Spatial Distribution Analysis of Landslide Deformations and Land-Use Changes in the Three Gorges Reservoir Area by Using Interferometric and Polarimetric SAR

Abstract

Landslides are geological events that frequently cause major disasters. Research on landslides is essential, but current studies mostly use historical landslide data and do not reflect dynamic, real-time research results. In this study, landslide deformations and land-use changes were used to analyze the landslide distribution in Fengjie County and Wushan County in the Three Gorges Reservoir Area (TGRA) by using interferometric and polarimetric SAR. In this study, the mean annual rate of landslide deformations was obtained using the small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) for the ALOS-2 (2014–2019) data. Land-use changes were based on the 2007 and 2017 land-use results from dual-polarization ALOS-1 and ALOS-2 data, respectively. To address the problem of classification accuracy reduction caused by geometric distortion in mountainous areas, we first used texture maps and pseudocolor maps synthesized with dual-polarization intensity maps to perform classification with random forest (RF), and then we used coherence and slope maps to run the K-Means algorithm (KMA). We named this the secondary classification method. It is an improvement on the single classification method, exhibiting a 94% classification accuracy, especially in rugged areas. Combined with land-use changes, GIS spatial analysis was used to analyze the spatial distribution of landslides, and it was found that the landslide rate was significantly correlated with the type after change, with a correlation coefficient of 0.7. In addition, land-use types associated with human activities, such as cultivated vegetation, were more likely to cause landslide deformation, which can be used to guide local land-use planning.

1. Introduction

From 1949 to 2010, China experienced 20,000 landslide and debris-flow disasters each year, causing more than 1000 casualties, affecting more than 900,000 people, and causing direct economic losses of RMB 2–6 billion. The Three Gorges Reservoir Area is one of the areas that is most prone to landslide disasters in China. This seriously affects the socio-economic development of the reservoir area and causes incalculable economic and human losses. Research on landslides currently focuses on landslide hazard assessment [1,2,3,4], forecasting and early warning [5,6], etc., mainly to minimize losses by predicting the occurrence of landslides in advance. However, because landslides often occur in areas with complex environments and dense forests, there are still major difficulties in identifying hidden hazards using remote sensing methods. As a result, more than 70% of landslide disasters that have occurred have not been predicted in advance [7]. Even when landslides can be accurately predicted, property such as fields and buildings in the disaster area will suffer damage. Therefore, it is of great importance to effectively reduce the losses from landslide disasters by understanding the distribution law of landslides and slowing down or preventing the occurrence of landslides at the source.

InSAR technology has developed rapidly in recent years. Time series InSAR techniques such as PS-InSAR [8], SBAS-InSAR [9] and SqueeSAR [10] have achieved good results in monitoring cities, mining areas and landslide areas [11,12]. When time-series InSAR technology was combined with land-use data to study the deformation in urban areas, plains, peatlands, flooded areas, and other areas, it was found that the average deformation rates corresponding to different land-use categories differed in the same study area [13,14], and that the type of land use is an important factor in deformation [15,16]. However, in deformation rate prediction using land-use types, the prediction results obtained after adding historical land-use data are more accurate [17], demonstrating that the influence of historical land use on the local deformation rate cannot be ignored. Changes in land use in urban areas are mainly reflected in changing building loads due to accelerated urban development [18]. The location of the main subsidence areas in Wuhan’s urban areas changed after 2010, gradually moving into areas near new metro lines [19]. In addition, at different spatial scales, the Moran index shows that there is always a significant correlation between urbanization indicators and subsidence [20]. Thus, results concerning the flood inundation area and subsidence rates obtained using historical land-use predictions have realistic value and can guide the future planning of urban areas to reduce human casualties and economic losses caused by disasters [21,22,23]. The joint analysis of land use type/change and surface deformation rate provides a new solution for the prevention and control of surface deformation, and also brings new ideas to the study of the spatial distribution of landslide deformation.

The formation of landslides is influenced by a variety of internal and external factors [24,25], such as distance from water systems, faults, slopes, slope aspects, undulations, elevations, rainfall, human activities, etc. Among them, the intensity of human activities is the only controllable factor that can be used as the core when studying the spatial distribution of landslide deformation.

