Early Detection of Geohazards in Alpine Regions Using Seasonally Partitioned InSAR: A Case Study of the Eastern Himalayan Syntaxis

Highlights

What are the main findings?

• A novel seasonal-partition Stacking-InSAR method is proposed, which separately processes winter and summer interferograms based on InSAR coherence variation to mitigate seasonal decorrelation in alpine regions.
• Applied in the eastern Himalayan syntaxis, the method identified over 26% more geohazards than conventional Stacking-InSAR, with less than 19% overlap between hazards detected in winter and summer, highlighting strong seasonal activity differences.

What are the implications of the main findings?

• The method enables more accurate and comprehensive early detection of both high-altitude concealed hazards (best detected in summer) and riverside landslides (best detected in winter), directly supporting disaster prevention in high, cold mountains.
• It demonstrates that incorporating qualified long-temporal-baseline interferometric pairs can improve the detection of slow-creeping slopes, and the framework is particularly suitable for regions with significant seasonal InSAR coherence variations.

Abstract

In alpine mountain regions, significant seasonal surface changes reduce InSAR coherence over long time spans, hindering geohazard identification. This study proposes a method for geohazard detection based on InSAR seasonal coherence variation. First, time-series interferograms and coherence maps are generated from Sentinel-1 imagery. Each year is then partitioned into summer, transition, and winter seasons by analyzing the spatial migration of high-coherence zones. Interferometric pairs from the transition season are further screened and reassigned to summer or winter groups according to their coherence characteristics. Stacking-InSAR is applied separately to the summer and winter datasets to derive seasonal deformation rates; long-temporal-baseline pairs (60–120 days) that maintain sufficient coherence are selectively incorporated to improve the detectability of slow-moving slopes. Finally, geohazards are identified by combining the summer and winter deformation results. Applied in the eastern Himalayan syntaxis, the method showed that less than 19% of geohazards were detectable in both seasons, indicating seasonal variations in geohazard activity. Moreover, it identified approximately 29% more geohazards on average than traditional Stacking-InSAR using all interferograms. Thus, the proposed approach enables more accurate and effective geohazard detection in cold mountains, supporting disaster prevention and mitigation.

1. Introduction

Geohazards such as landslides and debris flows are widespread in mountainous western China, posing significant threats to infrastructure and communities [1]. The eastern Himalayan tectonic syntaxis, at the collision front of the Indian and Eurasian Plates, provides the necessary tectonic, topographic, and climatic conditions for widespread geohazard occurrence [2,3,4]. In recent years, significant disaster events in the eastern tectonic junction area have included the Layue landslide (1967), the 102 Daoban landslide (1991), and the Yigong landslide (2000). These disasters caused severe damage to National Highway 318, nearby townships and residential areas, and even affected regions downstream, such as India [5,6,7]. Therefore, it is particularly urgent to conduct accurate and timely early identification of geohazards in these areas.

The eastern Himalayan tectonic syntaxis exhibits a typical alpine landscape—high relief, dense vegetation, volatile climate, and active tectonics—which makes it a hotspot for concealed geohazards and limits conventional surveys [8,9,10,11]. Spaceborne InSAR, with its all-weather, wide-coverage, and sub-centimeter capabilities, overcomes these accessibility constraints and has become a key tool for early hazard detection [1,12,13]. However, InSAR deformation retrieval fundamentally depends on interferometric coherence, and the long spatiotemporal baselines typical of alpine settings often cause severe decorrelation that obscures surface signals. Multi-temporal InSAR techniques alleviate this issue through spatiotemporal filtering and deformation modeling; among them, Stacking-InSAR is widely adopted for large-scale hazard screening because of its computational simplicity and its ability to suppress random phase errors [14,15].

Nevertheless, it is critical to emphasize that maintaining coherence quality remains a persistent challenge in InSAR data processing. When radar echo similarity between two satellite passes is low, the resulting interferometric decorrelation significantly compromises InSAR-based deformation analysis, thereby introducing substantial uncertainties in geohazard identification [12]. Low-coherence zones predominantly occur in areas with temporally dynamic surface scattering properties, typically corresponding to dense vegetation, high-altitude snow/ice cover, or wetland/grassland ecosystems under natural geomorphic conditions. In such regions, coherence fluctuations not only degrade time-series deformation accuracy but also increase risks of false positives and missed detections in geohazard identification.

Consequently, in alpine terrain where both vegetation phenology and snow/ice cover vary markedly with the season, the coherence available for InSAR is not only low but strongly time-dependent. A conventional Stacking-InSAR approach indiscriminately uses all available interferograms. Consequently, seasonal coherence perturbations propagate into the deformation estimates. This can mask active deformation signals in one elevation belt while generating false positives in another. The resulting systematic omission or misidentification of geohazards defines a critical scientific problem. The challenge is to characterize this seasonal coherence variability and design a processing strategy that mitigates these elevation-dependent blind zones. Such a strategy would improve the early detection rate of landslides and rockfalls in high-altitude, densely vegetated environments. Recent InSAR-based seasonal analyses have largely concentrated on permafrost deformation [16,17] or seasonal velocity variations in individual landslides [18,19], but none have systematically exploited the seasonal migration of coherence zones to partition interferograms for Stacking-InSAR. The resulting blind zones, where hazards are systematically missed or misidentified, have not been addressed.

