Introduction

The Jinsha River, located in the upper reaches of the Yangtze River, covering regions with significant topographical undulations and substantial river drops, has high potential for hydropower resources1,2. Currently, there are 12 hydropower stations in the Jinsha River basin that have been completed water storage. The Baihetan Hydropower Station is the second-level power station on the lower reaches of the Jinsha River, and its installed capacity ranks second in the world3. Large-scale hydropower projects, such as Baihetan, are of great significance for clean energy production, regional economic development, and flood control and disaster mitigation4. However, the construction and operation of such large-scale infrastructure projects also involve significant geological risks, particularly the potential threat of landslides. The initial impoundment of the Baihetan Reservoir may cause changes in the mechanical properties of rock and soil due to water level fluctuations, thereby affecting the stability of the slopes on both banks5, leading to potential landslide disasters. Such disasters could cause river blockages and surge waves, threatening the safety of the neighboring communities. In addition to their immediate impacts, landslides can disrupt the long-term operation of hydropower facilities, affecting power generation and the stability of the infrastructure6. The geological formations in the reservoir area from Hulukou to Xiangbilin are diverse, and the population density is high. Hence, continuous detection of the slopes in this reservoir section are crucial for disaster prevention and mitigation. Effective landslide detection contributes to the overall safety management of large-scale hydropower projects by providing early warnings, reducing economic losses, and preventing catastrophic failures that could disrupt power generation and pose severe threats to human life and the environment.

Several methods are available for detecting and monitoring reservoir bank landslides, including ground monitoring sensors7, optical remote sensing images8, drone imagery9, and airborne LiDAR images10. However, these methods have some limitations, for example, ground sensors are difficult to use for monitoring large-scale deformations, optical imagery is affected by weather conditions, and drone and LiDAR images struggle to reflect historical conditions11. InSAR technology offers wide-area and weather-independent monitoring with the ability to detect slopes millimeter-scale historical displacements12,13. Some scholars have conducted landslide InSAR detection and monitoring research in areas, such as the Freshwater River faults14, the Heifangtai area15, the Wushan-Fengjie section of the Three Gorges Reservoir16, and Gansu Zhouqu17. Their findings demonstrate the high feasibility of InSAR technology in regional landslide detection and individual landslide deformation monitoring18. Although these studies have validated the capability of InSAR in landslide detection, they often rely on single-source SAR datasets, which limits their ability to capture comprehensive deformation signals over extended periods. Research on using multi-source SAR data for reservoir landslide monitoring remains scarce. Therefore, the use of multi-source SAR data for reservoir landslide monitoring still needs further research.

Despite the numerous advantages of InSAR, atmospheric delay remains a significant source of error, particularly in mountainous areas with complex climatic conditions, such as the Baihetan reservoir area. For example, variations in humidity can introduce deformation errors of 10–14 cm, severely compromising the millimeter-level precision required for deformation monitoring in reservoir area19. Several atmospheric correction methods have been proposed, including traditional atmospheric delay field models20, mesoscale meteorological models21, and spatial–temporal filtering22. However, traditional atmospheric delay field models are often challenged by steep terrain and significant elevation variations, which lead to pronounced local atmospheric delay effects that cannot be accurately captured by low-resolution delay models23. Mesoscale meteorological models may fail to accommodate the nonlinear spatial and temporal variations in humidity and temperature, exceeding their effective range24. Spatial–temporal filtering methods rely on statistical assumptions, which make it difficult to effectively extract true atmospheric delay characteristics under locally complex climatic conditions25. These methods will be limited in areas with highly variable terrain and climate. Therefore, for the complex terrain and climate conditions in the Baihetan reservoir area, especially in the Baihetan reservoir area during impoundment, the use of a reasonable and effective atmospheric model deserves further study, which will help improve the accuracy of InSAR monitoring.

Given this motivation, we aim to use multi-source SAR data, combine time series InSAR technology with Generic Atmospheric Correction Online Service (GACOS) external atmospheric correction data, and thus, reduce the atmospheric impact in the reservoir area, and accurately detect and update the landslides in the Hulukou to Xiangbiling section of the Baihetan Reservoir before and during impoundment. Specifically, we used SAR images collected by L-band Advanced Land Observing Satellite (ALOS)-1, C-band Sentinel-1A and Sentinel-1B satellites to process and generate atmospherically corrected deformation maps in combination with the GACOS atmospheric correction model. By aligning surface deformation with geomorphological features observed in optical images, landslides were monitored and mapped before and during impoundment. The results were subsequently validated through field investigations to determine the final landslide count. The spatial distribution patterns of landslides were analyzed, and the impact of introducing GACOS external atmospheric correction data on InSAR monitoring results in the Baihetan Reservoir area were discussed. In addition, the correlation between landslide deformation in the reservoir area and changes in reservoir water levels and rainfall was also attempted. The research results will provide valuable guidance for disaster prevention and mitigation during and after the impoundment of the Baihetan Reservoir. Additionally, this research also provides a systematic approach to the safety management of hydropower projects and serve as a reference for future landslide detection and analysis in other hydropower reservoirs and mountainous regions.

Study area

The Baihetan Hydropower Station is located on the downstream segment of the Jinsha River. The dam site on the left bank is located in Paoma Township, Ningnan County, Liangshan Yi Autonomous Prefecture, Sichuan Province, while the right bank is situated in Dazhai Town, Qiaojia County, Zhaotong City, Yunnan Province. The station is 45 km upstream from Qiaojia County and 195 km downstream from the Xiluodu Hydropower Station. The dam type is a concrete double-curvature arch dam, with a crest elevation of 834 m and a maximum dam height of 289 m. The reservoir drainage area encompasses 430,300 km2, and the installed capacity of the power station is 16,000 MW. Its average annual electricity generation is 624.43 billion kW·h. The reservoir began impounding water on April 6, 2021, with a normal storage level of 825.0 m, a flood control level of 785.0 m, and a dead storage level of 765.0 m.

