Mechanisms, monitoring and modeling of shrub encroachment into grassland: a review

ABSTRACT

Shrub encroachment into arid and semi-arid grasslands has elicited extensive research attention worldwide under the background of climate change and increasing anthropogenic activities. Shrub encroachment may considerably impact local ecosystems and economies, including the conversion of the structure and function of ecosystems, the shift in ambient conditions, and the weakness of local stock farming capacity. This article reviews recent research progresses on the shrub encroachment process and mechanism, shrub identification and dynamic monitoring using remote sensing, and modeling and simulation of the shrub encroachment process and dynamics. These studies can help to evaluate the ecological effect of shrub encroachment, and thus, practically manage and recover the ecological environment of degraded areas. However, the lack of effective measures and data for monitoring shrub encroachment at a large spatial scale severely limits research on the mechanism, modeling, and simulation of shrub encroachment, and the shrub encroachment stages can hardly be quantitatively defined, resulting in insufficient analysis and simulation of shrub encroachment for different spatiotemporal scales and stages shift. Improvement in remote sensing-based shrub encroachment dynamic monitoring might be crucial for analyzing and understanding the process and mechanism of shrub encroachment, and multi-disciplinary and multi-partnerships are required in the shrub encroachment studies.

KEYWORDS:

• Shrub encroachment
• arid and semi-arid grassland
• mechanism
• dynamic monitoring
• modeling and simulation

1. Introduction

Grasslands are among the most important types of global vegetation land cover, occupying over 40% of the world's terrestrial area, excluding Greenland and Antarctica (Suttie, Reynolds, and Batello 2005). Over 50% of grasslands worldwide experience different stages of degradation under the background of climate change and increasing anthropogenic activities (White, Murray, and Rohweder 2000). Among these degradation stages, shrub encroachment (also called shrub invasion or woody thicketization) in arid and semi-arid grasslands and savannas has elicited extensive attention (Archer, Schimel, and Holland 1995; Eldridge et al. 2011; Naito and Cairns 2011). Shrub encroachment has been occurring worldwide, including in the savannas of southern Africa (Levick and Rogers 2011), the arid grasslands of Australia (Robinson, van Klinken, and Metternicht 2008), the rangelands of southern Argentina (Cipriotti and Aguiar 2012), the dry grasslands of the western United States (Goslee et al. 2003; McGlynn and Okin 2006), the Mediterranean regions (Maestre et al. 2009), the grasslands of Inner Mongolia in China (Zhou 1990; Tong et al. 2001; Xiong et al. 2003; Peng et al. 2013; Peng, Li, and Tong 2013; Zhang et al. 2017), and the alpine grassland of the Qinghai–Tibet Plateau (Zhang et al. 2006; Wang et al. 2013).

Shrub encroachment into arid and semi-arid grasslands can be observed through the increase in the density, coverage, and biomass of shrubs or woody plants (Van Auken 2009), which significantly impacts local ecosystems. Shrub encroachment results in the shift from a herbaceous-dominant ecosystem to a shrub-dominant ecosystem, along with changes in the ecological structure and function of an ecosystem, including vegetation productivity, biological diversity, carbon storage, and flux (Briggs, Blair, and Mccarron 2005; Eldridge et al. 2011; Peng, Li, and Tong 2013). The increase of shrubs in grasslands may alter the conditions of the environment, such as the physical structures and chemical properties of soil (Archer, Schimel, and Holland 1995; Zhang, Shi, and Zheng 2001; Li and Gao 2012) and the re-allocation of water (Peng, Li, and Tong 2014). These changes in ambient conditions may facilitate the growth of shrubs, form a positive feedback effect that can make the shrub encroachment process irreversible (Archer, Schimel, and Holland 1995; Scheffer and Carpenter 2003), and finally lead to the formation of sand dune shrublands (Xiong et al. 2003) or desertification (Kéfi et al. 2007). Shrub encroachment lowers the biomass of edible pasture, thereby decreasing stock farming capacity, which affects the local economy (Naito and Cairns 2011). Therefore, research on shrub encroachment into arid and semi-arid grasslands is important to improve the understanding of the response and adaptation of grassland ecosystems to climate change and anthropogenic activities. Such research can help properly evaluate the ecological effect of shrub encroachment, and thus, practically manage and recover the ecological environment of degraded areas.