This study proposes a seasonal-partitioning Stacking-InSAR method for alpine geohazard detection. In contrast to conventional Stacking-InSAR that uses all interferograms indiscriminately, our approach defines seasons by analyzing spatial migration of coherence zones, reassigns qualified transition-season interferograms to summer or winter groups based on coherence, and selectively incorporates long-temporal-baseline pairs (60–120 days) to improve sensitivity to slow-moving slopes. Interferometric pairs are stratified into winter and summer datasets and processed separately. Summer-dominant coherence in high-altitude areas reveals concealed high-elevation hazards, while winter coherence in sparsely vegetated valleys benefits riverside landslide identification. The framework was validated in the eastern Himalayan syntaxis using Sentinel-1 data, with results cross-checked against existing surveys and multi-temporal optical imagery, confirming improved reliability in mitigating seasonal decorrelation.

2. Overview and Data of the Study Region

2.1. Study Region

The eastern Himalayan tectonic syntaxis is situated in the southeastern Tibetan Plateau, at the forefront of the Indian Plate’s subduction beneath the Eurasian Plate (Figure 1). This region exhibits rugged terrain and significant vertical relief, with elevations ranging from 1200 to 7500 m and local height differences exceeding 5000 m. Glacial meltwater has fostered dense river networks, contributing to the predominance of deep canyon landforms. The bedrock is predominantly Proterozoic gneiss, extensively overlain by Quaternary surficial deposits that mainly consist of alluvial–pluvial gravelly soils, colluvial–talus gravelly soils, and morainic boulder–gravel soils. Major tectonic structures include the Yarlung Tsangpo, Jiali, Zhamu–Maniong, and Milin regional active fault zones, which not only govern the courses of the Yarlung Tsangpo, Yigong Tsangpo, and Parlung Tsangpo rivers but also fundamentally control the development and distribution of geohazards in the study area. Under the joint influence of these lithological conditions and active faults, slopes are typically high and steep, the bedrock is highly jointed and fractured, and the Quaternary deposits are thick and poorly consolidated—making slopes particularly prone to landslides, rockfalls, and other instability under rainfall, snowmelt, or seismic action [3,5,6,7,20]. Climatically, the area is characterized by complex seasonal variations, with temperatures and precipitation showing regular fluctuations. Based on daily records from meteorological station No. 56312 (94°22′E, 29°37′N) during the study period (2018–2021), the annual temperature ranges from −5 °C to 20 °C, and the snow cover period typically lasts from January to March. Winter brings extensive snow and ice cover, while rainfall is concentrated between June and September, with a mean annual precipitation of approximately 1715 mm. The combination of dense river systems and summer precipitation supports lush vegetation along valley slopes. In general, the typical deep canyon landform in the study area and the periodic freeze–thaw effect in the high cold region lead to frequent geohazards such as rockfalls, landslides, and debris flows, with some high-altitude hazards being particularly concealed and high-risk.

These frequent disasters have posed severe threats to National Highway 318, local county/township roads, and residential safety in surrounding towns. For instance, the 2000 Yigong landslide involved a high-altitude rock avalanche that rapidly descended the slope, destroying local topography and blocking the Yigong Tsangpo River, forming a barrier lake. Subsequent dam breaches and disaster chain effects caused widespread devastation downstream, destroying bridges and settlements and extending impacts as far as India [21,22,23]. Consequently, early detection of geohazards in this high-altitude alpine region is critical, particularly for identifying evolving high-altitude, high-risk hazards in their nascent stages.

2.2. Dataset

This study utilizes SAR imagery from the European Space Agency’s (ESA) Sentinel-1 satellite as the primary data source (

Table 1. Parameters of Sentinel-1A SAR data.

SAR Parameter Type	Basic Parameters
Orbit Type	Sun-synchronous orbit
Band Mode	C Band (wavelength 5.56 cm)
Satellite Revisit Period (days)	12
Satellite Imaging Mode	Interferometric Wide swath (IW)
Spatial Resolution (m)	2.3 × 13.9
Swath Width (km)	250
Orbit Direction	Ascending	Descending
Incidence Angle (°)	39.4	39.5
Heading Angle (°)	347.3	192.7
Number of Acquisitions	111	110
Monitoring Period	10 January 2018–20 December 2021	5 January 2018–27 December 2021

), covering the monitoring period from January 2018 to December 2021, with the spatial extent of the SAR data illustrated by the red dashed box in Figure 1. The Sentinel-1 images were acquired in Interferometric Wide swath (IW) mode from both ascending and descending orbits, with a revisit period of 12 days, and approximately 110 scenes were obtained for each orbit over the monitoring period (see Table 1 for further details). To address satellite orbital errors, Precise Orbit Determination (POD) data provided by the ESA were employed to refine and correct the Sentinel-1 orbital parameters [24]. For surface deformation retrieval, SRTM-DEM data were applied to simulate and remove the topographic phase from the interferograms. Furthermore, given the region’s complex climatic conditions and the potential for atmospheric disturbances to introduce errors in deformation calculations, tropospheric delay (ZTD) data with a 90 m spatial resolution from the Generic Atmospheric Correction Online Service (GACOS) were incorporated to perform atmospheric correction on the InSAR interferograms, thereby mitigating atmospheric delay artifacts and enhancing the accuracy of InSAR-based deformation monitoring [13,25].

