46,47]; methods of this kind are based on the premise that topography is strongly correlated with TDAD and aim to model it to reduce the influence of TDAD. Although the above methods have achieved certain results in different regions, there are some drawbacks to them. The first type, mathematical modeling, estimates the TDAD based on a certain model for all the interferograms of all conditions. It is clear that even in the same area, the interferograms with different SAR acquisitions affected by the TDAD are not the same [4,48]. The second type requires external data, which are constrained by low spatiotemporal resolution [1,23,49], and the third type, based on deep learning, has demonstrated successful TDAD correction in individual interferograms. However, currently, there is a lack of relevant cases and analysis regarding the improvement of time-series InSAR results influenced by TDAD, which deserves further research.

In this study, we propose a time-series InSAR method that incorporates TDAD correction using a deep neural network to enhance the accuracy of the interferometric phase in each interferogram. The corrected interferograms are then utilized in small baseline subset InSAR [50] (SBAS-InSAR) and stacking InSAR [51] time-series techniques. We apply this method to identify potential landslide hazards in the Baihetan reservoir area of southwestern China using Sentinel-1 data. The effectiveness and applicability of the proposed method in TDAD correction are evaluated from both qualitative and quantitative perspectives. The improved time-series InSAR algorithm can significantly reduce the influence of TDAD, enhance the reliability of time-series results, identify landslides affected by TDAD, and is of great significance for the early identification and monitoring of potential landslides.

2. Study Area and Datasets

As shown in Figure 1, the Baihetan hydropower station is located in the territory of Ningnan County, Sichuan Province, and Qiaojia County, Yunnan Province, downstream of the Jinsha River. It is the second cascade power station in the downstream mainstream of the Jinsha River, mainly for power generation, with additional functions of flood control, sediment interception, improvement of downstream navigation conditions, and development of the reservoir area for navigation purposes. The region is situated on the southeastern edge of the Qinghai–Tibet Plateau, with elevations ranging from 1000 to 3000 m. It belongs to the high mountains and plateau geomorphic unit of southwestern Sichuan and northeastern Yunnan, as well as the Transverse Mountain System. The area is primarily composed of fluvial erosion landforms, tectonic landforms, and glacial erosion landforms, with deep-cut valleys and significant weathering and erosion. This study focuses on the area of approximately 800 km² along the main channel of the Jinsha River, from Hulukou to Xiangbilin, where the Baihetan Hydropower Station is situated. The study area is densely populated, with residential areas along the river, multiple transportation routes, and intense human engineering activities. The area is characterized by poor stability, large elevation differences in slope, and sparse vegetation. Additionally, factors such as river erosion make it highly susceptible to geological hazards such as landslides, collapses, and debris flows. Therefore, it is a key section for geological hazard prevention and control in the Baihetan reservoir area.

This study utilizes Sentinel-1 SAR data for InSAR time-series analysis. The satellite data are acquired in the C-band with short revisit periods and wide coverage, and are widely applied in geological hazard monitoring. A total of 37 descending single-look complex (SLC) images were acquired for the study area (Figure 2) for the period from March 2017 to June 2018. The SLC images were obtained in VV polarization mode under the interferometric wide swath mode, with an incidence angle of 39.99°.

3. Methodology

3.1. Typical Time-Series InSAR Method Used in This Study

Differential stacking [51], based on D-InSAR technology, is a form of InSAR technique that involves the superimposition and weighted averaging of multiple-phase images generated through D-InSAR processing. Its primary objective is to reduce errors and obtain more precise surface displacement information. This method assumes that the noise phase in the interferograms is random and equal and that the surface within the region exhibits linear changes. Based on this assumption, it effectively reduces the generation of random errors. The main steps involve the selection of suitable interferometric pairs and the subsequent stacking of all interferometric images, weighted by time baseline. This process ultimately yields a displacement quantity that is accumulated based on time baselines. Consequently, in the final phase image, the quality of results is enhanced, resulting in more accurate displacement information.

PS-InSAR [52] (persistent scatterer interferometric synthetic aperture radar) technology generates interferometric pairs, and due to the selection of a single master image, it may include interferometric pairs with longer spatial baselines, making it susceptible to spatial decorrelation effects. SBAS (small baseline subset) technology, proposed by Berardino et al. [50] in 2002, is a technique that reduces the effects of temporal and spatial decorrelation and atmospheric delays in traditional InSAR techniques. The basic principle is to partition existing images into subsets by setting time and spatial baseline thresholds. This ensures that the images within a subset have small inter-image baselines, while the baselines between subsets are large. The least squares method is used to compute the subsidence sequence within each subset, and a singular value decomposition (SVD) method is employed to jointly solve all subsets, thereby obtaining the complete subsidence sequence over the entire period.

