Extraction of landslide morphology based on Topographic Profile along the Direction of Slope Movement using UAV images

Abstract

The landslide morphology is quickly and accurately extracted from Unmanned Air Vehicle (UAV) images. It is of great significance for emergency rescue and quantitative evaluation of landslide disasters. However, due to the complexity of landslide morphology, choosing the reasonable extraction thresholds is a challenging issue. A threshold selection method of Topographic Profile along the Direction of Slope Movement (TP-DSM) was proposed. Firstly, a hierarchical extraction rule sets for landslide morphology was constructed by integrating multi-feature information such as spectral, texture, geometry, topography and space of UAV images. Second, TP-DSM was proposed to select the optimal elevation thresholds for classifying different landslide morphology. Finally, the thresholds were introduced into the rule sets to achieve effective extraction of landslide morphology. This study uses Digital Orthophoto Map (DOM) and Digital Elevation Model (DEM) generated by UAV images as data sources, and the landslide in Luquan County, Yunnan Province, China as the Study area, the results show that the overall accuracy (OA) of landslide morphology extraction was 89.58%, and the Kappa coefficient was 0.88, which is effective and more consistent with the reality. The proposed method can also be applied to other potential locations.

Keywords:

• UAV images
• multi-feature information
• TP-DSM
• landslide morphology

1. Introduction

1.1. Background

Landslide morphology refers to the typical surface morphological elements unique of landslides, such as the cliff, source area, circulation area, accumulation area, etc. These recognizable morphologies can reflect the material composition and movement mode of the landslide. By obtaining the landslide morphology, valuable information can be provided for landslide inventory mapping, which is important for quantitative risk assessment, secondary hazard prediction and post-disaster reconstruction planning (Nichol and Wong 2005; Bui et al. 2018; Sameen and Pradhan 2019; Karantanellis et al. 2020; Viet-Ha et al. 2020). However, there are few studies on the fine extraction of landslide morphology nowadays, so it is of practical significance to explore landslide morphology extraction. Luquan County in Yunnan Province is one of the counties with the worst landslide disasters in China. It is recorded that two large-scale landslides occurred in this area in 1965, with a total volume of 170 million m³, resulting in the tragedy of five villages being flooded and hundreds of people leaving. The survey found 193 landslides affecting settlements and construction facilities as of 2015, landslides and other geological disasters directly threaten more than 20,000 people in 4,616 households, with potential economic losses of about 290 million yuan. A series of such disasters have become a major bottleneck constraining local socio-economic development. The landslide morphology is explored to provide a reference for landslide disasters prevention in local or other potential locations.

Unmanned Air Vehicle (UAV) has unique advantages due to its high timeliness, high imaging resolution and flexibility (Pellicani et al. 2019; Singh et al. 2022). The new levels of spatial details provided by UAV images (Blaschke 2010). Digital Orthophoto Map (DOM) and Digital Elevation Model (DEM) are generated using UAV images. DOM is an image generated by digital differential correction and mosaicking of aerial or aerospace photos, cropped according to a certain map range. Rich multi-feature information is included such as spectral, texture, geometry and space, and with high spatial resolution to capture localized details on the surface. DEM is a digital expression of the surface morphology. The UAV images have high spatial resolution, which means that the generated DEM is more accurate, can describe the micro-topographic features of landslide (Chen et al. 2017; Wang et al. 2018). Therefore, they are becoming increasingly popular in the study of landslide morphology extraction (Van Den Eeckhaut et al. 2012).

1.2. Related works

The extraction of landslide morphology is mainly based on DEM and derived data such as slope, curvature, roughness and profile (Guzzetti et al. 2012; Pawluszek 2019), and there are three extraction methods used. The first is visual interpretation based on expert knowledge and experience. After comprehensive analysis of topographic maps and geological maps, and superimposed or aligned with remote sensing images, landslide morphology can be visually interpreted according to their position relationship and topography features. Huang et al. (2021) compared visual interpretation of satellite images acquired shortly before and after the earthquake and field survey, created a new landslide database. Chong et al. (2009) combined multi-source remote sensing images, topographic maps, and DEM to visual interpretation the number of secondary hazards affected by the Wenchuan Ms8.0 earthquake. But this method has certain subjectivity, the accuracy is easily affected by the professional level of interpreters (Scaioni et al. 2014) and time-consuming, cannot meet the emergency needs. In this context, automated/semi-automated methods using computers began to rise. Pixel-based statistical analysis method is one of them. The slope, curvature and roughness of the landslide are clustered according to certain rules based on pixels, and the relatively homogeneous pixels are regarded as spatially independent entities, and then landslide morphology is judged according to various topography features (Van Den Eeckhaut et al. 2012). Shi et al. (2018) used the local Gi* statistical method based on curvature threshold to extract landslide morphology represented by concave-convex trends. Tarolli et al. (2012) proposed others based on variability statistical analysis to define objective thresholds of landform curvature for morphology extraction. However, due to the complexity of landslide morphological characteristics, the results of this method will be noisy by reduced homogeneity. It is more suitable for landslides with relatively obvious topographic change. To solve the above problem, object-oriented image analysis (OOA) method has emerged. Firstly, pixels are clustered into relatively homogeneous objects, and then thresholds are set based on topography features such as elevation, slope, surface roughness and profile (Lahousse et al. 2011; Blaschke et al. 2014; Chen et al. 2017; Feizizadeh et al. 2017) to extract the landslide morphology. Van Den Eeckhaut et al. (2012) mapped landslide morphology using only single pulse LiDAR data for both the segmentation and classification steps, without the support of any spectral information. Compared with the pixel-based statistical analysis method, this method works more logically by first delineating objects and subsequently classifying them (Drăguţ and Eisank 2011), and the landslide morphology characterized by various topographic expressions can be better quantified and extracted. However, due to the complexity of landslide environment and its morphology, the traditional single-level OOA is difficult to meet the needs. Therefore, it is necessary to construct a hierarchically organized landslide morphology extraction framework.

