Refined Landslide Susceptibility Mapping Considering Land Use Changes and InSAR Deformation: A Case Study of Yulin City, Guangxi

Abstract

Landslide susceptibility maps (LSMs) are valuable tools typically used by local authorities for land use management and planning activities, supporting decision-makers in urban and infrastructure planning. To address this, we proposed a refined method for landslide susceptibility assessment, which comprehensively considered both static and dynamic factors. Neural network methods were used for susceptibility analysis. Land use and land cover (LULC) change and InSAR deformation were then integrated into the traditional susceptibility zoning to obtain a refined susceptibility map with higher accuracy. Validation was conducted on the improved landslide susceptibility map using site landslide data. The results showed that the LULC were proven to be the core driving factors for landslide occurrence in the study area. The GRU model achieved the highest model performance (AUC = 0.886). The introduction of InSAR surface deformation and land use and land cover change data could rationalize the inappropriateness of traditional landslide susceptibility zoning, correcting the false positive and false negative areas in the traditional landslide susceptibility map caused by human activities. Ultimately, 12.25% of the study area was in high-susceptibility zones, with 3.10% of false positive and 0.74% of false negative areas being corrected. The proposed method enabled refined analysis of landslide susceptibility over large areas, providing technical support and disaster prevention and mitigation references for geological hazard susceptibility assessment and land management planning.

1. Introduction

Landslides, a frequent global geological hazard, can lead to considerable casualties and economic losses, as well as environmental harm [1,2]. These occurrences can be caused or triggered by various factors such as precipitation [3,4], earthquakes [5], and Anthropogenic factors [6,7]. Between 1940 and 2020, China experienced 1470 fatal landslide events, which led to the loss of 14,394 lives [8]. As a consequence of global climate change and human activities, the incidence of landslides has risen, along with the associated loss of life and property [2,9]. Consequently, researchers aim to improve the accuracy of landslide susceptibility assessments [10], which estimate the spatial distribution and likelihood of landslides [11], to reduce their impact on the economy, public safety, and local ecosystems [12]. Landslide susceptibility mapping (LSM) can be achieved through various methods. Initially, researchers used physically based models to create LSM [13,14]. These methods are straightforward and suitable for simple landslide types or cases with detailed mechanical parameters, but they are restricted to small areas. Later, statistical modeling approaches were adopted for LSM, which helped in comprehending landslide patterns and triggering mechanisms [15]. Statistical models are represented by General Linear Models (GLMs) [16], the Information Value method (IV) [17], and the frequency ratio method (FR) [18]. Recently, many advanced Machine Learning (ML) techniques have been created to simulate the response relationship between landslides and their predisposing factors [19]. Typical models encompass random forest (RF) [20], binary Logistic Regression (LR) [21], Support Vector Machine (SVM) [22], and Artificial Neural Networks (ANNs) [23]. With the rapid development of Deep Learning techniques (DL), Convolutional Neural Networks (CNNs) [24], Recurrent Neural Networks (RNNs) [25], and ensemble method models [26] have been used for landslide susceptibility mapping. Compared to ML models, they show great potential in powerful data mining and nonlinear relationship modeling [27,28]. Nevertheless, the model accuracy varies in specific cases. Due to the variability of landslide-inducing and triggering factors and the unique environmental conditions of landslide-prone areas [29], no single model is suitable for landslide susceptibility assessment in all regions [30].

Landslides are induced by both natural and human factors [31]. Natural factors encompass elevation, slope, aspect, lithology, and rainfall as fundamental terrain and environmental conditions, while human factors primarily involve land use and land cover (LULC) changes and improper land management [32]. Many studies also categorize landslide conditioning factors into static and dynamic factors [33]. LULC changes, as dynamic factors and results of human activities, are significant contributors to landslide occurrences [34]. The influence of LULC changes on landslide susceptibility has been investigated in numerous studies [35,36,37,38,39]. Xiong et al. developed a hybrid model named HMLP-LAF to analyze the impact of LULC and its changes on landslide susceptibility in Enshi City [35]. Rabby et al. assessed the effects of LULC changes on landslide susceptibility in Rangamati City, Rangamati District, and Bangladesh under three LULC scenarios using an ANN model [39]. However, these studies often use LULC from different periods (past, present, and future) as input factors, focusing on the temporal trends in landslide susceptibility [40]. They lack quantitative analysis of the relationship between change in LULC and landslide hazards, insufficiently addressing the landslide hazards induced by LULC changes. Additionally, appropriate LULC changes and proper land management do not lead to landslide occurrences and developments. Inappropriate LULC changes (deforestation, agricultural activities, and slope-cut housing) significantly affect surface stability, consequently influencing landslide occurrences and developments [36,41]. Yulin City is in a period of rapid development, and a large number of human engineering activities (cutting slopes to build houses, building roads, mining, etc.) have caused a large number of high and steep slopes, which are the main cause of landslides and other geological disasters. The city has more than 20,000 cut slopes for building houses, and even more cut slopes for road construction, and many slopes lack protective measures, resulting in prominent safety hazards. Previous research has primarily focused on the static impacts of LULC changes on landslide susceptibility, with less emphasis on the surface deformation caused by LULC changes and how such deformation further affects landslide occurrences and developments [42,43]. Traditional single-model methods struggle to comprehensively explain the complex mechanisms by which LULC, as a dynamic factor, affects landslide occurrences and developments.