3. Methodology

3.1. Data Preprocessing

Prior to time-series InSAR processing, D-InSAR differential interferometry (Figure 2) was applied to the collected SAR images, following a structured workflow: (1) Winter and spring SAR images were selected as the common master images, with all other images co-registered to them to ensure geospatial consistency across acquisitions. (2) A spatial baseline threshold of 200 m and a temporal baseline threshold of 120 days were set for differential interferometry to generate a sufficient number of candidate interferometric pairs (Figure 3). Notably, interferograms with a temporal baseline ≤ 60 days generally maintained high coherence, whereas those with 60–120 days showed reduced but still acceptable coherence for a subset of pairs. Consequently, in subsequent seasonal Stacking-InSAR processing, the default dataset consists of the 60-day short-temporal-baseline pairs, while a coherence-filtered selection of 60–120 day pairs is added only when needed to improve signal clarity (see Section 4.3). Both the short-temporal-baseline (<60 days) and the long-temporal-baseline (60–120 days) interferograms were screened using coherence thresholds defined in Section 4, ensuring that only pairs with adequate signal quality are retained for deformation calculation. (3) Topographic phase removal and atmospheric error correction were performed using external 30 m resolution DEM data and GACOS atmospheric data. (4) Goldstein adaptive filtering was applied to the differential interferograms with a 128 × 128-pixel window to enhance fringe quality and better reveal coherence trends (Figure 4). (5) The Minimum Cost Flow (MCF) method was employed for phase unwrapping. (6) Interferometric pairs with poor coherence or severe atmospheric noise were excluded. And (7) the interferometric pairs were stratified into winter and summer datasets following the quantitative criterion detailed in Section 4.1. Specifically, each transition-season pair is assigned to winter or summer according to the comparison of its mean coherence within the pre-defined winter and summer high-coherence zones (Figure 5a) [23,26].

3.2. Stacking-InSAR Deformation Extraction

Stacking-InSAR (phase-weighted stacking) is a sequential InSAR processing technology based on interferometric phase stacking. Firstly, InSAR differential interferometry is performed on multiple SAR images in the region in pairs within the set time baseline (see Section 3.1 for the process), and the unwrapping phase obtained by D-InSAR processing is further linearly superimposed. According to the weighted average of time span (Figure 2), the average annual deformation information of radar line-of-sight (LOS) in the study area is finally obtained (Formula (1)).

$$P_{rate}={\displaystyle \frac{{\sum }_{i=1}^N{\phi }_i\Delta t}{{\sum }_{i=1}^N\Delta t^2}}$$

(1)

In the equation, P_rate represents the annual average phase rate, φ_i denotes the unwrapped phase value of a single interferometric pair, Δt is the temporal baseline of the interferometric pair, and N is the number of interferometric pairs included in the calculation. Since atmospheric delay noise and orbital errors in InSAR typically manifest as high-frequency noise in the time domain while exhibiting spatial autocorrelation in the space domain, the stacking and weighted averaging of unwrapped phases effectively suppress these noise components, thereby enhancing the accuracy of deformation retrieval [27].

4. Stacking-InSAR Processing with Seasonal Coherence Difference

To acquire InSAR interferometric pairs with enhanced deformation signals, this study first performed full interferometric processing on the SAR image dataset using a 120-day temporal baseline (Figure 3). For the ascending track data, a total of 1012 interferometric pairs were generated, from which 113 pairs with low coherence or severe atmospheric noise were excluded. Interferograms were classified by acquisition date into winter (Jan, Feb, Mar), summer (Jul, Aug, Sep, Oct), and transition season (Apr–Jun, Nov–Dec). Transition-season pairs were then quantitatively assigned to the winter or summer group based on a zone-based coherence comparison criterion detailed in Section 4.1. After this assignment, the remaining pairs were stratified into 423 winter pairs and 476 summer pairs. Similarly, for the descending track data, 895 interferometric pairs were initially generated, with 143 pairs removed due to significant errors, and the remaining pairs were classified into 411 winter pairs and 341 summer pairs based on coherence. Subsequently, Stacking-InSAR processing was applied separately to the full dataset, winter subset, and summer subset for both ascending and descending tracks to extract landslide deformation information (Figure 2).

4.1. InSAR Coherence Characteristics in Cold Mountain Regions

InSAR coherence directly influences the retrieval of time-series deformation and the detection of potential hazards. This study first conducted a comparative analysis and seasonal stratification of regional InSAR coherence based on filtered coherence coefficient maps generated from consecutive 12-day Sentinel-1 interferograms, which were used to characterize seasonal coherence patterns across the study area (Figure 4).

Using the coherence characteristics of the study area (Figure 4), regions with coherence > 0.75 were classified as high-coherence zones, while those with coherence < 0.35 were designated as low-coherence zones. The high-coherence threshold was derived as follows. Using the 2018 winter coherence map, we delineated the valley-floor area (riverbanks and adjacent low-elevation terrain) based on optical satellite imagery, and computed the mean coherence within this zone (~0.72). Similarly, the 2018 summer coherence map was used to delineate the high-altitude region (above 4000 m a.s.l.) where snow-free bedrock and sparse vegetation maintain good coherence; the mean coherence in that zone was ~0.78. A single representative high-coherence threshold of 0.75 was then obtained by averaging the two seasonal means, rounding slightly for operational simplicity. The low-coherence threshold of 0.35 was defined as the average coherence of the remaining area after excluding both the perennial high-coherence zone and the seasonal high-coherence zones identified above. While these values are derived from the 2018 dataset, the temporal consistency of the seasonal coherence pattern (see Figure 4) justifies their application to the full 2018–2021 period. All subsequent analyses use 0.75 and 0.35 as the high- and low-coherence boundaries, respectively, including the mask applied in the deformation maps.