The uncertainty of thresholds selection may easily affect the extraction accuracy of landslide morphology. To solve the problem of threshold setting, multi-period DEM is usually used for topographic profile analysis, linear profile is extracted, and thresholds are determined by comparing the elevation fluctuation changes before and after landslide occurrence (the elevation of source area drops and the elevation of accumulation area rises after the slide), and then the various of landslide morphology are partitioned (Drăguţ and Eisank 2011; Tang et al. 2015). Tang et al. (2015) deduced the elevation change data of the source area and accumulation area by comparing the changes of the topographic profile before and after the landslide. Although this method has made some progress, it also suffers from two shortcomings: One is that a single linear profile is susceptible to the fluctuation of micro-geomorphology such as cracks, depressions, drumlins (Telbisz et al. 2013), which can easily cause sudden changes in elevation in local areas and lead to reduced accuracy. The other is that the suddenness of the landslide makes pre-disaster DEM data difficult to be obtained and only post-disaster DEM is available (Qi et al. 2022), which does not satisfy comparative analysis using multi-period DEM. Therefore, it is an urgent issue to be solved that landslide morphology is accurately quantified and extracted by post-disaster DEM.

1.3. Contributions

Aiming at the problems of scene complexity and thresholds selection in landslide morphology extraction, in this study, the landslide in Luquan County, Yunnan Province, China, was tested using DOM and DEM by post-disaster UAV images. Based on OOA (Blaschke 2010), a multi-feature hierarchical extraction rule sets of landslide morphology is constructed, and a Topographic Profile along the Direction of Slope Movement (TP-DSM) method is proposed, which screens the optimal elevation thresholds of different landslide morphology for hierarchical extraction rule sets, so as to achieve rapid and accurate extraction of landslide morphology. Our contributions and innovations are twofold: (1) the multi-feature information of UAV images was integrated to establish a hierarchical extraction rule sets, which had optimized the extraction method in the complex landslide environment to a certain extent. (2) The differences in elevation changes and concave-convex trends among various landslide morphologies were taken into account by the proposed TP-DSM method, and reasonable thresholds were eventually selected.

2. Study area and data

2.1. Study area

The study area is located in Luquan County, Kunming City, Yunnan Province, China (Figure 1). The geographical location is between 102°14′ − 102°56′ E and 25°25′ − 26°22′ N, with the mountainous area accounts for 98.4% of the total area. The terrain is undulating. The subtropical plateau monsoon climate has abundant and concentrated precipitation, with the annual average precipitation about 1000 mm. The strata are developed, and the neotectonic movement is active. In addition, under the influence of unreasonable human activities, environment deterioration, resulting in frequent geological disasters such as landslides.

Figure 1. Study area.

The target landslide is located near the road, with the average altitude of more than 2267.94 m, and the dip angle is nearly vertical. The landslide morphologies are prominent. The lithology of this landslide is mainly sandstone, with thick-layered and medium-layered structure, a rough texture and weak weathering resistance. The upper part of the landslide has obvious cracks, the middle and lower surfaces are rough and uneven, with debris falling off, and deep cracks exist at the edge.

2.2. Data source

The DJI Phantom 4 Pro as the UAV image acquisition equipment, inexpensive, stable, portable, the specifications are shown in Table 1. Data were collected on August 22, 2020, with an overall flight time of 8 min. Before the flight, 7 control points were evenly distributed in the study area, measured by means of GPS-RTK and their true coordinates were obtained, the WGS_1984_UTM projection coordinate system is selected, in which 4 control points are used for absolute positioning for spatial reference accuracy of UAV product data, and 3 control points are used as checkpoints for accuracy verification of UAV product data.

Table 1. The specific specifications of the DJI Phantom 4 Pro aircraft and camera.

Aircraft	Take-off weight	Wheelbase	Maximum flight time	Maximum take-off altitude	Maximum withstand wind speed	Control accuracy
	1375g	350mm	30min	6000 m	10 m/s	±0.03°
Camera	Effective pixel	Image sensor	Maximum photo resolution
	20 million	1 inch CMOS	5472 × 3648

Clear weather and no cloud cover on the day of data collection, with surface information unaffected by atmospheric factors. The trajectory of the UVA was predetermined; the images were set at 85% heading overlap and 75% side overlap, and the flight altitude was 150 m. A total of 324 images were acquired, including red, green and blue bands. Ground Sampling Distance (GSD) refers to the actual distance from the ground corresponding to a single pixel on the image. The higher the value is, the better the acquisition of landslide morphology information is. The GSD of this study is 5.75 cm.