Over the past 30 years, Interferometric Synthetic Aperture Radar (InSAR) technology has been extensively applied to the monitoring of seismic deformation [44], landslide deformation [45], urban deformation [46], mining deformation [47], volcanic deformation [48], glacier movement [49], and many other fields [50], providing a powerful tool for large-scale surface deformation monitoring. Combining LULC data with advanced satellite-based radar measurements, specifically SBAS-InSAR, provides a powerful and comprehensive approach for monitoring landslide activity and enhancing early warning systems. LULC data reveal the potential impacts of surface cover types and human activities on landslides, while SBAS-InSAR data provide high-precision surface deformation information. By integrating these two types of data, landslide risk areas can be more accurately identified, improving the accuracy and timeliness of landslide predictions.

This paper aims to develop a refined and high-resolution landslide susceptibility assessment method using multi-source data, deep learning, and InSAR technology. Guangxi Yulin City was selected as the study area, with the landslide susceptibility study conducted in three stages: (1) Deep learning models were used to obtain landslide susceptibility maps based on static factors. (2) Dynamic surface deformation was calculated using InSAR technology and correlated with LULC changes to detect potential landslides. (3) The model results and landslide hazard results were integrated to achieve a more refined landslide susceptibility analysis. A refined approach to evaluating landslide susceptibility with dynamic timeliness and high resolution is provided by this study, which not only provides a foundation for effective disaster prevention and mitigation but also supports the formulation of informed land use planning.

2. Study Area

The research area is Yulin City in Guangxi Zhuang Autonomous Region, located in the southeastern part of Guangxi. It spans from 109°39′E to 110°18′E and from 22°19′N to 23°01′N, covering a total area of 12,800 km² (Figure 1). For simplicity, we refer to the study area as Yulin. Yulin has complex geological conditions and a fragile geological environment. Due to its unique natural geography and geological conditions, Yulin’s geological disasters, although not large in scale, are characterized by their diversity, wide distribution, poor stability, and frequent occurrence.

Yulin is surrounded by mountains, with higher elevations in the center that slope towards the north and south. The vegetation is dense, primarily consisting of evergreen broad-leaved forests and some coniferous forests. The geological structure of Yulin is complex, with the main structural line trending northeast and being well developed. The strata exposed in the study area include four major types: sedimentary rocks, volcanic rocks, metamorphic rocks, and intrusive igneous rocks. Among these, sedimentary rocks are the most widely distributed, followed by igneous rocks, accounting for 48.5% and 47.4% of the study area, respectively.

Yulin has a subtropical monsoon climate, marked by hot summers and chilly winters and springs. The average annual temperature is around 21 °C, with a maximum of 38 °C and a minimum of 2.3 °C. The average annual rainfall ranges between 1400 to 1600 mm, with a maximum daily rainfall of 274 mm. Rainfall is abundant but unevenly distributed, significantly influenced by the terrain. As a result, mountain areas receive heavy rainfall, while river valley basins receive less. At the same time, rainfall distribution is controlled by the ocean monsoon, with the rainy season from April to August, accounting for 78% of the annual precipitation. The main rivers in Yulin include the Nanliu River, Jiuzhou River, and the Beiliu River and Wusi River of the Xijiang River system in the Pearl River Basin. In recent years, human engineering activities in Yulin have increasingly impacted the geological environment, particularly with the widespread phenomenon of slope cutting for construction and road building. All unstable slopes in the city and 85.2% of landslides are linked to slope cutting construction activities.

