Landscape Fragmentation and Spatial Autocorrelation of a Typical Watershed in the Wenchuan Earthquake-Affected Area—A Case Study in the Longxi River Basin

Abstract

The 2008 8.0 Ms Wenchuan earthquake caused serious ecological degradation in the northwest of China’s Sichuan Province. Elucidating landscape fragmentation and spatial pattern of post-earthquake landscape is of great interest in order to improve ecological restoration and predict the spatial distribution of damaged ecosystems in earthquake-affected areas. We used four real-time remote sensing images to analyze the landscape pattern characteristics and spatial autocorrelation of the Longxi River Basin before and after the Wenchuan earthquake during the period from 2005 to 2015. In the study period, the degree of landscape fragmentation increased 1 year after the earthquake in 2009 but decreased in 2011 and 2015. The spatial distribution of forest, farmland, and shrub-grassland had significantly positive spatial correlation and the characteristics of spatial aggregation from 2005 to 2015. Construction land had no significant spatial correlation before the earthquake in 2005, but significantly positive spatial correlation after the earthquake, while traffic land had no significant spatial correlation across the study period. Unlike the other landscape types, geological disasters changed from a significantly negative spatial correlation before the earthquake to a significantly positive spatial correlation after the earthquake. However, the positive autocorrelation of all landscape types decreased with the increase of space distance, but in different distance-decay rates. The High-high spatial aggregation areas of geological disasters and construction land clustered gradually upstream of the basin and distributed in Longchi town, respectively; those of farmland distributed in the southeast of the basin increased in 2009 but then decreased, while those of forests had the opposite pattern as the dominant landscape type after the earthquake. Our results elucidated the spatial structure and distribution features of the Longxi River Basin to give a theoretical foundation for assessing the effects of ecological restoration and reconstruction management in earthquake-affected areas.

1. Introduction

Major earthquakes can produce a large amount of geo-hazards from available land resources. This change was always accompanied by heavy economic losses [1], serious ecological degradation [2], and landscape fragmentation [3]. To improve the restoration process in earthquake-affected areas, many researchers used the integrated techniques of 3S technology combined with geography and landscape ecology theories to study the spatiotemporal dynamic characteristics of land use affected by secondary disasters induced by earthquakes [4,5,6,7,8]. They also reported the automatic detection of earthquake-triggered landscape changes [9], landscape evolution [10], and changes in ecosystem services in the earthquake-stricken area [11]. For example, the Jiuzhaigou earthquake occurred on 8 August 2017, with a magnitude of 7.0, which caused the overall landscape grade of Jiuzhaigou Nature Reserve to decrease by 15% (Luo et al., 2023) [12]. In the 2004 Mw 6.6 Chuetsu earthquake-affected areas in Japan, natural succession improved landscape structure, resulting in good vegetation recovery (85.5%) by 2021 (Xiang et al., 2023) [13]. Their works help us predict geomorphological evolution, landscape patterns, and ecosystem functioning and processes [14,15,16,17]. However, little is known about the consequence of landscape fragmentation caused by the earthquake on a small scale, leaving a great need for an ecological restoration process and landscape pattern optimization after an earthquake.

In order to describe the dynamic of landscape patterns and its control factors and predict possibilities of landscape change trajectory, researchers set up a dynamic model to establish robust diagnostics, prediction, and simulation through mathematical models with regard to logical systems and flow interactions among various spatial elements [18]. At present, cellular automata (CA) modeling, based on multi-criteria evaluation (MCE) and logistic regression (LR), is used widely to investigate landscape dynamics due to its ability to capture dynamic information and predict future scenarios [19,20]. Classic statistical methods including MCE and LR assumed individual observations to be statistically independent, while spatial landscape data have the tendency to be dependent, which was defined as spatial autocorrelation—a phenomenon that random variables affected by geographic proximity would have more similar or less similar values over distance than randomly associated pairs of observations [21,22].

Spatial autocorrelation can be used to quantify the spatial dependency in univariate data, including Moran’s index (Moran, 1950) and Geary’s coefficient (Geary, 1954), and multivariate data, such as a Mantel correlogram (Legendre and Legendre, 1998), across a geo-referenced space. Though some researchers debate the methodological disadvantage of the spatial autocorrelation analysis [21], it was still applied to describe the spatial variability of land use or landscape [21,23,24,25]. For example, to determine the law of landscape pattern evolution, scholars analyzed the geographical correlation of the landscape index [26], the change in land cover characteristics [27], the dynamic evolution characteristics of habitat quality [28], and the correlation features with interference elements [29] using spatial autocorrelation analysis. However, previous studies mainly concerned the spatial autocorrelation of landscape pattern change in a large-scale river basin but did not consider how landscape patterns respond to earthquakes in a small watershed.