2.3. Construction of DEM and DOM

Based on UAV images, the DEM and DOM of the study area were constructed by using Pix4D software, through aerial triangulation, multi-view image dense matching, irregular triangulation network construction, etc. According to the accuracy verification of the previously laid control points, the Root Mean Square Error (RMSE) was used as the evaluation standard, and the mean RMSE of the control points was 0.027 m, where the RMSE of x was 0.014 m, the RMSE of y was 0.011 m and the RMSE of z was 0.060 m. The DOM and DEM have high resolution and positional accuracy, providing a fundamental data source for the extraction of landslide morphology.

3. Methodology

The main purpose of this study is to extract landslide morphology by using post-disaster DOM and DEM. The contents include the following four aspects: (1) Data acquisition and processing; (2) Construction of landslide morphology hierarchical extraction rule sets; (3) Selection of optimal elevation thresholds for landslide morphology by TP-DSM. (4) Evaluation and analysis of the accuracy of the results. The technical process of this paper is shown in Figure 2, where the red dotted line boxes are the key steps in landslide morphology extraction.

Figure 2. Technical process.

3.1. Hierarchical extraction rule sets of landslide morphology

Based on the object-oriented hierarchical analysis technology (Martha et al. 2011; Kurtz et al. 2014; Karantanellis et al. 2020), a hierarchical extraction rule sets of landslide morphology based on multi-feature information is constructed by making full use of the spectral, texture, geometry, topography and space feature information in DOM and DEM. Combined with the internal properties and spatial relationship features of the ground objects, and the target ground objects are eliminated and extracted layer by layer through multi-feature differences.

Firstly, through multi-level and multi-scale segmentation (Drǎguţ et al. 2010), the optimal scale for the three layers of landslide background layer, landslide itself and its morphology (Martha et al. 2011; Aksoy and Ercanoglu 2012) are selected, and the target ground objects are constructed for each layer. Afterwards, the attribute features of the actual target ground objects in the UAV images are analysed, the extraction rule sets of landslide morphology is established by combining the multi-feature information.

For the hierarchical extraction rule sets, the first layer is built on the macro level of the landslide area, using the spectral as the main feature to distinguish between vegetation covering a wide area and concentrated distribution and other ground objects. The second layer is built on the non-vegetation level, and the bare land, road, building and landslide are distinguished by multi-feature information such as spectral, texture, geometry and space. The third layer is built on the landslide level, with topography as the main feature, supplemented by spatial features. The TP-DSM is used to determine the thresholds, and to achieve the extraction of landslide morphology such as source area, circulation area, accumulation area and the cliff.

Figure 3 shows the task and processing process at different layers. The DOM and DEM of the study area are input, and the final three segmentation/extraction results are output after multi-scale segmentation and combine with multi-feature information rule sets extraction. In each layer, the output objects will be used as input objects of the next layer (Kurtz et al. 2014; Karantanellis et al. 2020).

Figure 3. Schematic diagram of hierarchical extraction rule sets of landslide morphology.

3.2. Selection of landslide morphology thresholds based on TP-DSM

The interior of the landslide is mostly bare rock and soil, with similar spectral and texture features, making it is difficult to distinguish the landslide morphology. However, there are more obvious differences in topography features of them. Therefore, the optimal thresholds should be determined according to the topography features to realize the extraction of different landslide morphology (Telbisz et al. 2013; Hergarten et al. 2014; Li et al. 2016).

Typically, the landslide morphology is shown in Figure 4: 1) Source area: The height difference is large, and massive rock and soil have depression after sliding. 2) Circulation area: The height difference is relatively large, and the overall shape is straight. 3) Accumulation area: The height difference is small and gentle, and the material accumulation is convex terrain. 4) The cliff: It can be divided into main scarp and flank. The main scarp is basically vertical to the Unshedded Mountain, with the sharp drop in elevation and large height difference. The main direction of the flank is perpendicular to the main scarp.

Figure 4. The landslide morphology diagram is modified according to Highland and Bobrowsky (2008).

The morphological characteristics and spatial distribution patterns of the above landslide are clearly represented with DEM. Based on this, TP-DSM is proposed in this paper for extracting landslide morphology.

The main idea of TP-DSM is that the spatial and topographic differences of landslide morphology are utilized. By obtaining the relevant elevation information of the specific buffer, the macroscopic topography in the landslide is intuitively reflected. Then the variation trend of surface morphology within the landslide is judged, and the optimal elevation thresholds of different morphology are determined. The TP-DSM is applied to choose the optimal elevation thresholds, by the following process:

1. Definition of the swath profile

   The swath profile is a buffer with a certain range, along the sliding direction of the landslide and through the main landslide morphology (Telbisz et al. 2013). Figure 5 shows the location of the proposed swath profile.

