Bush encroachment dynamics and rangeland management implications in southern Ethiopia

Abstract Rangelands in southern Ethiopia have been undergoing a rapid regime shift from herbaceous to woody plant dominance in the past decades, reducing indigenous plant biodiversity, altering ecosystem function, and threatening subsistence pastoralism. Despite significant rangeland management implications, quantification of spatial encroachment extent and transitional pathways that result in encroachment remain largely under‐explored. This paper develops a phenology‐based approach to map rangeland vegetation states in southern Ethiopia, and examines transition pathways among states using the state‐and‐transition model. The results indicate that nearly 80% of landscape was dominated by woody plants in 2013. While stable encroached states have been established in both high and low lands through different transition pathways between 2003 and 2013, we identified spatial locations where bush encroachment occurred rapidly. The multiplicity in the transition pathways indicates opportunities for positive transformation in the entire rangeland system in southern Ethiopia and other semi‐arid regions of Africa.

In the Borana Zone of southern Ethiopia, bush encroachment has been intensifying over the past decades, which is becoming as a serious threat to both rangeland health and the livelihoods of millions of pastoralists (Angassa, 2014;Coppock, 2016). However, our poor spatial understanding of the extent and dynamics of bush encroachment in Borana continues to limit the effectiveness of rangeland management. Previous spatial evaluations produced only aggregate and seemingly contradictory estimates of encroachment extent within the Borana Zone: about 40% of Borana rangelands were estimated as encroached in the 1980s (Coppock, 1994), over 70% in the late 1990s (Oba, Post, Syvertsen, & Stenseth, 2000), and 52% in the 2000s (Gemedo-Dalle, Maass, & Isselstein, 2006). Such aggregate estimates revealed little about the complex spatial distribution of encroached rangelands or potential encroachment risk. Without a better spatial understanding of encroachment dynamics throughout the entire zone, favorable opportunities for rangeland management and encroachment mitigation may be missed, and substantial efforts and resources can be applied in areas where mitigation stands little chance of success. Filling this spatial information gap is, consequently, both crucial and long overdue.
The complexity of rangeland vegetation dynamics can be interpreted by the state-and-transition model, in which rangeland dynamics are described as a set of discrete "states" of vegetation at a specific site and changes between states that occur as discrete "transitions" (Briske, 2017;Milton & Hoffman, 1994;Westoby, Walker, & Noy-Meir, 1989). Transitions from one state to another often require a combination of climatic circumstances and management actions (e.g., fire or grazing) to bring them about (Mayer & Khalyani, 2011;Staver, Archibald, & Levin, 2011). Knowledge of the spatial distribution of vegetation states across the landscape is needed to appropriately fit intervention strategies to rangeland management practices where they will be most advantageous (D'Odorico, Okin, & Bestelmeyer, 2011). Yet, spatial knowledge of vegetation states alone is insufficient to guide effective management because efficacy also depends on the underlying mechanisms that bring about these states. Therefore, spatial knowledge of current vegetation states plus understanding of past and future transition pathways is needed to properly prescribe and apply efforts to mitigate undesirable processes such as bush encroachment.
The goal of this study was to provide pastoralists, rangeland managers, and policy makers with a spatial understanding of the past, current, and potential rangeland vegetation states in Borana, Ethiopia, which will then contribute to assessing spatial risk of bush encroachment and applying mitigation actions for rangeland management. Our specific objectives were to: (a) develop a phenologybased remote sensing approach to classify rangelands according to the state-and-transition theory, (b) investigate spatial rangeland states and distribution patterns, and (c) examine transition pathways among different states in order to enhance the prediction of vegetation dynamics. Our findings will aid in predicting directionality (and perhaps magnitude) of future vegetation dynamics and serve to improve the efficacy of bush encroachment mitigation efforts by identifying areas or situations to prioritize in southern Ethiopia and other similar semi-arid areas of Africa.