The Longxi River Basin located in Dujiangyan City, Sichuan Province is considered a typical basin in Wenchuan earthquake-affected areas [10,30]. The Ms 8.0 Wenchuan earthquake caused frequent secondary disasters, including debris flows, landslides, collapses, and floods, as well as subsequent deterioration in the ecological environment of the Longxi River Basin [31]. The changes caused by secondary disasters in the geomorphology and river system induced a large area of vegetation destruction and reshaped land use patterns [32]. Our previous study showed that unutilized land induced by secondary disasters replaced grassland as the second dominant land type, accounting for 15.99% of the Longxi River Basin’s area in 2009 [8]. However, to our knowledge, the spatial autocorrelation analysis in a typical basin after the earthquake is still poorly understood and undocumented. Moreover, the Longxi River Basin belongs to the Wolong Nature Reserve, which supports approximately 10% of the total wild giant panda population across the globe [3] and plays an important role in the safe operation of the Zipingpu Reservoir and water supply for the Chengdu Plain [31]. Therefore, it is necessary to elucidate landscape fragmentation and the spatial pattern of post-earthquake landscapes.

In this study, we collected four different high-resolution remote-sensing satellite images of interpreted land use types in the Longxi River Basin. The aim of this study was to determine how changes in the spatial structure and distribution features were affected by the earthquake and subsequent recovery. Here, we hypothesized that natural recovery could reduce the degree of landscape fragmentation and change spatial autocorrelation characteristics of landscape types after the earthquake. Our main objectives were to (1) analyze the characteristics of the basin’s landscape fragmentation before and after the earthquake, (2) determine the change of spatial correlation of landscape types under ecological recovery after the earthquake, and (3) demonstrate spatial aggregation characteristics of landscape types. This study could provide a theoretical foundation and scientific recommendations for the restoration and reconstruction of fragile habitats in the Longxi River Basin.

2. Materials and Methods

2.1. Study Area

The Longxi River Basin is located on the upper reach of the Minjiang River along the eastern margin of the Qinghai–Tibet Plateau in the northwest of Dujiangyan City, Sichuan Province, China (103°30′4.98″–103°36′2.08″ E, 31°1′28.06″–31°10′49.53″ N). Its elevation ranges from 822 m to 3291 m a.s.l. with a maximum east–west distance of 6.88 km and north–south distance of 17.32 km, and an area of 79.24 km². The topography of the basin is dominated by moderate and high mountains with an intensive surface incision from north to south, which leads to a “V”-shaped canyon [30]. According to WRB classification (Hempel et al.) [33], the basin has three soil types: ali-udic argosols are found in the zone at an elevation of 822 to 1600 m a.s.l., haplic acrisols are found in the zone at an elevation of 1600 to 2200 m a.s.l., and haplic luvisols are found in the zone at an elevation of 2200 to 3291 m a.s.l. The climate is tropical humid monsoon with concentrated precipitation (approximately 1100 mm per year), which occurs between May and October (Ding et al.) [34]. The length of the basin’s main channel is 15.7 km, with an average slope gradient of >20°, and the gully gradient is between 287‰ and 626‰ [30]. Due to frequent floods and debris flows after the Wenchuan earthquake, the basin is one of the three major demonstration sites for typical debris flow monitoring and early warning in Sichuan Province (Figure 1 and Figure 2).

2.2. Data

The remote sensing images used in this study were SPOT-5 in 2009, Quick Bird in 2005 and 2011, and Worldview-2 in 2015, supplied by the Institute of Mountain Hazards and Environment, Chinese Academy of Sciences (CAS). The spectral resolution and information of the images are shown in

Table 1. Satellite imagery data information in the study area.

Satellite Imagery	Date	Intervals	Resolution/m	Coverage	Data Quality
QuickBird	26 June 2005	Before the earthquake	0.61	Entire study area	Few Clouds
SPOT-5	10 February 2009	1 year after the earthquake	2.50	Entire study area	Cloudless
QuickBird	26 April 2001	3 years after the earthquake	0.61	Entire study area	Few Clouds
Worldview-2	15 April 2015	7 years after the earthquake	0.46	Entire study area	Cloudless

. All images were covered less than 2% by clouds. We simulated blue band of the SPOT-5 image in trained fitting model in Erdas 9.2 to receive its natural color image [35]. Then, we used ENVI 5.3 version software to correct spatial incompatibility across four images on the basis of World Geodetic System (WGS) 1984 coordinate system. The second-order polynomial model was used on the basis of primary scale Chinese national topographic maps (1:50,000) to ensure the accuracy of all images’ geometric correction within only one pixel with 24 ground control points. The method of human–computer interactive interpretation was used to interpret land use types with the reference of field survey data and unmanned aerial vehicle (UAV) images in 2018 in Longxi River Basin. Moreover, Shuttle Radar Topographic Mission (SRTM) with 30 m × 30 m resolution was used to produce the Digital Elevation Model (DEM) and extract topographic features.

2.3. Methodology

2.3.1. Landscape Types Classification System

According to the classification system created by Jia et al. (2003) [36], we divided the land use types of Longxi River Basin into six categories as cultivated land, forest, construction land, water, shrub-grassland, and unutilized land (mainly including mountains damaged by natural disasters, bare rock mass, bare land, etc.). The process of land use category identification, field observation, pixel-based supervised image classification, and classification accuracy assessment could be seen in Tian et al. (2022) [8]. The overall classification accuracy and the Kappa coefficient were 95.2% and 0.896, respectively.

Based on the classification of land use types, we classified them into nine landscape types, including forest land, farmland, shrub-grassland, construction land, traffic land, bare land, gully channel, geological hazard, and water according to the standard created by You et al. (2006) [37]. This helped us to analyze the change in each landscape in detail.