2. Elevations statistics of equal-interval zoning

   Within the swath profile, elevations are counted by equal-interval zoning. As shown in Figure 6, the swath is divided into several grids with equal area, denoted as S₁, S₂, ……, S_i, and generates the midpoint of each grid as M₁, M₂, M₃, ……, M_i, is used to record the position on the swath profile.

   For each grid S_i, statistical parameters of elevation values are calculated (including the maximum value H_i-max, minimum H_i-min, and mean H_i-mean), and a mapping relationship is established between the elevation value and the corresponding midpoint M_i. Finally, the elevation curves of the landslide swath profile are drawn separately: L_max, L_min and L_mean.

3. Calculate the elevation change rate

   Different landslide morphology has specific topography features, which will be represented by the elevation change rate to highlight, specifically subdivided into the first-order derivative and the second-order derivative of elevation. Let L′ be the first-order derivative of elevation, and Lʺ be the second-order derivative of elevation, L′ and Lʺ of L_max, L_min and L_mean, respectively, are given in Eqs. (1)–(6) as follows:

(1) $$L^{\prime }\max =\frac{d{H}_{i\text{-}\max }}{d{x}_i}$$(1) (2) $$L^{\prime }\min =\frac{d{H}_{i\text{-}\min }}{d{x}_i}$$(2) (3) $$L^{\prime }m\mathit{ean}=\frac{d{H}_{i-\mathit{mean}}}{d{x}_i}$$(3) (4) $$L^{″}\max =\frac{d^2{H}_{i\text{-}\max }}{d{x}_i{}^2}$$(4) (5) $$L^{″}\min =\frac{d^2{H}_{i\text{-}\min }}{d{x}_i{}^2}$$(5) (6) $$L^{″}m\mathit{ean}=\frac{d^2{H}_{i-\mathit{mean}}}{d{x}_i{}^2}$$(6)

In the equations, L′_max, L′_min, L′_mean are the first-order derivative curves of elevation curves L_max, L_min, L_mean respectively. Lʺ_max, Lʺ_min, Lʺ_mean are the second-order derivative curves of elevation curves L_max, L_min, L_mean respectively. The fluctuation tendency of the elevation curves within the swath profile is reflected, and the degree of topographic variation of landslide morphology is quantified. x_i is the distance from the starting point to point M_i on the swath profile.

• Determining landslide morphology extraction thresholds

  Elevation change rates (L′ and Lʺ) were used to quantify the difference of elevation change and concave-convex trend of different landslide morphology. The L′_i of point M_i represents the trend of elevation change within a certain range, and the relationship between L′_i and Lʺ_i is used to represent the convex (concave) characteristics (Shi et al. 2018). The specific description is shown in Eq. (7)(7) $$\{\begin{array}{c}L\prime i>0\text{and}{L}^{″}i>0,\text{convex terrain}\\ L\prime i<0\text{and}{L}^{″}i<0,\text{convex terrain}\\ L\prime i>0\text{and}{L}^{″}i<0,\text{concave terrain}\\ L\prime i<0\text{and}{L}^{″}i>0,\text{concave terrain}\end{array}$$(7) :

(7) $$\{\begin{array}{c}L\prime i>0\text{and}{L}^{″}i>0,\text{convex terrain}\\ L\prime i<0\text{and}{L}^{″}i<0,\text{convex terrain}\\ L\prime i>0\text{and}{L}^{″}i<0,\text{concave terrain}\\ L\prime i<0\text{and}{L}^{″}i>0,\text{concave terrain}\end{array}$$(7)

In the equation, L′_i is the value of point M_i on the first-order derivative curve of elevation, Lʺ_i is the value of the second-order derivative curve of elevation at point M_i.

Main scarp and source area: There is a sharp drop in elevation at the position where the rock and soil mass begin to fall off, and the elevation change has an accelerating trend, the elevation change rate is large, which is expressed as a large absolute value or a local extreme value of L′_i at M_i within the region. Due to the sliding of rock and soil mass, the terrain is mainly in a concave shape. It is worth noting that the fluctuation of L′ and Lʺ in the main scarp tends to be consistent (no mutation). In contrast, the rock and soil mass in the source area slides down irregularly, and the local concave and convex trend is changed, so there are local mutations in L′ and Lʺ.

Circulation area: The topography is gentler than that of the source area and steeper than that of the accumulation area. There are few deposits remaining in the circulation area of small landslides, which show that L′ and Lʺ fluctuate steadily and tend to be consistent. However, in large or medium-sized landslides, there are some debris in this area, which leads to the roughness of the surface, so there are local mutations in L′ and Lʺ.

Accumulation area: The topography is relatively flat and the elevation change rate is relatively slow, so the absolute value of L′_i at M_i is small and there is a local minimum value. The area is continuously receiving input materials and presents a convex terrain.

Based on the above rules, a semi-automatic threshold selection method is used to determine the elevation range of landslide morphology. l_i is noted as the separation line between the elevation ranges. Combined with the actual situation, the M_i on l_i is selected as the optimal threshold point of landslide morphology. The elevation values H_i-max, H_i-min and H_i-mean corresponding to this point are potential thresholds. H_i-mean is usually the optimal threshold.