The study area is located in the Baihetan Reservoir section from Hulukou to Xiangbiling, spanning 50 km upstream to 10 km downstream within Qiaojia County in Yunnan Province. It crosses five counties: Qiaojia County, Huize County, Ningnan County, Huidong County, and Dongchuan District, covering a total area of approximately 530.7 km2 (Fig. 1). According to field investigations, the study area has a high population density and intensive human engineering activities. It is also a key focus area for geological hazard prevention and mitigation within the reservoir area26. This region lies on the southeastern edge of the Tibetan Plateau. Due to the collision of the Indian and Asian plates, leading to the uplift of the Tibetan Plateau, the geological structure of the study area has experienced intense tectonic stress and deformation27, creating a unique mountainous canyon landscape (Fig. 2) and a complex geological environment.

Fig. 1
figure 1

SAR data coverage for the Baihetan Reservoir section from Hulukou to Xiangbiling. The white rectangles represent the ascending and descending coverage areas of Sentinel-1A, the purple rectangles denote the descending coverage area of Sentinel-1B, and the yellow rectangles indicate the ascending image of ALOS PALSAR-1. This map was processed using ArcMap version 10.5 (https://www.esri.com/).

Full size image
Fig. 2
figure 2

(Source: https://image.baidu.com/); (b) Broad, gentle valley in the Baihetan Reservoir area; (c) U-shaped valley in the Baihetan Reservoir area; (d) V-shaped valley in the Baihetan Reservoir area.((a), (b), (c) source: Drone aerial photography after impoundment).

Actual topography and landscape of the Baihetan Reservoir area: (a) Baihetan Hydropower Station

Full size image

The lithology of the study area is diverse, including the Quaternary System (mainly unconsolidated sediments such as gravel, sand, silt, and clay), Permian System (limestone, dolomite, and basaltic volcanic rocks), Carboniferous System (limestone, calcareous mudstone, and shale), Devonian System (limestone, quartzite, and sandstone), Silurian System (shale, siltstone, and argillaceous limestone), Ordovician System (limestone, dolomite, and calcareous mudstone), Cambrian System (black shale, sandy shale, and limestone), Sinian System (dolomite, quartzite, and siliceous rocks), and Presinian System (metamorphic rocks such as phyllite, schist, and quartzite). Simultaneously, the research area lies at the junction of the Kang-Dian Axis (II1) and the Upper Yangtze Fold Belt (II4), two secondary tectonic units, with a well-developed regional fault system (Fig. 3). The complex topography, lithology, and active tectonics have fostered favorable conditions for landslide occurrences.

Fig. 3
figure 3

Geological map of the study area. This map was processed using ArcMap version 10.5 (https://www.esri.com/).

Full size image

The study area is characterized by a subtropical monsoon climate. During the winter half-year (from November to April), the region is influenced by the southwesterly circulation of the Tibetan Plateau, resulting in clear, dry weather with scarce rainfall. Conversely, the summer half-year (from May to October) is dominated by subtropical westerlies and the southwest monsoon, leading to frequent precipitation, which is the peak period for landslides. During impoundment, fluctuating water levels might destabilize slopes, posing a significant threat to the safe operation of reservoirs.

Data and methodology

Data

SAR data

To ensure comprehensive and long-term detection of landslides before and during reservoir impoundment, and to overcome the limitations of single SAR datasets in capturing subtle and prolonged deformation signals, we used multiple sources of SAR data. SAR data from February 2007 to March 2022 were collected for the study area, which includes open-source SAR data from ALOS PALSAR-1 (2007–2010) and Sentinel-1 (Sentinel-1A and Sentinel-1B) from 2014–2022. There were no relevant SAR satellite capturing data for the study area between December 2010 and September 2014, resulting in a monitoring gap during this interval. Pre-impoundment landslides were detected using ALOS PALSAR-1 and for certain periods using Sentinel-1A SAR satellites, while landslides during reservoir water level fluctuations were detected using both Sentinel-1B and Sentinel-1A SAR satellites. In the InSAR data processing, coseismic deformation solutions were achieved by combining Sentinel-1A and Sentinel-1B ascending and descending orbit data, thereby minimizing the influence of mountainous topography on SAR images to the greatest extent possible. The basic parameters of the SAR data for the study area are presented in Table 1.

Table 1 Basic parameters of SAR data.
Full size table

Other data

Optical remote sensing imagery of the study area is sourced from the Google Earth platform, capturing images from various periods. The reservoir water level fluctuation data during impoundment were derived from the real-time water level fluctuation database for the Baihetan Reservoir area, developed by Zhejiang Huadong Construction Engineering Co., Ltd. As of March 13, 2022, the reservoir water level is 788 m. Rainfall data for the study area was provided by the Qiaojia County Meteorological Station in Yunnan Province, covering the historical monthly average rainfall from January 2016 to December 2021. For the processing of InSAR data, the topographic data utilized are the ALOS World 3D Digital Elevation Model (DEM) with a resolution of 12.5 m, provided by the Japan Aerospace Exploration Agency (JAXA). These data are incorporated as external information to eliminate the influence of the topographic phase. Using the 12.5 m ALOS-2 DEM, slope, aspect, and elevation information for the study area were extracted. Concurrently, a 1:25,000 geological map of the study area was acquired to extract information on lithology, faults, and water levels. During InSAR data processing, we used GACOS data, which is an online platform providing high-resolution atmospheric delay corrections for InSAR data. By using global atmospheric models and advanced algorithms, it effectively mitigates tropospheric delays, enhancing the accuracy of deformation monitoring28.

