Landslide Mapping and Susceptibility Assessment in the Middle and Lower Reaches of the Nujiang River (2017–2025) Using Satellite Embedding and Multidimensional Environmental Factors

Highlights

What are the main findings?

• A satellite embedding-driven framework was developed for annual landslide inventory mapping in the middle and lower reaches of the Nujiang River from 2017 to 2025.
• Landslides showed strong spatial clustering, marked interannual variability, and persistent hotspot regions, revealing distinct spatiotemporal patterns of landslide activity.

What are the implications of the main findings?

• The proposed workflow improves the efficiency of annual landslide identification and spatiotemporal characterization in complex mountainous terrain.
• The findings provide scientific support for regional landslide monitoring, hotspot tracking, hazard zonation, and risk-informed management.

Abstract

Landslide mapping and susceptibility assessment are essential for hazard identification, infrastructure protection, and risk management. The middle and lower reaches of the Nujiang River have high relief, rapid geomorphic change, and fragile landscape conditions, which increase landslide susceptibility and hinder timely detection. To improve the spatiotemporal characterization of landslide activity, we developed a multi-source Earth observation framework for annual landslide mapping and susceptibility assessment. First, interannual embedding-change intensity maps were generated to guide the visual interpretation of landslide-related surface disturbances. Second, annual landslide and non-landslide samples were collected through field validation and visual interpretation. Third, annual 10 m landslide maps for 2017–2025 were generated using random forest on Google Earth Engine. Finally, 24 multidimensional environmental factors were incorporated into landslide susceptibility modeling. Landslides were concentrated mainly along the Nujiang River corridor and adjacent high-relief canyon slopes, with marked interannual variability but relatively stable hotspot regions. SHAP analysis further identified BSI_mean as the most important predictor, with a mean absolute SHAP value of 0.116, followed by NDVI_mean and terrain-related variables, indicating that bare-surface exposure, vegetation condition, and terrain dissection were strongly associated with mapped landslide occurrence. This study provides annual landslide inventories and susceptibility information for hazard mitigation and infrastructure planning.

1. Introduction

Landslides, defined as the downslope movement of rock, soil, or unconsolidated debris under the influence of gravity, pose a significant threat to human life, critical infrastructure, and mountain ecosystems [1,2,3,4]. Between 2004 and 2016, an estimated 55,997 fatalities were recorded in 4862 non-seismic landslide events worldwide [5]. Under global change, landslide hazards are likely to intensify due to climatic shifts and increasing anthropogenic disturbances, thereby raising the frequency of landslides in susceptible terrain [3,6,7]. Therefore, the timely and accurate identification of existing landslides and landslide-susceptible areas is essential for hazard assessment, risk mitigation, and disaster prevention in mountainous regions.

Remote sensing has become a primary tool for landslide investigation, enabling repeated, large-scale observations across remote and inaccessible mountainous regions. Optical imagery is effective for identifying surficial deposits, vegetation disturbance, and spectral anomalies, although its utility often declines under persistent cloud cover, terrain-induced shadows, and pronounced seasonal surface variability [4,8]. Synthetic aperture radar (SAR) enables all-weather, day-and-night Earth observation and is particularly valuable in regions where optical data are frequently unavailable [4]. In mountainous terrain, however, its effective use is complicated by shadow, geometric distortion, and reduced interpretability under dense vegetation [9,10]. Spaceborne interferometric synthetic aperture radar (InSAR) provides direct measurements of ground deformation and is highly effective for detecting slow-moving or progressively destabilizing slopes [11,12,13,14]. Its capability is nevertheless reduced where temporal decorrelation is severe, vegetation cover is dense, or slope failures occur too rapidly to be captured effectively [10,15,16]. These limitations indicate that no single sensor can simultaneously meet the demands of timely detection, wide-area coverage, and temporally continuous monitoring for landslide investigation. In this context, satellite embedding datasets offer a promising alternative by integrating multi-source Earth observation data into a compact feature space, thereby enabling more effective characterization of complex land-surface disturbances than single-sensor observations alone [17]. For cloud-prone mountainous regions, annual embedding representations can reduce reliance on individual cloud-free optical observations and provide a more stable description of land-surface conditions under frequent cloud contamination. However, since these datasets have only recently emerged alongside large-scale geospatial foundation models, their application to landslide detection remains limited.

Landslide investigation methods can generally be classified into heuristic, statistical, physically based, and machine-learning approaches [18,19,20,21,22]. Among these approaches, machine learning has become one of the dominant strategies as its ability to flexibly integrate multi-source predictors and capture nonlinear relationships between landslide occurrence and environmental conditions [23,24,25]. The performance of supervised models, however, depends strongly on the quality and representativeness of the training data. In practice, existing landslide inventories are often affected by delayed updates, positional uncertainty, and the inclusion of relict or revegetated landslides [26,27]. These issues can introduce label noise and weaken the correspondence between current surface conditions and mapped landslide classes. Such deficiencies are particularly problematic for regional-scale annual landslide mapping, which requires training samples that are both temporally explicit and spatially precise.

The middle and lower reaches of the Nujiang River represent an important yet still insufficiently studied landslide-prone region in southwestern China. The study area is characterized by pronounced topographic relief and deeply incised high-mountain canyon terrain, both of which favor frequent slope instability [28]. Moreover, the main stem of the Nujiang River has remained undeveloped for large-scale hydropower projects owing to substantial geological risk and the difficulty of balancing development with environmental constraints [29,30]. Also, the region is affected by widespread anthropogenic disturbances, including small hydropower facilities, road networks, and local engineering activities along river valleys and transportation corridors, which may modify slope geometry and locally reduce slope stability. Under these conditions, timely and spatially comprehensive landslide investigations are essential for hazard mitigation, community resilience, and the sustainable planning of future large-scale infrastructure. Despite this practical significance, systematic landslide investigations in this region remain limited.