Currently, an increasing amount of machine learning and deep learning algorithms are being applied to classification, and the improved performance of these algorithms is improving classification accuracy. However, as compared with the problem of classification accuracy differences caused by different classification methods, the problem of accuracy degradation caused by a lack of training samples appears to be more serious [26]. Multispectral remote sensing images are widely used for land classification due to their high spatial resolution [27], but they are susceptible to cloud occlusion, which can lead to missing information and a failure to achieve effective classification. Synthetic aperture radar (SAR) has become an important resource for land use classification due to its ability to penetrate clouds and rain. The SAR’s multipolarization channel provides more adequate information for classification in many regions, such as urban areas, humid tropical areas, and forest areas, and achieves a high classification accuracy [28,29,30,31,32]. However, as a result of the side-look imaging method of SAR, geometric distortion tends to occur in areas with rough terrain, and the intensity information caused by overlapping or shadowing is confused with other land types, resulting in misclassification problems.

In this study, land use changes are used as a characterization of the intensity of human activities; we first combined land use changes and landslide deformation to analyze the spatial distribution using interferometric and polarimetric SAR. On the basis of the dual-polarization of ALOS-1 (2007) and ALOS-2 (2014–2019) data, we adopted a secondary classification method, that is, the random forest (RF) was used for the first classification and then the K-Means algorithm (KMA) was performed. We effectively reduced the classification error caused by geometric distortion and improved the accuracy of land-use classification results of Wushan County and Fengjie County in 2007 and 2017. Combined with the landslide rates obtained using SBAS-InSAR technology, the relationship between land-use changes and landslide deformations was quantitatively evaluated using spatial analysis method.

2. Methods

This chapter introduces the estimation of landslide rates using SBAS-InSAR technology, the secondary classification method based on machine learning algorithms, and the analysis of the landslide spatial distribution with land use. The main flow of our method is shown in Figure 1.

2.1. SBAS-InSAR Technology

SBAS-InSAR technology is a time-series technique that was proposed in 2002 [9]. Its basic principle is to select appropriate spatial and temporal baselines to form differential interference pairs to create multiple small baseline subsets. This can improve image exploitation and reduce the effects of spatial-temporal decorrelation.

In this study, HH polarization data from the ALOS-2 satellite were selected to implement SBAS-InSAR technology using GAMMA software-64-18.04. First, we set appropriate threshold of spatial-temporal baseline to generate interference pairs to attenuate the decorrelation noise. Second, we simulated the terrain-related phases according to the Copernicus DEM. Then, the coherence threshold was set and all pixels above the threshold were retained to complete the phase unwrapping using the minimum cost flow (MCF) method. In addition, we used DEM to calculate the height-dependent atmospheric phase delay. Finally, the linear rate of ground points from 2014 to 2019 was obtained using the singular value decomposition (SVD).

2.2. The Secondary Classification Method and Accuracy Assessment

2.2.1. Secondary Classification with Dual-Polarization SAR Data

There are five types of land use in the Fengjie and Wushan areas: forest, water, cropland, cultivated vegetation and buildings. Of these, the cultivated vegetation types are mainly fruit trees such as navel oranges and plums, and the cropland are rice, corn and cotton. Although both are highly affected by human activities, there are differences in their growing conditions, such as terrain and slope. Figure 2 shows comparisons of the pseudocolor image and the optical images of the five types.

It can be seen that some optical images are largely covered by clouds, making it difficult to distinguish features. SAR images are susceptible to geometric distortion, and errors will occur when used for classification. Therefore, we proposed a secondary classification method to solve this problem. One scene of single look complex (SLC) images from the same season (July and August) in both 2007 and in 2017 was selected, with two types of polarization, i.e., HH and HV, which comprised the original data source for the secondary classification. This method mainly uses the RF algorithm and KMA to complete the first and second classification steps, respectively, and then synthesize the two results to obtain the final classification results. The secondary classification process is shown in Figure 1.

The random forest algorithm was used in the first classification process. It is a typical supervised learning algorithm, i.e., an integrated algorithm based on a decision tree. Its “random” nature is reflected in two ways: first, there is put-back sampling from the original set of samples; second, there is put-back sampling from the original features and then the optimal features are selected among them. The subset of samples and features form a decision tree, and multiple decision trees form a “forest” in which all trees are integrated to vote on the classification results. The category with the highest number of votes is then output as the result [33]. The advantage of this is that it can handle multiple input variables and produce a high accuracy rate for multiple input elements.

The second classification process was performed using the K-means algorithm, which is an unsupervised learning algorithm. The principle is to randomly select K objects as initial cluster centers, calculate the distance between each object and each cluster center, and assign each object to the cluster center closest to it. The cluster centers and the objects assigned to them form a cluster. The method can combine classes based on distance without sampling, but it allows only one input quantity [34].

The secondary classification method was adopted to extract two classes when serious confusion occurred in the first classification result, and select the input quantity that had high discriminative power for these two classes to perform K-means clustering after achieving the RF. The essence of this method is to simplify the multiclassification task into a two-classification task, so that it can still maintain a high classification accuracy when there is a large number of categories.