| Study area
This study was conducted in the Borana Zone (~44,000 km 2 ) of southern Ethiopia (Figure 1). Elevation here ranges from 500 to 2,500 m above sea level. Terrain varies from flat, dry river and lake beds in the lowlands to steep hillslopes in the highlands. Climate is largely semi-arid with relatively cool annual temperatures (19-24°C) for this tropical setting. Mean annual rainfall ranges from 300 mm in the lowlands to 1,000 mm in the highlands. Precipitation distribution is bimodal, with 60% in the primary rainy season (April-May) and 30% in the secondary rainy season (October-November). In between the rainy seasons are warm dry season (December-March) and cool dry season (June to September). Generally, annual precipitation is positively correlated with elevation (Coppock, 1994).

| Borana rangeland dynamics and hypotheses
Vegetation in Borana is dominated by herbaceous and woody plants in varying composition ratios, with woody plants cover ranging from 5% to 75% (Angassa & Oba, 2008;Coppock, 1994;Tefera, Snyman, & Smit, F I G U R E 1 Geographic location and topography of the Borana Zone in southern Ethiopia 2007). Such variability gives rise to a complex and diversely-vegetated landscape typical of the Horn of Africa (Cossins & Upton, 1987). Like many semi-arid rangelands, much of the Borana vegetation occurs as transitional phases between stable states of grassland and savanna and between savanna and woodland (Mayer & Khalyani, 2011). Temperature, precipitation, fire, grazing, and soil have all been described as crucial in the origin, maintenance, and shift of vegetation on tropical rangelands (Lehmann, Archibald, Hoffmann, & Bond, 2011;Stevens, Lehmann, Murphy, & Durigan, 2017). While climatic and edaphic factors primarily determine broad-scale vegetation distribution, complex patches of open and closed canopy rangelands can exist within a single climate zone, suggesting that controls such as fire and herbivory are important at a finer spatial scale (Beard, 1953;D'Odorico et al., 2011;Kraaij & Milton, 2006).
The state-and-transition model (Briske, Fuhlendorf, & Smeins, 2003, 2005 sheds light on interpreting the evolving nature of Borana rangelands at the local scale. Before the 1970s, fire was a major factor that determined rangeland states and transitions (Angassa & Oba, 2008), which could transform the rangeland into a grass-dominated system. Subsequently, plant recruitment, herbivory, and low-intensity fire would contribute to maintaining an open woodland/scrubland state, until the next high intensity fire to bring the entire system back to grassland. After fire ban since the 1970s, however, grazing and plant recruitment gradually transformed grasslands and open woodland into dense thickets. Continued grazing could diminish forage in the understory, thus leaving the scattered woody plants free from competition and accelerating bush encroachment (Liao & Clark, 2018).

| Data collection and rangeland classification scheme
We conducted fieldwork in the Borana Zone 2013 to build the foundation for a rangeland classification scheme. The research team traveled to both highland and lowland areas of the zone to investigate the diversity of rangeland vegetation in terms of physiognomic features, floristic composition, and phenological characteristics. In addition, we conducted vegetation surveys in five study sites that represent a wide range of ecological zones in Borana (see details in supplementary materials). At each site, we surveyed 27 plots and collected data on species composition and vegetation cover.
Voucher specimens were collected, and their scientific names identified (Supporting Information Table S1).
A couple of existing classification approaches, which were developed to distinguish and map African vegetation classes, likely have applicability in our study area. The White classification system, which is primarily based on vegetation physiognomy and floristic composition to achieve an objective classification framework, consist of 80 major vegetation types (White, 1983). In East Africa, a combination of the growth-form type and relative contribution of woody and herbaceous plants can lead to the differentiation of six major vegetation types: (a) bushland, (b) woodland, (c) grassland, (d) bushed grassland, (e) wooded grassland, and (f) dwarf shrub grassland (Pratt, Greenway, & Gwynne, 1966). Based on field vegetation assessment and existing classification approach by Pratt et al. (1966), we developed a classification scheme that includes eight classes representing the vegetation states potentially present in Borana (Table 1).
In order to link remotely-sensed vegetation indices with contextualized vegetation states, we collected geo-referenced photos of representative rangeland vegetation classes throughout the entire Borana Zone. The sample we collected was distributed across the Borana Zone covering both low and high lands (Supporting Information Figure S1). The vegetation types recorded by these photographs were classified based on the criteria in Table 1 We referred to both spectral and temporal traits of satellite imageries to map the spatial distribution of existing vegetation states in the Borana Zone. Specifically, we used the Normalized Difference Vegetation Index (NDVI) images with a spatial resolution of 250 m and temporal resolution of 16 days, which has proven to be effective for rangeland classification (Piao, Mohammat, Fang, Cai, & Feng, 2006). Since the training points were collected in the year of 2013, we used the 23 images from that year in our classification. We also obtained the 23 images in 2003 to derive the vegetation states a decade prior and compare those with 2013 conditions (Supporting Information Table S2).