3.2. Time-Series InSAR with TDAD Deep Learning Correction

This paper takes into account the topography and displacement characteristics of the Baihetan area, as well as the ability of neural networks to approximate any function model, as indicated by the universal approximation theorem [53]. Therefore, we employ a neural network model to correct TDAD and obtain more reliable time-series InSAR results. The specific implementation process is outlined in Figure 3.

Firstly, SAR data preprocessing is conducted, involving processing N scenes of SLC images to obtain M pairs of differential interferometric phases. The main operations include registration, interferometry, filtering, phase unwrapping, and geocoding. After obtaining M pairs of differential interferometric results, considering the topography and displacement characteristics of the Baihetan reservoir area, a deep neural network model is constructed to model TDAD. The design of the deep neural network model is based on several aspects. Firstly, the characteristics of the problem to be solved are considered: (1) TDAD is related to topography, so the topography information is incorporated into the model during the modeling process; (2) long-scale delay signals often exist in the interferometric phase, so spatial location information, i.e., latitude and longitude information, needs to be considered during modeling; (3) due to the presence of decorrelation noise, the model needs to take into account the impact of noise signals. Thus, coherence data are integrated into the model. Furthermore, since different regions may have different dominant phase components (such as TDAD phase, long-scale dominant phase, etc. [47]), a channel attention mechanism module [54] is introduced in this study to address this issue. Moreover, considering the problem of underfitting and overfitting [55,56,57], appropriate network layer settings are needed. First, since the relationship between TDAD and topography is nonlinear and spatially heterogeneous, a sufficiently deep network layer is required to capture this complexity. Then, due to the limited extent of the Baihetan reservoir area, when the model is set to be too complex, its complexity may exceed the complexity of the problem. Taking these two points into consideration, this study sets the model layers to be 4096, 2048, 1024, 512, and 1. In addition, the main hyperparameter settings for training are as follows: A batch size of 8192 was used. The optimizer employed was adaptive moment estimation (Adam) [58], a widely adopted optimization algorithm in deep learning known for its merits of requiring fewer hyperparameters and delivering strong performance. The learning rate was set at 0.0002. The proposed neural network underwent 50 epochs of training. The loss function chosen was the mean squared error (MSE). The specific modeling process is as follows (1)–(9):

$$X_{4096}=RELU(FC(x_{dem},x_{lon},x_{lat},x_{cc}))$$

(1)

$$X{\prime }_{4096}=ECANet(X_{4096})$$

(2)

$$X_{2048}=RELU(X{\prime }_{4096})$$

(3)

$$X{\prime }_{2048}=ECANet(X_{2048})$$

(4)

$$X_{1024}=RELU(X{\prime }_{2048})$$

(5)

$$X{\prime }_{1024}=ECANet(X_{1024})$$

(6)

$$X_{512}=RELU(FC(X{\prime }_{1024}))$$

(7)

$$X{\prime }_{512}=ECANet(X_{512})$$

(8)

$$X_{tropo}=RELU(FC(X{\prime }_{512}))$$

(9)

Once the model construction is completed, each pair of geocoded unwrapped phases along with their corresponding elevation, latitude, longitude, and coherence data are input into the deep neural network to perform TDAD correction for each interferometric pair. After obtaining the simulated TDAD phase, it is subtracted from the original phase to obtain the corrected unwrapped phase. Finally, the corrected unwrapped phase is used for time-series analysis (i.e., stacking InSAR and SBAS-InSAR) to obtain time-series results, and a comparison is made between the pre- and post-correction time-series results to analyze the performance.

4. Results and Discussion

4.1. Corrected Time-Series Results and Verification

This paper presents a statistical analysis of the results of the time-series InSAR technique with TDAD correction for the Hulukou to Xiangbiling section in the Baihetan reservoir area, focusing on the identification of potential landslide hazards. Figure 4 illustrates the time-series InSAR displacement results in the Baihetan region after TDAD correction. Based on the InSAR displacement results corrected by the deep neural network, we successfully identified 26 potential landslide hazard points with significant displacement. All these potential landslides were validated by the optical remote sensing images and the local microtopography. After correcting TDAD, the InSAR technique demonstrates good reliability in high-precision time-series computation. Therefore, the landslides identified by InSAR exhibit high coherence and clear displacement, providing highly reliable monitoring results.