To address the limitations of existing remote sensing data and training samples, this study developed a satellite embedding-assisted workflow for annual landslide mapping and susceptibility assessment in the middle and lower reaches of the Nujiang River. By linking embedding-change analysis with year-specific HR-image interpretation, the workflow enables temporally matched sample construction and reduces label noise associated with outdated or mixed-temporal inventories. This design provides a consistent basis for annual 10 m landslide mapping, susceptibility assessment, and spatiotemporal analysis. Specifically, the objectives of this study were to: (1) generate annual 10 m landslide maps for 2017–2025 using satellite embedding data, field surveys, high-resolution (HR) optical imagery, and Google Earth Engine (GEE); (2) develop a landslide susceptibility model using multidimensional environmental factors and interpretable machine-learning methods; (3) identify the dominant controls on landslide occurrence and quantify their relationships with environmental gradients; and (4) characterize the spatiotemporal dynamics of detectable landslide disturbance.

2. Materials and Methods

2.1. Study Area

The study area is shown in Figure 1. Yunnan Province is located in southwestern China (Figure 1a), and the study area is situated in the western part of the province (Figure 1b). It encompasses ten county-level administrative units: Gongshan Derung and Nu Autonomous County, Fugong County, Lushui City, Yunlong County, Longyang District, Shidian County, Longling County, Mangshi City, Yongde County, and Zhenkang County, covering approximately 32,800 km² in total. The Nujiang River flows from north to south through the study area. With an elevation range exceeding 3000 m, the region forms a deeply incised alpine canyon landscape characterized by steep slopes and deep valleys (Figure 1c). These topographic conditions, together with frequent geological hazards and a fragile ecological environment, make the region highly susceptible to landslides and other slope-related disasters. Owing to both natural and socio-economic constraints, no large-scale hydropower stations have yet been constructed along the main stem of the Nujiang River, and hydropower development remains limited [29,31].

2.2. Data and Preprocessing

The multi-source datasets used in this study are summarized in

Table 1. Multi-source data used in this study.

Data Category	Data Source	Period	Factor Item	Description
Satellite Embedding	Google Satellite Embedding	2017–2025	Embedding product	Landslide identification
Topographic-Geological Factors	SRTM DEM	2000	DEM	Digital elevation model
			Slope	Terrain slope
			Aspect	Terrain aspect
			TotalCurvature	Terrain curvature
			TerrainRelief	Local elevation range
			RoughnessStddev	Terrain roughness
	Geological map	1999	GeologyReclass	Geological/lithological units
Climatic-Hydrological Factors	MERIT Hydro	Released in 2019	FlowAcc_UPA	Upslope contributing area
			FlowAcc_UPG	Upstream gradient-related flow accumulation
			HAND	Height above nearest drainage
			SPI	Stream power index
			TWI	Topographic wetness index
			RiverDensity	River density
			DistRiver	Distance to nearest river
	CHIRPS Daily	2000–2025	MeanAnnualPrecip	Mean annual precipitation
			MeanWetSeasonPrecip	Mean wet-season precipitation
			MeanAnnualMaxDaily	Mean annual maximum daily precipitation
			MeanExtremeFreq	Mean frequency of extreme precipitation events
Land-cover factors	WorldCover 2021	2021	WorldCover	Land cover class
	Sentinel-2 Surface Reflectance	2019–2025	NDVI_mean	Mean NDVI
			NDVI_amplitude	NDVI amplitude
			BSI_mean	Mean BSI
Human-activity factors	OpenStreetMap	Last accessed 15 Mar 2026	DistRoad	Distance to nearest road
	WorldCover 2021	2021	DistSettlement	Distance to settlements
Field surveys and HR imagery	Field surveys, GF-1/2/6, ZY-1/3	2022–2025	Landslide/non-landslide samples	Training and validation

. Google Satellite Embedding data were used primarily for annual landslide identification, while the Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM), MERIT Hydro, WorldCover 2021 data, climatic data, and anthropogenic indicators were used for spatiotemporal analysis and landslide susceptibility modeling. Based on these datasets, 24 multidimensional environmental factors were derived and grouped into four categories: topographic-geological, climatic-hydrological, land-cover, and human-activity factors. In addition, landslide samples for 2022–2025 were established through manual interpretation of field surveys and HR optical satellite imagery, and were subsequently used for model training and accuracy assessment.

2.2.1. Satellite Embedding Dataset

This study employed the annual Google Satellite Embedding dataset available in GEE. Produced by the large-scale Google AlphaEarth Foundation model, the dataset provides global geospatial embeddings at a spatial resolution of 10 m. Each pixel is represented by a 64-dimensional vector that summarizes annual surface conditions and temporal dynamics derived from multiple Earth observation datasets [17]. As the embeddings are generated as annual representations from multi-source observations, they are less dependent on individual cloud-free optical scenes than conventional image-based interpretation, which is particularly useful in monsoon-affected mountainous areas. Compared with conventional spectral bands, these learned embedding features can capture complex surface changes more effectively and are therefore well suited to classification and change-detection tasks [32]. Annual embedding images from 2017 to 2025 were used to derive interannual surface-change intensity layers, which served as auxiliary information for visual interpretation, sample selection, and landslide mapping.

2.2.2. Topographic-Geological Factors

Topographic and geological factors were used to characterize terrain morphology, surface roughness, and lithological background conditions associated with landslide occurrence. The topographic variables included elevation (Figure 1c), slope, aspect, total curvature, terrain relief, and terrain roughness, expressed as the local standard deviation of elevation. All of these variables were derived from the 30 m SRTM DEM, which provides near-global elevation data at a spatial resolution of 1 arc-second [33].

A reclassified geological map (GeologyReclass) was also included to represent the lithological setting of the study area. Specifically, scanned images of the original 1:50,000 regional geological maps were digitized and then reclassified into 12 geological types according to series, systems, formations, and lithological characteristics (Figure 1d). To ensure spatial consistency in subsequent modeling, all topographic and geological factors were resampled to a uniform spatial resolution of 30 m.

2.2.3. Climatic-Hydrological Factors

Climatic and hydrological factors were used to characterize the external triggering conditions and hydrological environment associated with landslide occurrence. The climatic variables included mean annual precipitation (MeanAnnualPrecip, Figure 1e), mean wet-season precipitation (MeanWetSeasonPrecip), mean annual maximum daily precipitation (MeanAnnualMaxDaily), and the mean frequency of extreme precipitation events (MeanExtremeFreq). In this study, the wet season was defined as May to October, and an extreme precipitation event was defined as a day with precipitation exceeding 50 mm. These variables were derived from the CHIRPS daily precipitation dataset for 2000–2025 and were used to represent long-term precipitation conditions, seasonal rainfall concentration, rainfall intensity, and extreme rainfall activity in the study area [34,35].