Analysis of Figure 4 reveals that the half-moon-shaped region on the western side of the study area exhibits a wide-valley geomorphology. Time-series InSAR coherence analysis indicates that this region is only affected by snow and ice cover from February to March each year, resulting in reduced coherence for some interferometric pairs. However, over 93% of the pairs maintain high coherence throughout the year, leading to its classification as a perennial high-coherence zone (Figure 5). As the river flows along the Yarlung Tsangpo River into the “thumb-shaped” meander around Namcha Barwa Peak [20], InSAR coherence begins to exhibit seasonal variations. From January to April, the deep valleys within the eastern Himalayan tectonic syntaxis maintain high coherence, while high-altitude regions show poor coherence (Figure 4). Starting around May, snow and ice melt in high-altitude areas significantly improves coherence, whereas dense vegetation in low-altitude valleys leads to a notable decline in coherence. This results in a spatial migration of high-coherence zones from low-altitude to high-altitude regions.

Based on the above analysis and integrated with regional topography and geomorphology (Figure 5b,c), the high-coherence zones in this region can be delineated as shown in Figure 5a. These include: (1) the perennial high-coherence zone, which maintains relatively high InSAR coherence year-round; (2) the winter high-coherence zone, primarily corresponding to low-altitude areas along river valleys; and (3) the summer high-coherence zone, associated with high-altitude regions.

Further analysis of regional meteorological data (Figure 6) reveals that the average temperature in the eastern Himalayan tectonic syntaxis from January to April is below 5 °C, with high-altitude regions (above 4500 m) experiencing even lower averages of −4 °C or less. During this period, precipitation primarily occurs as snow and ice, and historical optical remote sensing imagery confirms that high-altitude areas are predominantly snow-covered from January to April [20,28], leading to unstable radar scattering properties and poor InSAR coherence. In contrast, low-altitude valleys exhibit relatively sparse vegetation during winter (Figure 7), enabling better coherence maintenance (Figure 8) and facilitating the identification of landslides and collapses along riverbanks.

Starting in May, regional temperatures rise above 10 °C, causing snow and ice melt in high-altitude areas, which transition to bare rock and soil (Figure 7). This allows effective InSAR coherence maintenance (Figure 8), enhancing the detection of high-altitude concealed hazards. However, in low-altitude valleys, the combination of snowmelt recharge, increased precipitation, and rising temperatures leads to a rapid increase in vegetation density (Figure 7), significantly reducing InSAR coherence. This phenomenon persists until October, marking the onset of the next coherence transition phase.

Notably, during the transitional period between winter and summer, coherence fluctuations occur due to rapid and irregular changes in surface features (Figure 4). Interferometric pairs with excessively low coherence (mean coherence < 0.35 in both the winter and summer high-coherence zones) are excluded. The remaining transition-season pairs are quantitatively assigned to the winter or summer dataset as follows: for each pair, the mean coherence within the winter high-coherence zone ($\overline{\gamma }$win) and within the summer high-coherence zone ($\overline{\gamma }$sum), as delineated in Figure 5a, is computed. If $\overline{\gamma }$win ≥ $\overline{\gamma }$sum, the pair is assigned to the winter dataset; otherwise, it is assigned to the summer dataset.

The above analysis demonstrates that high-altitude regions within the eastern Himalayan tectonic syntaxis exhibit better InSAR coherence during summer, facilitating the identification of high-altitude concealed hazards. Conversely, valley regions maintain superior InSAR coherence during the winter, enabling more accurate interpretation of landslides and collapses along riverbanks.

4.2. Seasonal Partitioned Stacking-InSAR Processing in the Eastern Himalayan Tectonic Syntaxis

Based on the InSAR coherence characteristics and the two-stage baseline strategy described in Section 3.1, this study first selects the default 60-day temporal baseline interferometric pairs for both ascending and descending tracks, which constitute the primary dataset with generally high coherence. The complementary long-temporal-baseline pairs (60–120 days) are considered separately in Section 4.3. Using the seasonal Stacking-InSAR method, deformation calculations were performed under winter (Win), summer (Sum), and full-interferometric (Total) conditions (Figure 9 and Figure 10). In the figures, colors represent the magnitude of displacement, with blue indicating movement away from the satellite and red indicating movement toward the satellite. Masked areas correspond to topographic shadows and layover regions caused by radar side-looking geometry, as well as low-coherence zones (average coherence < 0.35).

Analysis of the ascending track results revealed that the short-temporal-baseline Win-Stacking method detected an annual deformation rate range of −130 to 25 mm/year, identifying 55 potential hazards (including 18 high-altitude concealed unstable slopes and 37 riverside landslides), while the short-temporal-baseline Sum-Stacking method showed a deformation rate range of −95 to 20 mm/year, detecting 78 hazards (comprising 43 high-altitude concealed unstable slopes and 35 riverside landslides), and the short-temporal-baseline Total-Stacking method yielded a deformation rate range of −90 to 23 mm/year, identifying 81 hazards (including 37 high-altitude concealed unstable slopes and 44 riverside landslides) (Figure 9).

Similarly, for the descending track, the short-temporal-baseline Win-Stacking method exhibited an annual deformation rate range of −98 to 27 mm/year, identifying 77 hazards (including 36 high-altitude concealed unstable slopes and 41 riverside landslides), the short-temporal-baseline Sum-Stacking method showed a deformation rate range of −125 to 30 mm/year, detecting 100 hazards (comprising 66 high-altitude concealed unstable slopes and 34 riverside landslides), and the short-temporal-baseline Total-Stacking method revealed a deformation rate range of −85 to 23 mm/year, identifying 103 hazards (including 63 high-altitude concealed unstable slopes and 40 riverside landslides) (Figure 10).