The secondary classification method has four kinds of characteristic data, which are pseudo-color map, texture map, coherence map and slope map. Among them, the pseudo-color map is synthesized from dual-polarization SAR images, and the discrimination of different land use types can be improved by displaying different colors. Texture information and coherence maps are derived from HH polarimetric SAR images, and slope maps are calculated from DEM.

2.2.2. Accuracy Evaluation Index

Overall accuracy (OA) is a commonly used accuracy evaluation index. If there are two types of features A and B in a region, the actual A and the classification result A is denoted as TP, the actual A but the classification result B is denoted as FN, the actual B and the classification result B is denoted as TN, and the actual B but the classification result A is denoted as FP; then, the OA is calculated as follows [28]:

$$\mathrm{OA}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

(1)

Another evaluation index is the Kappa coefficient, which is obtained on the basis of overall accuracy. It can be expressed as:

$$\mathrm{K}=\frac{\mathrm{OA}-\mathrm{P}}{1-\mathrm{P}}$$

(2)

among them, the expression of P is:

$$\mathrm{P}=\frac{\left(\mathrm{TP}+\mathrm{FN}\right)\times \left(\mathrm{TP}+\mathrm{FP}\right)+\left(\mathrm{FN}+\mathrm{TN}\right)\times (\mathrm{FP}+\mathrm{TN})}{{(\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN})}^2}$$

(3)

2.3. Spatial Analysis

2.3.1. Data Preprocessing for Spatial Analysis

Using the secondary classification method, the land-use results for 2007 and 2017 were obtained by inputting the ALOS-1 and ALOS-2 data, respectively. Then, using the ArcGIS 10.2 software, both land-use results were resampled to the same spatial resolution as the average landslide rates of ALOS-2. On the basis of the landslide rates in the study area, the typical landslide areas were vectorized, and the land-use and landslide rates corresponding to the typical landslide areas were separately extracted to perform the following analysis.

First, for the typical landslide areas, the land use in 2017 was overlaid with the average rates to analyze the relationship between the landslide deformations and land-use types. Second, for the typical landslide areas, land-use changes from 2007 to 2017 were overlaid with the average rates to obtain the influence of land-use changes on the landslide deformations. Finally, a spatial heterogeneity analysis was performed using the geographically weighted regression (GWR) for the landslide rates over the whole study area with influence factors, such as slope, aspect, elevation, undulation, land-use type, and distance from the water system.

2.3.2. Spatial Analysis Method

(1) Overlay analysis. Overlay analysis is a very important spatial analysis technique in GIS (the geographic information system). It refers to the process of generating new data under the same spatial reference system through a series of set operations on two data (Figure 3). The aim is to analyze the relationship between the spatial characteristics and properties of spatial objects that have some correlation in terms of spatial location.

(2) Transfer matrix. The land use transfer matrix [17] is an application of the Markov model in land-use change, which not only quantitatively represents the changes between land-use types, but also reveals the transfer rates between different land-use types. In addition, it can be obtained via the overlay analysis method using ArcGIS. The matrix calculation formula is as follows:

$$\mathrm{Area}\mathrm{Percentage}=\frac{{\mathrm{A}}_{\mathrm{i}}}{\mathrm{W}}\times 100\%$$

(4)

where A_i is the area corresponding to each type of change and W is the total area of the study area.

(3) Geographically weighted regression model. Traditional regression analysis was developed based on the stability of the relationship between the independent variable and the dependent variable in a given area. Only a unique spatial coefficient could be obtained with the following mathematical expression:

$$\mathrm{y}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{x}}_1+{\mathsf{\beta }}_2{\mathrm{x}}_2+\dots {\mathsf{\beta }}_{\mathrm{p}}{\mathrm{x}}_{\mathrm{n}}+\mathsf{\epsilon }$$

(5)

where y is the dependent variable, ${\mathrm{x}}_{1,2,\dots \mathrm{n}}$ is the explanatory variable (independent variable), ${\mathsf{\beta }}_{1,2,\dots \mathrm{p}}$ is the coefficient of each explanatory variable, and ε is the random error.