The hydrological variables included upslope contributing area (FlowAcc_UPA), upstream gradient-related flow accumulation (FlowAcc_UPG), height above nearest drainage (HAND) [36], stream power index (SPI) [37], topographic wetness index (TWI) [38], river density (RiverDensity), and distance to river (DistRiver). These variables were derived mainly from the MERIT Hydro dataset and were used to describe runoff accumulation, local moisture conditions, drainage influence, and potential fluvial incision effects [39]. Among them, UPA and UPG characterize the potential concentration of surface runoff; HAND reflects the relative vertical position of a location with respect to the nearest drainage channel; SPI represents the erosive power of concentrated flow; and TWI indicates the tendency for water accumulation and surface wetness. River density and distance to river further capture the influence of drainage-network development and river proximity on slope instability.

2.2.4. Land-Cover Factors

Land-cover factors were incorporated to characterize surface cover conditions, vegetation condition, and surface exposure. The variables considered in this study included land-cover type (WorldCover), mean normalized difference vegetation index (NDVI_mean) [40], NDVI amplitude (NDVI_amplitude), and mean bare soil index (BSI_mean) [41]. WorldCover was derived from the ESA WorldCover 2021 product and used to represent the spatial distribution of major land-cover types across the study area [42]. NDVI_mean, NDVI_amplitude, and BSI_mean were derived from Sentinel-2 surface reflectance data for 2019–2025.

2.2.5. Human-Activity Factors

Human-activity factors were included to characterize the influence of anthropogenic disturbances on landslide occurrence. Two variables were considered: distance to roads (DistRoad) and distance to settlements (DistSettlement). DistRoad was derived from OpenStreetMap road data, whereas DistSettlement was derived from the built-up class of the ESA WorldCover 2021 product. Both variables were converted into distance rasters to represent the proximity of each pixel to major human-activity features.

2.2.6. Field Surveys and HR Imagery

Sample and validation data were derived from field investigations and HR optical satellite imagery. Field surveys were conducted from 2022 to 2025 mainly along national and provincial roads across the counties in the study area, with particular attention to slopes adjacent to transportation corridors, river valleys, settlements, and engineering-disturbed areas. Field points were recorded near the central part of landslide bodies rather than along landslide margins, and the coordinates were stored to five decimal places, corresponding to an approximate coordinate precision of about 1 m. In total, 259, 240, 256, and 300 field observation sites were investigated in 2022, 2023, 2024, and 2025, among which 45, 55, 62, and 70 complete landslides were verified in the field.

To complement the field data, cloud-free HR satellite images acquired during 2022–2025 were obtained from the China Centre for Resources Satellite Data and Application and provided coverage of the entire study area. The imagery included Gaofen-1/2/6 (GF-1/2/6) and Ziyuan-1/3 (ZY-1/3) data and was pan-sharpened to a spatial resolution of 2 m. Cloud-free HR images acquired from January to June were used for each corresponding year, and field surveys were conducted in July and August. When multiple HR images were available for the same area, images with lower cloud cover, clearer terrain visibility, and acquisition dates closer to the field surveys were preferentially used. The field survey data and HR imagery served as the primary reference for sample interpretation and accuracy assessment. The detailed procedure used to construct landslide and non-landslide samples is described in Section 2.3.2. Detailed information on the HR satellite imagery used in this study is provided in

Table 2. High-resolution satellite imagery used in this study.

Year	Satellite	Number of Images	Panchromatic Resolution (m)	Multispectral Resolution (m)
2022	GF-1	15	2	8
	GF-6	22	2	8
	ZY-1	11	2.5	10
	ZY-3	4	2	5
2023	GF-1	17	2	8
	GF-6	22	2	8
	ZY-1	19	2.5	10
2024	GF-1	10	2	8
	GF-6	8	2	8
	ZY-1	15	2.5	10
2025	GF-1	11	2	8
	GF-2	5	1	4
	GF-6	3	2	8
	ZY-1	17	2.5	10

Note: GF = Gaofen; ZY = Ziyuan.

.

2.3. Methodology

The methodological framework of this study is illustrated in Figure 2. The framework was designed around four functional components: embedding-change analysis for interpretation guidance, year-specific sample construction, annual landslide mapping, susceptibility assessment, and spatiotemporal analysis. First, the satellite embedding dataset was used to generate interannual surface-change intensity layers, thereby providing auxiliary spatial guidance for subsequent manual interpretation. Second, annual landslide and non-landslide sample sets for 2022–2025 were constructed by integrating the embedding-change intensity maps, HR optical imagery, and field survey data. Third, the random forest (RF) model was trained using samples collected for 2022–2024 and then applied to generate annual landslide maps for 2017–2025. Fourth, the multi-year landslide maps were used to characterize the spatial distribution of mapped landslide surfaces, and 24 multidimensional environmental variables were incorporated into landslide susceptibility modeling to produce the susceptibility map. Finally, the resulting landslide maps and susceptibility results were evaluated and analyzed.

2.3.1. Embedding-Change Analysis for Visual Interpretation Guidance

Interannual surface-change intensity was measured in the satellite embedding feature space to provide auxiliary guidance for visual interpretation. Landslides typically cause abrupt and localized disturbances in surface morphology, material exposure, and vegetation condition. As a result, pixels affected by landslide activity are expected to exhibit greater interannual deviations in embedding features than pixels in stable areas [17,43,44]. Accordingly, embedding-based change analysis was used to highlight anomalous zones of land-surface disturbance, thereby narrowing the search space for manual interpretation in complex mountainous terrain.

As shown in Figure 3a, samples from stable areas in 2023 and 2024 largely overlap in the UMAP-projected embedding space [45], indicating that their embedding features remain relatively stable between adjacent years. By contrast, samples from landslide-affected areas show a clear displacement between the pre-event state in 2023 and the post-event state in 2024, suggesting that landslide occurrence produces a marked shift in embedding features. This contrast indicates that annual embedding features are sensitive to landslide-related surface disturbances.