3. Materials and Methods

The methodological framework of this study is shown in Figure 2, which mainly includes the steps of data collection, LULC change detection, SBAS-InSAR processing, model processing, model validation, and model integration. Finally, a refined method for landslide susceptibility assessment is presented, integrating static and dynamic factors to advance the accuracy and reliability of predictions.

3.1. Data Preparation

Landslide inventory: This study used address data of historical landslide occurrence points from 2014 to 2022 provided by the Guangxi Department of Natural Resources. Through verification with high-resolution Google Earth images and field surveys, an inventory of 84 historical landslides was compiled. Landslides in Yulin primarily occurred as shallow soil landslides on natural slopes (with slope cutting at the foot) and artificial slopes. Most landslides were small in scale, with 3 medium-sized landslides, 81 small-sized landslides, and no large-scale landslides recorded. These landslides were categorized into rock landslides and soil landslides. Currently, most landslides remain in an unstable state, with 68 (80.95%) landslides still undergoing deformation. The specific classification of landslides can be seen in

Table 1. Landslide inventory classification for Yulin City.

Classification Basis	Type	Description	Quantity	Percentage (%)
Cause	Natural Landslide	Landslides caused by natural geological processes	16	19.05
	Human-induced Landslide	Landslides triggered by human engineering activities	68	80.95
Material Composition	Rock Landslide	Sliding along a highly weathered structural plane on a slope facing an open space	10	11.9
	Soil Landslide	49 soil landslides occurred in granite residual soil and completely weathered soil layers. 25 soil landslides occurred in residual slope deposits of clastic rocks and metamorphic rocks	74	88.1
Scale	Small Landslide	<10 × 104 m3	81	96.43
	Medium Landslide	10 × 104 m3~100 × 104 m3	3	3.57
Thickness	Shallow Landslide	Thickness within 10 m	84	100
Current Stability	Stable	No signs of activity	16	19.05
	Relatively Stable	Slight signs of activity	45	53.57
	Unstable	Obvious signs of activity	23	27.38

. Simultaneously, 84 non-landslide points were randomly selected within the study area to construct a database for model training and validation. Historical landslide points were labeled as 1 and non-landslide points as 0. 70% of the samples were allocated as the training set, while the remaining 30% served as the validation set.

Based on a review of the existing literature and an analysis of historical landslide patterns within the study area, an eleven influencing factors evaluation system for landslide susceptibility was developed (Figure 3 and Figure 4). This system incorporated key topographic factors (elevation, slope, aspect), geological variables (lithology, distance to fault zones, rock absorption rate), hydrological influences (rainfall, distance to rivers), and environmental factors (distance to roads, LULC), as well as the Normalized Difference Vegetation Index (NDVI). To better investigate the preconditioning factors influencing landslide occurrence, the data were collected from a time period corresponding to the occurrence times of historical landslides (

Table 2. List of data for this study.

Data	Source	Data	Source
30 m SRTM DEM	NASA Earthdata Search	Lithology	China Geological Survey
Slope	30 m SRTM DEM	Faults	China Geological Survey
Aspect	30 m SRTM DEM	Roads	China Geological Survey
Rainfall	Han et al. [51]	Rivers	China Geological Survey
NDVI	Google Earth Engine	LULC	Google Earth Engine
Rock	China Geological Survey	Sentinel-1A	European Space Agency

).

Topographic factors include elevation, slope, and aspect. Elevation data with a 30 m resolution were obtained from the Shuttle Radar Topography Mission (SRTM). These data were then used to extract slope and aspect details within ArcGIS Pro. Slope angles within the study area range from 0° to 74°. Aspect was categorized into nine classes: flat, northeast (NE), east (E), southeast (SE), south (S), southwest (SW), west (W), northwest (NW), and north (N).

Geological factors included lithology, Rock Absorption (RA), and Distance to Faults (DF). Engineering geological and fault zone distribution maps were created at a scale of 1:500,000 based on geological maps provided by the China Geological Survey. Based on the engineering geological conditions of Yulin, ten rock groups were classified: dolomite intercalated with dolomitic limestone (I), limestone intercalated with dolomitic limestone and marl (II), conglomerate sandstone intercalated with mudstone (III), sandstone intercalated with mudstone (IV), mudstone intercalated with sandstone (V), mudstone (VI), siliceous rock intercalated with siliceous mudstone (VII), sandstone and soft shale (VIII), granite (IX), and clay (X). The fault zone increased the likelihood of soil and rock body sliding and the risk of landslides. Due to the influence of multiple tectonic movements, faults, folds, and anticline structures were highly developed.