Current studies have identified the main factors that induce shrub encroachment into arid and semi-arid grasslands, such as regional and global climate change (Tietjen et al. 2010), increasing atmospheric CO₂ concentration (Mooney and Cleland 2001), decreasing wildfire frequency (Roques, O’Connor, and Watkinson 2001; Lohmann et al. 2014), and increasing grazing pressure (Van Auken 2009). These exogenic factors and feedback mechanisms of an ecosystem (D’Odorico et al. 2010) mutually affect the process and status of shrub encroachment. However, debates continue on the scales, ranges, and relations of impact factors (D’Odorico, Okin, and Bestelmeyer 2012). The research on the shrub encroachment mechanism requires the long-term and dynamic monitoring of the shrub encroachment process (Roques, O’Connor, and Watkinson 2001; Van Auken 2009). Long-term monitoring methods include field observation within a limited region (Cipriotti and Aguiar 2012) and the dynamic interpretation of historic aerial photos (Goslee et al. 2003) or high spatial resolution (HSR) remote sensing images (Laliberte et al. 2004). At the regional scale or within a large geological range, moderate or coarse spatial resolution remote sensing data are useful sources for monitoring the shrub encroachment process and dynamics. Existing methods include vegetation indices (Xu et al. 2010), spectral mixture analysis (Sohn and McCoy 1997), multi-temporal or time series data analysis (Wilfong, Gorchov, and Henry 2009; Zhou et al. 2013), and multi-angle remote sensing (Chopping et al. 2006, 2008). However, accurately recognizing shrub and estimating its coverage from grass–shrub–soil mixture pixels are difficult due to the high similarity between the spectra of grass and shrub. The lack of effective measures and data for monitoring shrub encroachment at a large spatial scale severely limits research on the mechanism, modeling, and simulation of shrub encroachment. The improvement of shrub extraction, inversion, and monitoring methods, along with providing long-term and large spatial range shrub dynamic data, can help analyze the effects of various factors at different spatiotemporal scales. They can also help model and simulate the dynamics of the shrub encroachment process.

This article reviews recent progresses in the research on the shrub encroachment process and mechanism (Section 2), shrub identification and dynamic monitoring using remote sensing (Section 3), and modeling and simulation of the shrub encroachment processes and dynamics (Section 4). Several conclusions are drawn in Section 5.

2. Process and mechanisms of shrub encroachment

2.1. Stages of shrub encroachment

Shrub encroachment into grassland indicates a shift from the stable landscape state dominated by grass to another state dominated by shrubs (Anderies, Janssen, and Walker 2002; van Langevelde et al. 2003; Okin, D’Odorico, and Archer 2009). The shrub encroachment process can be observed through several demographic changes (D’Odorico, Okin, and Bestelmeyer 2012). Thus, researchers have divided this process into stages or phases. Xiong, Han, and Bao (2005) proposed three states of the shrub encroachment process based on the steady states of an ecosystem. State 1 or pristine grassland is identified as potential sandy desertification. It is dominated by perennial gramineous grasses, and the ecological function of shrub patches is weaker than the grass matrix. The main grass species are native high gramineous grasses inside and outside shrub patches. State 2 or thicketization grassland is identified as light sandy desertification. It is dominated by shrub patches that consist of shrubs and understory grasses. The main grass species are native high gramineous grasses inside shrub patches and small gramineous grasses or weed outside shrub patches. The ecological function of shrub patches is enhanced in the ecosystem during this state. State 3 or dune shrubland is characterized by heavy sandy desertification. Coppice dunes or nebkhas develop only on sandy soil, thereby indicating severe grassland degradation, soil erosion, and desertification. No continuous grass layer exists within shrub patches, and the basic ecological function of the ecosystem tends to disappear. The shift between states relies on ecological thresholds. The crossing of an ecological threshold indicates changes in the original ecosystem structure and the loss of the corresponding ecological functions, which suggests the overall degradation of the ecosystem (Xiong and Han 2005). D’Odorico, Okin, and Bestelmeyer (2012) divided the shrub encroachment process into four stages: (I) pure grassland with nearly no shrub, (II) grassland with limited shrub cover, (III) shrubland where most grass cover is absent and shrubs have increased in size and density, and (IV) shrub duneland. Stages II–IV can correspond to Stages 1–3 of Xiong, Han, and Bao (2005). Figure 1 shows the stages of Caragana microphylla Lam. encroachment in a steppe in Inner Mongolia, China based on the four stages proposed by D’Odorico, Okin, and Bestelmeyer (2012).

Figure 1. Four stages of C. microphylla Lam. encroachment into steppe in Inner Mongolia, China: (a) Stage I, (b) Stage II, (c) Stage III, and (d) Stage IV. (Photos taken by X. Cao in 2015–2016)

The determination of the states or stages of shrub encroachment is dependent on the experience or knowledge of ecologists rather than on quantitative measures or indicators. However, the relationship between the size and number of shrub patches can show the change in landscape pattern, which may be an ‘indicator’ of ecosystem response to climate change or human activities (Rietkerk et al. 2004; Kéfi et al. 2007; Scheffer et al. 2009). The size and number of vegetation patches were observed as a power law in southern Africa (Scanlon et al. 2007) and a truncated power law in Mediterranean arid ecosystems (Kéfi et al. 2007). Increasing grazing pressure may lead to a deviation from the power law (Kéfi et al. 2007). In a typical steppe region in the Xilin River Basin, Inner Mongolia, grassland degradation caused by overgrazing finally resulted in C. microphylla patches that dominated the shrub-thicketizated grassland (Xiong and Han 2006). The landscape pattern of shrub patches in grasslands can help to reflect the stages or phases of shrub encroachment by power law.