On this basis, pixel-wise comparisons were performed between annual embedding products for adjacent years from 2017 to 2025. Specifically, annual embedding vectors were compared for the year pairs 2017–2018, 2018–2019, …, and 2024–2025 using cosine similarity computed from the 64-dimensional embedding vectors. This metric was adopted because the embedding is designed as an integrated feature representation in which all dimensions jointly encode annual surface characteristics. For a given pixel i, let ${\mathrm{E}}_{i,t-1}=[e_1,e_2,\dots ,e_{64}]$ and ${\mathrm{E}}_{i,t}=[{e^{\prime }}_1,{e^{\prime }}_2,\dots ,{e^{\prime }}_{64}]$ denote the embedding vectors in two adjacent years. The cosine similarity is calculated as shown in Equation (1):

$$\mathrm{Cos}{\mathrm{Sim}}_{i,t}={\displaystyle \frac{{\mathrm{E}}_{i,t-1}\bullet {\mathrm{E}}_{i,t}}{‖{\mathrm{E}}_{i,t-1}‖‖{\mathrm{E}}_{i,t}‖}}$$

(1)

The corresponding embedding-change intensity was defined as ΔE:

$$\Delta E_{i,t}=1-\mathrm{Cos}{\mathrm{Sim}}_{i,t}$$

(2)

The statistical significance of ΔE is illustrated in Figure 3b. Compared with stable areas, landslide-affected areas exhibited significantly larger cosine distances between 2023 and 2024, confirming that annual embedding changes measured by cosine distance can effectively capture landslide-related surface disturbances. Repeating this procedure for all adjacent-year pairs produced a series of annual embedding-change maps, which were subsequently used as auxiliary evidence to identify areas potentially affected by landslide activity.

The continuous ΔE map was not converted into a binary landslide mask by thresholding. Instead, it was used as an auxiliary embedding-change layer to guide visual inspection of HR imagery. High ΔE values may also reflect non-landslide surface disturbances; therefore, landslide attribution relied on geomorphic evidence from HR imagery and field observations rather than on embedding change alone. Conversely, low ΔE values did not automatically indicate non-landslide conditions, and subtle landslides were interpreted when recognizable geomorphic evidence was present in HR imagery or field records. Nevertheless, low-contrast landslides that lack detectable surface expression in both ΔE layers and HR imagery may remain unidentified in the current workflow.

2.3.2. Construction of Training Samples from Satellite Imagery and Field Surveys

Training samples were constructed from field investigations and manually interpreted HR optical imagery, as existing landslide inventory datasets are not fully suitable for annual landslide mapping at a spatial resolution of 10 m. In some cases, available inventories contain positional inaccuracies, are not updated in a timely manner, and may include ancient or early-stage landslides that have already stabilized or revegetated [26,27]. The use of such inventories for supervised training may therefore introduce substantial label noise and lead to considerable mapping errors.

To address these limitations, a landslide sample dataset for 2022–2025 was specifically constructed using HR imagery and field data (Figure 4a). During sample interpretation, the continuous embedding-change intensity layers were used together with HR imagery and field survey records to guide visual inspection, rather than as a threshold-based screening mask. Final sample labels were determined primarily from HR imagery and field evidence, and areas with low ΔE values were not automatically excluded. Landslide samples were manually identified primarily from HR imagery and field evidence at locations showing recognizable landslide features, including arcuate or irregular scar boundaries, exposed bare material, disrupted vegetation, downslope movement traces, and runout or depositional patterns. Non-landslide samples were selected from surrounding terrain units without visual evidence of slope failure. Surface changes unrelated to landslides, such as road construction, agricultural activity, and exposed non-landslide surfaces, were excluded based on morphology, slope position, and field evidence. Non-landslide samples were interpreted separately for each year using HR imagery from the corresponding year and were not reused as fixed labels across years. Areas with ambiguous geomorphic evidence, including suspected old deposits, revegetated scars, road-cutting zones, and construction sites, were excluded from the non-landslide sample set. The final sample set was designed to represent contemporary landslide conditions rather than information inherited from outdated or mixed-temporal inventories.

A hexagonal grid-based sampling strategy was adopted to reduce spatial autocorrelation and improve the spatial representativeness of the sample set [46,47]. The entire study area was divided into hexagonal cells of 10 km², and random sampling was performed within each cell. For the non-landslide class, five sample points were randomly selected in each cell, whereas for the landslide class, two to five sample points were selected according to landslide scale (Figure 4b). This strategy allowed the samples to cover a broader range of terrain, land-cover, and disturbance conditions while reducing excessive clustering within a limited number of landslide hotspots. In total, 3368 landslide samples and 65,325 non-landslide samples were obtained, as shown in Figure 4c and

Table 3. Collected samples in this study.

Year	Landslide	Non-Landslide	Total
2022	812	16,402	17,214
2023	769	16,274	17,043
2024	864	16,344	17,208
2025	923	16,305	17,228
2022–2025	3368	65,325	68,693

.

2.3.3. Mapping Landslides Using GEE and RF Algorithm

Annual landslide mapping was implemented in GEE, which provides an integrated environment for multi-source data access and cloud-based computation, thereby enabling multi-year classification at the regional scale [48,49]. In this study, GEE was used to process candidate landslide areas, manage interpreted sample data, and train a RF classifier within a unified workflow. The mapped class represents remote-sensing-identifiable landslide surfaces in each annual image record, including newly exposed scars, active surfaces, and still-visible surfaces of earlier or reactivated landslides. The annual maps depict detectable landslide surfaces and do not constitute strict annual new-event inventories. Also, the mapped class was not further subdivided into specific landslide types, as type-specific classification was beyond the scope of this regional annual mapping study.

The annual sample sets for 2022–2024 were used for training, whereas the 2025 sample set was reserved for temporally independent but not spatially blocked validation. Applying the trained classifier to 2017–2021 involved temporal transfer within the same study area and annual embedding feature space. This transfer was supported by the common image characteristics of detectable landslide surfaces, including exposed material, vegetation disturbance, and scar morphology patterns. In the classification workflow, the annual satellite embedding image for each target year served as the predictor dataset, and the interpreted samples were used to extract the corresponding embedding features for RF training. The RF classifier was configured with 200 trees to balance classification stability and computational efficiency in multi-year GEE mapping. The tree-number sensitivity test indicated that both the F1-score and OOB score stabilized after approximately 170 trees, supporting the use of 200 trees as a stable setting. Other parameters were kept at their default values to maintain a consistent model structure across years. The trained classifier was then applied to the annual embedding images to classify pixels within the study area as landslide or non-landslide. The classifier output was a binary map in which landslide pixels were assigned a value of 1 and non-landslide pixels a value of 0.