Methodology

Time series InSAR processing

The majority of the study area features steep terrains, humid climate, and dense vegetation. These factors often lead to SAR geometric distortions, atmospheric effects, and decorrelation issues29. Relying solely on the SAR interferometry method for landslide identification presents challenges, making it difficult to comprehensively detect and delineate landslides. As such, we propose a comprehensive procedure for landslide detection, mapping, and analysis based on InSAR technology, designed for landslide detection, mapping, and analysis before and after impoundment under complex geological conditions, as illustrated in Fig. 4.

Fig. 4
figure 4

Flowchart illustrating the process for landslides detection, mapping, and analysis before and during impoundment.

Full size image

The SBAS-InSAR technology was employed for InSAR data processing using SARscape software30. To enhance the efficiency of data processing, the imported SAR data were trimmed according to the vector range of the study area prior to processing. Multiple short temporal baseline thresholds were set for pairs of master and slave images to overcome certain temporal coherence effects. Specifically, the spatial baseline thresholds for Sentinel-1A ascending and descending orbits and Sentinel-1B descending orbit were set at 3%, 4%, and 3% of the critical baseline, respectively. The temporal baseline thresholds for the Sentinel-1A ascending and descending orbits and Sentinel-1B descending orbit were set to 72, 72, and 60 days, respectively. Given the longer wavelength of the ALOS PALSAR ascending orbit satellite, it maintains coherence over an extended period31,32. The temporal baseline threshold was set at 400 days, and the spatial baseline threshold was set at 50% of the critical baseline value. To mitigate the influence of noise, SAR images were processed using multiview and Goldstein filtering algorithms prior to interferometry, The interference pairing of different data is shown in the Fig. 5. Due to the mountainous nature of the study area and its overall low coherence, coupled with areas of dense vegetation, the filtering threshold was appropriately enhanced to eliminate as much noise as possible. Phase unwrapping was conducted using the minimum cost flow method. The deformation inversion was carried out using the least squares method and singular value decomposition (SVD)33 to estimate the deformation rate, ultimately obtaining the temporal deformation amount34. The temporal InSAR processing workflow is illustrated in Fig. 6.

Fig. 5
figure 5

Distribution of interferometric pairs under temporal and spatial baselines: (a) before impoundment ALOS-1 ascending; (b) before impoundment Sentinel ascending; (c) before impoundment Sentinel descending; (d) during impoundment Sentinel ascending; (e) during impoundment Sentinel descending.

Full size image
Fig. 6
figure 6

Flowchart of time-series InSAR processing and GACOS atmospheric correction methodology.

Full size image

Satellite imaging systems induce certain distortions in SAR images over mountainous regions, leading to phenomena such as overlay, foreshortening, and shadow-induced geometric distortions35. The emergence of these geometric distortions can severely impede the capability of SAR in monitoring surface deformations. Moreover, due to the constraints of its intrinsic flight direction, the quality of north–south surface deformation monitoring is compromised36. Acquiring a 12.5 m DEM of the study area allowed for masking of geometrically distorted areas in each interferogram during interferometric processing, ultimately yielding valid deformation results in the line-of-sight direction.

GACOS atmospheric correction method

The GACOS data is an atmospheric delay field with high temporal and spatial resolution generated using existing global meteorological data and numerical elevation models. It can effectively consider the impact of topography and meteorology on atmospheric delay, especially in the Baihetan reservoir area during impoundment. In this study, we used the GACOS data for atmospheric correction, which was developed by37. This model is based on high-resolution numerical atmospheric products from the European Centre for Medium-Range Weather Forecasts with a grid resolution of 0.1° × 0.1°38. The iterative tropospheric decomposition (ITD) model is employed to separate terrain-related vertical stratified delay components and turbulence mixed delay components from the total tropospheric delay, generating high spatial resolution (default at 90 m) zenith total delay (ZTD) maps for InSAR atmospheric delay correction39. InSAR users can submit the location of their study area and the SAR image acquisition time on the website http://www.gacos.net to obtain the ZTD products for the corresponding period.

Based on the Zenith Tropospheric Total Delay (ZTDs) data from different dates, we used MATLAB to subtract the delay map of the earlier date from that of the later date, producing atmospheric delay differential data corresponding to the dates of the interferometric pair. A relatively stable reference point was selected, and its delay value was subtracted from the computed zenith delay difference, ultimately yielding spatial and temporal atmospheric delay differential data. Finally, by subtracting the atmospheric simulated phase from the original unwrapped phase, the unwrapped phase devoid of atmospheric delay effects was obtained. The atmospheric correction procedure is depicted in the “GACOS processing” section of Fig. 6.

Results and analysis

InSAR deformation results and analysis

Utilizing the SBAS-InSAR technique with atmospheric correction based on external GACOS data, we acquired deformation information in the satellite Line of Sight (LOS) direction from the Hulukou to Xiangbiling section of the Baihetan Reservoir area before and during reservoir impoundment, as shown in Fig. 7. Prior to reservoir impoundment, a total of 25 deformation areas were detected within the study region (Fig. 7a, b, and c). Among these, the ALOS-1 satellite ascending data (from February 2007 to November 2010) detected 15 deformation areas. The Sentinel-1A ascending data (from October 2014 to August 2020) detected 15 deformation areas, and the Sentinel-1A descending data (from October 2017 to August 2020) detected 11 deformation areas. Comparing the results from the Sentinel-1A ascending and descending data, six deformation areas (H04, H06, H11, H12, H17, H20) matched in location. Relative to the monitoring period of the ALOS-1 ascending data, the monitoring period of the Sentinel-1A ascending and descending data revealed 10 additional deformation areas (H16 to H25).