4.3. Stacking-InSAR Optimization with Long-Temporal-Baseline Interferometric Pairs

Due to the high density of geohazards in the study area, slow-creeping deformations in developing unstable slopes may exhibit subtle displacement signals over short temporal baselines, potentially leading to the omission of small-magnitude movements when relying solely on short-temporal-baseline interferometric pairs. To further enhance the detectability of slow-creeping slopes, this study incorporated long-temporal-baseline interferometric pairs (60–120 days) on top of the short-temporal-baseline dataset. It should be noted that the primary benefit of including long-temporal-baseline pairs is improving the signal clarity and boundary delineation of already-detected hazards, rather than substantially increasing the total number of identified hazards. Only long-temporal-baseline pairs with sufficient coherence were selected for the seasonal Stacking-InSAR processing.

The selection threshold for long-temporal-baseline pairs was determined quantitatively as follows. Using the 2018 long-temporal-baseline interferograms (60–120 days), we computed the mean coherence within the winter high-coherence zone (river valleys) for winter-season pairs, and within the summer high-coherence zone (high-altitude bare rock) for summer-season pairs. The average of these mean values across all long-temporal-baseline pairs was 0.64. Accordingly, a long-temporal-baseline interferogram was accepted for Stacking-InSAR processing only when its average coherence inside the corresponding seasonal high-coherence zone exceeded 0.64. This criterion ensures that the incorporated long-temporal-baseline pairs maintain acceptable signal quality while providing the extended temporal sensitivity needed to capture slow-creeping deformation signals.

This approach yielded Win-, Sum-, and Total-Stacking-InSAR deformation datasets incorporating long-temporal-baseline pairs (Figure 11 and Figure 12).

Figure 11 illustrates the ascending track deformation results after incorporating long-temporal-baseline interferometric pairs with sufficient coherence. The Win-Stacking results revealed an annual deformation rate range of −123 to 25 mm/year, identifying 59 potential hazards, including 19 high-altitude concealed unstable slopes and 40 riverside landslides. The Sum-Stacking results showed a deformation rate range of −70 to 21 mm/year, detecting 82 hazards, comprising 45 high-altitude concealed unstable slopes and 37 riverside landslides. The Total-Stacking results exhibited a deformation rate range of −67 to 29 mm/year, identifying 90 hazards, including 42 high-altitude concealed unstable slopes and 48 riverside landslides.

Similarly, the descending track results demonstrated that the Win-Stacking method yielded an annual deformation rate range of −75 to 25 mm/year, identifying 80 potential hazards, including 38 high-altitude concealed unstable slopes and 42 riverside landslides. The Sum-Stacking results exhibited a deformation rate range of −135 to 29 mm/year, detecting 102 hazards, comprising 67 high-altitude concealed unstable slopes and 35 riverside landslides. The Total-Stacking results revealed a deformation rate range of −80 to 31 mm/year, identifying 112 hazards, including 67 high-altitude concealed unstable slopes and 45 riverside landslides (Figure 12).

5. Results

In this study, a deformation feature is identified as a geohazard candidate when it satisfies the following jointly applied criteria: (1) it manifests as a contiguous cluster of pixels exhibiting coherent deformation in at least one seasonal stack; (2) the deformation anomaly is spatially associated with a recognizable geomorphological feature such as a steep slope, headscarp, lateral boundary, or debris accumulation zone; (3) the deformation pattern shows spatial consistency across neighboring pixels and a plausible kinematic signature (e.g., increasing deformation magnitude toward the slope toe or scarp crest); and (4) the anomaly is cross-verified against multi-temporal optical satellite imagery and/or existing documented geohazard inventories wherever available. Identification was performed by experienced interpreters through systematic visual inspection of the seasonal deformation maps, guided by the above criteria.

Although this visual interpretation approach introduces a degree of subjectivity, a single universal deformation rate threshold was not adopted for the following reasons. The seasonal-partitioning method inherently enhances deformation signals that are spatially consistent and context-dependent. The same deformation rate may represent a hazard in one seasonal zone (e.g., a slow-moving high-altitude slope in summer) but be masked by seasonal decorrelation in another. Defining a universal threshold would require extensive ground-truth data across different altitudes and seasons, which are currently unavailable for this remote and complex terrain. Moreover, the primary goal of this work is to demonstrate the complementary advantage of seasonal partitioning in improving detection sensitivity, rather than to establish a fully automated hazard delineation system. The coherence-based stratification reduces false positives by restricting analysis to zones of reliable signal quality, and the spatial consistency of deformation contours provides reliable qualitative indicators.

5.1. Seasonal Partitioned InSAR Results Based on Short Temporal Baselines

Based on the hazard identification criteria established above, this study first examines the deformation characteristics of typical landslide hazards using the seasonal Stacking-InSAR method with a 60-day short temporal baseline.

Figure 13 and Figure 14 present local magnification and comparison views for identifying typical geohazards (masked areas represent InSAR layover shadows and low-coherence zones). By integrating the deformation information from the aforementioned interferometric pairs, the following observations are made for the 60-day short-temporal-baseline Stacking-InSAR results:

(1)

Large-magnitude riverside landslides are effectively identified in the Win-Stacking results, with high consistency between deformation contours and slope geometries, indicating high reliability in hazard detection. However, these hazards exhibit weak deformation signals in the Sum-Stacking results, hindering accurate identification (Figure 13). Conversely, high-altitude hazards are more clearly detected in the Sum-Stacking results, whereas they show only subtle deformation signals in the Win-Stacking results.

(2)

Reservoir bank landslides with low deformation rates are well identified in the Win-Stacking results, but decorrelation effects make them difficult to detect in the Sum-Stacking results. Meanwhile, high-altitude hazards with low deformation rates are more easily identified in the Sum-Stacking results, while no significant deformation signals are observed in the Win-Stacking results (Figure 14).