2.2. Mechanisms of shrub encroachment

Understanding the mechanism of shrub encroachment is of considerable importance for managing and protecting arid grassland ecosystems (Briggs, Blair, and Mccarron 2005). Belayneh and Tessema (2017) summarized six theories for the mechanisms of shrub encroachment in semi-arid African ecosystems, including two-layer model, one-layer model, state-and-transition, equilibrium, disequilibrium, and non-equilibrium rangeland theories. The two-layer model (Walter 1939) uses moisture in top and lower layers of soil to describe the shift in competitive advantage from grass species to trees or shrubs, because heavy grazing pressure in dryland decreases grass cover and soil moisture becomes available to the trees or shrubs. The one-lay model (Adams 1967) states that grass fires can result in a constant nitrogen loss and limit grass growth, while benefit Acacia bushes due to its capacity of nitrogen-fixing. The state-and-transition model describes the rangelands in dryland environments as ecosystems with stable states, and transitions between stable states can be initiated by influence of biotic and abiotic factors (Westoby, Walker, and Noy-Meir 1989; Briske, Fuhlendorf, and Smeins 2003; van Langevelde et al. 2003). The equilibrium, disequilibrium, and non-equilibrium rangeland theories describes three types of ecosystems which are characterized by one stable state, two stable states and a fluctuation in tree-grass balance, respectively (Briske, Fuhlendorf, and Smeins 2003; Sankaran, Ratnam, and Hanan 2004; Sankaran et al. 2005), while the causes of shrub encroachment in these ecosystems may be different.

D’Odorico, Okin, and Bestelmeyer (2012) categorized the mechanisms that drive shrub encroachment into exogenic or endogenic factors. Exogenic mechanisms determine the shift in competitive advantage from grassland to shrubland (Archer, Schimel, and Holland 1995). The factors associated with exogenic mechanisms involve regional and global warming, increasing atmospheric CO₂ concentration, large-scale fire control, and increasing grazing pressure. A low temperature may damage shrub seedling, whereas an increasing temperature induced by global warming can decrease the mortality of shrub shoots and foster shrub growth (Tietjen et al. 2010). The water use efficiency of C₄ plants (mainly herbaceous plants) is higher than that of C₃ plants (shrubs), whereas increasing CO₂ concentration can enhance the water use efficiency of C₃ plants and reduce the competitive advantage of C₄ plants (Mooney and Cleland 2001). A wildfire can eliminate shrub shoots and reduce shrub regeneration. However, long-term and large-scale fire suppression in rangeland and grazing-induced fuel reduction will decrease fire frequency and enhance shrub survival rate (Roques, O’Connor, and Watkinson 2001; Xiong et al. 2003; Lohmann et al. 2014). Increasing grazing activities or overgrazing will accrete grazing by livestock, cut down grass cover (van de Koppel et al. 2002; Xiong et al. 2003), and promote the diffusion of shrub seeds (Archer 1989), which will also result in shrub encroachment (Van Auken 2009). Shrub encroachment may be the consequence of a combination of factors that are difficult to separate (Van Auken 2009).

Endogenic mechanisms include ecosystem feedback that may create bistability (D’Odorico, Okin, and Bestelmeyer 2012). A state shift in bistable ecosystem dynamics can be explained by four positive feedback processes, namely, vegetation–microclimate feedback, fire–vegetation feedback, soil erosion–vegetation feedback, and grazing of small mammals–vegetation cover feedback (Figure 2). At the landscape scale to meso-scale, shrub encroachment enhances the proportion of bare soil, thereby increasing diurnal soil heat fluxes and nocturnal upward long-wave radiation. These changes will modify microclimate conditions with an increase in near-surface air temperature and benefit shrub survival (D’Odorico et al. 2010). At the landscape scale, fire is sustained by contiguous fuels (dry grass); periodic fires kill shrub seedlings and prevent shrubs from expanding (Roques, O’Connor, and Watkinson 2001). Fire management will decrease fire frequency and disrupt the fire–grass feedback, thereby leading to shrub expansion. At the patch scale, erosion–vegetation feedback operates in an environment where grasses exhibit competitive advantage, whereas shrubs are more likely to obtain soil and water resources. High grass cover prevents resource redistribution by wind and water erosion, improves water infiltration, and maintains grass production. Surface soil resources are depleted and redistributed to shrubs when grass coverage is lost (Schlesinger et al. 1990; Zhang, Shi, and Zheng 2001; Xiong and Han 2005; Han et al. 2014). The small mammals–vegetation cover feedback tends to increase the ratio of shrub to grass cover because shrub patches provide nests and shelters for small mammals and help them nibble on grass roots (Kerley and Whitford 2009).