2.3.2. Landscape Pattern Index

The landscape pattern index is a simple quantitative indicator and one of the important indicators that highly summarizes the composition of landscape structure and spatial pattern characteristics. Landscape indices can describe the landscape pattern, establish the connection between landscape structure and processes or phenomena, and better explain and understand landscape functions. In this study, we used Fragstats 4.2 to assess the spatial pattern of landscape types [38]. To describe the characteristics of landscape fragmentation, the following landscape metrics, including edge density (ED), Shannon diversity index (SHDI), and landscape division index (DIVISION), were chosen according to the method described in You et al. (2006) [37]. ED represents the edge length of a patch per hectare, which determines the density of the patch. SHDI is a sensitive measurement indicator for the uneven distribution of various patch types in a landscape. If SHDI = 0, it indicates that the entire landscape is composed of one patch. The larger value indicates that there are more patch types with the uniform distribution of each patch type. DIVISION refers to the degree of fragmentation in the distribution of individuals with different patch numbers in a certain landscape type. Their formulas are as follows:

$$\mathrm{ED}=\frac{{\displaystyle \sum _{\mathrm{k}=1}^{\mathrm{m}}}{\mathrm{e}}_{\mathrm{ik}}}{\mathrm{A}}(10000)$$

(1)

where e_ik is the total edge length of patch type i, and A is the total landscape area. A smaller ED value indicates that the patch type is only adjacent to a few other types, while the higher value indicates the higher degree of landscape fragmentation, more significant edge effect, and stronger openness.

$$\mathrm{SHDI}=-{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{P}_i\ln \left(P_i\right)$$

(2)

where P_i is the proportion of patch type i in the landscape. The larger index indicates the richer the types and distribution of various patches in the landscape.

$$DIVISION=1-\sum _{j=1}^n\left(\frac{a_j}{A}\right)$$

(3)

where a_j is the area of a certain type of patch j in the landscape type. The larger division index indicates the more severe the fragmentation of landscape types.

2.4. Landscape Pattern Spatial Autocorrelation

Moran’s I, Moran scatter plot, and Local Indicators of Spatial Association (LISA) were selected in this study to analyze and test spatial autocorrelation at the global and local levels using ArcGIS 10.2 and Geoda 1.6.7 software [39]. In detail, global and local Moran’s I can assess the degree of spatial association, which was visualized by Moran scatter plot and LISA to reveal the correlations and difference among the attributes of study units of the Longxi River Basin [40]. In this study, we used the grid analysis method to divide the Longxi River Basin into 374 grid cells of uniform size of 500 m × 500 m based on the principle of optimal sample square described by Fu et al. (2017) [41]. Each landscape type was superimposed with the grid cells for analysis except water because it occupies a concentrated area, resulting in insufficient patch numbers for autocorrelation analysis.

2.4.1. Global Spatial Autocorrelation Statistics

The global Moran’s I could reveal the spatial structure, distribution pattern, and their causes of regional variables [36]. Its formula is expressed as follows [42]:

$${\mathrm{Moran}}^{’}\mathrm{s} \mathrm{I}=\frac{\mathrm{n}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=\mathrm{j}}^{\mathrm{n}}}{\mathrm{W}}_{\mathrm{ij}}({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})({\mathrm{x}}_{\mathrm{j}}-\overline{\mathrm{x}})}{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\displaystyle \sum _{\mathrm{j}=1}^{\mathrm{n}}}{\mathrm{W}}_{\mathrm{ij}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}(\mathrm{i}\ne \mathrm{j})$$

(4)

where n is the total number of spatial units in the Longxi River Basin. x_i and x_j represent the attribute values of spatial elements of sites i and j, respectively. x is the average attribute values of spatial elements. W_ij is a binary symmetric spatial weight matrix that indicates the distance or closeness between sites i and j. Distance-based spatial weight data were created using the queen’s standard with the distances of 0.5, 1, 2, 3, 4, 5, 6, 8, and 10 km, respectively. The value of Moran’s I generally ranges from −1 to 1. On this basis, the global spatial autocorrelation features and Moran’s I distance effects of each landscape type within the Longxi River Basin were examined between 2005 and 2015. Low negative values represent a negative spatial autocorrelation showing the clustering of dissimilar value, while high positive values represent a positive spatial autocorrelation showing the clustering of similar value. The significance of Moran’s I is usually assessed using the Z-score [43], and a Z-score > 1.96 indicates the significance at the level of p < 0.05.

2.4.2. Local Spatial Autocorrelation

Global autocorrelation could not determine the specific position of clustering locations or dispersed places due to the deficiency in aberrant local spatial pattern [44]. Therefore, we used LISA and the Moran scatter plot outlined by Anselin (1995) to assess the local spatial autocorrelation of attribute values and to identify the spatial distribution of anomalous regions. The formula calculating the local Moran’s I is as follows [45]:

$${\mathrm{I}}_{\mathrm{i}}=\frac{{(\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}\underset{\mathrm{j}=1}{\sum ^{\mathrm{n}}}{\mathrm{W}}_{\mathrm{ij}}\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)$$

(5)

where x_i and x_j are the normalized cellular observations from samples i and j; the other parameters are the same as the global Moran’s I.