| Data analysis
In order to capitalize on phenological differences existing among differing vegetation states for rangeland classification, we conducted supervised classification using the non-parametric random forest classifier. The classification was performed using the randomForest package (Liaw & Wiener, 2002) in the R software environment (R Development Core Team, 2014). In this study, the 23 NDVI images available throughout the year were considered 23 predictor variables. We randomly selected 75% of ground-truth data for use in training the classifier, and the remainder were used to assess the accuracy of resultant classification map by constructing a classification error matrix.
We also investigated the phenological features of each vegetation class. Generalized additive model (GAM) was used to investigate the phenological features of vegetation classes. The model can be represented as: where β 0 is the intercept; f(Date) is non-linear and is subject to smoothing splines; β 1 is coefficient of vegetation class; ε is the error and ε ~ N(0, s 2 ). The analysis was conducted at the pixel level.
However, there is a total of 688,100 pixels within Borana Zone, and this huge sample size was beyond the computational power of R software. Therefore, we randomly selected 1% pixels to perform GAM. Since the mean NDVI curve of open canopy woodland is the closest to sample mean, we used it as the benchmark factor in the model.
We used a digital elevation model (DEM) having a spatial resolution of 3-arc seconds (~90 m) to evaluate the topographic characteristics of each vegetation class. From this DEM raster image, we also estimated slope gradient in percent. Then, we overlaid the topographic raster images with the predicted rangeland map, and extracted the elevation and percent slope gradient for each classified vegetation at the pixel level. We conducted ANOVA to test whether there is any significant difference on the topographic characteristics among the eight vegetation classes.
In order to identify vegetation transition pathways from 2003 to 2013, we conducted a pixel-level change detection analysis (Tewkesbury, Comber, Tate, Lamb, & Fisher, 2015). We first constructed a vegetation change matrix from the two classified vegetation raster images in these two years, and then summarized the pixel-level transition from one vegetation class to another. We used the Gmisc package (Gordon, 2016) in the R statistical software to create a vegetation transition plot.