In terms of the overall distribution, the potential landslide hazard points are distributed with 9 points on the left bank and 17 points on the right bank of the Jinsha River (Figure 4). Among these, nine points exhibit significant displacements, and three of them are shown corresponding to Google Earth images in Figure 5b–d.

4.2. Comparative Analysis of Results before and after Correction

4.2.1. Comparison and Analysis of the Interference Phase before and after Correction

To evaluate the effectiveness of the model, we selected two typical interferograms from the Baihetan reservoir area for analysis. Figure 6 shows the correction results for these two interferograms, with the first three columns presenting the original interferogram, simulated interferogram, and corrected interferogram, respectively. It is evident from the figures that the selected interferograms exhibit a correlation with topography and are significantly affected by TDAD (Figure 6a,d), highlighting the necessity of TDAD correction. Comparing the original interferogram and the simulated interferogram (Figure 6a,b,d,e), we can observe a high degree of spatial similarity between them. This indicates that the model successfully captured the TDAD in the interferograms. After model correction (Figure 6c,f), the phase becomes smoother overall, and its spatial correlation with topography is significantly reduced. This suggests that the model effectively learned and mitigated the TDAD characteristics, resulting in a more reliable and accurate corrected phase, which benefits the generation of more reliable time-series InSAR results.

To validate the correction results of our method, we conducted a comparison with a traditional linear model atmospheric correction method. As shown in Figure 7, we present the results of the linear model and our deep neural network model corrections for two typical interferometric pairs in the Baihetan area. It is evident that after linear model correction, good correction results were obtained for some regions (Figure 7b,d) but the improvement is limited due to the spatial heterogeneity in atmospheric delay, which cannot be addressed using a specific model-based correction method [4,24,31], while clear global improvement was achieved (Figure 7c,f).

Furthermore, we conducted quantitative evaluations for the two typical interferograms in the Baihetan reservoir area (Figure 8). The evaluation includes the correlation between the phase and topography before and after correction and the distribution of the phase before and after correction.

In the first column, we calculated the correlation between the phase and topography before correction, and in the second column, we calculated the correlation after correction. The third column presents the distribution of the phase before and after correction. As the displacement phase is usually unrelated to topography, a reduced correlation between phase and topography indicates the successful removal of the TDAD. Figure 8a,d show that the original interferograms exhibit strong correlations with topography, with Pearson correlation coefficients of −0.87 and 0.85, respectively. After model correction, the correlation between topography and phase is significantly reduced, with Pearson correlation coefficients dropping to −0.23 and 0.19, respectively, representing an average reduction of 76% (Figure 8a–e). Moreover, after correction, the phase distribution becomes more concentrated, with its mean centered around zero, and the standard deviation (StdDev) significantly decreases from 4.81 and 4.74 to 1.39 and 1.17, representing an average reduction of 73% (Figure 8c,f).

In summary, through qualitative and quantitative evaluations of the two typical interferograms from the Baihetan reservoir area, we verified the effectiveness of the model correction. The correlation between phase and topography is significantly reduced after correction, and the phase distribution becomes more concentrated with reduced StdDev. These results indicate that the model successfully removes TDAD, thereby enhancing the reliability of the phase results and contributing to generating more reliable time-series InSAR results.

To assess the applicability of the model, we conducted a StdDev statistical analysis on all interferograms from the Baihetan reservoir area. The specific results are shown in Figure 9. Figure 9a illustrates the percentage of reduction in phase StdDev for the selected 127 differential interferograms, while Figure 9b represents the corresponding distribution.

From Figure 9a, it is evident that after correction, the StdDev of all interferograms significantly decreases, indicating the model’s excellent TDAD correction performance. However, it is essential to note that if the model lacks robustness or applicability, there might be substantial differences in the correction effects for different interferograms or regions. Therefore, when the StdDev of the percentage reduction in the interferogram StdDev is low, it indicates that the model exhibits good robustness and applicability. Figure 9b presents the distribution of the percentage reduction in interferogram StdDev. The results show that the mean percentage reduction is 0.70, with a StdDev of 0.11. This indicates that the model demonstrates excellent robustness and wide applicability, making it effectively suitable for TDAD correction.