Fig. 7
figure 7

Deformation rate maps of the study area before and during impoundment. (a) Deformation rate map from ALOS-1 ascending SAR satellite spanning from February 7, 2007, to November 18, 2010, before impoundment; (b) Deformation rate map from Sentinel-1A ascending SAR satellite spanning from October 19, 2014, to August 12, 2020, before impoundment; (c) Deformation rate map from Sentinel-1A descending SAR satellite spanning from October 19, 2017, to August 12, 2020, before impoundment; (d) Deformation rate map from Sentinel-1A ascending SAR satellite spanning from April 9, 2021, to March 11, 2022, during impoundment; (e) Combined deformation rate map from both Sentinel-1A and Sentinel-1B descending SAR satellites spanning from April 5, 2021, to March 13, 2022, during impoundment. This map was processed using ArcMap version 10.5 (https://www.esri.com/).

Full size image

During reservoir impoundment, fluctuations in the water level can impact the stability of banks. Accordingly, we employed Sentinel-1A ascending data (from April 9, 2021, to March 11, 2022), and combined Sentinel-1A and Sentinel-1B descending data (from April 5, 2021, to March 13, 2022) to detect 30 and 45 deformation areas within the study region, respectively, as illustrated in Fig. 7d and e. Comparing the results from the ascending and descending satellite data, 12 deformation areas aligned in their locations (H04, H11-H12, H17, H25, H29-H32, H34, H38-H39). During reservoir impoundment, Sentinel satellites detected a total of 63 deformation areas. Compared to the deformation areas detected before impoundment, 52 new deformation areas (H26-H77) emerged. In summary, a total of 77 deformation areas were detected in the study region before and during reservoir impoundment.

We calculated the mean and standard deviation of the GACOS atmospheric correction deformation rate before and during impoundment in the study area to reflect the reliability of the deformation results. The results are illustrated in Fig. 8. The standard deviations of the deformation rates derived from Sentinel-1 satellite images, both before and during reservoir impoundment, are relatively low, indicating the reliability of the deformation outcomes. In contrast, the standard deviation of the deformation rates from the pre-impoundment ALOS-1 satellite images is relatively higher than that from Sentinel-1. The discrepancies between the results from different satellite images may arise from two factors. First, there are fewer acquired scenes from ALOS-1 over the study area before impoundment, with larger and uneven intervals between each scene. This can impact the coherence and effectiveness of atmospheric corrections. Second, SAR images of different wavelengths have varying sensitivities to deformation signals. The sensitivity of Sentinel-1 images to surface deformation is higher than that of ALOS/PALSAR-1 images.

Fig. 8
figure 8

Histograms of deformation rate distributions. (a) ALOS-1 ascending SAR imagery before impoundment; (b) Sentinel-1A ascending SAR imagery before impoundment; (c) Sentinel-1A descending SAR imagery before impoundment; (d) Sentinel-1A ascending SAR imagery during impoundment; (e) Combined Sentinel-1A and Sentinel-1B descending SAR imagery during impoundment.

Full size image

Landslide detection results

For the 77 deformation zones detected by InSAR technology before and during the reservoir impoundment, combined with optical remote sensing imagery from different periods before and during reservoir filling, a total of 52 landslides were delineated based on the landslide deformation and small-scale topographic features (e.g., scarps, cracks, tension fissures, and minor slope undulations) displayed in the optical remote sensing imagery40. These landslides have been verified in the field. Among these, 21 landslides were detected before reservoir impoundment, and 41 were detected during reservoir impoundment. The distribution of the landslides is shown in Fig. 9. Compared to the period before reservoir impoundment, there were 31 newly detected landslides. The remaining 28 deformation zones in the remote sensing imagery are due to human engineering activities, with serious impacts from human activities such as migration and house construction, being evident during reservoir impoundment. As shown in Fig. 10, comparing the landslides identified before and during reservoir impoundment, landslides during impoundment showed patterns of persistent increase, weakening deformation, and new deformations. This finding indicates that the impact of reservoir water on landslide stability is multifaceted. The results of the detection of some typical landslides and their field verification are shown in Fig. 11, with specific deformation features of all detected landslides presented in Table 2. Some of the landslides detected in this study are entirely consistent with landslides mentioned in papers by other scholars, such as the Wangjiashan Landslide (shown as H25 in Figs. 10 and 11)7 and the Wulipo Landslide (shown in Figs. 10 and 11)41.

Fig. 9
figure 9

Distribution map of landslides in the study area before and during impoundment (The base image is from Google Earth dated December 1, 2020, with the water level of 788m recorded on March 13, 2022). This map was processed using ArcMap version 10.5 (https://www.esri.com/).

Full size image
Fig. 10
figure 10

InSAR deformation rates of typical landslides before and during impoundment.

Full size image
Fig. 11
figure 11

Optical remote sensing images of typical landslides before and during impoundment.

Full size image
Table 2 InSAR Deformation Information.
Full size table

As illustrated in Fig. 9, landslides in the study area are predominantly distributed in the middle and upper reaches of the Jinsha River, with a higher concentration on the left bank. Of the 52 detected landslides, 22 have foot slopes in direct contact with water (H11, H12, H77, H25, H48, H26, H50, H27, H28, H57, H19-1, H23, H16, H65, H01, H69, H71, H75, H09, H18, H45 and H47). Nine of these landslides (H11, H12, H25, H19-1, H23, H16, H01, H09 and H18) exhibited deformation signs before impoundment, and their deformation intensified during reservoir filling due to water level fluctuations. 13 landslides (H77, H48, H26, H50, H27, H28, H57, H65, H69, H71, H75, H45 and H47) were newly triggered as a direct result of the reservoir water influence during impoundment. The remaining 30 landslides are located at higher elevations, with their foot slopes not directly influenced by the reservoir water. Their deformation may be affected by other factors, such as rainfall.