2.3.4. Landslide Susceptibility Modeling Using Multidimensional Factors

Landslide susceptibility modeling was conducted using the multi-year landslide mapping results and the 24 multidimensional environmental factors described in Section 2.2. To construct the susceptibility training dataset, mapped landslide patches from the 2017–2025 annual landslide maps were used as the source of positive samples, whereas areas outside the mapped landslide patches were used as the source of negative samples. To reduce noise, fragmented mapped patches smaller than 1000 m² were removed before sample extraction. Positive samples were then selected from the central portions of the retained landslide patches rather than from patch boundaries, where mixed pixels and classification uncertainty are more likely to occur. In total, 5000 landslide samples and 10,000 non-landslide samples were selected, and the dataset was randomly divided into training and validation subsets at a ratio of 70% to 30%.

An RF classifier was adopted for susceptibility modeling, and its hyperparameters were optimized through grid search combined with five-fold cross-validation [24,50]. According to the optimization results, the RF model was configured with the following parameters: number of trees = 400, maximum depth = 20, minimum samples split = 2, and minimum samples per leaf = 1.

Correlations among the 24 candidate factors were examined to evaluate potential redundancy among the input variables. As shown in Figure 5a, several groups of variables exhibited relatively strong pairwise correlations, particularly among hydrological and precipitation-related variables, indicating that the direct inclusion of all factors could introduce redundancy. Therefore, recursive feature elimination with cross-validation (RFECV) [51] was applied, and the resulting optimal feature subset was used for final susceptibility modeling. According to the RFECV results (Figure 5b), three variables—FlowAcc_UPA, FlowAcc_UPG, and SPI—were removed, whereas the remaining factors were retained for model construction.

2.3.5. Performance Evaluation Metrics

The performance of annual landslide mapping and landslide susceptibility modeling was evaluated using confusion-matrix-based metrics. Specifically, producer’s accuracy (PA), user’s accuracy (UA), overall accuracy (OA), the kappa coefficient, and mean Intersection over Union (mIoU) were used to assess the annual landslide maps, whereas the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) were used to evaluate the discrimination ability of the susceptibility model. The formulas are given below:

$$\mathrm{PA}={\displaystyle \frac{TP}{TP+F{N}^{\prime }}}$$

(3)

$$\mathrm{UA}={\displaystyle \frac{TP}{TP+F{P}^{\prime }}}$$

(4)

$$\mathrm{OA}={\displaystyle \frac{TP+TN}{TP+TN+FP+F{N}^{\prime }}}$$

(5)

$$\mathrm{F}1={\displaystyle \frac{2TP}{2TP+FP+F{N}^{\prime }}}$$

(6)

$$P_o={\displaystyle \frac{TP+TN}{N}}$$

(7)

$$P_e={\displaystyle \frac{\left(TP+FP\right)\left(TP+FN\right)+\left(FN+TN\right)\left(FP+TN\right)}{N^2}}$$

(8)

$$\mathrm{kappa}={\displaystyle \frac{P_o-P_e}{1-P_e}}$$

(9)

$$\mathrm{mIoU}={\displaystyle \frac{1}{2}}\left({\displaystyle \frac{TP}{TP+FP+FN}}+{\displaystyle \frac{TN}{TN+FP+FN}}\right)$$

(10)

where TP, TN, FP, and FN denote the numbers of true positives, true negatives, false positives, and false negatives, while N is the total number of samples.

3. Results

3.1. Accuracy Assessment

3.1.1. Accuracy of Landslide Mapping

The accuracy of the embedding-based landslide mapping result was first evaluated using the 2025 validation sample set, which was temporally independent but not spatially blocked, as summarized in

Table 4. Confusion matrix and accuracy metrics for the 2025 landslide mapping results.

Referenced	Predicted		Total	PA (%)
	Non-Landslide	Landslide
Non-landslide	16,300	5	16,305	99.97
Landslide	92	831	923	90.03
UA (%)	99.44	99.40
OA (%)	99.44
F1-score (%)	94.49
Kappa	0.9419

Note: PA = producer’s accuracy; UA = user’s accuracy; OA = overall accuracy.

. A total of 16,300 non-landslide samples and 831 landslide samples were correctly classified, with 5 false positives and 92 false negatives. The resulting PA, UA, OA, F1-score, and kappa coefficient for the landslide class were 90.03%, 99.40%, 99.44%, 94.49%, and 0.9419, indicating strong point-based classification performance. Despite sample imbalance, the high landslide PA, UA, and F1-score suggest that the embedding-based classifier was not strongly biased toward the majority non-landslide class under the current validation setting.

To further examine the effectiveness of the satellite embedding dataset, we also trained and evaluated a Sentinel-2-based RF model using the original 10 multispectral bands. The same 2025 independent field sample set was used for validation. As shown in

Table 5. Confusion matrix and accuracy metrics for the 2025 Sentinel-2-based landslide mapping results.

Referenced	Predicted		Total	PA (%)
	Non-Landslide	Landslide
Non-landslide	16,153	152	16,305	99.07
Landslide	452	471	923	51.03
UA (%)	97.28	75.60
OA (%)	96.49
F1-score (%)	60.93
Kappa	0.5917

, the Sentinel-2-based result achieved an OA of 96.49% and a UA of 75.60% for the landslide class. However, its landslide PA and F1-score were only 51.03% and 60.93%, indicating substantial omission of landslide samples.

Nevertheless, point-based validation primarily reflects sample-level classification accuracy and does not fully characterize boundary delineation or patch-level spatial agreement. To address this limitation, a representative object-level comparison was conducted using manually interpreted ground-truth image blocks for 2017–2025. For each year, 20 representative blocks were selected to cover different counties, terrain settings, landslide sizes, and image conditions. The blocks were not selected solely from visually obvious landslides; each block included manually interpreted landslide surfaces and surrounding non-landslide areas to evaluate both omission and commission errors. For 2022–2025, the reference interpretation was supported by HR imagery and field evidence. For 2017–2021, owing to the lack of HR imagery, the reference blocks were interpreted mainly from Sentinel-2 imagery and geomorphic evidence. Therefore, the Sentinel-2-based metrics for 2017–2021 are not fully independent and are reported as comparative reference values rather than strict independent accuracy estimates.