In reality, performances will always differ depending on the spatial location, i.e., spatial heterogeneity. The GWR model [35] is a local regression analysis method that can obtain the respective spatial coefficients at each spatial location, thereby preserving the spatial heterogeneity of the data and expressing the relationship between the independent and dependent variables more accurately than traditional methods. Its mathematical expression is as follows:

$${\mathrm{y}}_{\mathrm{i}}={\mathsf{\beta }}_0\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)+\sum _{\mathrm{k}=1}^{\mathrm{p}}{\mathsf{\beta }}_{\mathrm{k}}\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right){\mathrm{x}}_{\mathrm{k}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(6)

where $\mathrm{i}$ is the sampling point serial number, $\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)$ is the geographic coordinate of sampling point i, ${\mathsf{\beta }}_{\mathrm{k}}\left({\mathrm{u}}_{\mathrm{i}}{\mathrm{v}}_{\mathrm{i}}\right)$ is the coefficient of the kth explanatory variable at sampling point i, and ${\epsilon }_i$ is the random error at sampling point i.

3. Study Area and Data

3.1. Overview of the Study Area

The Fengjie and Wushan districts are located in the eastern part of Chongqing, in the hinterland of the TGRA. The study area is bounded by east longitudes 109°25′30″ and 109°55′58″ and north latitudes 30°58′29″ and 31°5′25″, with the Yangtze River running east to west, a well-developed water system, and numerous river networks and reservoirs. The area is crisscrossed by ravines, and has an undulating terrain with the Yangtze River as the axis of symmetry. In addition, it has a gradually increasing elevation on both sides of the river, tectonic development, and a complex geological background. It has a humid subtropical monsoon climate with four distinct seasons, abundant rainfall, and sufficient sunshine. The rugged terrain and complex geographical environment have led to serious geological disasters in the region, resulting in significant losses each year.

By 2020, there were 2 roads, 11 cities, and 13 townships in Wushan County and 4 roads, 18 cities, and 11 townships in Fengjie County, with a population of 462,400 and 744,800, respectively. The area is rich in tourism resources, and with the completion and opening of the Yuyi Expressway, the level of economic development has gradually improved.

3.2. Data

One single-look complex (SLC) image from the ALOS-1 satellite in 2007 was acquired for both Fengjie and Wushan counties. In addition, 11 SLC images from the ALOS-2 satellite were acquired for the period 2014–2019. The SAR data bands used are all L-band and the polarization methods were HH and HV (see

Table 1. Data of ALOS-1 and ALOS-2.

Sensor	Date	Counts	SLC Thumbnails		Resolution (m)	Coverage
			HH	HV
ALOS-1	2007	1			9 (range) × 3 (azimuth)	Fengjie County
	2007	1				Wushan County
ALOS-2	2014–2019	11			4 (range) × 3 (azimuth)	Fengjie and Wushan County

). From the SLC thumbnails, it can be seen that the HH or HV image alone could not be used to discriminate land use, so it was necessary to combine data from these two channels for classification. A download of Copernicus DEM (30 m spatial resolution) [36] data was used as the basis for topographic phase removal and geocoding.

Two existing land-use classification sets were collected, namely, the Sentinel-2 satellite classification dataset published by ESRI [37] and the annual Chinese land-cover dataset (CLCD) obtained from Landsat data published by Wuhan University [38], with resolutions of 10 m and 30 m, respectively.

4. Results

4.1. SBAS-InSAR to Monitor Landslide Deformation Rates

We first obtained 22 interference pairs from 11 SLC by setting the temporal baseline to 400 days and the spatial baseline to 300 m. To obtain an accurate solution, we selected all pixels with a coherence value greater than 0.32. Figure 4 depicts the landslide rates in the study area from 2014 to 2019, which were derived by SBAS-InSAR and ranged from −9 to 9 cm/year. Despite the temporal span of the study, most of the pixels exhibited no deformation, indicating stable geological condition. However, representative landslide areas were found to be distributed along the Yangtze River, with a maximum deformation rate of 9 cm/year. Figure 4 (right) highlights prominent landslide features that can be used for analyzing the spatial distribution of landslides. These results provide valuable insights into the landslide patterns in this region, contributing to our understanding of the dynamics of landslide-prone areas.

4.2. Secondary Classification Results and Accuracy Assessment

Figure 5a,c,e, and Figure 5b,d,f show the data sources of the secondary classification method for the ALOS-1 and ALOS-2 data, respectively. There were two input data sources for the first classification, the pseudocolor map shown in Figure 5a,b, and the texture map shown in Figure 5c,d. Figure 5e,f and Figure 5g show the coherence and slope maps used for the second classification, respectively. The 2007 and 2017 land-use results were obtained from the secondary classification using these data sources (Figure 6).

Two validation areas and three comparison methods were selected for the secondary classification experiments. The first validation area contains five land-use types, mainly forested and rugged terrain. The second validation area contains only three land-use types, forest, and water, and mostly consists of buildings and flat terrain. Figure 7a,b and Figure 7g,h show the pseudocolor map coverage and validation samples, respectively. Figure 7c,i show the results of the secondary classification method performed for the two validation regions, respectively, with an overall accuracy of 94.44% and 95.52% and Kappa coefficient of 91.59% and 92.29%, and the confusion matrix as shown in

Table 2. Confusion matrix for the classification results.

