High-temporal-resolution slope-aspect deformation extraction and evolution analysis of the Xiongba landslide group based on time-series InSAR and unscented Kalman filter Authors: Chuangli Jing, Rui Li, Chengming Yang, Guoxiang Liu, Xiaowen Wang, Rui Zhang, Wenfei Mao, Nasim Khonsari Published: 2026 ABSTRACT The Xiongba landslide group, characterized by continuous movement and distinct signs of reactivation, necessitates continuous monitoring to mitigate potential hazards. While Interferometric Synthetic Aperture Radar (InSAR) offers wide-area coverage and high precision, its limitation to line-of-sight (LOS) measurements and constrained temporal resolution hinders the accurate characterization of true kinematic directions and dynamic evolutionary patterns. To address these constraints, this study proposes a method that first projects LOS deformations onto the slope-aspect direction to reveal actual displacement vectors. Subsequently, an Unscented Kalman Filter (UKF) is employed to fuse ascending and descending Sentinel-1A time-series slope-aspect deformations, thereby enhancing temporal resolution and capturing the landslide’s dynamic evolution. Monitoring results based on Sentinel-1A data from January 2018 to June 2024 indicate that the Xiongba, Sela, Maiba, and Gongba landslides remain in a state of continuous activity. Notably, the Xiongba landslide exhibits maximum LOS deformation rates of 94.78 mm/yr (ascending) and 82.57 mm/yr (descending). Quantitative validation demonstrates that the UKF method significantly improves monitoring accuracy, reducing the Root Mean Square Error (RMSE) by 32.1% compared to the traditional Kalman filter, while shortening the average observation interval to 7.18 days. The evolutionary analysis reveals that the deformation is primarily driven by gravitational forces and river erosion traction, and triggered by seasonal rainfall events. Introduction The Tibetan Plateau is among the most tectonically active regions worldwide. Its distinctive geological framework and extreme climatic conditions have sculpted a landscape marked by dramatic topographic relief and deeply incised valleys1–3. This geological background provides natural conditions for the formation of landslides. In recent years, due to tectonic activities, extreme climates and human activities, landslides have occurred frequently in the Tibetan Plateau region, resulting in significant casualties, severe property losses and ecological damage4–6. The Xiongba landslide group stands as a typical representative of large-scale reactivated landslides on the southeastern margin of the Tibetan Plateau7,8. In recent years, it has remained continuously active, exhibiting pronounced signs of reactivation, and its dynamic evolution poses a significant threat to the safety of the regional geological disaster chain9. Furthermore, the two Baige landslide events in October and November 2018, located 75 km upstream, triggered a cascading disaster chain involving river blockage, dam failure, and debris flows, which caused severe erosion at the toe of the Xiongba landslide. Consequently, the implementation of high-precision and continuous deformation monitoring in this area is of paramount significance. Traditional geodetic techniques, such as precise leveling, total stations, and the Global Navigation Satellite System (GNSS), can be utilized to monitor the dynamic evolution of the Xiongba landslide group and assist in preventing potential disaster chains10–12. Although these methods offer high-accuracy measurements for localized key points within known landslide areas, they are subject to limitations such as high deployment costs, intensive labor requirements, significant susceptibility to external environmental interference, and the difficulty in achieving comprehensive monitoring of large-scale landslide groups13–15. Consequently, traditional geodetic methods are insufficient to meet the demands for the dynamic monitoring of the Xiongba landslide group. In contrast, Interferometric Synthetic Aperture Radar (InSAR), as an emerging remote sensing technology, exhibits significant advantages over traditional methods16–19. Owing to its all-weather, day-and-night, and broad-area coverage capabilities, InSAR has been widely used to monitor and assess the stability of the Xiongba landslide. For instance, Guo et al. employed the Small Baseline Subset InSAR (SBAS-InSAR) technique to map surface deformation of the Xiongba landslide between October 2017 and June 2020, subsequently analyzing the landslide’s deformation characteristics and mechanisms through remote sensing interpretation. Similarly, Zhu et al. utilized multi-orbit InSAR data to estimate the active thickness of the landslide, facilitating stability simulations. Yan et al. monitored the Xiongba-Sela landslide group via SBAS-InSAR and integrated field surveys, geological data, and rainfall records to reveal its deformation trends and underlying mechanisms. Nevertheless, despite its proven efficacy, InSAR encounters notable technical constraints in this specific context. First, the side-looking imaging geometry of SAR satellites restricts InSAR to measuring only line-of-sight (LOS) deformation20–23. This presents a fundamental constraint for landslide monitoring, as slope failure is primarily governed by movement in the downslope direction. Moreover, the temporal resolution of current SAR satellites is often insufficient (typically exceeding ten days), hindering dynamic and high-precision monitoring of hazardous landslides24–27. In recent years, researchers have proposed fusing multi-source InSAR data to enhance temporal resolution, utilizing methods such as Least Squares fusion, Bayesian estimation, the Extended Kalman Filter (EKF), and the traditional Kalman Filter (KF). However, these methods generally assume that the deformation process is approximately linear or rely on model linearization, making it difficult to effectively handle the non-linear deformation during the landslide acceleration phase. In particular, the traditional KF exhibits an insufficient response to strong non-linear deformation, thereby limiting its application in the time-series fusion of complex landslides28–30. To address the above limitations of InSAR and conventional Kalman filter-based fusion, this study proposes a novel approach. The proposed method involves transforming the time-series InSAR-derived LOS deformation into the slope-aspect direction and employs an Unscented Kalman Filter (UKF) instead of the conventional Kalman filter to fuse the slope-aspect deformation measurements. Note that projecting the LOS deformation onto the slope-aspect under the assumption of creep-type kinematics. The UKF overcomes non-linear deformation challenges through dynamic adjustment of noise covariance mdatrices and implementation of dual Unscented Transformations (UT), thereby breaking the technical bottlenecks of conventional Kalman filter method. This approach enables high-precision, high-temporal-resolution monitoring, with results expected to support dynamic evolutionary analysis, reactivation monitoring, and early warning systems for the Xiongba landslide group. Results and analysis InSAR deformation results and analysis of the Xiongba landslide group Figure 1 presents the LOS deformation velocity results for the Xiongba landslide group derived from the ascending (Fig. 1a) and descending (Fig. 1b) SAR images during January 2018 to June 2024.Fig. 1Ascending and descending annual mean deformation rate map of the Xiongba landslide group. (a) Ascending LOS deformation rate, (b) descending LOS deformation rate, (c) enlarged view of the ascending LOS deformation rate, (d) enlarged view of the descending LOS deformation rate, (e) deformation rates result for the ascending orbit9, (f) deformation rates result for the descending orbit9. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/). Ascending and descending annual mean deformation rate map of the Xiongba landslide group. (a) Ascending LOS deformation rate, (b) descending LOS deformation rate, (c) enlarged view of the ascending LOS deformation rate, (d) enlarged view of the descending LOS deformation rate, (e) deformation rates result for the ascending orbit9, (f) deformation rates result for the descending orbit9. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/). As shown in Fig. 1, both the ascending (Fig. 1a and c) and descending (Fig. 1b and d) InSAR results detected significant deformation signals within the Xiongba landslide group area (delineated by the black rectangle box in Fig. 1). The spatial extents of the detected deformation are generally consistent, with the most prominent deformations observed in the northernmost and southernmost sectors of the landslide complex. The ascending InSAR results identified four distinct deformation zones, while the descending data revealed only two. Notable differences exist in the spatial coverage of deformation between the two platforms. For instance, at the southern extension of the northernmost deformation zone, the descending dataset exhibits more complete coverage with fewer data gaps. In terms of deformation magnitude, the ascending data indicate a maximum deformation rate of 94.78 mm/yr in the northernmost section. In contrast, the descending track shows a comparatively lower maximum rate of 82.57 mm/yr, albeit at the same location. It is noteworthy that the detected InSAR deformation directions in the northernmost and southernmost areas are opposite between the ascending and descending tracks. This discrepancy can be attributed to both the actual differential slope movements in these regions and the distinct imaging geometries of the ascending and descending SAR systems31,32. This also explains the significant differences in the deformation zones detected by ascending and descending InSAR. To validate the reliability of the SBAS-InSAR results in this study, a detailed analysis was conducted against previously published InSAR deformation monitoring results for the same area by Yan et al. (see Fig. 1e and f). The comparison demonstrates a high degree of consistency between our results and those of Yan et al. in terms of spatial distribution, deformation magnitude, and identification of internal creep zones within the major active landslide, such as Xiongba landslide (H03) and Gongba landslide (H15). In particular, the core area of high deformation rate within Xiongba landslide (H03) identified in this study exhibits strong agreement with Yan et al.’s results in location, geometry, and spatial extent. Additionally, similar movement patterns were observed in the toe region of the Sela landslide (H04) and on several southern landslides (e.g., H08, H09). This consistency not only cross-validates the accuracy of our data but also establishes a reliable data foundation for subsequent high-temporal-resolution deformation analysis. Note that we can also observe slight differences in the maximum deformation rates. This may be attributed to differences in the temporal coverage of SAR images or the large deformation gradient regions with low coherence. In this study, we masked the areas with low coherence, such as large gradient deformation regions, potentially excluding some areas with large deformation rates. In contrast, the results in Yan et al. exhibit more extensive spatial coverage, which may result from the use of a lower coherence threshold. To further elucidate the relationship between InSAR-derived deformation, landslide boundaries, and associated hazards to local residents and infrastructure, a high-resolution optical satellite image was incorporated for analysis (Fig. 2). Comparative analysis of Figs. 1 and 2 clearly indicates that the four deformation zones, from north to south, correspond to the Xiongba, Sela, Maiba, and Gongba landslides, respectively. This strongly suggests that the Xiongba landslide group shows signs of reactivation, especially within the Xiongba and Gongba landslides, where InSAR LOS deformation rates exceed 60 mm/yr. It should be emphasized that multiple residential structures are located within the Xiongba and Maiba landslide areas, as well as near the northern extents of the Gongba and Maiba landslides. Failure of these landslide bodies would pose a direct threat to human lives and property. In addition, the Jinsha River flows along the toe of these landslides in a north–south direction. The entry of landslide-generated debris into the river could trigger a cascading hazard chain, such as river blockage and dam failure, leading to secondary debris flows or flood surges. Such processes would exacerbate toe erosion at downstream landslide sections, potentially accelerating their motion and resulting in progressively severe consequences.Fig. 2High-resolution optical satellite image of the Xiongba landslide group area and the locations of the landslides and Residential area. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/). High-resolution optical satellite image of the Xiongba landslide group area and the locations of the landslides and Residential area. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/). Analysis of slope-aspect deformation time-series Characteristics for the Xiongba landslide group As previously mentioned, although the spatial distribution of deformation zones derived from ascending and descending InSAR data over the Xiongba landslide group is generally consistent, notable discrepancies exist in both deformation magnitude and direction. These discrepancies are attributable to factors such as terrain and SAR satellite imaging geometry. To address this issue, we projected the ascending and descending InSAR LOS time-series deformation from January 2018 to June 2024 onto the landslide slope-aspect according to (16). We defined the downward movement along the slope-aspect as the positive direction. To investigate the temporal evolution of deformation within the Xiongba landslide group, a total of 12 deformation target points were selected for analysis, where five in the Xiongba landslide area, two in the Sela landslide area, three in the Maiba landslide area, and the remaining two in the Gongba landslide area, as shown in Fig. 3. In addition, the projection parameters of each feature point are shown in Table 1.Fig. 3LOS deformation rate of the Xiongba landslide group and the distribution of deformation target points. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/).Table 1Projection parameters for the ascending slope-aspect deformation time-series of feature points in the Xiongba landslide group.Point IDAspect (°)Slope (°)Incidence angle (°)Projection coefficientLOS deformation (cm)Slope-aspect deformation (cm)OrbitXB-165.0320.2137.12− 0.8290− 25.7031.00AscendingXB-253.4925.6037.06− 0.8423− 14.8617.64AscendingXB-327.9827.2837.12− 0.7154− 29.0540.61AscendingXB-421.5018.5537.17− 0.5759− 20.9436.36AscendingXB-543.6722.3437.14− 0.7683− 15.9420.75AscendingSL-1126.7231.3937.21− 0.7503− 13.0217.35AscendingSL-2146.7231.3937.15− 0.5960− 23.5639.53AscendingMB-190.3722.6037.12− 0.8490− 16.4019.32AscendingMB-221.1021.1437.14− 0.6012− 14.1123.47AscendingMB-365.9519.7037.05− 0.8253− 10.1812.33AscendingGB-1123.0221.1537.09− 0.6800− 29.2643.03AscendingGB-2108.2123.1837.14− 0.7897− 23.2429.43AscendingXB-165.0320.2142.690.757317.2622.79DescendingXB-253.4925.6042.780.71697.3510.25DescendingXB-327.9827.2842.68− 0.4952− 15.9932.29DescendingXB-421.5018.5542.63− 0.3318− 12.7238.34DescendingXB-543.6722.3442.67− 0.6017− 11.6019.28DescendingSL-1126.7131.3942.530.91119.4210.34DescendingSL-2146.7131.3942.590.799218.1722.74DescendingMB-190.3722.6042.560.893011.0212.34DescendingMB-221.1021.1442.53− 0.3574− 5.5515.53DescendingMB-365.9519.7042.650.75885.767.59DescendingGB-1123.0221.1542.580.857631.7437.01DescendingGB-2108.2123.1842.510.908514.5115.97Descending LOS deformation rate of the Xiongba landslide group and the distribution of deformation target points. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/). Table 1Projection parameters for the ascending slope-aspect deformation time-series of feature points in the Xiongba landslide group.Point IDAspect (°)Slope (°)Incidence angle (°)Projection coefficientLOS deformation (cm)Slope-aspect deformation (cm)OrbitXB-165.0320.2137.12− 0.8290− 25.7031.00AscendingXB-253.4925.6037.06− 0.8423− 14.8617.64AscendingXB-327.9827.2837.12− 0.7154− 29.0540.61AscendingXB-421.5018.5537.17− 0.5759− 20.9436.36AscendingXB-543.6722.3437.14− 0.7683− 15.9420.75AscendingSL-1126.7231.3937.21− 0.7503− 13.0217.35AscendingSL-2146.7231.3937.15− 0.5960− 23.5639.53AscendingMB-190.3722.6037.12− 0.8490− 16.4019.32AscendingMB-221.1021.1437.14− 0.6012− 14.1123.47AscendingMB-365.9519.7037.05− 0.8253− 10.1812.33AscendingGB-1123.0221.1537.09− 0.6800− 29.2643.03AscendingGB-2108.2123.1837.14− 0.7897− 23.2429.43AscendingXB-165.0320.2142.690.757317.2622.79DescendingXB-253.4925.6042.780.71697.3510.25DescendingXB-327.9827.2842.68− 0.4952− 15.9932.29DescendingXB-421.5018.5542.63− 0.3318− 12.7238.34DescendingXB-543.6722.3442.67− 0.6017− 11.6019.28DescendingSL-1126.7131.3942.530.91119.4210.34DescendingSL-2146.7131.3942.590.799218.1722.74DescendingMB-190.3722.6042.560.893011.0212.34DescendingMB-221.1021.1442.53− 0.3574− 5.5515.53DescendingMB-365.9519.7042.650.75885.767.59DescendingGB-1123.0221.1542.580.857631.7437.01DescendingGB-2108.2123.1842.510.908514.5115.97Descending Projection parameters for the ascending slope-aspect deformation time-series of feature points in the Xiongba landslide group. Figure 4 presents the deformation time-series of the 12 deformation target points from January 2018 to June 2024. It can be clearly seen that the landslide slope-aspect deformations derived from both ascending and descending InSAR data are consistently positive, aligning with typical landslide kinematics33. Although slight numerical differences remain between the ascending and descending time-series after projection, their overall deformation trend show strong agreement. The deformation generally follows a near-linear trend over time, with intermittent fluctuations observed during certain periods. The discrepancies between the two datasets are primarily attributed to the different observation geometries of SAR satellite systems, while these fluctuations likely result from residual atmospheric errors and seasonal precipitation influences34.Fig. 4Ascending and descending LOS and slope-aspect cumulative deformation time-series of the 12 deformation target points within the Xiongba landslide group. Ascending and descending LOS and slope-aspect cumulative deformation time-series of the 12 deformation target points within the Xiongba landslide group. Among the 12 deformation target points, XB-3, XB-4, SL-2, GB-1, and GB-2 exhibit the most significant cumulative deformation, with exceeding or approaching 40 cm over the 6.5-year monitoring period. In contrast, the target point MB-3 shows the smallest cumulative deformation, which is likely associated with local landslide stabilization measures implemented in that area. Comparative analysis of UKF and traditional KF of the Xiongba landslide group To validate the advantages of the Unscented Kalman Filter (UKF) proposed in this study in handling nonlinear landslide deformation, we conducted a comparative experiment between the UKF algorithm based on a nonlinear model and the traditional Kalman Filter (KF) based on a linear constant velocity assumption. By analyzing the slope-aspect deformation of 12 deformation feature points, we identified distinct differences in the dynamic response characteristics of the two algorithms.From a qualitative perspective, the deformation trajectory estimated by the traditional KF appears overly smooth or even rigid, As explicitly highlighted by the detail-enlargement inset boxes and annotations for subplots SL-2 and XB-3 in Fig. 5. When the landslide is in a uniform creep stage, the traditional KF is capable of fitting the main trends. However, when the landslide enters an accelerated deformation stage triggered by rainy season precipitation, the traditional KF curve fails to adjust its state in a timely manner, exhibiting an insufficient response to sudden changes and an inability to effectively track the rapid changes in observational data. In contrast, the UKF trajectory demonstrates higher dynamic sensitivity. It not only fits the overall trend of the observation data closely, similar to the traditional KF, but also tracks the real fluctuation details within the data keenly, much like a “spring.” Particularly in the nonlinear fluctuation intervals triggered by seasonal rainfall, the UKF is able to rapidly adjust displacement and velocity estimates, effectively overcoming the insufficient response to changes inherent in the linear model and more authentically reconstructing the evolutionary characteristics of the landslide.Fig. 5Fused deformation curves of the Xiongba landslide group using Unscented Kalman Filter and Kalman Filter. The inset boxes and annotations in SL-2 and XB-3 highlight key fluctuations, demonstrating the UKF’s higher dynamic sensitivity in capturing non-linear deformation details compared to the over-smoothed trajectory of the traditional KF. Fused deformation curves of the Xiongba landslide group using Unscented Kalman Filter and Kalman Filter. The inset boxes and annotations in SL-2 and XB-3 highlight key fluctuations, demonstrating the UKF’s higher dynamic sensitivity in capturing non-linear deformation details compared to the over-smoothed trajectory of the traditional KF. In terms of quantitative evaluation, this study used the mean sequence of the ascending and descending observation data as the reference truth. The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were selected as accuracy evaluation metrics, and the improvement in temporal resolution was simultaneously examined (see Table 2). The statistical results show that the RMSE and MAE of the traditional KF are 0.265 cm and 0.205 cm, respectively. In contrast, after applying the UKF algorithm, these two indicators decreased to 0.180 cm and 0.134 cm, respectively, representing accuracy improvements of 32.1% and 34.6%. Furthermore, in the temporal dimension, affected by satellite orbit planning and data quality, the actual average revisit periods for the raw ascending and descending data were as long as 16.58 days and 12.58 days, respectively, leaving substantial observational blind zones. Conversely, the average time interval of the fused deformation sequence was shortened to 7.18 days, effectively doubling the monitoring frequency. In summary, while ensuring higher fitting accuracy, the UKF method effectively overcomes the dual limitations of sparse single-orbit data and linear model lag through densified temporal sampling, thereby enabling a more precise reconstruction of the landslide’s nonlinear dynamic evolutionary process.Table 2Accuracy evaluation and temporal resolution comparison between Unscented Kalman Filter and traditional Kalman Filter.Evaluation dimensionAverage RMSE (cm)Average MAE (cm)Average revisit interval (days)Traditional KF0.2650.2057.18UKF0.1800.1347.18Original temporal resolution((ascending/descending))16.58/12.58Performance improvement32.1%34.6%50.8% Table 2Accuracy evaluation and temporal resolution comparison between Unscented Kalman Filter and traditional Kalman Filter.Evaluation dimensionAverage RMSE (cm)Average MAE (cm)Average revisit interval (days)Traditional KF0.2650.2057.18UKF0.1800.1347.18Original temporal resolution((ascending/descending))16.58/12.58Performance improvement32.1%34.6%50.8% Accuracy evaluation and temporal resolution comparison between Unscented Kalman Filter and traditional Kalman Filter. The fundamental reason for the aforementioned performance discrepancy lies in the different mathematical modeling assumptions regarding the landslide movement mechanism. The traditional KF is based on a linear state transition equation, which essentially assumes that the landslide body moves in a uniform straight line between adjacent moments. While this strong constraint effectively suppresses noise, it also smooths out real nonlinear deformation signals, leading to underfitting. The UKF designed in this study introduces a nonlinear term, allowing the model to describe velocity changes over time, which aligns more closely with the variable-speed motion characteristics of landslides driven