Moran scatter diagram shown as Moran scatter plot can be used to analyze the spatial agglomeration characteristics between the attribute value in an observation and the average of its neighbors [46,47]. Moran scatter plot yields five types of spatial clusters, including not significant, High-high (HH), Low-low (LL), Low-high (LH), and High-low (HL) [48]. The type of “High-high” means that the attribute values of a spatial unit and its neighbors are both high, while that of “Low-Low” has the opposite pattern. Moreover, the type “High-Low” means that the attribute value of a spatial unit is high but its neighbors have low values, while that of “Low-High” has the opposite pattern [44,45,46,47,48,49,50,51].

3. Results

3.1. The Changes in Landscape Types

According to the actual situation of the Longxi River Basin, we divided the basin into nine landscape types, including forest, farmland, shrub-grassland, construction land, traffic land, bare land, gully channel, geological hazards, and water. As seen in

Table 2. Changes in landscape area from 2005 to 2015.

Year		Forest	Farmland	Shrub-Grassland	Construction	Traffic Land	Bare Land	Gully Channel	Geological Hazards
					Land
2005	area/hm2	7253.63	136.43	336.98	41.01	22.62	62.17	51.30	6.18
	percentage/%	91.70	1.72	4.26	0.52	0.29	0.79	0.65	0.08
2009	area/hm2	6213.55	85.11	269.94	36.00	38.05	75.18	175.94	1016.76
	percentage/%	78.55	1.08	3.41	0.46	0.48	0.95	2.22	12.85
2011	area/hm2	6864.10	66.15	320.56	29.95	41.06	44.04	175.78	377.42
	percentage/%	86.68	0.84	4.05	0.38	0.52	0.56	2.22	4.77
2015	area/hm2	7186.13	49.67	220.70	30.57	48.64	37.51	165.97	179.25
	percentage/%	90.75	0.63	2.79	0.39	0.61	0.47	2.10	2.26

, forest dominated all landscape types but decreased after the earthquake in 2009 and then recovered from 2009 to 2015, while shrub-grassland had a similar pattern but decreased during the period from 2011 to 2015. The area of farmland continually decreased, while that of traffic land had an opposite pattern during the study period. Bare land, gully channel, and geological hazards had similar patterns, in that their areas increased after the earthquake in 2009 but decreased from 2009 to 2015.

3.2. Landscape Fragmentation in the Longxi River Basin

Three landscape pattern indices, including edge density (ED), Shannon diversity (SHDI), and landscape isolation index (DIVISION), were selected to quantify the landscape fragmentation in the Longxi River Basin.

3.2.1. The Change in ED

ED is used to reveal the degree which a landscape or type is segmented by boundaries, showing a direct reflection of the degree of landscape fragmentation.

The change in ED can be intuitively seen from Figure 3. High-value ED areas were distributed along the Longxi River channel in 2005, while they spread over the basin after the Wenchuan earthquake in 2009. After natural recovery, the high-value ED area gradually reduced in 2011 and 2015, but it was still higher than in 2005.

3.2.2. The Change in SHDI

SHDI is a measurement index based on information theory and is widely used in ecology, reflecting landscape heterogeneity. It is particularly sensitive to the uneven distribution of various types in the landscape, emphasizing the contribution of rare types to information. Therefore, it is also used as a sensitive indicator when comparing and analyzing the diversity and heterogeneity changes of different landscapes or different periods of the same landscape.

As seen in Figure 4, the changes in SHDI in the Longxi River Basin also showed a similar trend from 2005 to 2015. The SHDI value along both sides of the Longxi River main channel was comparatively greater than the other parts in 2005, while it clustered in the tributary of the Longxi River after the Wenchuan earthquake in 2009. After natural recovery, the high-value SHDI area also gradually reduced in 2011 and 2015, but it was still higher than in 2005.

3.2.3. The Change in DIVISION

The DIVISION reflects the separation degree of patches in the landscape, indicating the complex degree of the landscape.

As seen in Figure 5, the high-value DIVISION areas increased dramatically after the Wenchuan earthquake in 2009 compared with those in 2005. After natural recovery, the high-value DIVISION areas still clustered in the upper reach of the Longxi River during the period from 2011 to 2015. In general, the high-value DIVISION areas changed significantly on the slopes of both sides of the Longxi River gully and its tributaries across the study period.

Overall, the degree of landscape fragmentation had a tendency of “increasing-reducing-reducing” in the whole basin across the study period, but it remained comparatively stable in the southwest of the basin.

3.3. Global Spatial Autocorrelation of Landscape Types

As shown in

Table 3. Global spatial autocorrelation characteristics of various landscapes in Longxi River Watershed from 2005 to 2015 at 0.5 km spatial aggregation distance. z-score is the critical value of the test statistic; p-value indicates the significance level; |z| > 1.96, p < 0.05, i.e., confidence level > 95%; |z| > 2.56, p < 0.01, i.e., confidence level > 99%.