According to the results of SBAS-InSAR landslide detection, while the SBAS-InSAR method has proven effective for monitoring and updating landslide inventories in reservoir areas, several limitations were encountered. First, the method relies heavily on the coherence of SAR signals, which can be significantly reduced in areas with dense vegetation or water body interference, leading to potential data gaps. Second, the temporal resolution of SBAS-InSAR is constrained by the satellite revisit cycle, which may not capture rapid or episodic landslide events, particularly during the initial stages of reservoir impoundment. In steep terrain, residual errors due to insufficient DEM resolution may further impact the accuracy of deformation estimates. To address these limitations, future studies could integrate multi-source datasets to enhance the reliability and completeness of landslide monitoring. Furthermore, combining SBAS with persistent scatterer interferometry (PSI) methods could improve monitoring in low-coherence areas.

Note: S1A-A, S1A-D, S1AB-D, and A1-A respectively represent Sentinel-1A Ascending, Sentinel-1A Descending, Sentinel-1A and 1B Descending, and ALOS PALSAR 1 Ascending. T1, T2, T3, and T4 respectively represent 2007/2-2010/11, 2014/10-2020/8, 2017/10-2020/8, and 2021/4-2022/3.

Discussion

Influence of GACOS atmospheric correction on landslide detection results

After utilizing the external GACOS data for atmospheric correction, the original unwrapped phase, the unwrapped phase after GACOS atmospheric correction, and the phase change rates were calculated separately for Sentinel-1 ascending and descending orbit data during four distinct periods: before impoundment (April 2017 to March 2018, April 2018 to March 2019, and April 2019 to March 2020) and during impoundment (April 2021 to March 2022) (Fig. 12). Concurrently, for each year from 2017 to 2021, typical relative raw unwrapped phases and those after GACOS atmospheric correction from similar periods were selected for comparison (Fig. 13). As shown in Fig. 12, the comparison of unwrapped phase results over different periods, using Sentinel-1 ascending and descending orbit data, shows that the annual phase change rates are consistently lower before impoundment compared to during impoundment, both before and after GACOS atmospheric correction. Figure 13 also clearly shows that the original unwrapped phase during impoundment, without GACOS atmospheric correction, is significantly greater than the original unwrapped phase before impoundment. Additionally, both the unwrapped phase values after GACOS atmospheric correction before and during impoundment are relatively low. This finding demonstrates that the extensive water storage in the Baihetan reservoir area brought atmospheric effects to the slopes on both sides of the Jinsha River, severely affecting the InSAR detection and monitoring results. The results of atmospheric correction using external GACOS data before and during impoundment suggest that the GACOS atmospheric correction model is highly suitable for the application of InSAR technology in reservoirs, with the most evident effects during impoundment.

Fig. 12
figure 12

(a) Mean and variation statistics of the Sentinel-1 ascending orbit original unwrapped phase and the unwrapped phase after introducing GACOS atmospheric correction for different years before and during impoundment; (b) Mean and variation statistics of the Sentinel-1 descending orbit original unwrapped phase and the unwrapped phase after introducing GACOS atmospheric correction for different years before and during impoundment.

Full size image
Fig. 13
figure 13

Comparison of the Sentinel-1 ascending and descending orbit original unwrapped phase and the unwrapped phase results after incorporating GACOS atmospheric correction for selected time periods before and during impoundment in the study area.

Full size image

To quantify the impact of reservoir level on the original phase and the atmospheric-corrected phase, we calculated the mean annual phase change rates (before and during impoundment) and the rate ratios for both ascending and descending satellite tracks, under conditions with and without GACOS atmospheric correction (Fig. 14). The rate ratio was used to quantify the relative change in phase between the pre- and post-impoundment periods. Before applying GACOS atmospheric correction, the phase change rates were relatively low during the pre-impoundment period (ascending: 0.1375 rad/year; descending: 0.025 rad/year), while the rates increased significantly during the impoundment period (ascending: 2.06 rad/year; descending: 2.73 rad/year). This indicates that the accumulation of water vapor and atmospheric delays during the impoundment period caused significant interference, particularly for ascending track data. From the perspective of the rate ratios, it is evident that atmospheric noise effects were significantly amplified during the impoundment period, overwhelming the real deformation signals. This is especially apparent in the descending track data, where the extremely high rate ratio indicates a more severe impact on descending observations. After applying GACOS atmospheric correction, the phase change rates during the impoundment period and the rate ratios for both ascending and descending tracks were substantially reduced. This demonstrates that GACOS correction effectively mitigated atmospheric delay effects, bringing the corrected data closer to the true ground deformation signals. The phase change rates during the pre-impoundment period were also improved, approaching near-zero values. It should be noted that both the mean annual phase change rates and the rate ratios were analyzed using their absolute values. This approach was adopted to eliminate the influence of directionality (positive or negative changes) and to emphasize the magnitude of phase change intensity between the pre- and post-impoundment periods. The occurrence of positive or negative values may be attributed to residual atmospheric effects and the staged variations in ground deformation within the reservoir area.

Fig. 14
figure 14

Comparison of Mean Annual Phase Change Rates and Rate Ratios Before and During Impoundment With and Without GACOS Atmospheric Correction.