by external forces at a physical level. Moreover, through Unscented Transformation (UT) and a deterministic sampling strategy, the UKF transmits the mean and covariance of the nonlinear system more accurately than the linearization approximation of the traditional KF, thereby maximizing the retention of high-frequency deformation signals reflecting landslide dynamic characteristics while suppressing random observation noise. High-temporal-resolution slope-aspect deformation monitoring via multi-platform InSAR fusion Given the inherent limitations of SAR satellite systems in temporal resolution and the constrained ability of conventional Kalman filter in capturing nonlinear deformation signals, this study used the UKF instead of conventional Kalman filter to fuse ascending and descending InSAR-derived slope-aspect deformation time-series over the Xiongba landslide group area. The UKF-based fusion reconstructed high-temporal-resolution slope-aspect deformations for the 12 target points distributed across the Xiongba, Sela, Maiba, and Gongba landslides from January 2018 to June 2024. To further investigate the deformation characteristics and triggering mechanisms of these target points, precipitation data were acquired from the ERA5-Land reanalysis dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), featuring a spatial resolution of 0.1° × 0.1° and a monthly temporal resolution. Statistically, the study area exhibits an annual average precipitation of approximately 480 mm. The period from June to September is classified as the rainy season, as the cumulative precipitation during these months accounts for about 70% of the annual total. These rainfall sequences were then jointly analyzed with the slope-aspect deformation time-series. The deformation time-series of the five deformation target points on the Xiongba landslide (Fig. 6a) generally exhibits a linear trend superimposed with weak periodic fluctuations. Interestingly, following each rainy season (typically June to September annually), the slope-aspect deformation rate exhibits a noticeable increase, indicating rainfall triggered the accelerated movement of the landslide. This is primarily because precipitation increases water infiltration into the superficial soil layers on the landslide, which elevates the sliding mass weight and significantly reduces the shear strength through decreased cohesion and internal friction angle at the basal slip surface. As a result, the basal resistance becomes inadequate to balance the driving forces, leading to accelerated downslope movement35,36. Moreover, we can observe a significant acceleration after December 2018 at target points XB-1, XB-3, and XB-4. This is likely associated with the two catastrophic river-blocking landslide dam events that occurred in Baige in October and November 2018. Subsequent dam breaches and flooding caused intense toe erosion of the Xiongba landslide, reducing slope stability and inducing retrogressive failure mechanisms37,38.Fig. 6The relationship between rainfall conditions and slope-aspect deformation time-series for the five deformation target points on the Xiongba landslide group. (a) The relationship between rainfall conditions and slope-aspect deformation time-series for the five deformation target points on the Xiongba landslide; (b) The relationship between rainfall conditions and slope-aspect deformation time-series for the two deformation target points on the Sela landslide; (c) The relationship between rainfall conditions and slope-aspect deformation time-series for the two deformation target points on the Maiba landslide; (d) The relationship between rainfall conditions and slope-aspect deformation time-series for the two deformation target points on the Gongba landslide. The relationship between rainfall conditions and slope-aspect deformation time-series for the five deformation target points on the Xiongba landslide group. (a) The relationship between rainfall conditions and slope-aspect deformation time-series for the five deformation target points on the Xiongba landslide; (b) The relationship between rainfall conditions and slope-aspect deformation time-series for the two deformation target points on the Sela landslide; (c) The relationship between rainfall conditions and slope-aspect deformation time-series for the two deformation target points on the Maiba landslide; (d) The relationship between rainfall conditions and slope-aspect deformation time-series for the two deformation target points on the Gongba landslide. Figure 6b shows the slope-aspect deformation time-series of the two target points, SL-1 and SL-2, on the Sela landslide, as well as monthly average rainfall data in the area. As illustrated, both deformation target points exhibit a near-linear trend from January 2018 to June 2022. The deformation rate at SL-2, situated near the southern margin of the landslide, is significantly higher than that at SL-1. This discrepancy may be attributed to enhanced fluvial erosion at the river bend where SL-2 is located, leading to reduced slope toe stability and consequently accelerated sliding. In contrast to the periodic fluctuations observed in the five deformation target points of the Xiongba landslide, neither SL-1 nor SL-2 exhibits clear seasonal accelerations correlated with rainfall. This suggests that precipitation has limited influence on the kinematic behavior of the Sela landslide. Its deformation is likely dominated by continuous gravitational driving forces rather than hydrologically triggered mechanisms. Compared with the other landslides in the Xiongba group, the Maiba landslide exhibits the smallest deformation magnitude, which can be primarily attributed to landslide mitigation measures implemented by local authorities. Over the 6.5-year monitoring period, the cumulative deformation at the deformation target points MB-1 to MB-3 within the Maiba landslide ranged from 8 to 16 cm (Fig. 6c). Among these points, MB-1 and MB-3 exhibit roughly linear deformation patterns with no significant fluctuations. In contrast, the deformation curve of MB-2 shows both a linear trend and periodic fluctuations, primarily occurring after the rainy season. These periodic deformation fluctuations are likely related to rainfall. Figure 6d displays the slope-aspect deformation time-series of the Gongba landslide. It reveals that the cumulative deformation trends at the deformation target points GB-1 and GB-2 are generally consistent, but the former exhibited a higher displacement rate. This is mainly because the former is located in an area with a steeper topographic slope. This phenomenon suggests, to some extent, that the deformation of the Gongba landslide may be primarily caused by gravity. Notably, according to Fig. 6, we found that the slope-aspect deformation of the selected 12 deformation target points showed significant fluctuations between June 2022 and April 2023, characterized by initial acceleration in the upslope direction followed by downslope movement. This fluctuating pattern is inconsistent with the known deformation pattern of landslides in this region. The comparison of the ascending and descending InSAR slope-aspect deformation time-series of these 12 points shows that the ascending InSAR data was almost completely missing between June 2022 and April 2023. Moreover, the descending InSAR slope-aspect deformation time-series were systematically lower than those of the ascending track. These two factors result in an initial decrease followed by a subsequent increase in the fused ascending-descending slope-aspect cumulative deformation, generating a “false appearance” that mimics significant uphill movement of the landslide mass. Based on the above results and analyses, the Xiongba landslide group is in a long-term state of motion with a trend of reactivation. To further explore the driving mechanisms behind the deformation of the Xiongba landslide group, we conducted a comprehensive analysis building upon previous studies. According to the rainfall-triggering model established by Yan et al.9 (Fig. 7d), rainwater infiltration not only increases the slope