Year	Parameters	Forest	Farmland	Shrub-Grassland	Construction Land	Traffic Land	Bare Land	Gully Channel	Geological Hazards
2005	Moran’s I	0.234	0.300	0.260	0.085	−0.021	0.375	0.535	−0.748
	z	5.148	2.592	3.510	0.879	−0.033	3.108	5.934	−7.397
	p-value	<0.001	<0.01	<0.001	0.380	0.974	<0.01	<0.001	<0.001
2009	Moran’s I	0.478	0.494	0.474	0.412	−0.103	0.087	0.187	0.482
	z	9.849	3.294	7.454	4.355	−0.862	1.014	2.325	8.680
	p-value	<0.001	<0.001	<0.001	<0.001	0.389	0.310	0.020	<0.001
2011	Moran’s I	0.229	0.516	0.271	0.327	0.088	0.267	0.132	0.458
	z	4.865	3.982	4.399	3.090	0.998	2.979	1.632	7.408
	p-value	<0.001	p < 0.001	<0.001	<0.01	0.318	<0.01	0.103	<0.001
2015	Moran’s I	0.243	0.365	0.364	0.376	0.036	0.352	0.182	0.531
	z	5.167	2.731	5.913	3.490	0.465	2.777	2.282	7.889
	p-value	<0.001	<0.01	<0.001	<0.001	0.642	<0.01	0.023	<0.001

, Moran’s I values of forest, farmland, and shrub-grassland in the Longxi River Basin were all positive during the period from 2005 to 2015, indicating significantly positive spatial correlation characteristics (p < 0.01). Construction land had no significant spatial autocorrelation before the earthquake but had significantly positive spatial autocorrelation after the earthquake (2009–2015). Across the study period, traffic land had no significant spatial autocorrelation (p > 0.05). Bare land had significantly positive spatial autocorrelation except in 2009. Gully channel only had significantly positive spatial autocorrelation in 2005 before the earthquake. Geological hazards had significantly negative spatial autocorrelation in 2005 before the earthquake but had significantly positive spatial autocorrelation after the earthquake (2009–2015).

The global Moran’s I of landscape types except the distance of 0.5 km and geological hazards in 2005 in the Longxi River Basin gradually decreased with increasing distance from 2005 to 2015 (Figure 6). Within the lag distance of 10 km, the spatial autocorrelation of different landscape types had different characteristics across the lag distances before and after the earthquake. Before the earthquake in 2005, unlike the other landscape types, geological hazard had a significant negative spatial autocorrelation with the lowest Moran’s I (p < 0.01) at a distance of 0.5 km but had no significant spatial autocorrelation with the highest Moran’s I (p > 0.05) at a distance of 2 km. In addition, Moran’s I of construction land and traffic land had a lower value at a distance of 0.5 km than at a distance of 1 km. In 2009 (1 year after the earthquake), Moran’s I of traffic land, bare land, and gully channel had a lower value at a distance of 0.5 km than at a distance of 1 km. In 2011 (3 years after the earthquake), Moran’s I of traffic land and gully channel had a lower value at a distance of 0.5 km than at a distance of 1 km. In 2015 (7 years after the earthquake), only gully channel had a lower Moran’s I at a distance of 0.5 km than a distance of 1 km. Across the study period, Moran’s I of all landscape types had no significant spatial autocorrelation when the lag distance was >3 km.

3.4. Local Autocorrelation Analysis of Landscape Types

3.4.1. Moran Scatter Plot of Local Spatial Autocorrelation

Moran scatter plots were drawn using GeoDa, as seen in Figure 7 (in 2005), Figure 8 (in 2009), Figure 9 (in 2011), and Figure 10 (in 2015). In 2005, only geological hazards had a negative slope, and most of their points were located in quadrants II (LH) and IV (HL) in the scatter plot, indicating a negative spatial correlation. However, geological hazards had positive slopes and most of their points were located in quadrants I (HH) and III (LL), indicating positive spatial correlation after the earthquake (from 2009 to 2015). Across the study period, traffic land had no significant spatial correlation. Construction land had a low positive spatial correlation before the earthquake but a high positive spatial correlation after the earthquake, while bare land had the opposite pattern. The other four landscape types, forest, farmland, shrub-grassland, and gully channel, had comparatively stable Moran scatter plots with aggregation in HH and LL quadrants during the period from 2005 to 2015.

3.4.2. LISA Distribution of Local Spatial Autocorrelation

As seen in

Table 4. Statistics of spatial clustering and abnormal area of various landscapes in Longxi River Watershed.