Figure 2. Positive feedback for shrub–grass bistable state shift at different scales (adapted from (D’Odorico et al. 2010)).

Certain problems remain in the mechanism analysis of shrub encroachment. Factors that lead to shrub encroachment are controversial in terms of scale, size, and interactions (D’Odorico, Okin, and Bestelmeyer 2012). Studies on the mechanism of shrub encroachment are mostly aimed at a specific scale rather than at multiple scales (Goslee et al. 2003; Laliberte et al. 2004; Briggs, Blair, and Mccarron 2005; Robinson, van Klinken, and Metternicht 2008; Levick and Rogers 2011). However, shrub encroachment is likely due to multiple factors and their interactions at multiple scales (Archer, Schimel, and Holland 1995; Peters et al. 2006; Belayneh and Tessema 2017). During different stages of shrub encroachment, influencing factors may vary (D’Odorico, Okin, and Bestelmeyer 2012), and no relevant research has yet been conducted on the stage classification and mechanism analysis of shrub encroachment at different stages. Only a few studies have addressed the quantitative analysis of the influencing factors of shrub encroachment due to the lack of a wide range of dynamic monitoring data. For example, Roques, O’Connor, and Watkinson (2001) utilized aerial photographs taken from 1947 to 1990 and survey data from 1997 to study shrub encroachment in the lowland meadows of Swaziland using the multivariate regression analysis method. The influencing factors included fire frequency, grazing pressure, precipitation, soil type, and initial coverage of shrubs. Therefore, the long-term, large-scale, and multi-scale observations of shrubs and encroaching processes are the bases for studying the shrub encroachment mechanism.

3. Shrub identification and encroachment dynamic monitoring with remote sensing data

The spatiotemporal monitoring of shrub encroachment is the primary requirement for shrub encroachment research. Long-term field observations or in situ experiments are the traditional measures for studying shrub invasion or grassland degradation. However, these methods are constrained by intensive labor, high cost, limited spatial and temporal ranges, and difficulty in explaining spatial heterogeneity at a large spatial scale (Naito and Cairns 2011). In recent years, aerial photos and remote sensing data have been widely used to monitor shrub encroachment in grasslands, and various methods have been developed to identify shrubs in grass–shrub–soil co-existing landscapes. We summarized remote sensing-based techniques according to the spatial resolutions of data. The methods developed for identifying shrub and estimating its coverage were also introduced.

3.1. HSR data

Very high spatial resolution (VHSR) remote sensing images are preferred for detecting shrubs in grasslands and monitoring their dynamics. HSR data include satellite images, such as those taken by IKONOS, QuickBird and WorldView; aerial and unmanned aerial vehicle (UAV) photographs; light detection and ranging (LiDAR) data; and airborne hyperspectral data. Figure 3 shows the shapes of shrubs (C. microphylla Lam. in a steppe in Inner Mongolia, China) on the ground and in HSR images. In the VHSR remote sensing images (< 4 m), shrubs typically present circular or elliptical shapes with different sizes, which are clearly distinguished from background herbaceous plants and soil. Methods for identifying shrubs include human interpretation (Briggs, Blair, and Mccarron 2005; Xu and Li 2009), unsupervised classification (Frank and Tweddale 2006; Robinson, van Klinken, and Metternicht 2008), and object-oriented classification (Laliberte et al. 2004; Levick and Rogers 2011).

Figure 3. Shapes of C. microphylla Lam. patches at different observation scales: (a) field observation on the ground, (b) UAV image with a resolution of 0.08 m, (c) Google Earth image with a resolution of approximately 0.3 m; and (d) WorldView-2 multispectral fusion image with a resolution of 0.5 m.