Full size image

To evaluate the impact of the GACOS atmospheric correction on landslide detection, the H11 long landslide in the study area was selected as a case study. During impoundment, field investigations revealed evident signs of deformation and destruction in the landslide. The rear edge crack was well developed and length ranges from 250 ~ 365 m. The right side of the rear edge of the slope exhibited a downward displacement, and rock blocks in the central area fell sporadically, as shown in Fig. 15d and e. Without the GACOS atmospheric correction, the influence of reservoir impoundment made it difficult to accurately determine the landslide’s exact deformation location and displacement magnitude, as shown in Fig. 15b. By integrating the external GACOS data for atmospheric correction, the atmospheric effects induced by reservoir impoundment were mitigated, highlighting the landslide’s deformation location, and reliable deformation measurements were obtained, as illustrated in Fig. 15a. This finding demonstrates that reservoir impoundment influenced the surrounding atmospheric environment, thereby affecting InSAR’s capability to detect landslides. GACOS atmospheric correction can enhance the accuracy of InSAR landslide detection during the reservoir impoundment period.

Fig. 15
figure 15

(a) Deformation results of the Changdi landslide during impoundment with GACOS atmospheric correction applied; (b) Deformation results of the Changdi landslide during impoundment without incorporating GACOS external data for atmospheric correction; (c) Remote sensing imagery of the Changdi landslide during impoundment; (d) Fissures; (e) Local collapse.

Full size image

Distribution pattern of the detected landslides

The study area is characterized by complex geological conditions. Topographical features (slope, aspect and elevation), faults, lithology of strata, and distance to the reservoir water level are crucial factors influencing the formation and development of landslides within the reservoir area42. Using the 12.5 m ALOS-2 DEM, slope, aspect, and elevation information for the study area were extracted. Concurrently, a 1:25,000 geological map of the study area was acquired to extract information on lithology, faults, and water levels. Quantitative statistical analyses were conducted on the development and distribution characteristics of 52 landslide hazards detected before and during the impoundment period, as illustrated in Figs. 16 and 17.

Fig. 16
figure 16

Distribution patterns of landslide occurrences. (a) Relationship between landslides and slope gradient; (b) Relationship between landslides and slope aspect; (c) Relationship between landslides and distance to river; (d) Relationship between landslides and fault density; (e) Relationship between landslides and elevation; (f) Relationship between landslides and lithological stratigraphy.

Full size image
Fig. 17
figure 17

(a) Statistical relationship between landslide distribution and slope gradient; (b) Statistical relationship with slope aspect; (c) Statistical relationship with distance to river; (d) Statistical relationship with fault density; (e) Statistical relationship with elevation; (f) Statistical relationship with lithological stratigraphy.

Full size image

Drawing on the slope categorization criteria established by previous researchers43, the entire study area’s slope angles were divided into six classes at intervals of 10°: < 10°, 10°–20°, 20°–30°, 30°–40°, 40°–50°, and > 50°. As observed in Figs. 16a and 17a, among the 52 landslides detected, the majority lie within the 20°–40° slope range, accounting for approximately 67.3% of the total. Within the six divisions, the 30°–40° range harbors the highest landslide count, constituting approximately 40.4% of the total. The quantity of landslides in both gentle and steep slope intervals (0°–20° and > 40°) is less, with the least being in the extremely gentle and very steep categories (0°–10° and > 50°), making up approximately 3.8% of the total. This distribution pattern is consistent with the granular avalanching stability theory proposed by44. Slopes in the range of 30°–40° may be close to the limit of natural creeper stability. The statistics suggest that excessively gentle slopes possess lower gravitational potential energy, insufficient to trigger landslide failures. Conversely, overly steep slopes can result in shadow areas during SAR imaging, impeding the acquisition of valid deformation points and thus the detection of landslides. From Figs. 16b and 17b, it is evident that the 52 landslides primarily develop in the northeast, east, west, and northwest directions, comprising 96% of the total. The northeast and northwest directions dominate, accounting for 53.7% of all landslides. Landslides in the north–south direction are less prevalent, mainly because satellite sensitivity to north–south ground deformations is diminished by its own flight direction.

The distance relationship between the front edge of the 52 landslides and the water level of 788 m on March 13, 2022, was analyzed. Using an interval of 500 m, the distance between the landslides and the water level was categorized into six major classes, as presented in Figs. 16c and 17c. Notably, 42.3% of the landslides are situated within a 0–500 m distance from the reservoir water level. The stability of landslides within this range is directly influenced by the reservoir water. As the distance from the water level increases, the number of landslides decreases. Landslides located farther from the reservoir water level may have their stability influenced by factors such as rainfall and the strength of the soil and rock mass. This finding is consistent with research on the stability of granular slopes under wet conditions, which demonstrated that water infiltration reduces the effective stability angle of slopes, as developed by45. Therefore, slopes with toes submerged by the river are more prone to failure.

The fault lines in the study area were categorized by density using GIS, and divided into six classes at intervals of 0.2: 0–0.2, 0.2–0.4, 0.4–0.6, 0.6–0.8, 0.8–1, and > 1, as illustrated in Figs. 16d and 17d. Statistics indicate that a significant number of landslides, 63.4%, are distributed within a fault density range of > 0.4. These landslides show heightened activity heavily influenced by the fault lines, predominantly being rock landslides with visibly fragmented surface rocks at the site. Conversely, a minor number of landslides, 26.9%, are found within the 0–0.2 fault density range. These landslides are relatively distant from the faults and are predominantly located within the Quaternary strata, being considerably impacted by rainfall and reservoir water.