• Supplementing and mapping landslide morphology

  Local details are further corrected according to the actual situation using space feature information. For example, the landslide flank can be judged and supplemented by the spatial position relationship, according to its position on both sides of the landslide and roughly perpendicular to the main scarp.

Figure 5. Schematic location of swath profile: (a) the selected location of the swath profile, which takes into account the elevation information of the various morphology within the landslide, (b) the swath profile mapped on the DEM of the real landslide, reclassified the DEM using the natural breaks method, which shown the swath profile spans different elevation ranges.

Figure 6. Statistical elevations in equal-interval zones.

4. Results

4.1. Extraction results of landslide morphology

The results were divided into two parts: (1) The result of optimal elevation thresholds selection for different landslide morphology by TP-DSM; (2) The result of hierarchical extraction rule sets based on multi-feature information using OOA. The thresholds selected will be used to construct the rule sets.

4.1.1. Results of optimal thresholds selection based on TP-DSM

Based on the DEM of the landslide, the TP-DSM was used, the swath profile parallel to the sliding direction was drawn, which was 66 m long and 3 m wide. Figure 7 shown the swath profile mapped. 22 3 × 3 m grids were generated by equal-interval zoning, and the midpoint of each grid was obtained, the H_i-max, H_i-min and H_i-mean in each grid were counted to generate L_max, L_min and L_mean (Figure 8(a)). M₀ and M₂₁ were the starting point and the end point of the swath profile respectively, according to the elevation of them, the approximate sliding direction of the landslide was judged as EN-WS.

Figure 7. Zoning strategy for swath profile.

Figure 8. TP-DSM calculation results: (a) elevation curve, (b) the first-order derivative curves of elevation, (c) the second-order derivative curves of elevation, (d) the location of the optimal threshold point for landslide morphology in the actual landslide.

Based on the Eqs. (1)–(6), the first-order derivative curves of elevation L′_max, L′_min, L′_mean (Figure 8(b)) and the second-order derivative curves of elevation Lʺ_max, Lʺ_min, Lʺ_mean (Figure 8(c)) were calculated.

The values of the first-order derivative of elevation were all greater than 0. Along the sliding direction, the first-order derivative of elevation rises rapidly in a short distance and the value of L′_max reached the peak near the distance of 55 m, which was considered in the range of the landslide main scarp. Meanwhile, the values of the second-order derivative of elevation in the same distance were roughly from negative to positive, showing a concave terrain, so l_c was set to divide the main scarp. The overall fluctuation range of L′_max, L′_min and L′_mean between 30 and 50 m was small but irregular, combined with Lʺ_max, Lʺ_min and Lʺ_mean, it can be seen that the area was concave; However, there are some mutation values, that is, the local concave and convex trend was changed, it was considered to be in the source area, divided by l_b and l_c. The distance between 10 and 30 m, the overall fluctuation range of the elevation derivative curve was large but consistent, considered as circulation area, divided by l_a and l_b. The distance between 0 and 10 m showed a decreasing trend in L′_max, L′_min and L′_mean, indicating that the topography in the area gradually gentle; At the same time the values of Lʺ_max, Lʺ_min and Lʺ_mean were all greater than 0, the terrain was roughly convex and was judged to be the accumulation area, divided by l_a. The value of the first-order derivative of elevation of L_max at the three points of M_a, M_b and M_c was still higher than that of L_min and L_mean, while the value of the second-order derivative of elevation became smaller, indicating that the concave and convex trend changed.

In summary, the dark green dotted lines l_a, l_b and l_c were used as separation lines for the elevation range of each morphology, The M_a, M_b and M_c circled by the green rectangles were the points for the optimal thresholds, the corresponding H_a-mean, H_b-mean and H_c-mean were used as references for the optimal thresholds of landslide morphology. The main scarp, source area, circulation area and accumulation area of landslide near 2272.17, 2279.14 and 2289.45 m were considered. The position of the optimal threshold points in the actual landslide was shown in Figure 8(d). On this basis, combined with the specific characteristics of landslide, the slope and spatial features were used to further extract the landslide flank.

4.1.2. Hierarchical extraction results of landslide morphology

Based on OOA, during segmentation, three different segmentation parameters were selected to represent three interrelated object hierarchy levels. The scale factors of Level 1, Level 2 and Level 3 were gradually refined. Level 1 was the parent of Level 2 and Level 3. Level 3 and Level 2 also had a sub-relationship. The shape factors and compactness factors were combined with the actual situation of the target ground objects at each level, and adjusted by the trial and error until the most suitable.

During extraction, the Visible-Band Difference Vegetation Index (VDVI) (Wang et al. 2015) was chosen for distinguish vegetation at Level 1. At Level 2, the bare land is mostly soil, reddish brown or yellow, scattered between vegetation. Non-vegetation was first distinguished using VDVI, and then identified to bare land based on elevation and spatial distribution. The road is mostly concrete or asphalt, and the hue is different from that of bare land, which can be distinguished by spectral features. Meanwhile, the slope of the road generally changes little, showing a regular shape of slender and approximate rectangle, which can be identified by geometry and topography features. After eliminating non-landslides one by one, the real landslide was finally extracted. At Level 3, the thresholds of landslide morphology were determined by TP-DSM, and the rule sets were constructed. Then, source area, circulation area, accumulation area and main scarp were extracted. Next, spatial position relationship was adopted to further distinguish the flank, which was located on the right side of the landslide.