mass weight but, more critically, elevates the pore water pressure in the potential slip zone, substantially reducing the shear strength of the rock and soil. This serves as the direct trigger for post-rainy-season landslide acceleration. Additionally, Guo et al. visually documented7, via on-site UAV image (Fig. 7a–c), the large-scale collapses and scouring at the frontal edge of the landslide caused by erosion during this period. Such rapid undercutting disturbed the stress equilibrium of the slope, thereby inducing the “traction-type reactivation” mechanism described by Yan et al. (Fig. 7d)9. The instability initiating at the slope toe propagated upslope, leading to an overall accelerated creep stage. The high-temporal-resolution deformation time-series obtained in this study precisely captures the response process of this remote triggering effect (Fig. 6).Fig. 7Remote sensing image of the field survey in the revival area of the front edge of the Xiongba ancient landslide. (a) Deformation characteristics of the H front edge (based on UAV images); (b) erosion characteristics of rivers in the H2 front; (c) three-dimensional image of deformation characteristics of H1 front7. (d) Reactivation of ancient landslides triggered by rainfall infiltration9. Remote sensing image of the field survey in the revival area of the front edge of the Xiongba ancient landslide. (a) Deformation characteristics of the H front edge (based on UAV images); (b) erosion characteristics of rivers in the H2 front; (c) three-dimensional image of deformation characteristics of H1 front7. (d) Reactivation of ancient landslides triggered by rainfall infiltration9. Discussion Although this study utilized topographic parameters to uniformly project the ascending and descending Line-of-Sight (LOS) deformations onto the landslide slope-aspect, the deformation sequences from the two orbits still exhibited certain divergence trends in the long-term monitoring results, with discrepancies in the final cumulative deformation values. These discrepancies are not merely data processing errors but are collectively caused by the geometric characteristics of InSAR observations, the asynchronous nature of data acquisition, and the complex three-dimensional deformation mechanism of the landslide. Specifically, first, the slope-aspect projection model adopted in this paper is based on the “uni-directional translation assumption,” which assumes that the landslide body moves only along the steepest slope direction. However, actual landslide movement is often accompanied by soil compaction settlement (vertical direction) or terrain-constrained lateral creep. Since ascending and descending satellites have different flight directions and radar incidence angles, their sensitivities to vertical and north–south deformations are distinctly different. Consequently, unmodeled vertical settlement components are erroneously mapped onto the slope-aspect results, leading to systematic bias between the results of the two orbits. Second, the asynchronous acquisition of ascending and descending satellite data is another significant cause of the discrepancies. Landslide deformation is a continuous dynamic process, whereas the acquisition of ascending and descending images occurs at fixed time intervals. During this time lag, the actual displacement changes occurring in the landslide body lead to intrinsic inconsistencies in the deformation values observed by different orbits, which is particularly significant during the accelerated deformation stage triggered by rainfall. Furthermore, the Xiongba landslide group features dramatic topographic relief. Different slope-aspect regions may be in states of foreshortening, layover, or minor geometric distortion under the ascending and descending radar perspectives, respectively. These variations in geometric distortion result in different signal-to-noise ratios for deformation signals in the same region across different orbits. These factors combined ultimately result in inconsistencies after projecting ascending and descending data onto the slope-aspect. Regarding the non-linear dynamic process of a landslide transitioning from creep to an accelerated deformation stage, the Unscented Kalman Filter (UKF) introduced in this paper demonstrates methodological advantages over the traditional linear Kalman Filter (KF). The traditional KF, based on a linear state transition equation and a low process noise assumption, tends to over-smooth the deformation trajectory. This often leads to a certain lag when facing sudden rate changes triggered by rainfall or floods. In contrast, the UKF designed in this study introduces a second-order non-linear term and appropriately increases the process noise, endowing the model with higher dynamic sensitivity. Experimental results indicate that the UKF can track the real fluctuation details in the observation data keenly, like a “spring,” accurately reflecting the landslide’s response to environmental factors. Statistical results further confirm that, under the same observation noise weights, the accuracy of the UKF is higher than that of the traditional KF. This improvement in accuracy proves that the Unscented Transformation strategy of the UKF transmits the mean and covariance more accurately than linearization approximations or pure linear assumptions when dealing with non-linear systems, thereby preserving real deformation signals while suppressing random noise. Although the multi-platform InSAR fusion method based on UKF proposed in this study significantly improves the temporal resolution of landslide monitoring and effectively suppresses noise, it must be acknowledged that the method still has certain limitations regarding physical model assumptions and validation. First, the one-dimensional slope-aspect projection adopted in this study simplifies the complex kinematics of landslides, neglecting deformation components in the vertical direction (such as soil compaction or uplift) and lateral direction. As noted in related research39,40, a more rigorous approach is to perform a three-dimensional deformation solution based on the “surface-parallel flow” assumption by combining ascending and descending data. In the absence of multi-dimensional observation conditions in this study, slope-aspect projection is an effective trade-off, but neglecting multi-dimensional components may introduce systematic bias in areas with complex terrain or significant vertical settlement. Second, limited by the harsh natural environment and extremely low accessibility of the Xiongba landslide group, this study lacks continuous in-situ monitoring data (such as GNSS or inclinometers) for absolute accuracy validation. Current accuracy assessment mainly relies on cross-validation with existing InSAR research results and consistency checks between orbits. Future work will aim to overcome these limitations by combining multi-track SAR data with Pixel Offset Tracking (POT) technology and introducing surface-parallel flow constraints to reconstruct the full three-dimensional velocity field of the landslide, thereby further enhancing the scientific rigor of the deformation analysis. Materials and methods Study area overview The Xiongba landslide group is located on the western bank of the Jinsha River in Gongjue County, Tibet Autonomous Region, within the geographic coordinates 30°32′00″ N–30°37′00″ N and 98°52′30″ E–98°58′00″ E. It lies within the active fault zone of the Jinsha River, in the extremely high mountain and gorge region on the southeastern margin of the Tibetan Plateau. The climate exhibits complex characteristics of an alpine mountain-valley environment, with an annual average temperature of 4–6 °C and a diurnal temperature range exceeding 15 °C, which accelerates the freeze–thaw weathering of the rock mass. The annual precipitation ranges from 400 to 600 mm, with 70% concentrated between June and September. Heavy rainfall events (e.g., hourly rainfall > 30 mm) significantly elevate the pore water pressure along the slip zone. This is superimposed on a warming and humidification trend, with a warming rate twice