(3)

Some high-altitude concealed hazards, due to their maintained coherence during winter, can also be identified in the Win-Stacking results. Similarly, certain reservoir bank landslides are detectable in the summer InSAR results. Overall, the winter and summer Stacking results exhibit complementary advantages in hazard identification.

(4)

Due to seasonal coherence variations, some hazards with relatively small deformation magnitudes, including high-altitude hazards and riverside landslides, show weak signals in the conventional Total-Stacking results, making them challenging to identify.

5.2. Seasonal Partitioned InSAR Results Based on Long Temporal Baselines

Applying the same identification criteria to enhance the detection of slow-creeping deformations in the study area, this study proposes an optimized baseline combination incorporating long-temporal-baseline interferometric pairs with temporal baselines of 60 to 120 days for InSAR deformation analysis. Figure 15 illustrates the comparative identification results of typical slow-deformation hazards after integrating these long-temporal-baseline pairs, demonstrating improved detection capabilities for subtle and long-term deformation signals.

Analysis of the long-temporal-baseline results reveals that the Stacking-InSAR deformation and hazard identification outcomes incorporating long-temporal-baseline interferometric pairs with sufficient coherence exhibit the following characteristics:

(1)

The hazard identification patterns align with those of the 60-day short-temporal-baseline results, confirming that riverside landslides are more easily identified in the Win-Stacking results, while high-altitude concealed hazards are more detectable in the Sum-Stacking results. The increased number of interferometric pairs contributes to noise suppression and a smoother overall deformation pattern.

(2)

For the seasonal Stacking-InSAR processing, the addition of long-temporal-baseline pairs enhances the detection of hazards with low deformation rates, such as small-magnitude riverside landslides and high-altitude concealed hazards. The deformation signals become more easily identifiable, and the boundary contours of these hazards are clearer, improving the accuracy of hazard detection. As expected, the total number of identified hazards remained largely unchanged, confirming that the main contribution of long-temporal-baseline pairs is improving signal quality and boundary delineation rather than expanding the detection inventory.

(3)

Compared to the short-temporal-baseline Total-Stacking results, slow-creeping deformations are more easily identified in the long-temporal-baseline Total-Stacking results. Statistical data in Tables in Section 6 further demonstrate that incorporating long-temporal-baseline pairs significantly improves the hazard identification rate of the conventional Total-Stacking method.

(4)

It is worth noting that a small number of deformations with weak signals exhibit reduced quality after the inclusion of long-temporal-baseline pairs, primarily due to decorrelation effects and atmospheric noise.

5.3. Reliability Analysis of Seasonal Partitioned InSAR Results

This study employs the seasonal Stacking-InSAR method for early identification of geohazards in the eastern Himalayan tectonic syntaxis, revealing that over 50% of the hazards are concentrated along the Parlung Tsangpo River and the Bomi-Motuo Highway. This distribution pattern aligns closely with field survey data from previous studies [29,30,31], demonstrating the reliability and accuracy of the proposed method.

To further validate the reliability of landslide identification, this study analyzed multi-temporal historical optical remote sensing images of the study area. The results reveal that 60% of the identified hazards exhibit recognizable slope deformation features or potential movement trends in the time-series optical imagery (Figure 16). For instance, the typical riverbank landslide hazard in Figure 16 shows increasingly prominent collapse at the slope front, along with significant development and gradual expansion of tensile cracks at the slope rear, indicating a state of continuous movement characteristic of slow-moving landslides. Similarly, the typical high-altitude slope hazard displays progressive surface spalling at the rear rock mass and ongoing accumulation at the front gully, suggesting a slow-moving high-altitude slope with the potential to block rivers.

Additionally, no obvious movement traces were observed in the time-series optical imagery for the remaining 40% of hazards, which are typically in the early stages of slow slope deformation. These hazards are primarily classified as potential slow-moving landslides or high-altitude concealed slope hazards with small deformation magnitudes detected by InSAR [15,32]. Overall, the hazards identified in this study show strong consistency with existing field investigations and historical optical imagery, validating the reliability of the proposed method for hazard identification.

To quantitatively assess the reliability of the identified geohazards, the detection results were compared with an existing inventory of documented landslides provided by the Guiyang Engineering Corporation, Power China. This inventory contains all geohazards recorded within the study area before 2020 that had been formally registered in the local geological hazard database. Of the total geohazards identified by the seasonal Stacking-InSAR method, 57.2% were successfully cross-validated against this inventory, indicating a high consistency between the detections and previously known hazards. This validation ratio is close to the 60% ratio previously estimated using optical remote sensing imagery (Figure 16), confirming the reliability of the qualitative visual interpretation approach.

Among the remaining 42.8% of hazards that were not matched to the inventory, approximately 25% are located along the Parlung Tsangpo River valley. This region is characterized by extremely limited road access and sparse ground survey coverage; consequently, few entries exist in the historical inventory for this corridor. The remaining ~18% of unmatched detections likely correspond to geohazards that have developed after the inventory cutoff year (2020), i.e., new deformation signals that were not yet documented at the time of database compilation.

These results demonstrate that the majority of the identifications are consistent with existing ground-recorded hazards, and that the unverified detections are plausibly explained by either survey gaps in remote terrain or post-2020 hazard initiation.