Name	Method	Classification Accuracy of Each Type(%)					OA (%)	Kappa (%)
		Buildings	Forest	Water	Cropland	Cultivated Vegetation
Validation Area 1	Comparison Method 1	89.63	76.03	100.00	84.70	83.70	82.29	74.72
	Comparison Method 2	99.67	78.88	100.00	91.16	88.58	86.12	80.33
	Comparison Method 3	96.57	64.56	99.94	96.88	89.32	78.97	71.50
	Secondary Classification	84.45	95.80	99.94	98.07	89.00	94.44	91.59
Validation Area 2	Comparison Method 1	81.35	40.18	99.92	——	——	87.38	78.14
	Comparison Method 2	97.21	70.44	99.92	——	——	96.02	93.02
	Comparison Method 3	81.74	35.33	99.92	——	——	87.04	77.46
	Secondary Classification	90.62	89.83	99.96	——	——	95.52	92.29

The bold in the table are the results of the method proposed in this paper.

.

Comparison Method 1: Only the first classification was performed and only the pseudocolor map was input. The results for the first and second validation areas are shown in Figure 7d,j, respectively, with overall accuracies of 82.29% and 87.38%, and Kappa coefficients of 74.72% and 78.14%. The classification results show that most of the areas could be well distinguished using the pseudocolor map alone, but some of the forest types were misclassified as building types due to the severe geometric distortion in the mountainous areas. Thus, more input information was needed to distinguish between them.

Comparison Method 2: All four types of data sources were input for the first classification, and no second classification was performed. When more data sources were input, the overall accuracy greatly improved in flat areas, reaching 96.02% (Figure 7k). However, the accuracy did not significantly improve in rough areas, reaching only 86.12% (Figure 7e). There were also large differences in their Kappa coefficients, 93.02% and 80.33%, respectively (Table 2).

Comparison Method 3: The classification correction step using coherence was removed from the secondary classification framework. This comparison method achieved the lowest overall accuracy of all experiments in both validation areas, i.e., 78.97% and 87.04%, respectively (Figure 7f,l). The Kappa coefficients are also inaccurate.

The above comparison methods show that, for flat areas, adding data sources can effectively improve the classification accuracy, but not well in all types. The secondary classification method can be used to mitigate the effect of geometric distortion with coherence, showing good accuracies in both rugged and flat areas.

The CLCD and ESRI datasets selected for the rugged areas are shown in Figure 8a,b and Figure 8c,d, respectively. The results of the ESRI datasets (Figure 8d) show that forests and buildings were misclassified in the rugged areas, and cropland was not subdivided in the CLCD results (Figure 8b). Therefore, the land-use data used in this study were the results of the secondary classification process.

4.3. Land-Use Changes from 2007 to 2017

On the basis of the 2007 and 2017 land-use results, GIS spatial analysis was used to derive land-use changes from 2007 to 2017 (Figure 9). This method of first classifying and then calculating the change can directly show the type of change, making it easier to understand. Figure 9 (right) shows the land use changes in the five landslide areas marked in Figure 4.

In Figure 4, the rates in the A-D are large, while those in E are relatively small. From the type changes in Figure 9, it can be seen that the rates are not directly related to the type change or not. There is a large area of unchanged type in B, but the landslide deformation rate reaches 9 cm/year. Area E has a large number of change types, but the rates are relatively small.

Transfer matrixes for the study area and typical landslide areas were constructed according to the results of land-use change (

Table 3. Land-use transfer matrix for the whole area from 2007–2017. The area is color-coded from green (small) to red (large).

2007	2017
	Buildings	Water	Forest	Cropland	Cultivated Vegetation	Total
Buildings	0.15	0.01	0.17	0.04	0.01	0.39
Water	0.00	4.88	0.17	0.13	0.04	5.23
Forest	0.41	0.83	53.65	3.99	9.46	68.33
Cropland	0.13	0.10	3.42	3.11	2.67	9.43
Cultivated Vegetation	0.05	0.02	6.65	2.29	7.61	16.62
Total	0.74	5.85	64.06	9.55	19.79	100.00

and

Table 4. Land-use transfer matrix for the typical landslide areas from 2007 to 2017. The area is color-coded from green (small) to red (large).