Hydrological factors encompassed rainfall and Distance to Water (DW). Rainfall was the primary trigger for concentrated outbreaks of landslides. Rainfall data were collected based on daily precipitation observations from 2839 stations within and around China [51]. In Yulin, rainfall was concentrated in the summer, often leading to clusters of landslides during periods of heavy precipitation. Rainwater infiltration weakened the soil, increasing its weight, reducing its shear strength, and eroding slope surfaces, which decreased support and ultimately accelerated slope instability. Rapid river flow eroded riverbeds, forming narrow valleys. It weakened slope stability through lateral erosion and scouring, triggering landslides similarly to the cutting effects of human engineering activities.

Environmental factors included NDVI, LULC, and Distance to Roads (DR). Landsat 8 satellite images were used to calculate the NDVI distribution map, which reflected the vegetation growth status in the study area. Generally, lower NDVI values indicate lower vegetation coverage. LULC data were sourced from the global LULC map of ESA Sentinel-2 images with a resolution of 10 m. The main land use types in Yulin included water, flooded vegetation, built area, trees, crops, rangeland, and bare ground. Road data were sourced from the National Fundamental Geographic Information Center. Mountainous infrastructure construction required techniques such as slope cutting, which altered the initial geological characteristics and affected foundation stability.

InSAR data: Two sets of C-band Sentinel-1A ascending images from January 2018 to December 2022 were used, with a satellite incidence angle of 34°. All these factors were rasterized to a 30 m pixel size and projected to the UTM_WGS_1984_49N zone coordinate system. In addition, landslide influencing factors were normalized to account for variability and data magnitude, as neural network models can be highly sensitive to such factors.

3.2. Susceptibility Mapping

In this study, three Recurrent Neural Network models were used to model the probability of landslide occurrence: a Gated Recurrent Unit (GRU) [52], Long Short-Term Memory (LSTM) [53], and a Recurrent Neural Network (RNN) [54]. The model structures are shown in Figure 5. The RNN is the most basic recurrent neural network model, but its ability to handle long sequences is limited due to the issue of diminishing gradients. LSTM addresses the issue of diminishing gradients by introducing gating mechanisms, which can effectively capture semantic relationships in long sequences. The GRU, as a simplified version of LSTM, maintains similar performance while having a simpler structure and computation [55]. These models have been widely applied in the spatial modeling of landslide susceptibility [56]. Through training and evaluating these models, the best-performing model was identified. Multicollinearity among landslide influencing factors can significantly impact model performance; therefore, it was essential to perform a multicollinearity analysis between the selected landslide influencing factors before model construction. The Variance Inflation Factor (VIF) and Pearson’s correlation analysis were used to test the collinearity among the 11 landslide influencing factors. The Receiver Operating Characteristic (ROC) curve was used to quantitatively measure each model’s performance. The variable importance function in biomod2 [57,58] was used to model the Variable Importance (VI) of each landslide factor through the permutation of fitted and predicted values of the models. Final VI scores, representing the average importance across 100 permutation iterations, were derived for each factor. VI scores range from 0 to 1, with higher scores indicating a more pronounced influence on landslide susceptibility. The LSM calculated by the models were classified into five categories using the natural breaks classification method: very low, low, moderate, high, and very high, representing the susceptibility of the zones affected by landslides.

3.3. Potential Landslides Detection

3.3.1. LULC Change Detection

Changes in LULC constitute a significant conditioning factor within landslide susceptibility models. Deforestation, soil degradation, slope cutting, road building, and hillside building construction frequently decrease slope stability, increasing the precariousness of landslides [59]. This study combined the LULC2018 with the LULC2022 of Yulin to detect LULC changes, quantify the area of change, and investigate the influence of LULC changes on landslides. Only changes associated with increased susceptibility to landslides were considered in this study, categorized into seven types: Water to built-up area (WTB), rangeland to built-up area (RTB), Trees to built-up area (TTB), Trees to Bare ground (TTBG), Flooded Vegetation to built-up area (FTB), Crop to built-up area (CTB), and Bare land to built-up area (BTB). Built areas here included houses, towns, cities, densely populated villages, asphalt, and paved roads. Due to the distinct micro-geomorphological features of slopes such as free faces, gullies, cracks, and slip surfaces, only areas with slopes were considered. This excluded broad, flat plains and highly urbanized city centers to mitigate their influence on the identification results.