6. Discussion

This study first statistically analyzed the overall hazard identification results of Stacking-InSAR using both short-temporal-baseline and long-temporal-baseline scenarios. The analysis aimed to quantify the impact of seasonal interferometric pair combinations on landslide hazard identification in the eastern Himalayan tectonic syntaxis, as well as to evaluate the optimization effects of different baseline combinations. Based on the statistical data, the additional hazards identified compared to the conventional Total-Stacking-InSAR results were evaluated, and the proportions of different hazard types across various baseline combinations were examined. This analysis provides insights into the applicability and reliability of the seasonal-partition InSAR deformation extraction method in the high-altitude alpine regions of the eastern Himalayan tectonic syntaxis.

Here, the winter–summer hazard repeat identification rate and the new hazard identification rate are defined by the following equations. The specific calculation results are presented in

Table 2. Statistics of geohazard identification using short-temporal-baseline Stacking-InSAR.

Geohazard Identification in Different Seasons	Number of Identified Geohazards (Ascending)	Number of Identified Geohazards (Descending)
Win-Stacking Identifications	55	77
Sum-Stacking Identifications	78	100
Total-Stacking Identifications	81	103
Total Identifications in Win and Sum	115	149
Duplicate Identifications in Win and Sum	18	28
Additional Identifications Compared to Total	36	46

,

Table 3. Proportion of geohazards identified by short-temporal-baseline Stacking-InSAR.

Stacking-InSAR Result Types	Proportion of Riverside Landslides Identified by Win-Stacking (%)	Proportion of High-Position Hidden Slope Hazards Identified by Sum-Stacking (%)
Short-Temporal-Baseline Stacking—Ascending	67.3	55.1
Short-Temporal-Baseline Stacking—Descending	53.2	66.0

,

Table 4. Newly identified geohazards in short-temporal-baseline Stacking-InSAR.

Stacking-InSAR Result Types	Winter & Summer Geohazard Repeat Identification Rate (%)	New Geohazard Identification Rate Compared to Total-Stacking (%)
Short-Temporal-Baseline Stacking—Ascending	15.6	29.6
Short-Temporal-Baseline Stacking—Descending	18.7	30.8

,

Table 5. Statistics of geohazard identification using long-temporal-baseline Stacking-InSAR.

Geohazard Identification in Different Seasons	Number of Identified Geohazards (Ascending)	Number of Identified Geohazards (Descending)
Win-Stacking Identifications	59	80
Sum-Stacking Identifications	82	102
Total-Stacking Identifications	90	112
Total Identifications in Win and Sum	125	152
Duplicate Identifications in Win and Sum	16	30
Additional Identifications Compared to Total	37	40

,

Table 6. Proportion of geohazards identified by long-temporal-baseline Stacking-InSAR.

Stacking-InSAR Result Types	Proportion of Riverside Landslides Identified by Win-Stacking (%)	Proportion of High-Position Hidden Slope Hazards Identified by Sum-Stacking (%)
Long-Temporal-Baseline Stacking—Ascending	67.8	54.8
Long-Temporal-Baseline Stacking—Descending	52.5	65.7

,

Table 7. Newly identified geohazards in long-baseline Stacking-InSAR.

Stacking-InSAR Result Types	Winter & Summer Geohazard Repeat Identification Rate (%)	New Geohazard Identification Rate Compared to Total-Stacking (%)
Long-Temporal-Baseline Stacking—Ascending	12.8	28.0
Long-Temporal-Baseline Stacking—Descending	19.7	26.3

and

Table 8. Comparative analysis of geohazard identification with short and long temporal baselines.

Stacking-InSAR Result Types	Geohazard Repeat Identification Rate Between Short and Long Temporal Baselines (%)	New Geohazard Identification Rate Achieved by Long Baseline over Short Temporal Baseline (%)
Win-Stacking—Ascending	86.9	9.8
Sum-Stacking—Ascending	84.3	10.1
Total-Stacking—Ascending	85.7	10.9
Win-Stacking—Descending	86.9	8.3
Sum-Stacking—Descending	83.6	9.1
Total-Stacking—Descending	86.9	10.4

.

$$Winter\&Summergeohazardrepeatidentificationrate={\displaystyle \frac{thenumberofhazardsidentifiedinbothwinterandsummer}{thetotalnumberofhazardsidentified}}$$

(2)

$$Newgeohazardidentificationrate={\displaystyle \frac{thetotalnumberofhazardsidentified-thenumberofhazardsidentifiedinTotal-Stacking}{thetotalnumberofhazardsidentified}}$$

(3)

6.1. Statistical Analysis of Short-Temporal-Baseline Seasonal Partitioned InSAR

Table 2 shows that the total number of hazards identified in the descending track exceeds that in the ascending track. However, the statistical data in Table 3 and Table 4 reveal consistent patterns between the ascending and descending track results. Further analysis of Table 3 indicates that riverside landslides exhibit a higher identification rate in the winter Stacking-InSAR results for both tracks, while the summer Stacking-InSAR results are more effective for identifying high-altitude unstable slope hazards. According to Table 4, the repeat identification rates of hazards in the winter and summer Stacking results account for 15.6% and 18.7% of the total identified hazards, respectively, indicating a relatively low overlap. Moreover, compared to the conventional Total-Stacking-InSAR results, the short-temporal-baseline seasonal Stacking-InSAR method achieves a new geohazard identification rate exceeding 29%. This clearly demonstrates that seasonal Stacking-InSAR processing is more effective for precise hazard detection in the high-altitude alpine regions of the eastern Himalayan tectonic syntaxis.