2007	2017
	Buildings	Water	Forest	Cropland	Cultivated Vegetation	Total
Buildings	0.00	0.00	0.74	0.29	0.00	1.03
Water	0.00	1.86	0.26	0.19	0.13	2.45
Forest	0.00	0.97	11.18	12.32	9.64	34.11
Cropland	0.00	0.36	7.49	17.99	7.87	33.72
Cultivated Vegetation	0.00	0.00	4.34	5.76	18.59	28.70
Total	0.00	3.20	24.02	36.55	36.24	100.00

, respectively).

During this decade, the main land type in the study area was forest, covering almost 70% of the land, and more than three-quarters of the forest area was in stable condition. The major land-use changes in both the study area and the typical landslide areas were between forest, cropland, and cultivated vegetation, but the area composition of these three types differed greatly. For the whole area (Table 3), the area of forest was absolutely dominant, whereas in the typical landslide areas (Table 4), the area percentage of the three land types did not differ significantly. The area of cultivated vegetation was higher in typical landslide areas than for the whole area, indicating that the intensity of the influence of human activity was greater in the landslide areas.

4.4. GIS Spatial Analysis with Landslide Deformations and Land-Use Changes

4.4.1. For Typical Landslide Areas with No Change in Land Use from 2007 to 2017

In practical production, it is difficult to determine normality of data, so when comparing the differences between the means of several independent datasets, nonparametric test methods are generally used. The K–W test (Kruskal-Wallis test) can compare the differences between more than two independent datasets. The null hypothesis is that all data distributions are the same and there is no difference.

Forest, cropland and cultivated vegetation account for more than 90% of the total landslide area, so this paper only considered these three main types and the conversion between the three, and did not consider information on other types. All pixel points in typical landslide areas with no change in land-use type between 2007 and 2017 were extracted, and the mean rates corresponding to these points were overlaid and analyzed. The K–W test was used to test the nonparametric test of the overlay analysis results. The results are shown in

Table 5. Significance analysis of the corresponding rates of land-use types in landslide areas.

Analysis Items	Variables	Sample Size	Mean Velocity (cm/year)	Statistics	p-Value	Effect Amount
Velocity	Cultivated Vegetation	1308	3.47	1362.433	0.000 ***	0.019
	Forest	572	1.35
	Cropland	1322	1.54
	Total	3202	-

*** represents 1% significance levels, respectively. Those in bold in the table were mentioned in this paper.

, with a p-value of 0.000 *** and a significance level of 1%. The null hypothesis was rejected, i.e., there were significant differences between the landslide deformation rates corresponding to different land use types. The conclusion showed that in the landslide area, different land use types had different effects on the landslide rate, i.e., human intensity had different effects on landslide deformation. When the intensity of human activities was higher, the resulting landslide rate was significantly higher than in other areas.

The average rates of the three types were calculated separately, and it was found that the rate of cultivated vegetation (3.47 cm/year) was significantly higher than that of the other two types (forest: 1.35 cm/year and cropland: 1.54 cm/year). It showed that for the landslide area, if the land use type does not change, the cultivated vegetation type has a promoting effect on the landslide deformation, while the other two types have relatively little promoting effect on the landslide. The different facilitating effects of these three types on the landslide rate are related to the intensity of human activities they bear, which has important implications for future land use planning.

4.4.2. For Typical Landslide Areas with Land-Use Changes from 2007 to 2017

The pixels of land-use changes in typical landslide areas and the corresponding land-slide rates were extracted. Then, the two were used to perform an overlay analysis to obtain the data in

Table 6. Correlation between land-use changes and landslide deformation rates.

Type 1	Type 2	Type 1 Rate	Type 2 Rate	2017 Rate
Forest	Cropland	1.35	1.54	1.58
Forest	Cultivated Vegetation	1.35	3.47	3.36
Cropland	Forest	1.54	1.35	1.59
Cropland	Cultivated Vegetation	1.54	3.47	2.73
Cultivated Vegetation	Forest	3.47	1.35	2.79
Cultivated Vegetation	Cropland	3.47	1.54	2.30
Correlation Coefficient		0.15	0.70	1.00

Type 1 denotes the type before the change; type 2 denotes the type after the change; type 1 rate denotes the average rate corresponding to type 1; type 2 rate denotes the average rate corresponding to type 2. 2017 rate shows the average rate after the changes in the area. And the bold in the table was our concern.

.

During the period 2007–2017, there were six land-use changes. The correlation coefficient between the actual rates after the changes and the other rates was calculated, and it was found that the actual landslide rates in the changed area were significantly influenced by the land-use type after change, with a correlation coefficient of 0.7. If the land use change was of a type that can promote landslide deformation, the deformation rate would increase, and if the land use change was of a type that slows down the deformation rate, the deformation rate would have a downward trend. This is a good illustration of the effect of land use changes on the landslide deformation rates. Therefore, in order to achieve the purpose of slowing down and preventing landslide deformation from the perspective of land use planning, it is necessary to fully consider whether land use change has a promoting effect on landslide deformation.