Parameters	Area/km2 (Proportion/%)
	Forest	Farmland	Shrub-Grassland	Construction Land	Traffic Land	Bare Land	Gully Channel	Geological Hazards
2005
HH	15.64(21.56)	0.29(21.55)	0.74(22.10)	0.05(11.05)	0.04(15.72)	0.27(44.18)	0.13(24.85)	-
LL	0.89(1.22)	0.05(3.50)	0.09(2.72)	0.001(0.31)	0.001(0.38)	0.01(2.04)	0.01(2.71)	-
HL	2.51(3.46)	-	0.08(2.35)	0.005(1.11)	0.02(9.71)	-	-	-
LH	-	0.05(3.55)	0.08(2.49)	0.01(2.40)	-	0.01(0.87)	0.003(0.60)	-
NS	53.50(73.76)	0.97(71.41)	2.37(70.34)	0.35(85.13)	0.17(74.20)	0.33(52.91)	0.37(71.85)	0.06(100)
2009
HH	7.04(11.33)	0.32(37.69)	0.72(26.69)	0.15(40.92)	-	0.08(10.18)	0.37(20.82)	2.29(22.57)
LL	0.49(0.79)	0.05(6.33)	0.05(1.77)	0.01(2.42)	-	0.02(2.02)	0.03(1.93)	0.29(2.82)
HL	2.07(3.33)	0.03(3.42)	0.01(0.43)	-	0.02(6.40)	0.03(3.45)	-	-
LH	0.35(0.57)	0.02(2.31)	0.03(1.21)	0.02(4.56)	0.01(2.21)	0.01(1.19)	0.02(1.40)	0.16(1.61)
NS	52.19(83.99)	0.43(50.26)	1.89(69.91)	0.19(52.10)	0.35(91.40)	0.62(83.16)	1.33(75.84)	7.42(73.00)
2011
HH	10.76(15.67)	0.22(33.74)	0.54(16.75)	0.13(42.93)	0.01(2.75)	0.06(13.71)	0.24(13.86)	1.17(31.02)
LL	1.00(1.45)	0.03(4.59)	0.06(1.91)	0.002(0.80)	0.005(1.18)	0.001(0.28)	0.03(1.45)	0.09(2.32)
HL	2.14(3.11)	-	0.04(1.12)	-	0.02(5.99)	-	-	0.03(0.79)
LH	-	0.01(1.37)	0.10(3.19)	0.004(1.43)	0.01(3.19)	0.004(0.90)	0.01(0.66)	0.03(0.81)
NS	54.76(79.77)	0.40(60.30)	2.47(77.03)	0.16(54.84)	0.36(86.88)	0.37(85.11)	1.48(84.03)	2.45(65.05)
2015
HH	12.71(17.69)	0.04(8.04)	0.52(23.61)	0.15(47.74)	-	0.02(5.21)	0.34(20.49)	0.65(36.11)
LL	0.81(1.12)	0.03(6.61)	0.04(1.82)	0.004(1.33)	0.01(1.77)	0.002(0.48)	0.03(1.70)	0.04(2.17)
HL	2.28(3.17)	0.02(3.16)	-	-	0.02(4.78)	0.02(4.88)	-	0.04(2.47)
LH	-	0.02(3.19)	0.05(2.20)	0.004(1.30)	0.01(2.88)	0.003(0.85)	0.03(1.64)	0.03(1.58)
NS	56.07(78.02)	0.39(79.00)	1.60(72.37)	0.15(49.63)	0.44(90.57)	0.33(88.58)	1.26(76.17)	1.03(57.67)

and Figure 11, Figure 12, Figure 13, Figure 14, Figure 15, Figure 16, Figure 17 and Figure 18, the spatial aggregation or anomaly distribution of landscape types in the Longxi River Basin changed dramatically during the period from 2005 to 2015. As the substrate landscape in the Longxi River Basin, forest had the highest and fluctuating HH aggregation area across the study period (15.64 km² in 2005 accounting for 21.56%, 7.04 km² in 2009 accounting for 11.33%, 10.76 km² in 2011 accounting for 15.67%, and 12.71 km² in 2011 accounting for 17.69%, respectively), but comparatively stable LL aggregation, LH anomaly, and HL anomaly area (Table 3 and Figure 11). Farmland had stable LL aggregation, LH anomaly, and HL anomaly area across the study period, except that the area of HH aggregation decreased to 0.04 km², accounting for only 8.04% of the total area in 2015 (Table 3 and Figure 12). However, the HH aggregation area of construction land, gully channel, and geological hazards increased in 2009 after the earthquake, while the LL aggregation, LH anomaly, and HL anomaly area of those changed comparatively slightly (Table 3; Figure 14, Figure 17 and Figure 18). In addition, shrub-grassland and bare land had continuously decreasing HH aggregation area but also comparatively stable LL aggregation, LH anomaly, and HL anomaly area across the study period (Table 3; Figure 13 and Figure 16). Compared to the other landscape types, the spatial aggregation or anomaly distribution of traffic land slightly changed across the study period (Table 3 and Figure 15).

4. Discussion

4.1. Fragmentation of Landscape Patterns

Major earthquakes and their subsequent geological hazards caused huge damage to natural landscapes and ecosystems, resulting in heavy loss of life and property, ecological degradation, as well as landscape fragmentation [12,52,53]. It has been proven that landscape fragmentation had a high importance degree in assessing the extent of earthquake-induced landscape destruction due to its negative effects on biodiversity and the ecological process of species [12]. Our result showed that human activities and gully erosion across the Longxi River Basin were the main factors controlling landscape fragmentation in the basin before the earthquake in 2005. They caused high-value areas of edge density (ED), Shannon diversity (SHDI), and landscape isolation index (DIVISION), distributed along the banks of the Longxi River and concentrated in the southeast of the downstream due to human settlements. However, high-value areas of ED, SHDI, and DIVISION increased and spread over the basin due to landslides, collapses, and debris flows after the earthquake in 2009, indicating that a high degree of landscape fragmentation occurred in the Longxi River Basin. This study showed that high-value areas of ED, SHDI, and DIVISION decreased after 3 and 7 years of natural recovery (in 2011 and 2015), suggesting that the landscape of the basin had a great recovery potential due to the dominant natural extant forests managed strongly by the Longxi Hongkou National Nature Reserve administration [8]. This result may prove the hypothesis that natural extant forests are easier to restore after disturbances in comparison to plantations because of the even distribution of species, complicated community structure, and stable ecological equilibrium [54].