3.3.2. SBAS-InSAR

The surface deformation was determined using the Small Baseline Subset (SBAS)-InSAR method [60]. The SBAS-InSAR processing steps included connection graph generation, interferogram formation, phase unwrapping, orbital refinement, inversion, atmospheric correction, geocoding, and deformation projection. Temporal and spatial baseline thresholds were selected to be 36 days and 50 m, respectively, obtaining 548 interferometric pairs with a coherence threshold of 0.25. InSAR-derived deformation represents the actual surface deformation projected in the line of sight (LOS) direction. Direct application of LOS deformation is challenging. Landslides moving along a slope can result in LOS deformation that is either positive (approaching the satellite) or negative (moving from the satellite). Consequently, for practical purposes, it is necessary to convert LOS deformation into actual movement direction. Some techniques allow for the conversion to actual deformation through the combination of ascending and descending data. However, due to data acquisition constraints, we utilized single-track data, which constrained us to project deformation exclusively in either the slope or vertical direction [61,62]. Therefore, we chose a slope of 5° as the threshold. For areas with slopes less than 5°, Equation (1) was used to project the deformation direction vertically. For areas with slopes greater than 5°, deformation was projected in the slope direction using Equation (2).

$$V_{VERT}=V_{LOS}/\mathrm{c}\mathrm{o}\mathrm{s}\theta$$

(1)

where $V_{VERT}$ represents the vertical deformation and $V_{LOS}$ denotes the LOS deformation, while $\theta$ denotes the incident angle.

$$\begin{array}{c}{V}_{SLOPE}=V_{LOS}/C\\ C=n_{Los}.n_{slope}\end{array}$$

(2)

where $V_{SLOPE}$ indicates the slope deformation, $n_{Los}$ refers to the three-dimensional LOS vector, and $n_{slope}$ represents the three-dimensional slope vector.

Specifically, we used the following three conditions to identify potential landslides caused by land use and land cover changes:

(1)

LULC change: Using multi-temporal optical remote sensing images, we identified areas where LULC changes had occurred in the study area and created a buffer zone with a radius of 200 m. Rapid development of LULC changes could have led to geological instability, so these areas were given special attention.

(2)

Surface deformation: Based on the ground deformation maps obtained from SBAS-InSAR processing, we identified possible landslide areas. Areas with deformation were marked as potential landslide candidate areas.

(3)

Slope: Slope could help us identify the topographic features of landslides. Areas with slopes were given priority consideration.

Based on the above data, we first excluded areas where there were no land use changes, and then further screened the remaining areas based on the distribution of slopes and surface deformations to identify potential landslide risk areas. Considering the actual surface deformation in the study area, the classification was as follows: V1 (0–5 mm), V2 (5–10 mm), V3 (10–15 mm), V4 (15–20 mm), V5 (>20 mm).

3.4. Susceptibility Integration

To integrate the landslide susceptibility model results and the potential landslide dataset, a 30 m resolution grid was uniformly created. Each individual cell within the grid corresponded to the results of both the model and the identified deformation related to potential landslides. The final results relied on the integration of the susceptibility of each cell (30 × 30 m) with the LULC change related surface deformation rates based in that area (

Table 3. Correction matrix for LSM (1: very low, 2: low, 3: moderate, 4: high, 5: very high).

	Deformation (mm)
Susceptibility		0–5	5–10	10–15	15–20	>20
	1	1	2	3	4	5
	2	2	2	3	4	5
	3	3	3	3	4	5
	4	3	3	4	5	5
	5	3	4	5	5	5

).

4. Results

4.1. Landslide Susceptibility Map

4.1.1. Correlation Analysis

Correlation analysis (Figure 6) was performed on the 11 selected predisposing factors, calculating the Pearson correlation coefficient and VIF. All 11 factors registered VIF values below 5, with a maximum correlation coefficient of 0.45. This suggested no multicollinearity among the factors and a generally low degree of correlation.

4.1.2. Importance Analysis

According to the Variable Importance (VI) analysis (Figure 7), LULC were the most important factors influencing landslide occurrence in the Yulin. Distance to roads (DR) emerged as the second most significant factor. Rainfall and slope exhibited moderate VI values. In contrast, aspect and rock absorption rate (RA) were the least important. These findings suggest that human activities played a significant role in triggering landslides in the Yulin area.