the global average. The terrain in this area is steep, with ground elevations ranging from 2600 to 4000 m. The Jinsha River valley is deeply incised by over 2000 m, forming a typical “V”-shaped gorge. The slope of the landslide’s frontal edge ranges from 10° to 70° and is subject to intense scouring by the Jinsha River flow, leading to localized collapses. Geologically, the central and rear parts of the landslide body are transected by the Xiongsong-Suwalong fault (NNW-trending), a branch of the Jinsha River fault zone. The rock mass is intensely fragmented by tectonic activity. The main exposed strata are the mica quartz schist of the Permian-Lower Triassic Gangtuoyan Formation and Quaternary fragmental rock deposits7–9. Due to these unique topographic, geological, and climatic conditions, the Xiongba landslide group remains in a constant state of motion, making its deformation monitoring of great significance. Datasets This study utilizes C-band Sentinel-1A ascending and descending data to monitor the Xiongba landslide group. Sentinel-1 SAR data has a temporal resolution of 12 days and a spatial resolution of approximately 5 m (range) × 20 m (azimuth). We acquired 140 Sentinel-1A ascending SAR images spanning from January 12, 2018, to May 4, 2024, and 187 Sentinel-1A descending SAR images between January 7, 2018, and June 4, 2024. Both ascending and descending SAR images provide complete coverage of the Xiongba landslide group, as illustrated in Fig. 8. The Xiongba landslide group is indicated by the green rectangular box in Fig. 8. Table 3 provides the basic information of the acquired SAR images. To accurately co-register the SAR image pairs and remove the topographic phase in the InSAR interferograms, this study adopted the ALOS World 3D-30m (AW3D30) Digital Elevation Model (DEM) released by the Japan Aerospace Exploration Agency (JAXA). This dataset has a horizontal resolution of 30 m and a nominal vertical height accuracy (RMSE) better than 5 m, satisfying the requirements for topographic feature extraction in complex mountainous terrain.Fig. 8The study area and the coverage of SAR imagery. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/).Table 3Basic information of the Sentinel-1 images used for landslide hazard detection.Platform IDSatellite platformOrbit directionPathFrameNumber of imagesAcquisition periodIncidence angle∘(°)1Sentinel-1AAscending99128014020180112–2024050437.12Sentinel-1ADescending3349218720180107–2024060442.6 The study area and the coverage of SAR imagery. The map was prepared by the authors using QGIS version 3.40.0 (QGIS Development Team, https://qgis.org/). Table 3Basic information of the Sentinel-1 images used for landslide hazard detection.Platform IDSatellite platformOrbit directionPathFrameNumber of imagesAcquisition periodIncidence angle∘(°)1Sentinel-1AAscending99128014020180112–2024050437.12Sentinel-1ADescending3349218720180107–2024060442.6 Basic information of the Sentinel-1 images used for landslide hazard detection. SBAS-InSAR technique Time-series InSAR (InSAR) technology can retrieve millimeter-level ground deformation information by processing a stack of multi-temporal SAR images covering the same area and performing spatiotemporal adjustments. Classical time-series InSAR techniques primarily include Permanent Scatterer InSAR (PS-InSAR) and SBAS-InSAR. The former is sensitive to linear deformation but struggles to accurately capture the characteristics of accelerated deformation. Additionally, in areas with significant vegetation coverage, PS points are often sparse, resulting in poor spatial coverage. In contrast, the measurement accuracy of SBAS-InSAR is slightly lower than that of PS-InSAR, although it can effectively capture both linear and non-linear deformation information and offers better spatial coverage in vegetated areas. Given the significant vegetation cover in the Xiongba landslide group area, PS targets are likely to be sparse, using the PS-InSAR technique would make it difficult to obtain a high density of measurement points. This could lead to the omission of typical deformation information in the study area. Therefore, this study employs the SBAS-InSAR technique to monitor deformation in the Xiongba landslide group area. The SBAS-InSAR technique employs spatial and temporal baseline thresholds to flexibly select high-quality interferometric pairs from multi-temporal SAR images, which are subsequently used to calculate time-series deformation of ground targets. The basic principles and technical details of SBAS-InSAR can be found in41–43. The main workflow of SBAS-InSAR is shown in Fig. 9.Fig. 9The main workflow of high-temporal-resolution landslide deformation monitoring using SBAS-InSAR and UKF. The main workflow of high-temporal-resolution landslide deformation monitoring using SBAS-InSAR and UKF. In this study, we employed the open-source time-series InSAR processing package Mintpy to derive surface deformation measurements using the SBAS-InSAR method. The time-series analysis incorporated 140 ascending and 187 descending Sentinel-1A images acquired over the Xiongba area from January 2018 to June 2024. Considering the complex mountainous terrain of the study area, temporal and spatial baseline thresholds were set at 360 days and 150 m, respectively, to maximize the number of interferograms for the time-series analysis while minimizing the impact of spatiotemporal decorrelation. In addition, to ensure the connectivity of the interferometric network, some interferograms with good coherence were retained even if their spatiotemporal baselines exceeded the given thresholds. Ultimately, we generated 555 interferograms for the ascending track and 746 for the descending track. The spatiotemporal baseline network structures of the interferograms are illustrated in Fig. 10. To suppress speckle noise in the SAR images, interferograms were generated using a multi-look ratio of 4:1 (azimuth:range), corresponding to a ground range resolution and an azimuth resolution of approximately 14 m and 17 m, respectively. We also considered the severe atmospheric disturbances in the high-mountain and gorge terrain, and therefore, the GACOS atmospheric product were employed to correct the tropospheric delay in the InSAR data. Subsequently, a Goldstein filter were used to enhance the signal-to-noise ratio (SNR) of the interferograms. Afterward, the layover and shadow areas in interferograms were masked, followed by phase unwrapping using the Minimum Cost Flow (MCF) method to retrieve the absolute deformation phase.Fig. 10Spatiotemporal baseline plots for the ascending (top) and descending (bottom) orbits used for landslide group detection. Spatiotemporal baseline plots for the ascending (top) and descending (bottom) orbits used for landslide group detection. High-temporal-resolution slope-aspect deformation extraction from multi-platform InSAR based on the unscented Kalman filter To address the challenge that the insufficient temporal resolution of spaceborne SAR satellites poses in accurately capturing the dynamic processes of landslides, Cai et al. proposed a multi-source InSAR data integration method based on the conventional Kalman filter44. However, the conventional Kalman filter is often inadequate for accurately capturing the nonlinear deformation characteristics of a landslide body during its acceleration phase. To overcome this limitation, we utilizes the UKF with an excellent nonlinear information capture capability instead of the conventional Kalman filter45–47. As a nonlinear filtering method, the UKF retains the classic statistical framework of the Kalman filter while incorporating the Unscented Transformation (UT) to achieve accurate propagation of the mean and covariance through nonlinear systems. In this study, landslide deformation is modeled as a nonlinear function of time. Specifically retrieve, a nonlinear functional relationship between landslide deformation d and time t is established as follows,1d = x(t) Additionally, external factors such as rainfall and temperature can induce periodic displacements within a landslide body. However, such periodic signals are often overlooked in conventional deformation analysis. To incorporate these periodic influences into the deformation model, this study performs a second-order Taylor expansion on the nonlinear landslide deformation-time model. Assuming the landslide deformation is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\left( {t_{k} } \right)$$\end{document} at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{k}$$\end{document}, the second-order Taylor expansion for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\left( {t_{k} } \right)$$\end{document} at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{k - 1}$$\end{document} can be expressed as,2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\left( {t_{k} } \right) = x(t_{k - 1} ) + f^{\prime}(t_{k - 1} )(t_{k} - t_{k - 1} ) + \frac{1}{2!