Earlier photos are mostly panchromatic; hence, the typical method for extracting shrub individuals or patches is human interpretation or threshold selection to extract the darkest areas (Hutchinson, Unruh, and Bahre 2000; Lahav-Ginott, Kadmon, and Gersani 2001). To avoid errors that might arise from the single thresholding of a panchromatic HSR image, Goslee et al. (2003) manually marked the center points of all shrub patches and then adopted an iterative algorithm to expand outward from the marked point until an edge difference value was reached. This parameter was calculated based on the contrast between dark shrub canopies and lighter background areas in each image. Through this method, Goslee et al. (2003) extracted the population dynamics and spatial pattern of all shrubs (Prosopis glandulosa) with a diameter greater than 2 m in the Chihuahuan Desert, southern New Mexico by using 10 high-resolution aerial and satellite images taken from 1936 to 1996. For multispectral or colored HSR images, object-oriented classification can improve the accuracy of shrub extraction. Laliberte et al. (2004) used image segmentation and object-based classification to monitor shrub changes from 1937 to 1996 in the Jornada Basin of southern New Mexico based on aerial photos taken during this period and QuickBird images from 2003. Both of these studies found an increase in the shrub coverage of the desert grasslands in southern New Mexico. Robinson, van Klinken, and Metternicht (2008) collected aerial photos from 1943 to 2004 in the semi-arid grassland of western Australia and used an iterative self-organizing clustering procedure (ISODATA) to extract patches and monitor dynamics of invasive mesquite trees and shrubs. Garrity et al. (2008) used spatial wavelet analysis to automatically detect the location and crown diameter of shrub individuals ranging from 0.25 to 1 m in HSR multispectral images. Shrubs are frequently higher than background soil and herbaceous species, and thus, LiDAR data can provide information that is useful for identifying shrubs and estimating their height and structure (Levick and Rogers 2011). Hellesen and Matikainen (2013) combined LiDAR data with a point density of 13 points/m² and color infrared orthoimages with a spatial resolution of 0.5 m; they then used an object-oriented classification and regression tree classifier to extract shrubs and trees. Their results indicated that using LiDAR data could considerably increase the accuracy of shrub and tree extraction in complex environments in Denmark.

HSR images can be used directly to identify shrubs from background grassland and to monitor shrub population and coverage at the individual and patch scales. However, HSR data are characterized by high costs for covering large regions and historical data are lacking in many regions. Therefore, HSR data are difficult to use in large-area shrub mapping and dynamic shrub monitoring. HSR images may miss shrubs that are smaller than the spatial discriminability of the data (Frank and Tweddale 2006). In many studies, HSR images are only used as references or validation data for large-area shrub monitoring and mapping (Chopping et al. 2006; Olthof and Fraser 2007; Selkowitz 2010).

3.2. Moderate to coarse spatial resolution data

In shrub encroachment research or dynamic monitoring at the regional scale, moderate to coarse spatial resolution (30 m–1 km) data are frequently used to identify shrubs and estimate their coverage. These data include images taken by the Systeme Probatoire d’Observation de la Terre (SPOT) (Duncan et al. 1993), Landsat Thematic Mapper (TM) (Olthof and Fraser 2007; Selkowitz 2010), moderate-resolution imaging spectroradiometer (MODIS) (Selkowitz 2010), and multi-angle imaging spectroradiometer (MISR) (Chopping et al. 2008). Synthetic aperture radar (SAR) data, such as Advanced Land Observing Satellite (ALOS) Phased Array Type L-Band SAR (PALSAR) (Mitchard et al. 2009, 2011; Skowno et al. 2017), are also used to monitor woodland expansion in grasslands in central and southern Africa. These technologies include spectral mixture analysis (SMA) for multi-spectral data, phenology or sensitivity analysis for multi-temporal data, and geometric-optical (GO) modeling for multi-angle data.

3.2.1. SMA for multi-spectral data

SMA or spectral unmixing can provide land cover information at subpixel levels for practical applications. The concept of SMA is to resolve a linear or nonlinear combination of mixed spectrum through its constituent spectral components or endmembers and to obtain their subpixel fractional cover (Somers et al. 2011). Linear SMA (LSMA) assumes that a landscape is composed by several materials on the ground, i.e. endmembers, and that each incident photon interacts with only a single pixel component, thereby disregarding multiple scattering among endmembers (Camacho-De Coca et al. 2004; Song 2005). LSMA can linearly model the mixed spectral information and estimate the fraction of each endmember. LSMA has been extensively used in estimating vegetation coverage (Chen et al. 2001; Small 2001; Wu and Murray 2003; Niu, Liu, and Wang 2005; Pan and Li 2010). Sohn and McCoy (1997) simulated endmember spectra in a desert grassland environment at the TM pixel scale by using measured spectra and applied LSMA and the constrained least squares method to obtain shrub coverage.