The study area’s elevation is divided into six categories at 200 m intervals: 580–800 m, 800–1000 m, 1000–1200 m, 1200–1400 m, 1400–1600 m, and > 1600 m, as depicted in Figs. 16e and 17e. Approximately 67.3% of landslides in the study area are distributed within the elevation range of 800–1200 m, the primary altitude range for landslide concentrations. Specifically, the 1000–1200 m range holds the most, constituting approximately 34.6% of the total. Within the 580–1200 m elevation span, the number of landslides increases with altitude. However, in regions exceeding 1200 m, the landslide count decreases as the elevation rises.

The study area showcases a comprehensive development of strata, ranging from the Precambrian of the early Proterozoic eon to the Quaternary of the Cenozoic era. Given the complexity in lithological characteristics of the strata, based on their hardness, the different developed strata can be divided into five categories: Quaternary (colluvium, landslide deposits, flood deposits) soft rock, relatively soft rock, semihard rock, and hard rock, as depicted in Figs. 16f and 17f. The statistics reveal that among the 52 detected landslides, 57.7% are situated within the Quaternary strata. For landslides found in other strata, as the rock layers transition from soft to hard, there is a gradual decline in their number. It is evident that most landslide disasters develop in less hard strata, which inherently possess a lower geotechnical strength. Such formations, under external factors, such as rainfall, reservoir water, earthquakes, etc., readily undergo stability degradation, leading to disasters.

Causative factors of landslide activities

Based on the time-series InSAR results incorporating the GACOS atmospheric correction model, a substantial number of landslides were detected in the Hulukou to Xiangbiling reservoir section of the Baihetan reservoir area. Among them, some landslides were not effectively detected before impoundment. During impoundment, due to their front edges being submerged and fluctuations in the reservoir water level, deformations occurred in the slope, posing serious threats to nearby assets. Taking the two landslides depicted in the figures as examples, the causes of deformation in landslides in the Baihetan reservoir area are examined. Figure 18a displays the Wulipo landslide, located approximately 1.1 km upstream of the resettlement area of Menggu Town, Qiaojia County, Yunnan Province, and 69.5 km away from the Baihetan dam site. This landslide has a longitudinal length of approximately 300 m and a transverse width of approximately 250 m and slides in the 270° direction. With slope angles ranging between 40° and 50°, the rear edge has an elevation of 920 m, a relative height difference of 190 m, and a volume of approximately 105 × 104 m3. Notably, substantial deformation appeared at the front edge of this landslide less than a month after the reservoir began impoundment. Figure 19a portrays the Shimenkan landslide. This massive ancient landslide is located on the right bank of the Jinsha River, west of the Jindong Bridge in the Dongchuan District of Yunnan Province, and is 74 km from the Baihetan dam site. The landslide extends approximately 1800 m longitudinally, is approximately 900 m wide transversely, and slides overall in the 300° direction. With slope angles ranging from 40° to 55°, on-site investigations revealed significant deformations at the front edge of this landslide during impoundment.

Fig. 18
figure 18

Relationship between Wulipo landslide deformation and reservoir water level and rainfall. (a) Average deformation rate from February 19, 2017, to March 13, 2022. (b) Temporal displacement of Point P1 in relation to rainfall and reservoir water level. (c) Temporal displacement of Point P1 during impoundment in relation to reservoir water level fluctuations and rainfall. (d) Temporal displacement of Point P2 in relation to rainfall and reservoir water level. (e) Temporal displacement of Point P2 during impoundment in relation to reservoir water level fluctuations and rainfall.

Full size image
Fig. 19
figure 19

Relationship between Shimenkan landslide deformation and reservoir water level and rainfall. (a) Average deformation rate from February 19, 2017, to March 13, 2022. (b) Temporal displacement of Point P1 in relation to rainfall and reservoir water level. (c) Temporal displacement of Point P1 during impoundment in relation to reservoir water level fluctuations and rainfall. (d) Temporal displacement of Point P2 in relation to rainfall and reservoir water level. (e) Temporal displacement of Point P2 during impoundment in relation to reservoir water level fluctuations and rainfall.

Full size image

We calculated the annual average deformation rate of the Wulipo and Shimenkan landslides from February 19, 2017, to March 13, 2022. As depicted in Fig. 18a and 19a, the entire Wulipo landslide body exhibited significant deformation, with a maximum deformation rate of approximately 55 mm/year. Additionally, a high deformation zone appeared to the upper right, reaching a maximum deformation rate of approximately 60 mm/year, which could indicate localized instability due to specific geological conditions. For the Shimenkan landslide, the highest deformation rate was approximately 30 mm/year. Due to the steep front edge of this landslide, the SAR imagery presented some shadowing, resulting in fewer deformation points at the front edge.

As illustrated in Fig. 18b, c, 19b, c, the frontal deformation of the Wulipo and Shimenkan landslides was strongly influenced by changes in reservoir water levels. Before impoundment (April 6, 2021), no evident deformations were observed in either landslide. The deformation in the Wulipo landslide began to change abruptly 23 days after impoundment (April 29), correlating with the rapid rise in reservoir water level. Similarly, the front edge of the Shimenkan landslide exhibited a sharp increase in deformation 17 days after impoundment (April 23), suggesting different response times likely caused by variations in the landslide material or slope geometry. The timing of these deformations in both landslides aligns reasonably well with the actual observed on-site landslide deformation events. These events indicate that water level changes act as a primary trigger for deformation. Notably, the different response times suggest that slope geometry, lithology, and pre-existing conditions play a critical role in determining the sensitivity of the landslide to water level fluctuations. Both landslides exhibited accelerated deformation during periods of rapid water level rise (April 6 to June 28 and August 16 to October 1). Conversely, during short periods of water level decrease (June 29 to August 15 and October 2 to November 21), the deformation rate gradually slowed but did not cease entirely. This implies that residual deformation processes remain active even with stable or decreasing water levels, likely due to internal stress redistribution or delayed pore water pressure dissipation. After November 22, as water level changes became moderate, the deformation continued to grow steadily. This steady deformation underscores the prolonged and cumulative impact of water level fluctuations on landslide stability, where even smaller changes can sustain deformation over extended periods.