4.2. Landscape Spatial Global Autocorrelation

Our previous study showed that all vegetation types had significantly positive spatial autocorrelation in the most severely earthquake-affected areas of west Sichuan Province, China [55]. However, the spatial autocorrelation analysis was seldom used to characterize the spatial structure of landscape types on a small scale, such as a watershed or a basin.

In this study, whether the earthquake occurred or not, forest, farmland, and shrub-grassland had a significantly positive spatial correlation, showing spatial aggregation characteristics. Forest, as the first dominant landscape type, accounted for 91.53% before the earthquake in 2005, and still 78.42% after the earthquake in 2009. Its large area and wide-spread range caused positive spatial correlation across the study period. Shrub-grassland, as the second dominant landscape type, had a similar pattern. Farmland is concentrated in the area with flat terrain, an abundant source of water, and a dense population, which is mainly located in the southeast of the downstream of the Longxi River, so it also had a positive spatial correlation. In addition, though secondary disasters caused the reduction of farmland area, the demand for crop harvest and economic development with convenience resulted in the cluster of farmland.

Construction land had a significantly positive spatial correlation after the earthquake because the local government carried out the building reconstruction for the relocation of victims, resulting in an intensive distribution of settlements in the flat downstream alluvial fan, which is suitable for reconstruction in the basin. Though the construction of residential areas and the laying of cement land interrupted the natural environment and caused ecological damage, the centralized reconstruction after the earthquake greatly reduced the damage to the surrounding environment compared to the scattered distribution before the earthquake, resulting in the improvement of the local environment. This result is consistent with the findings of Shi et al. (2020) [56], who suggested the positive effect of reconstruction on construction land. It also proved that flat land in a mountainous area had the convenience of supporting a dense population (Zou et al., 2021) [57].

However, there is no significant spatial autocorrelation in traffic land across the study period because it is mainly affected by nonrandom human activities. Bare land had significantly positive spatial autocorrelation across the study period except 1 year after the earthquake in 2009. This may be due to the transition from bare land to geological disasters 1 year after the earthquake that caused the independent tendency over distance, resulting in less similar values for randomly associated pairs of observations [21]. This result is consistent with the findings of Li et al. (2012) [58], who observed that earthquake-induced disasters caused the random distribution of bare land.

Gully channels had a significantly positive spatial correlation before the earthquake in 2005 but not after the earthquake. This was due to frequent debris flows burying the original channels and forming headfall after the earthquake, resulting in the dispersed distribution of channels. In addition, the changes in river channels also caused road realignment and even flush back into roads and buildings. Therefore, an early warning system should be set up for disaster prevention and control. However, geological disasters changed from significantly negative spatial correlation before the earthquake to significantly positive spatial correlation after the earthquake, suggesting that they had high spatial dependency controlled by topographical factors (including elevation, slope gradient, distance to river) and seismic factors (including seismic intensity and distance to fault). This result is consistent with other studies demonstrating the positive relationship between topographical and seismic factors and geological disasters after an earthquake [59].

The Moran’s I of all landscape types decreased with distance and had no significant spatial autocorrelation when the lag distance was > 3 km. This means that the spatial autocorrelation of all landscape types could not be detected by aggregating the data when their aggregation level exceeded the level of increasing spatial scale (Overmars et al., 2003) [21]. This result is consistent with other studies demonstrating significant impacts of spatial scale on spatial structure [55].

4.3. Landscape Spatial Local Autocorrelation

Cluster and significance maps of LISA can show the interactions between sites and their neighbors in close proximity to identify the local spatial structure and instability by comparing their values [44]. In this study, forest had a high HH aggregation area in 2005 but the proportion of forest aggregation area decreased after the earthquake in 2009, then increased in 2011 and 2015, indicating that the spatial aggregation degree of forest decreased due to the disturbance of geological disasters but gradually recovered after natural restoration and succession.

Topographical factors are the main controlling factors of farmland distribution, while socioeconomic factors, including human tillage management, fertilization, and construction control the direction, quantity, and speed of farmland transition [17]. In this study, the distribution of farmland showed a trend of HH and LL aggregation but with a low area before the earthquake, suggesting that farmland resources in the Longxi River Basin had comparatively fewer spatial aggregation areas, distributed discretely. The HH aggregation area of farmland changed slightly one year after the earthquake but then decreased in 2011 and 2015, indicating that the spatial dependency of farmland diminished. This may be due to the following two reasons. First, frequent geological disasters such as landslides and debris flows induced by heavy rainfall in rainy seasons buried farmland distributed along the river [30,58]. Second, the local government converted farmland to forest and shrub-grassland or made a requisition of farmland due to its convenience for transportation, construction, and development for reconstruction for the relocation of victims, resulting in the reduction of the HH aggregation area of farmland [8].