4.1.3. Susceptibility Assessment

LSMs (Figure 8) were generated from model outputs using ArcGIS Pro 3.2 software. Susceptibility levels were classified using the natural breaks method into five categories: very low, low, moderate, high, and very high. Each model was applied to landslide susceptibility modeling with 100 different training and testing combinations, and the average AUC over the 100 iterations was summarized (Figure 9). The AUC values for all three models were greater than 0.88, and overall, all models were highly accurate. Specifically, the GRU model showed the highest AUC (0.886), followed by the LSTM model (0.885) and the RNN model (0.880). The models’ performance was validated using a landslide dataset that was not involved in the model construction. The GRU model outperformed the others in predicting landslide susceptibility, with 95% of landslides correctly identified within high-susceptibility zones. The LSTM and RNN models demonstrated relatively poorer performance compared to the GRU model, consistent with their AUC results.

4.2. LULC Change Analysis

Given the observed LULC changes from 2018 to 2022 (Figure 10), in areas of the study region with slopes greater than 5 degrees, the LULC change area reached 541.366 km², accounting for 4% of the study area. These regions contained 80% of the landslide hazard and risk points. Overall, from 2018 to 2022, built-up areas expanded by 540.522 km². The main types of LULC changes in Yulin were forest conversion to built-up areas and farmland conversion to built-up areas, with change areas of 316.737 km² and 173.095 km², respectively. This was mainly caused by human activities, primarily reflected in urban development, rural residential construction, and transportation and water conservancy construction. Improper slope cutting easily disrupted the natural balance of the slope, making landslides more likely to occur under the combined effects of heavy rainfall, vibrations, ground loading, and continuous weathering and erosion of the slope. Field surveys found that landslides frequently occurred on steep artificial slopes created by slope cutting for building construction. The continuous transformation of land use types towards more concentrated human activities might indirectly lead to geological disasters to some extent. This proved that dynamic changes in LULC had a major effect on the spatial distribution of landslides.

4.3. SBAS-InSAR Analysis

The vertical and slope deformation derived from SBAS-InSAR technology is presented in Figure 11a. Its effective coverage area accounted for 74% of the study area. During 2018–2022, deformation values ranged from −66 to 50 mm. Dense vegetation in the study area caused SAR image decorrelation, leading to the absence of deformation data in some regions. Figure 11b,d show the deformation distribution in two local areas (P1 and P2), respectively. P1 is a typical area of artificial slope cutting for building construction, and slope cutting caused deformation. Figure 11c illustrates the deformation of the triangle points in Figure 11b. Unreasonable excavation of the natural slope in this area has disrupted the slope balance formed during the geological history, causing slope deformation and instability. P2 is a typical area of artificial slope cutting due to road reconstruction. Figure 11e illustrates the deformation of the triangle points in Figure 11d. Human activities disrupted the slope integrity and the natural balance of the rock and soil mass, creating an unsupported face at the lower part of the slope. Stress concentrated at the excavation face, causing deformation under external factors such as rainfall. From May to September, Yulin experiences a rainy season with frequent heavy rain. Rainfall erodes slope cutting units and infiltrates rock and soil masses through cracks and pores, causing saturation and softening. This reduces the shear strength, cohesion, and internal friction of the rock and soil, increasing weight and leading to slope sliding.

The box plots in Figure 12 illustrate the correlation between surface deformation and both the transformation type and slope gradient within the LULC buffers. Surface deformation was more rapid on gentler slopes, specifically those within the 5° to 25° range, with deformation rates decreasing as slope gradients increased. In areas where forests were converted to built-up areas, subsidence trends were more pronounced and covered a larger overall extent. The conversion of forests to bare ground resulted in the most widespread deformation patterns, exhibiting the largest magnitudes of deformation.

4.4. LSM with Integrated InSAR, LULC Change, and GRU

Figure 13a shows the landslide susceptibility classification detected through surface deformation, slope, and LULC change. Using a correction matrix, this classification was integrated with the susceptibility results from the GRU model (Figure 13b) to generate the final LSM (Figure 13c). The very high-susceptibility zone accounted for 4.17%, high susceptibility 8.08%, moderate susceptibility 16.54%, low susceptibility 28.77%, and very low susceptibility 42.33% of the study area. Validation with the landslide dataset showed that all listed landslides fell into high-susceptibility zones (Figure 13d). Some high-susceptibility zones were distributed in the built-up areas of mountainous regions with steep slopes, but susceptibility in some areas was significantly overestimated. In Yulin, 3.10% of regions were identified by the GRU model as high-risk zones, but surface deformation results showed slight deformation, insufficient to classify them as landslide-prone areas. These false positives were well corrected through the correction matrix. Meanwhile, 0.74% of Yulin area shifted from moderate to high susceptibility through the correction matrix, attributed to regions where human activities caused slope instability detected by surface deformation. Landslide evolution is a dynamic process. Combining dynamic factors (human activities) with other static factors (e.g., topography and geological conditions) allows for more accurate landslide risk assessment and effective preventive measures.