The SMA model exhibits several advantages over conventional empirical models for estimating vegetation coverage. This model can build multiple representative endmembers to simulate various backgrounds, such as soil and dry grass, in a landscape with low vegetation coverage, thereby enhancing the accuracy of shrub coverage estimation. Empirical models that suffer from background noise can barely provide an accurate estimation (Henry and Hope 1998; Rogan and Franklin 2001). The SMA model can be independent of field measurement if pure pixels are found in a remote sensing image and used as endmembers (Vila and Barbosa 2010; Veraverbeke et al. 2012). However, the number of endmembers are constrained by the number of spectral bands in remote sensing data, e.g. a maximum of four endmembers are generally recommended for Landsat TM/ETM+ data (Theseira, Thomas, and Sannier 2002). In a shrub and grass mixed landscape, SMA may encounter a collinearity problem because of the spectral similarity between shrub and grass endmembers (Chen et al. 2011); such problem becomes more serious when a virtual endmember-based nonlinear SMA (NLSMA) model is used. Appropriate seasonal remote sensing data have been suggested to reduce the influence of grass (Selkowitz 2010). Another important issue is the nonlinear effect of multiple scattering from the 3D structure of shrubs. Wang et al. (2015) simulated the relationship between the multiple scattering effect and shrub parameters (e.g. density and height) by using a Monte Carlo ray tracing model, namely, Forest LIGHT (FLIGHT). Their results indicated that the contribution of multiple scattering to landscape spectra increases linearly with the fractional cover of shrub and crown height. Multiple scattering exhibits a stronger effect in the near-infrared range than in the visible bands, thereby implying that selecting LSMA or NLSMA significantly affects shrub coverage estimation. In arid and semi-arid grasslands, low vegetation coverage makes selecting endmembers from remote sensing images difficult. Possible means to obtain endmember spectra include simulation from field-measured spectra (Sohn and McCoy 1997) and purification from high vegetation coverage pixels (Ma et al. 2015). Through the assumption that the reflectance of pure vegetation is identical among all impure vegetated pixels within a target image, the concept of spectral purification involves simultaneously resolving shrub endmember spectra and all endmember fractions using multiple NLSMA equations (Ma et al. 2015). The results showed that the purified shrub endmember spectrum was close to the measured spectrum in a semi-arid steppe area in Inner Mongolia. The endmember purification method does not rely on field-measured spectra. Thus, this method is practical for large-area shrub mapping.

3.2.2. Multi-temporal remote sensing data

Multi-temporal remote sensing data can indicate the discrepancy of phenology and sensitivity to climate variance among different vegetation types. Target species can be detected using multi-temporal remote sensing data based on the phenological differences of various species. Wilfong, Gorchov, and Henry (2009) used three Landsat TM/ETM+ images taken in January, June, and November. They found that the difference in the normalized difference vegetation index (NDVI) between January and November could best distinguish invasive shrubs (Amur honeysuckle) from native woody species based on the phenological difference among invasive and native species. Selkowitz (2010) compared shrub canopy mapping accuracies in Arctic Alaska by using multi-temporal Landsat ETM+, multi-temporal MODIS, and multi-spectral and multi-angle MISR data and applying the regression tree method, and found that the Landsat image from July exhibited the highest mapping accuracy. Combined with meteorological data, multi-temporal remote sensing data can help determine sensitivity difference to climate variance among shrub and grass species, and then effectively identify shrubs. Zhou et al. (2013) assumed that during a dry year with slight precipitation, the growth of leguminous shrub (C. microphylla) will be influenced less than that of grass species, and the difference in the growth status of C. microphylla between a dry and a wet year will be smaller than that of grass species. Two indicators were designed based on MODIS NDVI/enhanced vegetation index (EVI) according to the preceding assumption that considered the difference between maximal or accumulative NDVI/EVI values of dry and wet years, as follows:(1) $\text{DM}=\frac{\text{Ma}{\text{x}}_{\mathrm{wet}}-\text{Ma}{\text{x}}_{\mathrm{dry}}}{\text{Ma}{\text{x}}_{\mathrm{wet}}+\text{Ma}{\text{x}}_{\mathrm{dry}}}$(1) (2) $\text{DA}=\frac{\text{Ac}{\text{c}}_{wet}-\text{Ac}{\text{c}}_{\mathrm{dry}}}{\text{Ac}{\text{c}}_{\mathrm{wet}}+\text{Ac}{\text{c}}_{\mathrm{dry}}}$(2) where DM and DA respectively denote the difference in maximum NDVI (or EVI) and the difference in annual accumulated NDVI (or EVI), Max_wet and Acc_wet respectively denote the maximum and annual accumulated NDVI (or EVI) in a wet year, and Max_dry and Acc_dry indicate the same definition in a dry year. The two indicators were then used to identify C. microphylla in a steppe in Inner Mongolia (Zhou et al. 2013).