As shown in Fig. 18d and e, the deformation zone in the upper right area of the Wulipo landslide starts to enter the flood season in June each year. The deformation of the slope is closely related to the monthly rainfall, with intense precipitation leading to an acceleration in deformation. This highlights rainfall as a significant contributor to deformation, independent of water level fluctuations in areas distant from the reservoir. In contrast, as shown in Fig. 19d and e, the rear edge of the Shimenkan landslide, situated at a higher elevation and further from the Jinsha River, exhibited different deformation dynamics. Before impoundment, its deformation was closely related to rainfall, with acceleration observed during heavy rainfall seasons. After impoundment, the deformation at the rear edge became influenced by both rainfall and the deformation occurring at the front edge. This interplay between the front and rear edges underscores the complexity of landslide deformation dynamics, where both external (rainfall) and internal (stress transmission) factors are at play. During impoundment from April 23 to May 4, despite minimal rainfall, an intensification in the deformation of the front edge led to a sudden increase in the deformation of the rear edge starting from May 5. This response suggests a delayed transmission of stress or pore water pressure from the front to the rear. This delayed effect is critical in understanding the overall dynamics of landslide deformation, as it demonstrates that the rear edge deformation is not only influenced by external rainfall or reservoir levels but also by internal processes initiated at the front edge. During periods of rapid water level rise outside the rainy season (May 5 to May 31 and September 1 to October 1), both the front and rear edges experienced accelerated deformation. This observation highlights the compounded effect of water level fluctuations across the landslide body, where rapid changes can synchronize deformation responses in both the front and rear edges, amplifying overall instability. Notably, after November 22, the reduction in rainfall coincided with intensified deformation at the front edge, while the rear edge continued to exhibit steady growth. This highlights that intense rainfall and water-affected deformation at the front edge collectively act as critical factors driving overall landslide instability, particularly through delayed stress propagation to the rear edge. Moreover, these findings emphasize the interconnectedness of landslide segments, where destabilization in one area can propagate and exacerbate conditions in another, necessitating holistic landslide management strategies.

Conclusions

This study integrates multi-source SAR data (L-band ALOS-1 and C-band Sentinel-1A/B) and introduces the GACOS atmospheric correction method into time-series InSAR technology, enabling the precise monitoring and mapping updates of landslides in the Hulukou to Xiangbiling section of the Baihetan Reservoir before and during impoundment. The study not only accurately map the spatial distribution of landslides in the Baihetan Reservoir area but also reveals the multifaceted impacts of water level fluctuations and rainfall on landslide deformation patterns. The findings of this study provide a scientific basis for the precise monitoring and prevention of reservoir landslide disasters and lay a foundation for improving landslide monitoring methods and technological applications in similar complex environments. The main conclusions are as follows:

  1. 1.

    The integration of multi-source SAR data has improved temporal and spatial coherence, enabling the detection of 52 landslides in the study area. Among these, 22 have toe slopes in direct contact with the water, with 9 showing intensified deformation during impoundment and 13 induced by reservoir water fluctuations. The remaining 30 landslides are located at higher elevations. Pre- and post-impoundment comparisons revealed three deformation behaviors: continuous increase, weakening, and newly emerged deformation. These results update the landslide mapping before and during impoundment and highlight the multifaceted impact of reservoir water on landslide stability.

  2. 2.

    The GACOS atmospheric correction significantly improves the accuracy and reliability of InSAR landslide monitoring. By comparing the phase change rates before and after GACOS correction, the GACOS data effectively mitigates the atmospheric effects caused by reservoir impoundment, bringing the monitoring results closer to the true ground deformation signals. After correction, the phase change rate shows a significant decreasing trend, especially when water vapor accumulation has a greater impact on the atmospheric phase during impoundment, the correction effect is particularly significant. The applicability of the GACOS model was further verified through a case study of site landslides. This provides precise corrections for localized atmospheric disturbances caused by steep terrain and water vapor accumulation during impoundment.

  3. 3.

    Landslides in the reservoir area are mainly distributed on the left bank of the Jinsha River, concentrated in areas with slopes between 30° and 40° and altitudes of 800 to 1200 m. Most of these landslides are associated with fault zones, and their distance from the reservoir water level indicates that water level fluctuations have a significant impact on slope stability. Moreover, the distribution of landslides exhibits distinct patterns in terms of topography, slope aspect, and lithology. Most landslides are located on the northeast and northwest slopes, and are common in Quaternary loose deposits and relatively soft strata. These strata are more prone to instability under the action of external factors such as rainfall and water level fluctuations due to their low geological strength.

  4. 4.

    Intense summer rainfall and rapid reservoir water level rise are two critical factors that may accelerate deformation in active landslides and reactivate unstable slopes. The deformation of the Wulipo landslide intensified approximately 23 days after the start of impoundment, whereas the Shikanmen landslide showed significant deformation about 17 days after impoundment began. These differences may be closely related to the geometric characteristics of the slopes, lithological properties, and initial conditions of the landslides. During impoundment, water level fluctuations exerted compound effects on the landslide, where rapid changes synchronized deformation responses at the front and rear edges, thereby amplifying overall instability.