Eighty percent of the existing landslide hazard points in Yulin City are related to human engineering activities, primarily consisting of small, shallow soil traction landslides, with most landslides being in an unstable state. We will briefly discuss the details of three typical cases, all occurring in areas where slope cutting was performed for building construction. Figure 14 shows the process from slope deformation results to preliminary susceptibility results, and finally to corrected susceptibility results for three different areas. In the first case, the GRU model predicted low susceptibility (Figure 14b). However, the slope deformation results (Figure 14a) clearly revealed slope deformation due to cut-and-fill operations. The correction matrix effectively captured this discrepancy, updating the susceptibility map to reflect the increased risk (Figure 14c). In the second case, the area also showed significant deformation due to slope cutting (Figure 14d), with substantial subsidence. The GRU model results indicated high susceptibility for this area, but some local areas did not reflect the impact of ground deformation (Figure 14e). The corrected results (Figure 14f) showed a significant increase in susceptibility in the ground deformation areas, more accurately reflecting the actual risk compared to the initial assessment. In the third case, although the area had slope cutting for building construction, most deformation was minor and remained stable (Figure 14g). However, the GRU model overestimated susceptibility in some locations (Figure 14h), resulting in false positives. Under these conditions, the LSM was updated through the correction matrix (Figure 14i), enhancing the reliability of the results. Misclassification can lead to social and economic losses.

5. Discussion

5.1. Mapping Unit Selection

In landslide susceptibility evaluations, the choice of evaluation units significantly impacts the precision and dependability of the results. The primary units commonly employed are grid units and slope units [63,64]. For areas with plains and mountains, grid and slope units are most effective, each offering unique benefits. Slope units offer a holistic view of the slope’s overall stability. However, they suffer from uncertainties due to their discontinuous nature and lower refinement, which can lead to overly generalized results, lacking in necessary detail and precision. Conversely, grid units emphasize local details affecting landslide mechanisms. By utilizing smaller unit sizes, grid units can more accurately reflect geological structures and topographic variations of the slope, effectively integrating dynamic factors from human engineering activities. This method produces more realistic and dependable assessment outcomes, enhancing landslide prediction and disaster warning capabilities. While grid units may not depict surface undulations as effectively as slope units, their focus on local detail makes them more suitable for regions where such factors critically influence landslide occurrences. Therefore, grid units are the preferred choice for evaluating landslide susceptibility in plains and mountainous areas.

5.2. The Response Relationship of Influencing Factors

Figure 15 summarizes the correlation between landslide susceptibility and a variety of influencing factors. Landslide occurrence is influenced by multiple factors. Statistical analysis of landslide distribution patterns in the study area is important for selecting influencing factors and understanding landslide occurrence. Despite the study area’s relatively low elevation, landslides are significantly more frequent at elevations above 800 m, with the incidence peaking around 1000 m. The influence of slope gradient on landslides is profound, with slopes ranging from 5 to 25 degrees being particularly susceptible to such occurrences. This propensity is attributable to the fact that natural slopes with gradients below 25 degrees typically possess thicker layers of weathered soil, encompassing both residual soil and fully weathered strata. Intensive human engineering activities are prevalent in these regions, with artificial slope cutting frequently observed at the front of natural slopes. This, combined with rainfall which enhances water infiltration and runoff, significantly increases the likelihood of landslides.