3.2.3. GO model and multi-angle remote sensing data

Shrubs exhibit a distinct 3D structure that differs from background herbaceous species in arid and semi-arid grasslands. Therefore, using a GO model and multi-angle remote sensing data to identify shrubs in grassland and estimate their coverage is feasible. Franklin et al. (1994) built a spatial discrete GO radiative transfer model for shrubs to simulate reflectance at a large spatial scale using herbaceous vegetation and soil in a savanna as background and disregarding shaded canopies and shadows on the soil. Chopping et al. (2006) used multi-angle Compact High-resolution Imaging Spectrometer/Project for On-board Autonomy (CHRIS/PROBA) data and a GO model to inverse shrub coverage in the experiment range in Jornada, New Mexico. CHRIS/PROBA data use 62 bands within the 415–1050 nm spectral range and 5 observation angles, with a spatial resolution of 17 m. The measured shrub density and average radius were used to parameterize the Walthall bidirectional reflectance distribution function (BRDF) model, and an isotropic-LiSparse-RossThin kernels-driven BRDF model was used to predict the Walthall background model coefficients. In the same experimental region, Chopping et al. (2008) used MISR data with nine observation angles and a spatial resolution of 275 m with a hybrid GO model for inverse shrub coverage inversion. The background models included volume scattering kernel weight from a Li–Ross BRDF model, diffuse brightness from a modified Rahman–Pinty–Verstraete BRDF model, other Li–Ross kernel weights (e.g. isotropic, geometric), and MISR near-nadir bidirectional reflectance factors (BRFs) in the blue, green, and near infrared bands. The forward modeling BRFs are in agreement with the MISR data in terms of shape and magnitude. However, the model inversion results of shrub coverage exhibit a relatively low agreement with HSR IKONOS data. Multi-angle, multi-temporal, and multi-spectral data can be used in shrub mapping; for example, Selkowitz (2010) used multi-temporal MISR data, IKONOS data, and the regression tree model to estimate shrub coverage in an Arctic area.

The advantages of moderate to coarse spatial resolution remote sensing data include low cost, large spatial coverage, and over 30 years of available dynamic data. However, their disadvantages are also evident. Shrub is difficult to precisely recognize from a shrub–grass–soil mix because of the spectral similarity between shrub and grass. Thus, vegetation indices, SMA, and other typical methods can be applied only to estimate shrub coverage under a scenario without grass (Xu et al. 2010; Vanselow and Samimi 2014). When multiple temporal remote sensing data are used, the 16-day revisiting interval of Landsat may miss the key phenological date. Meanwhile, remote sensing data with high revisiting frequency, such as MODIS, are characterized by low spatial resolution, which limits shrub identification or coverage estimation accuracy (Zhou et al. 2013). The major difficulty in applying GO models and multi-angle remote sensing data to desert grasslands is the spatial variation in combined soil and understory background reflectance; such variation cannot be modeled as a Lambertian scatterer or by using a fixed BRDF model (Chopping et al. 2008).

4. Dynamic modeling and simulation of shrub encroachment

4.1. Modeling of population structure

The growth pattern of shrubs, including regeneration and mortality, and the determination of shrub life span are of considerable value to the understanding of the shrub encroachment process and the prediction of possible changes in land cover and biodiversity (Wiegand et al. 2000). A precise population age structure model is an effective surrogate for understanding the growth pattern and reflecting the regeneration and mortality of a population. Crisp and Lange (1976) analyzed the influence of grazing on shrub population growth based on the shrub age structures of multiple regions under different grazing pressures. Their research predicted that shrubs in certain regions would be extinct after 100 years if grazing pressure would remain unchanged. Seligman and Henkin (2000) obtained shrub age structures after a fire by sampling several grasslands in the eastern Mediterranean area; they used the age structures to analyze shrub recovery mechanism after a fire. Kuuseoks, Dong, and Reed (2001) quantified the relationship between shrub regeneration/mortality rate and environmental factors, such as illumination and upper vegetation canopy density, to improve the understanding of the growth mechanism of understory shrubs. Myers-Smith et al. (2011) designed an age evaluation model for a red willow and then estimated the life span of the willow. The relationship between the shrub encroachment process and these disturbance events is analyzed by investigating historical natural disturbance events.

The shrub population structure is typically constructed by collecting field samples and then measuring the ages of the samples (Kuuseoks, Dong, and Reed 2001; Myers-Smith et al. 2011). A precise model of shrub population structure requires abundant shrub age samples to accurately estimate regeneration and mortality rates. However, one of the main problems in the construction of shrub population structure is the difficulty in collecting a sufficient number of field samples (Crisp and Lange 1976; Wiegand et al. 2000). The destructive measurement method for shrub age is another factor that limits the number of field samples. Cao et al. (2018) recently proposed a shrub age structure estimation model that depended on the relationship among shrub age, canopy size, and precipitation. Canopy size can be obtained from HSR remote sensing images. This technique can help build a precise shrub age structure model using a small number of field samples and a less destructive method.

4.2. Modeling of shrub growth

The simulation of shrub encroachment mainly relies on plant growth and spatial expansion models. Plant growth models describe the dynamic process of plant growth and hydrology with a series of functions. For example, the logistic function can be used to describe the dynamic process of shrubs and herbaceous plants, and it can involve influencing factors such as grazing pressure, climatic factors, and fire (D’Odorico, Okin, and Bestelmeyer 2012). The EcoHydrology in Drylands (EcoHyD) model (Tietjen et al. 2010) comprises a vegetation process model and a hydrological process model by calculating the annual growth process of shrubs, perennial herbs, and herbaceous plants in 2 weeks and the daily soil moisture contents of the two soil layers. EcoHyD can be used to simulate the influence of fire management measures on shrub encroachment (Lohmann et al. 2014). Chiou et al. (2013) used boosted regression trees to identify and quantify the potential factors influencing the likelihood of shrub invasion in Taiwan. The influencing factors include climatic conditions (temperature and precipitation), landscape features (altitude; slope ratio and aspect; percentages of natural or secondary forest, agriculture land, developed area, and water bodies; and distances to the nearest forest edge and river), and anthropogenic factors (length of forest edge, and distances to the nearest road and agriculture land), and result indicated the most influential variables are average annual temperature, altitude, precipitation, and slope.