As shown in

Table 6. Representative object-level comparison between Sentinel-2-based and embedding-based landslide mapping results from 2017 to 2025.

Year	Prediction Data	PA (%)	UA (%)	OA (%)	F1-Score (%)	Kappa	mIoU (%)
2017	Sentinel-2	27.13	73.44	78.39	39.62	0.2971	50.72
	Embedding	86.73	77.17	98.09	81.67	0.8066	83.51
2018	Sentinel-2	15.82	100.00	95.95	27.32	0.2635	55.87
	Embedding	75.32	67.61	97.07	71.26	0.6972	76.16
2019	Sentinel-2	32.04	98.51	90.20	48.36	0.4445	60.81
	Embedding	55.62	94.26	93.16	69.96	0.6639	73.19
2020	Sentinel-2	46.89	100.00	96.78	63.84	0.6238	71.78
	Embedding	65.98	91.38	97.56	76.63	0.7537	79.78
2021	Sentinel-2	39.13	50.94	96.41	44.26	0.4244	62.39
	Embedding	78.26	45.76	95.82	57.75	0.5572	68.15
2022	Sentinel-2	46.48	99.77	93.17	63.42	0.6020	69.59
	Embedding	83.74	82.61	95.68	83.17	0.8069	83.18
2023	Sentinel-2	36.60	87.65	95.18	51.64	0.4955	64.93
	Embedding	61.68	72.82	95.69	66.79	0.6450	72.82
2024	Sentinel-2	42.44	97.33	96.72	59.11	0.5767	69.30
	Embedding	67.44	82.86	97.40	74.36	0.7301	78.24
2025	Sentinel-2	39.26	100.00	95.86	56.38	0.5464	67.50
	Embedding	68.70	98.15	97.78	80.83	0.7969	82.75

Note: mIoU = mean Intersection over Union.

, the embedding-based results consistently outperformed the Sentinel-2-based results in terms of F1-score, kappa coefficient, and mIoU across all years. The mean F1-score increased from 50.44% for Sentinel-2 to 73.60% for the embedding-based method, while the mean mIoU increased from 63.65% to 77.53%. These results indicate that satellite embedding features provide a more effective representation of landslide-related surface disturbances than raw Sentinel-2 multispectral bands, particularly for object-level landslide delineation. The relatively low UA and F1-score of the embedding-based result in 2021 indicate increased commission errors in that year, which may be related to fragmented reference patches, vegetation recovery, and confusion with other exposed or disturbed surfaces. The 2021 outlier also highlights the year-to-year variability of object-level mapping uncertainty. Despite the limited independence of the Sentinel-2 metrics for 2017–2021, the comparison provides useful evidence that satellite embedding features improve the delineation of remote-sensing-identifiable landslide surfaces, particularly for the HR-supported years of 2022–2025. This block-based assessment improves the evaluation of patch-level spatial agreement but remains a representative object-level comparison rather than a fully random validation design.

3.1.2. Performance of the Landslide Susceptibility Model

Figure 6 shows the ROC curve and precision–recall curve for test-set validation of the landslide susceptibility model. A total of 1500 landslide samples and 3000 non-landslide samples were used to evaluate the performance of the model. The model achieved an AUC of 0.9697, indicating strong discrimination between landslide and non-landslide samples within the inventory-derived susceptibility dataset. The average precision derived from the precision–recall curve reached 0.9426, suggesting that the model maintained high precision under class-imbalanced conditions. In addition, the model retained strong precision as recall increased, indicating robust performance in identifying landslide samples. These metrics reflect discrimination within the mapped-inventory-based susceptibility dataset, with mapping-related uncertainty discussed in Section 4.4.

3.2. Landslide Mapping Results for 2017 to 2025

Figure 7 shows the spatial distribution of mapped landslide surfaces at 10 m resolution in the study area. The mapped landslide surfaces exhibit pronounced spatial clustering rather than a uniform distribution across the basin. According to the longitudinal and latitudinal distribution curves, landslide area is concentrated mainly between 98.5° and 99.0°E and between 25.6° and 26.4°N. These high-density zones are concentrated mainly within the Nujiang Grand Canyon, particularly in Gongshan, Fugong, and Lushui. These spatial statistics were derived from the mapped landslide surfaces and are therefore reported as descriptive mapping-based estimates rather than error-adjusted area estimates.

The representative examples in Figure 7b–f further illustrate the mapping performance across different landslide surfaces. Visual comparison with Sentinel-2 imagery, 2 m HR imagery, and manually interpreted ground truth shows that the embedding-based results better preserve the spatial extent and morphological characteristics of landslide patches than the Sentinel-2-based results. In contrast, the Sentinel-2-based results show more fragmented patches and greater omission in several examples. These comparisons support the ability of the embedding-based results to delineate remote-sensing-identifiable landslide surfaces and highlight the added value of satellite embedding features for regional landslide mapping.

3.3. Landslide Susceptibility Results

Figure 8 shows the spatial distribution and county-level composition of landslide susceptibility in the middle and lower reaches of the Nujiang River. Landslide susceptibility was classified into five relative levels using fixed equal-interval RF susceptibility-score thresholds: very low (0–0.2), low (0.2–0.4), medium (0.4–0.6), high (0.6–0.8), and very high (0.8–1.0). These classes represent relative susceptibility levels rather than exact absolute probabilities. The equal-interval scheme was adopted to provide a transparent and reproducible classification rule and to avoid the distribution-dependent thresholds produced by natural breaks or quantile classification. At the regional scale, the very low susceptibility class occupies the largest proportion of the study area (87.8%), whereas the low, medium, high, and very high classes account for 5.0%, 2.5%, 2.2%, and 2.5%. This distribution indicates that the RF susceptibility scores are strongly skewed toward low values and that most pixels in the study area have relatively low modeled susceptibility. The extensive very low susceptibility zones are located mainly in sparsely inhabited mountainous areas. By contrast, although the high and very high susceptibility classes occupy relatively small areas, they are distributed primarily along the main stem and tributaries of the Nujiang River. These valley corridors are more closely associated with settlements, roads, and other human activities, implying greater potential exposure of populations and infrastructure to landslide hazards.