Finally, the hierarchical extraction rule sets of landslide morphology were constructed, as shown in Table 2.

Table 2. Landslide morphology extraction rule sets.

Segmentation level	Target ground object	Segmentation parameter	Extraction rule
Level 1	Vegetation	Scale 71	VDVI^a ≥ 0.0418
		Shape 0.1
		Compactness 0.5
Level 2	Bare land	Scale 62 Shape 0.3 Compactness 0.5	VDVI < 0.0418, Y Max^b ≤ 1340 Pxl or Y Max > 3060 Pxl, X Center^b> 3330 Pxl, Elevation < 2268 m or Elevation > 2312 m
	Road		B_Mean^c ≥ 132, Length/Width^d > 3.345, Slope < 27°
Level 3	Main scarp	Scale 56 Shape 0.3 Compactness 0.6	Elevation ≥ 2289.45 m, Slope ≥ 30°, B_Mean > 100
	Source area		Elevation > 2279.14 m
	Circulation area		2272.17 m ≤ Elevation ≤ 2279.14 m
	Accumulation area		Elevation < 2272.17 m
	Flank		X Center < 3400 Pxl, X Min^b ≥ 2270 Pxl

^aVDVI is a vegetation index based on three visible light bands (red, green and blue), with a value range of [−1, 1], which is used to distinguish the vegetation in UVA images.

^bY Max is the maximum y-orientation of the segmented object, X Center is the x-orientation of the centre point of the segmented object, and X Min is the minimum X-orientation of the segmented object. All three are used to represent the spatial position relationship of the landslide and its morphology.

^cB_Mean is the average values of the pixel values that make up the blue band, used to distinguish eroded surfaces in landslides or ground objects with specific spectral features.

^dLength/Width is the length of an object divided by its width, used to distinguish narrow ground objects or cracks similar to source area.

The areas of statistical landslide and its morphology were shown in Table 3. Figure 9 was shown that target landslide was mainly located on mountains with relatively broken surface and low vegetation coverage, and near road. There was partial misclassification between bare land and low vegetation coverage area and landslide, and the small area of bare land between vegetation was prone to be missed. In general, the ground objects extracted by the method in this paper were consistent with the actual distribution, and the boundary of all kinds of ground objects was clear and the shape was complete. There was no case where the extraction result was too broken to affect the distribution.

Figure 9. Extraction results of various ground objects and landslide morphology in the study area: (a) DOM, (b) extraction results of landslide, (c) extraction results of landslide morphology, (d) extraction results of various ground objects.

Table 3. Landslide morphology area statistics.

	Total area of landslide/m^2	Source area /m^2	Circulation area /m^2	Accumulation area /m^2	Main scarp /m^2	Flank /m^2
Luquan	5332.03	1543.44	1547.33	1084.98	489.18	667.10

4.2. Accuracy evaluation

The extraction results were evaluated by confusion matrix, with Producer Accuracy (PA), User Accuracy (UA), Overall Accuracy (OA) and Kappa coefficient as specific evaluation indexes (Liu et al. 2007). OA and Kappa coefficient describe the overall accuracy of landslide morphology extraction results in UAV images, representing the proportion of correctly extracted objects; PA and UA describe the extraction accuracy of a certain type of ground object, specifically, PA represents the ratio between the correct extraction result of a certain type of ground object and the real condition of the ground object, and UA represents the ratio between the correct extraction result of a certain type of ground object and the extraction result of this type ground object. The value range of the four evaluation indexes is [0, 1], and the larger the value, the higher the extraction accuracy.

According to the confusion matrix, the expressions of PA, UA, OA and Kappa coefficient are shown in Eqs. (8)–(11): (8) $$PA=\frac{x_{jj}}{x_{+j}}$$(8) (9) $$UA=\frac{x_{ii}}{N}$$(9) (10) $$OA=\frac{x_{+j}}{N}$$(10) (11) $$\mathit{Kappa}=\frac{OA-p_e}{1-p_e}$$(11)

In the equations, x_ij is the number of objects belonging to class j in the reference sample data while a certain type of ground object in the inspection sample is classified into class i, N is the total number of sample objects, P_e is the proportion of objects that the accidental opportunity leads to the consistency of the correct extraction result with the real situation.

Ensure that all kinds of ground object inspection data was random and evenly distributed. 265 objects were taken as inspection data in the study area, of which 130 were landslide objects (30 source area, 30 circulation area, 20 accumulation area, 30 main scarp and 30 flank). The OA and Kappa coefficient of all kinds of ground objects were 95.75% and 0.93, and the PA and UA of landslide were 98.46% and 96.24% respectively.