4.4.3. For the Typical Landslide Areas from 2007 to 2017

On the basis of the contents of the previous two subsections, all types of changes that occurred in typical landslide areas (including changes from forest to forest in the same type) were organized, as shown in

Table 7. Land-use change in landslide areas. The area is color-coded from green (small) to red (large).

Velocity (cm/year)		Type after Change
		Forest	Cropland	Cultivated Vegetation
Type before Change	Forest	1.35	1.58	3.36
	Cropland	1.59	1.54	2.73
	Cultivated Vegetation	2.79	2.30	3.47
	Mean	1.91	1.81	3.19

. It was found that, regardless of the type of change, the maximum landslide rate occurred in areas after changes to cultivated vegetation. This indicates that the change type has a significant influence on the landslide rate.

From the analysis of typical landslide areas, the cultivated vegetation area, which is a land-use type that is heavily influenced by human activities, was shown to be more prone to landslides. When the land-use type changed from cultivated vegetation to forest and cropland, the landslide rates decreased more significantly, but they remained higher than the rates observed after other changes to forest land and cropland. This indicates that the landslide rates are not only influenced by the land type after change, but also by the land type before change. It can be seen that, although changing the land-use type is effective to slow down the occurrence of landslides, the historical land-use type still has a certain influence for a short period of time. Therefore, when selecting sites for populated areas, such as areas for construction, both the current land-use type and the historical type should be considered in order to avoid landslide disasters.

4.5. Spatial Heterogeneity Analysis with GWR

On the basis of typical landslide areas, we found that the landslide rate was higher in cultivated vegetation areas; however, from the perspective of land-use type for the whole area, cultivated vegetation areas did not all experience landslide phenomena. This is due to spatial heterogeneity, i.e., different spatial locations are affected by the same factors, but react differently. In reality, the occurrence of landslides is not only influenced by land-use change. It is, also influenced by factors such as elevation, undulation, distance from the water system, aspect, and slope (Figure 10a–d and Figure 5, respectively).

GWR was used to analyze the spatial heterogeneity of the study area, and the overall process is shown in Figure 11. Firstly, the ArcGIS 10.2 software was used to create a raster network for the study area, with a total of 31,344 raster points being obtained. Then, the elevation, undulation, land-use change, distance from the water system, slope, aspect, and landslide rate data were extracted according to the grid points, which were used as the input for GWR. Finally, the output from the GWR software was visualized in ArcGIS (Figure 12).

Figure 12 shows the influence of land use change on the landslide rate at different spatial locations, in which the base map is the deformation rate. The yellow irregular area is the landslide area, and the blue and red circles represent the influence of the landslide rate, which is in the range of (0.160, 0.210) and (0.210, 0.302), respectively. It can be seen that the landslide vector range has a high degree of overlap with the blue and red circles. This indicates that in the landslide area, the landslide deformation rate is more affected by land use change than in other areas.

Land use type/change had an effect on landslide deformation rate, but it was not the only factor in landslide occurrence. It is only when the slope meets the geological environment that the effect of land use on landslides becomes apparent. Therefore, not all areas heavily influenced by human activities will have landslide deformation, reflecting the spatial heterogeneity of landslide distribution.

After analyzing the laws of typical landslide areas and the whole study area, it was concluded that land-use change has an important impact on landslide occurrence, which can be used to provide guidance for local land-use planning. However, when land use planning was used to slow down or prevent the occurrence of landslides, it was necessary not only to consider the type of land use and its effect on deformation, but also to judge whether it has the basic conditions for landslides according to the geographical environment. This paper combined the analysis of spatial heterogeneity theory, which can avoid the blind reduction or increase of certain types in planning, resulting in the instability of the local ecological environment.

5. Discussion

5.1. Landslide Rates Monitored by SBAS-InSAR

SBAS-InSAR technology is a time-series observation technology that enables large-scale, high-precision measurements. It can be used to monitor landslides, earthquakes, urban areas, etc., providing a scientific basis for scientific disaster prevention and post-disaster reconstruction [39]. In this work, only 11 images from the ALOS-2 satellite were available between 2014 and 2019. Considering the stability of results, we set the spatial baseline at 300 m and the temporal baseline at 400 days. With these thresholds, 22 interferograms were generated that retained good coherence. By setting the coherence threshold to 0.32, the landslide rates in Fengjie and Wushan counties were obtained, in a range from −9 to 9 cm/year.