6.2. Statistical Analysis of Long-Temporal-Baseline Seasonal Partitioned InSAR

Table 5 presents the hazard identification statistics for the Stacking-InSAR method incorporating long-temporal-baseline interferometric pairs. The number of identified geohazards is comparable to the results obtained with a short temporal baseline. Furthermore, the ascending and descending track results based on long-temporal-baseline Stacking exhibit similar patterns to the short-temporal-baseline results: riverside landslides show a higher identification rate in winter, while high-altitude slope hazards are more frequently identified in summer (Table 6). The geohazard repeat identification rate remains below 20%, and the new hazard identification rate compared to the conventional Total-Stacking-InSAR results exceeds 26% (Table 7). These findings further validate the effectiveness and reliability of the proposed hazard identification method for early detection of geohazards in the eastern Himalayan tectonic syntaxis.

Table 8 compares the geohazard identification results of the seasonal-partition Stacking-InSAR method using short temporal baseline- versus long temporal baseline-enhanced combinations. The comparison reveals that the repeat identification rate between the different temporal baseline scenarios exceeds 83%, while the new hazard identification rate for the long temporal baseline-enhanced results ranges between 8% and 10% compared to the short-temporal-baseline results. Notably, the new identification rate is more pronounced in the Total-Stacking results, indicating that the overall hazard identification performance is largely consistent between the two approaches. Moreover, the inclusion of the long-temporal-baseline result significantly enhances the hazard identification rate of the conventional Total-Stacking-InSAR method.

To provide a single representative metric for the improvement achieved by the seasonal-partitioning method, we calculated the average of the four “newly identified geohazard rates compared to Total-Stacking” reported in Table 4 and Table 7 (i.e., ascending/descending × short temporal baseline/long temporal baseline). This yields an overall average improvement of 28.7%, which we round to approximately 29% in the Abstract and Conclusions for readability. The individual values range from 26.3% to 30.8%, and the detailed breakdowns are preserved in the respective tables.

6.3. Statistical Analysis of the Detected Geohazards Inventory Based on Frequency–Area Distribution

To evaluate whether the geohazard polygons identified by the seasonal Stacking-InSAR method constitute a statistically complete sample for landslide susceptibility modeling, we performed a frequency–area distribution (FAD) analysis following the approach of Malamud [33]. The areas of the 277 detected geohazards were calculated in a projected UTM coordinate system. The probability density of landslide areas was computed using logarithmically spaced bins and plotted against area on log-log axes (Figure 17).

The FAD plot exhibits a clear power-law decay for landslides larger than approximately 2.6 × 10⁴ m², with a fitted exponent β ≈ 1.99 and a coefficient of determination R² = 0.92. This indicates that the medium-to-large landslides in the inventory follow the well-known scale-invariant behavior typical of landslide populations. A distinct rollover occurs at ~2.6 × 10⁴ m², below which the probability density drops steeply, signifying that smaller landslides are substantially under-sampled in this dataset.

The relatively large rollover area compared to inventories of rainfall-triggered landslides (typically 10²–10³ m²) is physically consistent with the characteristics of this study. The detection limit is jointly controlled by the spatial resolution of Sentinel-1 (~100–225 m² per pixel) and the fact that the original field-based hazard survey primarily targeted large, high-consequence slope instabilities. Furthermore, the eastern Himalayan syntaxis is dominated by deep-seated, large-scale gravitational slope deformations, which naturally shift the size distribution toward larger areas.

The well-defined power-law tail above the rollover demonstrates that the inventory is statistically complete for landslides larger than approximately 2.6 × 10⁴ m². Therefore, this subset can be regarded as a representative sample of large landslides in the study area and is suitable for use as training or validation data in regional landslide susceptibility or hazard models. For applications that require the characterization of smaller landslides, higher-resolution SAR data (e.g., TerraSAR-X or airborne LiDAR) or targeted field mapping would be necessary to fill the gap below the detection limit.

Overall, the FAD analysis confirms that the geohazard inventory derived from the seasonal Stacking-InSAR method is of sufficient quality for large-landslide susceptibility mapping, despite the expected undersampling of small events in this high-altitude, densely vegetated terrain.

7. Conclusions

This study presents a seasonal-partitioning strategy for Stacking-InSAR, which separates interferograms into winter and summer subsets based on the seasonal migration of high-coherence zones in alpine terrain. Applied to Sentinel-1 ascending and descending tracks across the eastern Himalayan syntaxis, the method increased the number of identified geohazards by approximately 29% on average compared to conventional Total-Stacking (Table 4 and Table 7). The overlap between winter- and summer-detected hazards was below 19%, confirming that the two seasons reveal largely complementary deformation signals. The addition of long-temporal-baseline (60–120 day) interferometric pairs with adequate coherence further improved the clarity and boundary definition of slow-moving deformation features, although it did not substantially enlarge the overall detection inventory. Furthermore, the frequency–area distribution of the detected geohazards exhibits a well-defined power-law tail above ~2.6 × 10⁴ m², with a detection rollover controlled by the spatial resolution of Sentinel-1, demonstrating that the inventory is statistically complete for large landslides and suitable for use in regional susceptibility modeling.

It should be noted that Stacking-InSAR does not provide time-series deformation and is therefore best suited to detecting slopes undergoing steady, linear motion. The seasonal-partitioning approach is specifically designed for high-relief, cold mountain environments where InSAR coherence exhibits a pronounced seasonal cycle, such as the eastern Himalayan syntaxis; in regions with persistently good coherence, conventional Total-Stacking may remain adequate. Additionally, coherence preservation depends on the radar sensor: longer-wavelength systems are more effective in densely vegetated areas. Consequently, the optimal combination of regional topography, SAR data source, and processing strategy should be evaluated on a case-by-case basis to maximize the reliability of early geohazard detection. Future work should incorporate automated detection algorithms and quantitative validation against reference landslide inventories to further increase objectivity and reproducibility.