Spatial expansion models typically simplify complex growth process models as a series of sub-models and use methods based on discrimination rules. Cellular automaton (CA) model was popularly used to simulate the spatial grass-shrub dynamics. Jeltsch et al. (1996, 1997) simulated dynamic changes in shrub and grass by considering precipitation, soil moisture, grazing, and fire within 20,000 5 × 5 m grids based on the CA model. Zhou, Istanbulluoglu, and Vivoni (2013) built a Cellular Automata Tree-Grass-Shrub Simulator (CATGraSS) to simulate a tree-grass-shrub coexistence environment in a semiarid climate. In CATGraSS model, topography determines spatial distribution of solar radiation and future the spatial variations in evapotranspiration, soil moisture, and plant water stress. Each cell has a single type, i.e. tree, shrub, tree seedlings, shrub seedlings, grass, or soil. Plant growth is driven by evapotranspiration, and plant establishment and mortality are linked to water stress. Caracciolo, Noto, and Istanbulluoglu (2014) and Caracciolo et al. (2016) used CATGraSS to simulate shrub encroachment in the northern Chihuahuan desert with new algorithms concerning fire, grazing, and seed dispersal caused by animals and wind effect. Cipriotti et al. (2012) designed a spatial simulation model–DINVEG to describe the long-term dynamics of dominant grass and shrub species in the semiarid steppe of Argentina. The vegetation dynamic is linked with water balance model, and a 50 × 50 m steppe plot with a spatial resolution of 0.2 × 0.2 m was modeled. Komac et al. (2013) built a state transition probability model of shrubs, grass, and barren land by considering influencing factors such as fire and mechanized shrub removal management measures; they simulated the shrub encroachment process at the patch scale (1 m grid). However, current simulations of shrub encroachment are mostly conducted at the patch and landscape scales, and shrub encroachment simulation remains lacking at the regional scale and during stage transformation.

5. Conclusion and discussion

Shrub encroachment into arid semi-arid grasslands has elicited research attention for decades. However, challenges remain in the quantitative analysis of shrub encroachment mechanisms, simulation and prediction of shrub encroachment, and dynamic monitoring at the landscape and regional scales. First, present studies on the shrub encroachment process and mechanism have involved integrated factors, such as water use, temperature, soil, and biological and anthropogenic activities. However, these results are mainly dependent on the observation and analysis of the natural shrub encroachment process over small areas. They lack support from large-scale observations or long-term field control experimental evidence. Second, no consensus has been reached on whether or under which conditions shrub encroachment can be restored due to the incomplete understanding of its process and mechanism in grasslands. Meanwhile, the shrub encroachment stages can hardly be quantitatively defined, and no indicator or method is available for the classification of shrub encroachment stages based on vegetation information from remote sensing inversion. A possible means to describe the status or stages of shrub encroachment is by combining vegetation coverage information with patch characteristics at the landscape scale. Third, multi-scale remote sensing presents challenges to the dynamic monitoring of shrub encroachment in arid/semi-arid grasslands. HSR remote sensing data cannot directly identify small shrubs. Furthermore, several disadvantages exist, such as the high cost of large range coverage and the lack of historical data. Thus, monitoring the spatial distribution of shrubs at the regional scale and the dynamics of shrubs for a long period is difficult. The use of moderate and coarse spatial resolution multi-spectral remote sensing data to identify shrubs and accurately estimate their coverage is difficult due to spectral mixture, lack of spatiotemporal continuous data, and limited spatial resolution. Further studies are required for methods that consider phenological differences or sensitivity to precipitation differences between shrubs and grasses. Finally, the quantitative analysis and simulation of shrub encroachment at different spatiotemporal scales and during the stage shift process remain insufficient due to the lack of understanding of the shrub encroachment mechanism and process. In summary, shrub encroachment dynamic monitoring based on remote sensing at the landscape or regional scale can play a crucial role in analyzing and understanding the process and mechanism of shrub encroachment into arid/semi-arid grasslands. The research on shrub encroachment should be multi-disciplinary and should involve multi-partnerships.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [grant number 41571406], the Fund for Creative Research Groups of National Natural Science Foundation of China [grant number 41621061], and the State Key Laboratory of Earth Surface Processes and Resource Ecology at Beijing Normal University [grant number 2015-ZDTD-011].