North-facing slopes are more prone to landslides because most of the mountainous areas in the northeast have north-facing slopes, which receive less sunlight. The distribution patterns of roads, faults, and rivers exhibit a notable correlation with landslide occurrences; proximity to these features significantly increases landslide susceptibility. Lithology curves reveal that the susceptibility to landslides varies significantly across different lithological types, demonstrating the diverse impact each lithology has on the stability of slopes. The overall trend indicates that landslide susceptibility is notably higher in regions characterized by fine sand, sandy clay, clayey soil, and loose gravel soil layers. These soil types are predominantly distributed within karst basins and river terraces. The thickness of the overlying soil layers varies significantly, and the lithological composition is heterogeneous. The structure of these soils is typically loose and highly permeable, which considerably increases the likelihood of soil landslides, especially in steep sections affected by artificial slope cutting. The rock absorption rate curve illustrates that rocks with higher absorption rates are more susceptible to water uptake and subsequent expansion. This process undermines their structural stability, thereby elevating the risk of landslides. Landslide susceptibility is highest when rainfall ranges from 50 to 100 mm. As rainfall further increases, landslide susceptibility gradually stabilizes. This indicates that moderate to heavy rainfall has the most significant impact on triggering landslides. With the increase in vegetation coverage, landslide susceptibility fluctuates. Landslides occur in regions with both sparse and dense vegetation coverage, suggesting that a high density of vegetation does not inherently prevent landslides. The response curve for LULC shows significant differences in the impact of various land use types on landslide susceptibility. Landslide susceptibility increases markedly in built-up areas and croplands, which accounts for the persistently high landslide risk even in regions with dense vegetation coverage. Steep slopes created by slope cutting frequently evolve into landslide-prone areas. Thus, investigating the dynamic interplay between human activities and landslide susceptibility is of paramount importance.

5.3. Uncertainties and Advantages

Although high ROC values indicate the reliability of the predictions, there are still uncertainties in the results. This uncertainty stems from two factors. First, the model requires a large amount of data for training and learning, and the landslide data may be incomplete or inaccurate, as these landslide inventories are from 2018 to 2022. Landslides, being ubiquitous and not confined to specific areas, particularly in remote or inaccessible regions, may lead to biased input data for the landslide prediction model. Secondly, the study area has dense vegetation, which can cause SAR data to exhibit shortening, overlapping, and shadowing, leading to decorrelated or missing deformation data in some areas of the study region. SAR data are limited by time constraints and cannot fully cover the landslide evolution process. Additionally, the resolution of the datasets may have influenced the assessment results. In large-scale landslide susceptibility assessments, the lack of finer resolution datasets, including climatic and geological data, due to limited data availability and accessibility, could impact the neural network modeling process. Therefore, in future work, it is necessary to use more consistent and higher-resolution data, which can enhance the performance of the model over larger areas.

Compared to traditional susceptibility models, our method is capable of achieving significantly higher accuracy. This method employed human engineering activity disturbances (LULC) as dynamic factors and combined them with InSAR time-series surface deformation to correct landslide susceptibility maps, thereby producing maps that closely correlate with the actual geological hazards of the area. The accuracy was well controlled through validation with disaster and hazard points.

6. Conclusions

In this paper, a new method for detecting LULC changes and integrating InSAR data was proposed to improve the accuracy of LSM. Based on high-precision dynamic and periodic remote sensing data, this method scanned and detected multi-period surface cover and land use changes. It combined surface cumulative deformation data to identify significant deformation areas in LULC change regions, enabling rapid scan-type identification of hazards over large areas using remote sensing big data. Moreover, by integrating LULC changes and InSAR surface deformation as dynamic factors with the results from traditional landslide susceptibility models, this method generated a corrected landslide susceptibility map, thereby aiming to significantly reduce the incidence of false positives and false negatives. The following conclusions were drawn:

(1)

The performance and validation results of the model indicated that the GRU model, with the highest AUC value of 0.886, exhibited the best efficacy. Furthermore, it was demonstrated that LULC and distance to roads are closely correlated with landslide occurrence.

(2)

Based on the detection of LULC changes and surface deformation results, it was found that deformation occurred most rapidly on slopes ranging from 5 to 25 degrees, with human activities such as deforestation and deforestation identified as the primary drivers of this slope deformation.

(3)

Following the integration of deformation results and landslide susceptibility maps using the correction matrix, 3.10% of false positive areas and 0.74% of false negative areas in the study area were corrected. An area of 12.25% of Yulin was in high-susceptibility zones. The validation against actual landslide occurrences indicated that the corrected results were consistent with real-world conditions.

This method overcame the limitations of relying solely on static factors for landslide susceptibility mapping, highlighting the effectiveness of integrating dynamic factors and surface deformation data to enhance landslide susceptibility mapping and explore the dynamic response relationship between human engineering activities and landslide occurrence. The proposed method holds great potential for application in geohazard prevention and control, supporting informed land use planning and risk mitigation strategies.