}f^{\prime\prime}(t_{k - 1} )(t_{k} - t_{k - 1} )^{2} + r_{{_{k - 1} }}^{(0)}$$\end{document} From (2), the first derivative of the landslide deformation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\left( {t_{k} } \right)$$\end{document} at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{k - 1}$$\end{document} is the landslide deformation velocity, and the second derivative is the landslide deformation acceleration. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{{_{k - 1} }}^{(0)}$$\end{document} is the residual term. Suppose \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{k - 1}$$\end{document} represents the landslide displacement, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{k - 1}$$\end{document} is the landslide deformation velocity, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{k - 1}$$\end{document} is the landslide deformation acceleration, (2) can be rewritten as,3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} x_{k} = x_{k - 1} + v_{k - 1} (t_{k} - t_{k - 1} ) + \frac{1}{2}a_{k - 1} (t_{k} - t_{k - 1} )^{2} + r_{{_{k - 1} }}^{(0)} \hfill \\ v_{k} = v_{k - 1} + a_{k - 1} (t_{k} - t_{k - 1} ) + r_{{_{k - 1} }}^{(1)} \hfill \\ a_{k} = a_{k - 1} + r_{{_{k - 1} }}^{(2)} \hfill \\ \end{gathered} \right.$$\end{document} The matrix form of (3) can be expressed as,4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ \begin{gathered} x_{k} \hfill \\ v_{k} \hfill \\ a_{k} \hfill \\ \end{gathered} \right] = \left[ {\begin{array}{*{20}c} 1 & {t_{k} - t_{k - 1} } & {\frac{1}{2}(t_{k} - t_{k - 1} )^{2} } \\ 0 & 1 & {t_{k} - t_{k - 1} } \\ 0 & 0 & 1 \\ \end{array} } \right]\left[ \begin{gathered} x_{k - 1} \hfill \\ v_{k - 1} \hfill \\ a_{k - 1} \hfill \\ \end{gathered} \right] + \left[ \begin{gathered} r_{{_{k - 1} }}^{(0)} \hfill \\ r_{{_{k - 1} }}^{(1)} \hfill \\ r_{{_{k - 1} }}^{(2)} \hfill \\ \end{gathered} \right]$$\end{document} The specific expression for the residual term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{{_{k - 1} }}^{{}}$$\end{document} in (3) is5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{{_{k - 1} }}^{{}} = \left[ \begin{gathered} r_{{_{k - 1} }}^{(0)} \hfill \\ r_{{_{k - 1} }}^{(1)} \hfill \\ r_{{_{k - 1} }}^{(2)} \hfill \\ \end{gathered} \right] \approx \left[ \begin{gathered} \frac{1}{6}(t_{k} - t_{k - 1} )^{3} \hfill \\ \frac{1}{2}(t_{k} - t_{k - 1} )^{2} \hfill \\ t_{k} - t_{k - 1} \hfill \\ \end{gathered} \right] \cdot g$$\end{document}where g is the third derivative of the landslide deformation at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{k - 1}$$\end{document}. By introducing the residual term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{{_{k - 1} }}^{{}}$$\end{document} of the second-order Taylor expansion of the landslide motion equation into the process error covariance matrix, the following is obtained,6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{k - 1} = \left[ {\begin{array}{*{20}c} {(\frac{1}{6}g(t_{k} - t_{k - 1} )^{3} )^{2} } & 0 & 0 \\ 0 & {(\frac{1}{2}g(t_{k} - t_{k - 1} )^{2} )^{2} } & 0 \\ 0 & 0 & {(g(t_{k} - t_{k - 1} ))^{2} } \\ \end{array} } \right]$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{k - 1}$$\end{document} is the model process noise covariance matrix. By dynamically updating the process noise and measurement noise covariance matrices at each observation epoch based on the prediction residuals and reweighting the contributions, a fused time-series of slope-aspect deformation with high-temporal-resolution can be achieved. Combining with (4)-(6), the state equation of the landslide deformation can be obtained as7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{k} = f(x_{k - 1} ,Q_{k - 1} )$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q{}_{k - 1}$$\end{document} is the process noise. Similarly, the observation equation for the landslide deformation is established as8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_{k} = h(x_{k} ,T_{k} )$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_{k}$$\end{document} is the measurement at time k, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{k}$$\end{document} represents the measurement error, and f(x) and h(x) are non-linear functions. Given that InSAR is inherently limited to measuring deformation only along the radar LOS direction, it often fails to accurately represent the true kinematics of landslides. Hence, it is generally necessary to integrate multi-platform InSAR datasets and project the LOS deformations onto the actual slope-aspect of the landslide, and the mathematical formulation as follows,9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} a = - \cos \varphi \sin \theta \cos \alpha \sin \beta + \sin \varphi \sin \theta \cos \alpha \cos \beta - \cos \theta \sin \alpha \hfill \\ z_{k} = d_{k} /a \hfill \\ \end{gathered} \right.$$\end{document}where α is the slope angle of the landslide, β represents the aspect of the landslide, φ is the heading angle, and θ is the incidence angle. The slope and aspect are calculated from the AW3D 30 m DEM. The satellite heading angle is obtained from the SAR data parameters, and the incidence angle is obtained from the local incidence angle for each image. Based on the UT concept of the UKF, a set of Sigma points is generated at time k - 1,10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{k - 1\left| {k - 1} \right.}}^{{^{i} }} = \left[ {\begin{array}{*{20}c} {\hat{x}_{{k - 1\left| {k - 1} \right.}} } & {\hat{x}_{{k - 1\left| {k - 1} \right.}} + \sqrt {(n + \lambda )P_{{k - 1\left| {k - 1} \right.}} } } & {\hat{x}_{{k - 1\left| {k - 1} \right.}} - \sqrt {(n + \lambda )P_{{k - 1\left| {k - 1} \right.}} } } \\ \end{array} } \right]$$\end{document} Then, a prediction is made for the 2n + 1 Sigma points, which can be written as:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i,k|k - 1} = f\left( {x_{i,k - 1|k - 1} ,u_{k - 1} } \right),\quad i = 0, \cdots ,2n$$\end{document} According to the predicted Sigma points from (11), a weighted average is obtained to form the final state prediction, and the corresponding covariance is as follows:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} \hat{x}_{k|k - 1} = \sum\limits_{i = 0}^{2n} {\omega_{i}^{(m)} } x_{i,k|k - 1} \hfill \\ P_{k|k - 1} = \sum\limits_{i = 0}^{2n} {\omega_{i}^{(c)} } \left( {x_{i,k|k - 1} - \hat{x}_{k|k - 1} } \right)\left( {x_{i,k|k - 1} - \hat{x}_{k|k - 1} } \right)^{T} + Q_{k - 1} \hfill \\ \end{gathered} \right.$$\end{document} Once the priori estimation from time k - 1 to k is established, another UT is performed to generate a new set of Sigma points:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{k\left| {k - 1} \right.}}^{{^{i} }} = \left[ {\begin{array}{*{20}c} {\hat{x}_{{k\left| {k - 1} \right.}} } & {\hat{x}_{{k\left| {k - 1} \right.}} + \sqrt {(n + \lambda )P_{{k\left| {k - 1} \right.}} } } & {\hat{x}_{{k\left| {k - 1} \right.}} - \sqrt {(n + \lambda )P_{{k\left| {k - 1} \right.