At the county scale, the largest areas of very high susceptibility are observed in Yunlong (216.18 km²), Lushui (165.40 km²), Longyang (146.16 km²), and Gongshan (108.49 km²), followed by Fugong (75.89 km²). By comparison, Mangshi, Zhenkang, and Longling show relatively small extents of very high susceptibility. Overall, the susceptibility pattern is characterized by the dominance of low-susceptibility background areas and the concentration of higher-susceptibility zones along river valleys, where potential landslide impacts are likely to be greater.

To further examine the spatial relationship between landslide susceptibility and human activity, Figure 9 shows the areal composition of susceptibility classes across successive distance bands from roads and settlements. In both cases, the proportions of the high and very high susceptibility classes are relatively low within 0–500 m, increase markedly over 500–1000 m, and then gradually decline with increasing distance. By contrast, the very low susceptibility class becomes progressively more dominant farther from roads and settlements. This pattern suggests that landslide susceptibility is closely related to human activity, while also indicating that the highest susceptibility does not occur immediately adjacent to roads and settlements. Instead, the most susceptible zones are concentrated in intermediate distance bands, whereas slopes closest to roads and settlements may be better protected or stabilized by engineering measures.

4. Discussion

4.1. Major Controlling Predictors of Landslide Occurrence

Identifying the major controlling predictors of landslide occurrence is essential in susceptibility analysis, as it helps determine whether the model’s predictive behavior is consistent with underlying environmental processes. A substantial body of evidence indicates that landslide occurrence is influenced by topography, hydrology, geology, land cover, climate, and anthropogenic disturbance [52,53,54,55]. Rigorous interpretation of these factors is important for linking statistical or machine-learning models with process-based understanding [23,25,56]. Recent studies have applied explainability methods such as Shapley Additive Explanations (SHAP) to identify the major factors associated with landslide occurrence, because these methods can quantify both the relative importance of predictive variables and the direction of their effects on model output [57,58].

The SHAP analysis indicates that landslide occurrence in the study area is mainly associated with surface-condition and terrain-related variables. As shown in Figure 10, BSI_mean and NDVI_mean had the largest mean absolute SHAP values, followed by RoughnessStddev, Aspect, and TerrainRelief, whereas variables such as RiverDensity and TotalCurvature contributed comparatively little. Higher BSI_mean values increased model output, while higher NDVI_mean values reduced it, indicating strong associations with bare-surface exposure and vegetation loss. However, BSI_mean and NDVI_mean may reflect both predisposing surface conditions and post-failure, including exposed material, vegetation disturbance, and subsequent recovery; thus, their SHAP contributions indicate predictive associations rather than purely causal controls. These associations represent the combined mapped landslide class rather than type-specific controls for individual failure mechanisms. Terrain roughness, relief, slope, and proximity to roads and rivers also increase susceptibility, suggesting the influence of terrain dissection and anthropogenic disturbance. Aspect showed a relatively high contribution, and mapped landslide pixels were concentrated mainly on south- to southeast-facing slopes, possibly reflecting aspect-related differences in soil moisture, vegetation condition, weathering intensity, and local valley configuration. By contrast, precipitation-related variables contributed weakly, likely because they were represented by long-term statistical indicators rather than event-scale triggering rainfall. The relatively coarse resolution of CHIRPS may further smooth localized orographic rainfall signals in deeply incised canyon terrain, thereby weakening the apparent contribution of rainfall-related variables.

The findings of this study are broadly consistent with previous literature, but they also reveal a distinctive feature of the present landslide inventory. The positive contribution of bare-surface indicators and the negative contribution of vegetation-related variables are consistent with previous findings [16,55]. Likewise, the importance of slope, relief, and proximity to roads and rivers is consistent with earlier studies showing that steep terrain, road cutting, and fluvial undercutting are recurrent controls on landslide occurrence [59,60]. However, unlike many conventional susceptibility studies in which rainfall, lithology, or slope emerges as the dominant predictor, the present study ranked BSI_mean and NDVI_mean highest. This difference is likely attributable to the spatiotemporal specificity of the training dataset, which was derived from 10 m resolution landslide mapping conducted for 2017–2025. As a result, the inventory consists predominantly of recent and geomorphologically active landslides, which increases the diagnostic sensitivity of surface-exposure and vegetation-loss indicators relative to more slowly varying background variables.

4.2. Spatial Distribution Characteristics of Landslides

Landslide occurrence shows clear spatial selectivity rather than random distribution. Previous studies have shown that its spatial distribution is typically controlled by multiple environmental factors, including topography, hydrology, vegetation, lithology, and anthropogenic disturbance [10,22,52]. Accordingly, analyzing the gradient-based response of landslides across these factors is important for identifying the environmental conditions under which slope failure is more likely to occur.

To examine the spatial patterns of landslides, gradient analyses were conducted for the 15 most influential factors identified by the SHAP analysis, as shown in Figure 11. Landslides were concentrated within relatively narrow ranges of several variables. Most landslide pixels occurred at elevations of 1000–3000 m, on slopes of 15–45°, and in terrain-relief classes of 200–600 m. In terms of aspect, south- to southeast-facing slopes accounted for about 60% of all landslide pixels. In addition, landslide pixels were strongly concentrated near rivers (0–800 m) and roads (0–1000 m). About 86% of landslide pixels occurred in grassland and tree-cover classes, with mean NDVI values of 0.2–0.6 and NDVI amplitude values of 0.2–0.4. Landslide pixels were also concentrated mainly in areas with mean annual precipitation of 800–1600 mm and river density values ranging from 0.04 to 0.12. Geological and curvature classes also showed uneven distributions, indicating that landslides were preferentially associated with specific lithological and geomorphic settings across the study area.

These results indicate that landslides in the study area are concentrated mainly on mid-elevation slopes characterized by moderate to steep gradients and strong surface dissection. This spatial pattern is consistent with the combined influence of fluvial incision, slope-toe erosion, road cutting, and drainage-related disturbance, all of which are well-established controls on slope stability in mountainous terrain [61,62]. In addition, the contrast among geological classes suggests that lithological background exerts an important control on landslide occurrence under steep and highly dissected topographic conditions.