The area under study is stable, with only five significant landslides in the region. These landslides develop close to the water and cover a large area, so when they collapse they not only cause local damage but also affect the area upstream. Therefore, it is worth exploring the mechanisms that affect their development.

5.2. Land-Use Classification with the Secondary Classification Method

ALOS-1 and ALOS-2 both have two channels, i.e., HH and HV, which can provide more information than single-polarization, which aids in distinguishing different land use types. Using these two types of SAR data, the land-use results of the region in 2007 and 2017 were obtained using the secondary classification method, with the classification accuracy reaching 94%. The secondary classification method was performed by first using RF and then KMA with the dual-polarization SAR data. This addressed the problem of distortion. In addition, it produced a higher accuracy than the other methods, which only perform single classification, especially in rugged areas.

Thereafter, land-use changes from 2007 to 2017 (Figure 9) were obtained from the land-use information from 2007 and 2017. The types of land-use changes were directly obtained by classifying and then calculating the change types, which avoid the labeling of a large number of change types. The majority of land-use types in the region were in a stable state, and changes between three land types, i.e., forest, cropland, and cultivated vegetation, were our focus.

The combination of RF and KMA exhibited a good performance in the classification of rugged areas, which demonstrates that it can be used to convert complex classification problems into binary classification problems.

5.3. Relationship between Landslide Deformation Rates and Land-Use Changes

Previous studies analyzed the relationship between deformation rates and land-use changes in urban areas, peatlands, and other areas [13,14,15,16,17,18,19,20]. They concluded that land-use changes have an impact on the deformation rates. At present, the majority of research on landslide areas uses historical landslide data to evaluate and analyze the hazards and risks, with the use of real-time landslide monitoring being limited. However, it is necessary to incorporate time-series InSAR technology into the field of landslide monitoring. To the best of our knowledge, this study is the first to analyze the spatial distribution by linking landslide deformation rates from SBAS-InSAR with land-use changes.

First, we analyzed five typical landslide areas. After an overlay analysis of land-use changes and landslide rates, it was found that landslide rates were significantly related to the rates of the change types for a short period of time (with a correlation coefficient of 0.7), but they were not only influenced by the change type. First, no matter whether the change was from forest, cropland, or cultivated vegetation, as long as the change type was to cultivated vegetation, the landslide rate in the corresponding area was greater than the rate of other change types. Second, when the rate of the change type was smaller than that of the before-change type, the landslide rate in the corresponding area reduced to some extent; however, this reduction was still affected by the before-change type.

Thereafter, we used the GWR model to ensure spatial heterogeneity over the whole area. The results showed that slope, aspect, elevation, undulation, land-use changes, distance from the water system, etc., had different effects on the landslide rate in different geographical locations. In the study area, land-use change had a significant effect on the landslide rate, and the affected area coincided with the landslide area monitored by SBAS-InSAR, which further confirmed the conclusion proposed in this paper that land-use change affects landslide distribution.

However, there are still some limitations in this study. First, the lack of dual-polarization SAR data is not conducive to the analysis of long-term land use changes. Second, the combination of random forest and K-means clustering is not suitable for all regions due to algorithmic limitations.

In the future, more and more SAR satellite missions will be successfully launched, providing more polarimetric SAR data for long-term research. Machine learning algorithms with higher accuracy and applicability will also be applied to mountain classification. This will complement the research in this paper.

6. Conclusions

From the perspective of landslide mitigation and prevention, this paper combined polarimetric SAR and InSAR for the first time to complete the study of the spatial distribution of landslide deformation. To solve the problem that dual-polarization SAR images are easily affected by geometric distortion, this paper proposed a secondary classification method that first performed random forest classification and then K-means clustering, and the accuracy was greatly improved. Using spatial analysis methods, the landslide deformation rates and land use data were combined for analysis, and finally the following findings were obtained:

• Land use classification in rugged areas cannot improve accuracy just by increasing feature data.
• It was possible to complete the classification by simplifying the multi-classification task into a binary classification task, and the classification accuracy obtained was higher and more reliable than that of only one classification.
• The land use types strongly influenced by the intensity of human activities can promote landslide deformation, such as cultivated vegetation.
• The landslide deformation rates were affected not only by the current land use type, but also by the historical land use type.
• Landslides were influenced by various internal and external factors. It was only when the geological conditions of landslides were met that the effects of land use and land use changes became apparent.

This paper comprehensively analyzed the spatial distribution of landslide deformation from both local and overall perspectives, which can provide a reasonable basis for local land use planning to slow down and inhibit the occurrence of landslides.