}} } } \\ \end{array} } \right]$$\end{document} By substituting the obtained Sigma points from (13) into (8), we can get:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{Z}}_{k|k - 1}^{i} = h\left( {x_{k|k - 1}^{i} } \right),\quad i = 0, \ldots ,2n$$\end{document} Performing a weighted summation of the results derived from (14), the mean and covariance of the system’s prediction values can be obtained,15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} \hat{z}_{k|k - 1} = \sum\limits_{i = 0}^{2n} {\omega_{m}^{i} } z_{k|k - 1}^{i} \hfill \\ P_{{z_{k} z_{k} }} = \sum\limits_{i = 0}^{2n} {\omega_{c}^{i} } \left( {{\mathcal{Z}}_{k|k - 1}^{i} - \hat{z}_{k|k - 1} } \right)\left( {z_{k|k - 1}^{i} - \hat{z}_{k|k - 1} } \right)^{T} + R_{k} \hfill \\ P_{{x_{k} z_{k} }} = \sum\limits_{i = 0}^{2n} {\omega_{c}^{i} } \left( {{\mathcal{X}}_{k|k - 1}^{i} - \hat{x}_{k|k - 1} } \right)\left( {z_{k|k - 1}^{i} - \hat{z}_{k|k - 1} } \right)^{T} \hfill \\ \end{gathered} \right.$$\end{document} From (15), the Kalman gain for the UKF can be obtained,16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K = P_{{x_{k} z_{k} }} P_{{z_{k} z_{k} }}^{ - 1}$$\end{document} Finally, the non-linear system is updated as follows,17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} \hat{x}_{k|k} = \hat{x}_{k|k - 1} + K(z_{k} - \hat{z}_{k|k - 1} ) \hfill \\ P_{k|k} = P_{k|k - 1} - KP_{{z_{k} z_{k} }} K_{{}}^{T} \hfill \\ \end{gathered} \right.$$\end{document} In the equations above, (10) and (13) are the two UTs, with the scaling factor defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda = \alpha^{2} (n + \kappa ) - n$$\end{document}, Specifically, α determines the spread of the Sigma points around the mean, and is set to 0.001 in this study to avoid non-local effects; κ is set to 3 - n to minimize higher-order errors for Gaussian distributions; and β = 2 is adopted as the optimal weight selection for Gaussian distributions. (12) represents the prediction step of the UKF, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{x}_{k|k - 1}$$\end{document} is the a priori estimate, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{k|k - 1}$$\end{document} is the a priori error covariance matrix, ω represents the weights corresponding to the Sigma points, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_{k - 1}$$\end{document} is the process noise. (16) is for the Kalman gain, and (17) is the update step; they are used to calculate the Kalman gain K, the optimal estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{x}_{k|k}$$\end{document}, and the a posteriori error covariance matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{k|k}$$\end{document}, respectively. Figure 11 shows a schematic diagram of the Kalman-filtered time-series InSAR data fusion. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t_{1}$$\end{document} represents the minimum time interval from a single satellite’s orbital observation results. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t_{2}$$\end{document} represents the minimum time interval after fusing the results from dual-satellite orbits. With Sentinel-1A, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t_{2}$$\end{document} is 6 days. When one orbit is unavailable due to malfunctions or other reasons, such as during the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t_{3}$$\end{document} period, the deformation sequence in that time interval can be supplemented by another orbit, preventing the omission of anomalous landslide deformation signals.Fig. 11Schematic diagram of multi-platform time-series InSAR data fusion. Schematic diagram of multi-platform time-series InSAR data fusion. Assuming \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{1} ,\;k_{2} ...k_{n}$$\end{document} are the number of observation periods for different satellite platforms, theoretically, the optimal slope-aspect deformation monitoring period for the UKF can reach,18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K = k_{1} + k_{2} + \cdots + k_{n}$$\end{document} This study ultimately establishes a nonlinear fusion model for slope-aspect deformation based on the UKF. Through the fusion of multi-platform InSAR observation sequences, the model achieves high-temporal-resolution retrieval of deformation curves along the slope-aspect. It not only establishes a near-real-time dynamic monitoring framework for multimodal landslide movement, but also capitalizes on the strengths of multi-platform InSAR data integration. This fusion framework offers theoretical support for detailed characterization of displacement gradients throughout multi-stage landslide evolution. The proposed model first processes the SAR data from each platform using the SBAS-InSAR to acquire their respective LOS deformation time-series. Note that a common deformation reference point is uniformly established during subsequent deformation inversion calculations to mitigate systematic bias introduced by datum discrepancies. Subsequently, leveraging the key physical characteristic that the primary sliding direction of creeping landslides aligns with the slope orientation, the acquired LOS deformation time-series from each platform are projected onto a unified slope direction using a constructed simplified slope-orientation deformation extraction model. Finally, based on the principles of the UKF, and by performing two UTs and dynamically adjusting the covariance matrices of the observation noise and process noise48,49, the robustness of the model is ensured. The high-temporal-resolution nonlinear deformation curve of the slope-aspect is extracted through iterative computation. The main workflow is shown in Fig. 9. Conclusions To investigate the dynamic evolution and deformation mechanisms of the Xiongba landslide group, this study utilized ascending and descending Sentinel-1A SAR images covering the study area acquired between January 2018 and June 2024, along with the SBAS-InSAR technique, to retrieve long-term deformation time-series information. Given InSAR can only acquire SAR LOS deformation and its low temporal resolution, which is challenging to accurately obtain and interpret the landslide’s kinematic state. This study proposed an InSAR slope-aspect deformation fusion method. This approach first projects the InSAR LOS deformation onto the landslide’s slope-aspect and then, fuses the ascending and descending InSAR-derived slope-aspect deformation data using the UKF technique, enabling the extraction of high-temporal-resolution landslide slope-aspect deformation. Through analysis, the following conclusions can be drawn. The fusion of ascending and descending InSAR slope-aspect deformation data using the UKF significantly improved the monitoring accuracy over the traditional Kalman filter. Specifically, the Root Mean Square Error (RMSE) decreased from 0.265 cm with the traditional Kalman filter to 0.180 cm with the UKF. The achieved high-temporal-resolution monitoring significantly enhanced the interpretation of the InSAR data, improving the average revisit period from 14 to 7 days. Analysis of the fused slope-aspect deformation data from the 12 selected target points reveals that the deformation consists of both linear and periodic components. The linear trend is primarily driven by gravitational forces, whereas the periodic signal is strongly correlated with rainfall. The InSAR-derived deformation rates successfully identified key deformation zones, revealing that the Xiongba, Sela, Maiba, and Gongba landslides within the Xiongba landslide group are undergoing ongoing deformation. The maximum SAR LOS deformation observed in the Xiongba and Gongba landslides exceed 80 mm/yr and 60 mm/yr, respectively. Also, the slope-aspect deformations derived from ascending and descending InSAR data of the selected 12 deformation target points distributed across these four landslide areas are consistently oriented downslope. These findings have significant implications for landslide hazard monitoring, providing robust evidence that the Xiongba landslide group is currently in a reactivated state and highlighting the necessity of continuous monitoring to prevent potential landslide hazards. It should be noted that although the UKF technique can effectively enhance the temporal resolution of InSAR monitoring data by fusing multi-platform SAR data, the accuracy of the fusion results remains highly dependent on the quality of the InSAR observations and the quantity of available SAR images.