In the landslide morphology (Table 4), the PA of the landslide source area was lower than the UA, suggesting that there was a partial miss of extraction; the UA is lower than PA in circulation area, accumulation area and main scarp, reflecting the presence of some wrong extraction; The PA and UA of the flank were both low, and there were missing and wrong parts. The reason is that the elevation range of the flank was wide, which made it easy to be confused with the source area, the circulation area, the accumulation area and the main scarp. However, the OA of landslide morphology in the study area was still 89.58%, with a Kappa coefficient of 0.88. The evaluation results were shown that the thresholds selected by the TP-DSM can effectively extract the landslide morphology. Although there are still some over-extraction or under-extraction, the overall extraction accuracy is high.

Table 4. Extraction accuracy of landslide morphology.

	Source area	Circulation area	Accumulation area	Main scarp	Flank
PA/%	60.00	100.00	85.00	96.67	90.00
UA/%	94.74	78.95	85.00	90.63	72.00
OA/%			89.58
Kappa			0.88

5. Discussion

5.1. Hierarchical extraction of landslide morphology

The characteristics of landslide background, landslide and its morphology are diverse spatial distributions, complex structures and significant scale differences (Aksoy and Ercanoglu 2012). Given the challenge of accurately extracting landslide morphology in such complex environments using a single scale, in this study a hierarchically organized rule sets are constructed to extract target ground objects layer by layer (Akcay and Aksoy 2008; Kurtz et al. 2014). In the landslide area with dense vegetation, we are committed to eliminating the vegetation with wide coverage and significant spectral difference from the landslide, so as to obtain non-vegetation objects such as road, bare land and landslide. Of course, the spectral features of some non-vegetation objects such as road and bare land are extremely similar to those of the landslide. To solve this problem, texture, geometry, space, topography and other multi-feature information are combined in this paper. On the basis of the previous segmentation/extraction, the object is further detailed, after eliminating the non-landslide objects one by one; the spatial range of the landslide is obtained. Subsequently, TP-DSM was proposed to extract landslide morphology based on its topography features.

Compared with the results of visual interpretation of landslide (Figure 10), there is a phenomenon that partial omission of the source area or confusion with the main scarp and flank, and partial over-extraction of bare land into landslide. The main reason is that different landslide morphology on the image may have similarities such as hue and elevation, or differences in surface coverage and light intensity in the same morphology. But overall landslide morphology can be identified more completely and has high consistency with the actual distribution, which is a valuable useful supplement to landslide inventory mapping (Guzzetti et al. 2012).

Figure 10. Comparison between visual interpretation and extraction results of the proposed method.

5.2. Influence on the selection of landslide morphological thresholds

It is a challenging process to select the thresholds for landslide morphology. In previous topographic profile method, multi-period DEMs were usually used to draw linear profiles and compared elevation changes to determine the thresholds (Wang et al. 2021). However, most landslides are sudden, which makes it difficult to obtain high-precision DEM before the landslide occurred, limiting the use of this method. Therefore, only single-period DEM can be used for profile analysis. But the profile elevation curves of landslides tend to have relatively gentle topographic fluctuation, leads to trouble finding reasonable thresholds for delineating their internal morphology. In order to enhance the extent of elevation curve fluctuation changes and identify more appropriate thresholds, the TP-DSM method was adopted to draw the elevation change rate curve in the paper, elevation range of different landslide morphology are divided to determine the optimal thresholds.

A linear profile was drawn along the direction of landslide movement (Figure 11(a)), cracks, depressions or dumped vegetation inside the landslide may cause abnormal values in the elevation curve (Figure 11(b)). These values are easily misinterpreted as optimal thresholds. In addition, as shown in Figure 12(a) and (b), after calculating the elevation change rate (the first-order derivative and the second-order derivative of elevation) of the linear profile, the elevation change rate curve has no obvious threshold characteristics, and there are a lot of abnormal changes, which makes it impossible to judge the landslide morphology thresholds. The range of swath profile was defined by the proposed TP-DSM method, based on landslide size information (widths and lengths) in the study area. A buffer was mapped (Figure 11(a)), dividing it into multiple equal-interval grids; the elevation values (maximum, minimum, average) in each grid were counted, and the corresponding elevation curves (Figure 11(c)) were drawn. Topography features such as the highest, lowest and average elevation in a region are reflected in the elevation curves, which can either described the local micro-topographic changes or the macro-geomorphic features of the region. The elevation change rate was calculated, and there were clear threshold characteristics to reflect the elevation change and concave-convex trend of landslide morphology. The landslide morphology thresholds can be easy to determine. In this way, the influence of local topographic changes in complex landslide can be overcome to a certain extent, and the topographic fluctuation characteristics of the landslide itself are more comprehensively characterized.

Figure 11. Comparison between linear profile and swath profile: (a) the profiles were drawn, where subfigures a1, a2 and a3 shown the area with dumped vegetation or cracks in the landslide, (b) the elevation curve of the linear profile, (c) the elevation curve of the swath profile.

Figure 12. Elevation change rate curve of the linear profile was obtained: (a) the first-order derivative curves of elevation, (b) the second-order derivative curves of elevation.