The HH aggregation area of construction land increased from 2005 to 2009 but changed slightly during the period from 2009 to 2015. This may be due to the immediate reconstruction in disaster-affected areas after the earthquake and the stable state of reconstruction in subsequent years. This result is consistent with a previous study, which highlighted the positive effects of reconstruction on the stability of construction land after the earthquake [31]. Unlike its spatial global autocorrelation, the HH aggregation area of bare land continually decreased from 44.18% to 5.21% during the study period, indicating that natural disasters induced dispersed distribution of bare land, which were mainly distributed in the banks of the middle and lower stream of the Longxi River. In addition, the HH aggregation area of geological disasters increased intensively after the earthquake in 2009 but decreased with a higher percentage from 2011 to 2015, demonstrating a concentrated distribution of geological disasters in the north of the Longxi River Basin after natural recovery (Figure 18). This may be due to steep landforms and difficult human accessibility to the north of the basin that caused no treatment and frequent occurrence of geological disasters around the Longchi Lake distributed there [10,30].

4.4. The Boundedness and Prospect

The overall landscape pattern of the Longxi River Basin has undergone strong changes, resulting in fragmented landscapes. After 7 years, the damaged ecology has gradually recovered. The formation and evolution of landscape patterns are influenced by multiple factors. For example, from 2005 to 2009, due to the impact of the Wenchuan earthquake, the forest landscape area significantly decreased, and geological disasters and gully landscape area significantly increased, resulting in serious damage to the ecological environment, the decrease in wildlife habitats and resources, the reduction of water conservation capacity, and a high incidence of debris flow. In the future, closing mountains to facilitate afforestation; protecting natural forest, animal, and plant resources; and vigorously developing ecotourism will be important ways to reduce the impact of disasters and protect the environment of the basin. From 2009 to 2015, although the population in the watershed increased and geological disasters such as debris flows and landslides occurred frequently under heavy rainfall conditions, the ecological environment in the watershed gradually recovered. This is due to the policy of returning farmland to grassland and forest, as well as ecological and natural restoration. In addition, migration and resettlement, new town construction, and industrial structure adjustment also have a certain impact on human life and production activities in the basin. Affected by urbanization and human activities, the area of farmland and construction land in the Longxi River Basin has decreased, while the area of transportation land has increased.

In this study, we used four different remote-sensing satellite images to analyze the landscape pattern characteristics and spatial autocorrelation of the typical earthquake-affected basin. However, we only conducted the analysis using the 500 m × 500 m grid as a research cell. It is well known that the occurrence of spatial autocorrelation in geological elements can disappear or emerge going from one scale to the other [21]. Therefore, elucidating multi-scale effects on spatial autocorrelation of landscape would require further research. In addition, we were not able to consider potential driving factors of landscape pattern (e.g., elevation; slope gradient and aspect; distances to road, river, and fault; and rainfall). A spatial autoregressive model combining potential driving factors should be carried out to explain the correlation between landscape patterns and their driving factors. Human intervention, including reconstruction, disaster treatment, reforestation, and so on, can also change landscape patterns [60]. Understanding how human interventions combined with other driving factors control the process of landscape types could improve our knowledge of the ecological restoration of earthquake-affected areas.

5. Conclusions

We used the landscape index method to analyze the landscape pattern characteristics and the succession characteristics of landscape patterns in the Longxi River Basin. The following results are obtained:

(1)

In the study period, due to the earthquake, the degree of landscape fragmentation increased in 2009 compared with 2005 but decreased after 3 and 7 years of natural recovery.

(2)

Forest, farmland, and shrub-grassland in the basin had significantly positive spatial autocorrelation across the study period, while construction land only had significantly positive spatial autocorrelation after the earthquake. Bare land had significantly positive spatial autocorrelation except in 2009, while gully channel only had significantly positive spatial autocorrelation in 2005. Geological hazards shifted from significantly negative in 2005 to positive spatial autocorrelation from 2009 to 2015.

(3)

The Moran’s I of all landscape types decreased with the increase of space distance but in different distance-decay rates. In general, all landscape types had no significant spatial autocorrelation when the lag distance was >3 km.

(4)

Forest had the highest and fluctuating HH spatial aggregation area across the study period. Geological disasters had the clustered HH spatial aggregation areas upstream of the basin, while construction land was gradually distributed in Longchi town. The HH spatial aggregation areas of farmland distribution in the southeast of the basin increased in 2009 but then decreased. In addition, shrub-grassland and bare land had a continuously decreasing HH aggregation area.

Despite some limitations, our research contributes to improving landscape patterns, human intervention, and ecological restoration in earthquake-affected areas. Especially for post-disaster ecological restoration and reconstruction work, data support can be provided. For example, based on the results in this article, the degree of disaster in ecological areas can be graded and evaluated, and different ecological protection measures can be implemented for different regions, such as closing mountains to facilitate afforestation and returning farmland to forests. For post-disaster reconstruction, various reconstruction and relocation strategies can be implemented according to local conditions. Future research should extend this work by combining more influencing factors, and using more high-resolution remote-sensing satellite images to predict the spatial distribution of landscape types during ecological recovery.