4.2.2. Comparison and Analysis of Time-Series Results before and after Correction

As shown in Figure 10 and Figure 11, this study utilized two time-series InSAR techniques to calculate the annual average displacement rates in the Baihetan reservoir area. Figure 10A and Figure 11A show the displacement results before correction, while Figure 10B and Figure 11B show the displacement results after correction.

Firstly, a comparative analysis was conducted on the stacking InSAR TDAD corrected results in the Baihetan area. From Figure 10A, it can be observed that although phase stacking can effectively reduce random noise in the interferograms and estimate average displacement rates, the TDAD signals in the interferograms, which are non-stationary over large spatial scales, cannot be simply removed through phase stacking, leading to the presence of systematic errors in the results. Conversely, after TDAD correction, most of the systematic errors are eliminated, leaving only local small-scale residual signals (Figure 10B).

We selected two representative areas to show the performance of TDAD correction in the stacking InSAR results. As shown in Figure 10C,E, the displacement center was affected by TDAD, and this signal covered the entire region, resulting in an overestimation of the displacement results, causing unclear displacement boundaries, and thus making it difficult to accurately locate displacement boundaries (Figure 10(c2,e1)). After correction (Figure 10B), the TDAD signal was effectively removed, and the landslide boundaries appeared more distinct and accurate (Figure 10(d2,f1)). Therefore, it is concluded that the time-series InSAR results after TDAD correction can accurately characterize surface displacements, which is of great significance for the early identification of landslide hazards in the Baihetan area.

Next, the analysis was extended to the displacement results of SBAS-InSAR. As shown in Figure 11 A, in the original displacement results, long-wavelength signals did not dominate, but there existed spatially distributed signals related to topography, i.e., TDAD signals. In this case, the non-stationarity of the atmospheric delay signals in the interferograms was mainly caused by TDAD. As mentioned in Section 3, the deep neural network effectively reduces the errors caused by TDAD noise. Comparing Figure 11A and B, it can be observed that after TDAD correction, the local-topography-dependent component was effectively removed.

We selected two representative areas to show the performance of TDAD correction in the SBAS-InSAR results. From Figure 11(c1,c3,d1,d3), it is evident that within a small range, the time-series results were influenced by TDAD, resulting in overall accelerated displacement results. Based on the previous analysis of c1 and d3 (potential landslide hazard areas), it can be observed that after correction, the values in the displacement area were overestimated, and the contours became more evident. The main reason was the existence of error signals, causing an overall overestimation in that region while relatively reducing the values in the displacement area, thus disrupting the displacement contours. Consequently, the area was not identified in the pre-correction time-series results. Similar analyses apply to E(e1–e3) and F(f1–f3), and are thus not repeated here.

In conclusion, through the comparative analysis of the displacement results in SBAS-InSAR, it was found that the DNN effectively removes the TDAD signal and can identify potential landslide hazard points after TDAD correction.

5. Conclusions

This study focuses on the correction of TDAD in time-series InSAR for potential landslide identification and monitoring. We present a novel time-series InSAR algorithm based on deep learning for correcting TDAD. We used Sentinel-1 image data from 2018 to 2019 and applied the proposed method to identify the potential landslide hazards in the Baihetan reservoir area. A comparative analysis was conducted between the time-series results obtained using the new algorithm and original time-series methods. Our major findings can be summarized as follows:

1. Comparing the results of interferometric pairs considering TDAD effects, the StdDev of the interferometric phases was reduced by an average of 70% and up to 90% among 127 interferograms in the study area. The proposed method was compared with traditional approaches (GACOS), demonstrating the superiority of TDAD correction. The proposed method demonstrates certain applicability, with a StdDev of the StdDev reduction ratio being only 0.11.

2. The comparison of the time-series InSAR results before and after TDAD correction reveals that the corrected time-series displacement results show clearer displacement boundaries and can identify potential landslide hazards affected by TDAD signals. This finding is of significant importance for the early identification and warning of potentially unstable landslides.

3. The improved time-series InSAR results successfully identified 26 potential landslide hazard points in the Baihetan area. This finding provides robust support for early identification and warning of landslides, making it highly valuable for mitigating and responding to potential geological disasters.