In order to verify the feasibility of the TP-DSM method, the landslide morphology extraction results were compared with those obtained by other methods. Based on the available landslide spatial results, the K-MEANS method (Kurtz et al. 2014) was chosen. When four clusters of landslide morphology were extracted as main scarp, source area, circulation area and accumulation area (Figure 13(a)), the average elevation values of each cluster centre were 2294.83, 2282.72, 2278.08, and 2271.96 m, respectively, which were consistent with the proposed TP-DSM method to delineate the elevation ranges of these four morphologies. When five clusters were extracted (Figure 13(b)), the results showed that the K-MEANS appears to be incorrectly clustered, making it difficult to further differentiate the flank. Contrasting visual interpretation results (Figure 13), it can be seen that the approximate range of different landslide morphology can be distinguished by the K-MEANS based on elevation, but it was hard to effectively extract the morphology with special spatial distribution. Conversely, the results of the TP-DSM method proposed in the paper were more consistent with the actual, has the flexibility to adapt better to different landslide scenarios and their morphological compositions.

Figure 13. Comparison of landslide morphology extraction results by different methods: (a) K-MEANS extraction four types of morphology comparing visual interpretation; (b) K-MEANS extraction five types of morphology comparing visual interpretation; (c) TP-DSM method selecting the thresholds construction rule to extract five types of morphology comparing visual interpretation.

Finally, in the current method, the landslide flank are also not purely derived from geomorphologic segment features (Van Den Eeckhaut et al. 2012), because for some landslides, the traces are blurred (For example, there is no clear boundary between part of the landslide and the adjacent unaffected slopes). It still needs to be combined with spatial information to extract. Therefore, it is necessary to expand the proposed method according to the topography feature of the flank.

5.3. Application of landslide morphology extraction in other areas

The method was tested in a landslide area in Qiaojia County, Yunnan Province (Figure 14). The total area of the extracted landslide is 3886.17 m², including 862.34 m² of source area, 1969.88 m² of circulation area and 1053.95 m² of accumulation area. In addition, OA of 90.32% and Kappa coefficient of 0.87 were obtained for all types of ground objects (vegetation, bare land, building, road, landslide), with PA of 81.90% and UA of 98.85% for landslide. The accuracy evaluation results of landslide morphology extraction were shown in Table 5, with OA of 87.46% and Kappa coefficient of 85.14. In this area, the similarity of spectral and texture features between landslide and bare land was high, and the transition between the two was not obvious, which lead to some landslide objects being missed or wrong extracted as bare land, so that the PA of the landslide source area and the UA of the circulation area were poor. In contrast, due to the restoration of vegetation in the accumulation area, a number of accumulation area were extracted as vegetation, and there were a lot of omissions, resulting in low PA. In general, the landslide morphology can be effectively extracted by the proposed method, and has excellent applicability.

Figure 14. Extraction results of various ground objects and landslide morphology in Qiaojia: (a) DOM, (b) extraction results of landslide, (c) extraction results of landslide morphology, (d) extraction results of various ground objects.

Table 5. Extraction accuracy of landslide morphology in Qiaojia.

	Source area	Circulation area	Accumulation area
PA/%	82.86	74.29	65.71
UA/%	93.55	81.25	95.83
OA/%	87.46
Kappa	85.14

6. Conclusions

Luquan County, Yunnan Province, China, was chosen as the study area, and UVA images were used as the data source. The TP-DSM method was proposed to screen the optimal elevation thresholds for landslide morphology extraction. From the perspective of method, the multi-feature information of UAV images was integrated to establish a hierarchical extraction rule sets, which had optimized the extraction method in the complex landslide environment to a certain extent. That reduces the problems caused by missing some details of data. Topography features of landslide morphology were finely analyzed by TP-DSM. On the basis of the original elevation curve, the elevation change rate was calculated, that was, the first-order derivative and the second-order derivative of elevation. The differences in elevation changes and concave-convex trends among various landslide morphologies were taken into account, and reasonable thresholds were eventually selected. From the perspective of application, the TP-DSM method was used to achieve effective extraction results in two sample areas and had favourable applicability.

Author contribution

Conceptualization: [Jia Li]; methodology: [Yujie Zhang]; formal analysis and investigation: [Ping Duan], [Jiajia Liu]; writing – original draft preparation: [Yujie Zhang]; writing – review and editing: [Jia Li]; funding acquisition: [Jia Li]; resources: [Jiajia Liu], [Wenbin Xie]; supervision: [Jia Li]; accuracy evaluation: [Yujie Zhang], [Wenbin Xie].

Data availability statement (DAS)

The data that support the findings of this study are available from the corresponding author Jia Li upon reasonable request.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This research was funded by the National Natural Science Foundation of China (NO.41961061) and the Yunnan Fundamental Research Projects (NO. 202301AT070061) and 'Revitalizing Yunnan Talents Support Program' project funding support(No.YNWR-QNBJ-2020-048, NO.YNWR-QNBJ-2020-103) and Yunnan Academician and Expert Workstation (No.2017IC063) and Yunnan Provincial Basic Research Project-Key Project (202201AS070024).