4.3. Temporal Dynamics of Mapped Landslide-Affected Slope Units

Analysis at the slope-unit scale provides important insights into the persistence and spatiotemporal variability of mapped landslide-affected slope units in complex mountainous environments. Compared with pixel-based approaches, slope-unit-based analysis aligns more closely with geomorphological units and therefore provides a more physically meaningful representation of the spatial organization and temporal evolution of landslide processes [63,64]. Accordingly, the r.slopeunits tool was applied to analyze the spatiotemporal dynamics of landslides larger than 1000 m² in the middle and lower reaches of the Nujiang River Basin at the slope-unit scale [65]. As shown in Figure 12, a total of 103,736 slope units were delineated within the study area. This procedure aggregates pixel-level landslide mapping results to the object scale, thereby facilitating assessment of slope stability. The annual counts and ratios of landslide-affected slope units were calculated directly from the mapped results and therefore represent mapping-based descriptive indicators.

Figure 13 shows pronounced interannual variability in mapped landslide-affected slope units among the 10 counties during 2017–2025. In terms of absolute counts, Gongshan, Lushui, Yunlong, and Fugong consistently exhibited the highest numbers of landslide-affected slope units, whereas Mangshi, Zhenkang, and Yongde remained at comparatively low levels. Several counties showed distinct interannual peaks of mapped landslide, indicating non-monotonic temporal variation. Lushui and Fugong maintained the highest landslide ratios across most years, and Gongshan likewise showed persistently elevated ratios. By contrast, Longling, Shidian, Mangshi, Zhenkang, and Yongde consistently showed lower ratios throughout the study period. Yunlong displayed pronounced interannual variability, with peak ratios in 2017 and 2019 followed by a marked decline in subsequent years.

At the slope-unit scale, mapped landslides exhibited both spatial persistence and interannual variability. In high-activity counties, the spatial distribution of landslide-prone zones remained relatively stable over time, indicating that these zones are controlled primarily by enduring geomorphological and environmental factors. At the same time, pronounced interannual variability in the number of mapped landslide-affected slope units suggests that detectable landslide disturbance remained dynamic despite the persistence of spatial hotspots. Overall, the mapped landslide-affected slope units were concentrated predominantly in a limited number of high-relief canyon counties, although the mapping-based annual ratios in these counties still exhibited pronounced interannual variability.

4.4. Limitations and Future Improvements

Despite the high accuracy of the landslide mapping and susceptibility assessment results, several limitations remain. The annual samples were interpreted primarily from HR optical imagery and field evidence. Low-contrast landslides, such as slow-moving, partially vegetated, reactivated, or morphologically subtle failures, may remain under-detected when they lack recognizable surface expression in both ΔE layers and HR imagery. The current susceptibility model is also inherently static. The SRTM-derived topographic variables and WorldCover 2021 land-cover layer provide static representations of terrain and land-cover conditions and therefore cannot fully capture local terrain modification, land-cover change, or vegetation recovery during 2017–2025. This may introduce temporal inconsistency and uncertainty into susceptibility modeling. Moreover, landslide occurrence in the study area is influenced by dynamic and time-varying controls, such as rainfall anomalies, short-term surface disturbances, and progressive slope deformation, which cannot be adequately captured by multi-year mean environmental variables. Uncertainty may also propagate from landslide mapping to susceptibility zonation. The removal of patches smaller than 1000 m² and center-based positive-sample extraction reduced the influence of mapping noise but did not fully eliminate mapping-related uncertainty. Formal confidence intervals for annual mapped area, slope-unit counts, and hotspot persistence were not estimated, as such uncertainty quantification would require probability-based validation and error-adjusted area estimation.

Although InSAR provides valuable information on pre-failure and slow-moving slope deformation, it was not integrated into the present workflow as the primary objective of this study was annual inventory mapping of landslide-related surface disturbances. In the middle and lower reaches of the Nujiang River, steep terrain, dense vegetation, temporal decorrelation, and rapid slope failures pose additional challenges for large-scale InSAR application. Future work will benefit from integrating HR optical imagery, InSAR, field surveys, updated DEMs, and annual land-cover information to construct a more complete landslide inventory and better characterize dynamic terrain and surface changes. A dynamic landslide susceptibility framework is also needed to characterize both long-term environmental backgrounds and short-term triggering conditions [66,67]. Furthermore, modeling strategies that incorporate both uncertainty awareness and interpretability could be introduced to improve predictive reliability and strengthen process-based interpretation [68,69].

5. Conclusions

(1)

This study developed a satellite embedding-assisted workflow for annual landslide mapping and susceptibility assessment in the middle and lower reaches of the Nujiang River. Annual 10 m maps of remote-sensing-identifiable landslide surfaces for 2017–2025 and an inventory-derived susceptibility map were generated by integrating satellite embedding data, HR imagery, field evidence, GEE, and RF modeling. The results demonstrate the potential of satellite embeddings for regional landslide mapping in complex mountainous terrain.

(2)

The mapped landslide surfaces showed clear spatial clustering along the Nujiang River corridor and adjacent high-relief canyon slopes, especially in Gongshan, Fugong, Lushui, and Yunlong. At the slope-unit scale, mapped landslide-affected units exhibited pronounced interannual variability, while the main hotspot regions remained broadly stable.

(3)

The susceptibility analysis identified BSI_mean, NDVI_mean, RoughnessStddev, Aspect, and TerrainRelief as the dominant predictors for the combined mapped landslide class. These results indicate that bare-surface exposure, vegetation condition, terrain roughness, slope orientation, and relief are strongly associated with mapped landslide occurrence. However, as BSI_mean and NDVI_mean may partly reflect post-failure spectral responses, their SHAP contributions indicate predictive associations rather than purely causal controls.

(4)

The results are framed within the scope of remote-sensing-based annual landslide mapping. The annual maps characterize detectable landslide surfaces rather than strict new-event inventories, while the susceptibility map represents an inventory-derived background susceptibility result. Applications of the results therefore need to account for uncertainties related to post-failure spectral responses, static environmental predictors, and mapped-inventory-based labels. Further integration of deformation monitoring, updated terrain and land-cover data, and uncertainty-aware modeling would improve dynamic landslide monitoring and susceptibility assessment.