| Spatial vegetation classes and phenological features
The

| Vegetation state transitions
In

| D ISCUSS I ON
Although earlier research has investigated the diversity of rangeland vegetation and assessed the condition of bush encroachment in southern Ethiopia and other semi-arid areas of Africa (Angassa & Baars, 2000;Gemedo-Dalle et al., 2006;Pratt et al., 1966), little empirical work has been conducted to quantify the spatial extent of different vegetation classes and interpret vegetation dynamics.
Our analysis of rangeland vegetation states and transition pathways in Borana, while confirming the overall trend of encroachment as highlighted in literature (Coppock, 2016;Gemedo-Dalle et al., 2006), reveals the spatial distribution of different vegetation classes and the transition pathways that led to the proliferation of woody plants.
The findings refuted the perception of Borana rangelands as homogeneously encroached. Instead, the vegetation was highly diverse and varied substantially in its composition of woody and herbaceous plants. In addition, the findings increase our ability to predict which areas are at the nexus of being at a high risk of further encroachment, and shed light on developing spatially explicit strategies for mitigating bush encroachment and improving rangeland management in the arid and semi-arid environment.
Our integrated approach, which combined field-based assessment, time-series NDVI data, and random forest algorithm, represented a significant contribution to mapping bush-encroached rangelands (Reed, Schwartz, & Xiao, 2009) Table S4 NDVI is associated with better forage (Chantarat, Mude, Barrett, & Carter, 2013), we find that rangelands with higher NDVI values typically corresponded to more encroached vegetation states with low foraging value for cattle, while valuable open grasslands demonstrated consistently lower NDVI at any time of year than encroached vegetation states. In addition to revealing the phenological features of different vegetation types, the integrated approach has a potential to be widely adopted by local researchers where research infrastructure is relatively poor, because this approach relied completely on free data and analytical tools, and the entire research process is straightforward to replicate.
The application of state-and-transition model to hypothesize rangeland vegetation dynamics in Borana allows the identification of variability in the rates and pathways of encroachment among different vegetation states (Briske, 2017;D'Odorico et al., 2011), which differs from previous conceptions of homogenous encroachment (Angassa & Oba, 2008). Another primary transition pathway is from bushland to dense scrubland, representing a shift from the second dominant class into the most dominant class. Although the average elevation of bushland is about 900 m, it has a wide altitudinal range, and those at the higher end of the range is shifting into dense scrubland. Given the ongoing climate change that is favorable to woody plant establishment (Kulmatiski & Beard, 2013), the plant recruitment process is facilitating the thickening of woody layer of bushland.
Although dense scrubland represents what is likely a very stable state, it could be converted into other classes under human-imposed intervention (Gowda, Iams, & Silber, 2018). Dense scrublands, along with other minor classes such as closed and open canopy woodlands, that are situated at above 1,200 m are being converted to cultivated areas, which allows the practice of rain-fed agriculture. Even until the 1950s, crop cultivation throughout the Borana Zone was banned by indigenous rules (Tache & Oba, 2010). Changes started in the 1980s, when a severe drought hit the zone. Pastoralists then began to fence rangelands and cultivate maize and teff to make ends meet.
In recent years, commercial farming has become more prevalent, which has also contributed to cropland expansion at the expense of other rangeland vegetation classes.
One minor transition pathway in the highlands is from dense scrubland to closed canopy woodland. Given higher precipitation and absence of fire, tree seedlings in dense scrubland could grow into mature trees and gradually close the canopy (Mayer & Khalyani, 2011). Additionally, the government prohibited grazing in such forested areas for conservation purposes, which facilitated the transition into a closed vegetation state.
In the relatively dry lowlands of the Borana Zone, the primary vegetation transition is from grassland to bushland. Our changedetection analysis revealed that over a quarter the grassland extent was lost over the 2003-2013 decade, making it the least extensive vegetation state among the eight identified. Once the dominant landscape in the lowlands (Cossins & Upton, 1987), it is being encroached by invasive species such as A. mellifera and A. reficiens, which could potentially change soil features and make it more difficult for grasses to grow (Rundel et al., 2014). Without fire, the encroaching woody species can continue to establish and eventually dominate the landscape, unless manual bush clearing intervention programs are implemented to thin the woody plant layer (Liao & Fei, 2017). The sparsely vegetated state occurs on lands where the unfavorable soils and other conditions severely limit vegetation presence and productivity. Consequently, transitions from this state to any of the others are highly improbable, as evidenced by its negligible change in extent over the [2003][2004][2005][2006][2007][2008][2009][2010][2011][2012][2013] decade.
The identified transition pathways suggest potential opportunities for bush encroachment mitigation and rangeland management. In the highlands where the major pathways are from open canopy woodland and bushland to dense scrubland, a combination of browsing and fire can potentially reduce the woody plant layer and reverse the trend of encroachment (Staver, Bond, Stock, Rensburg, & Waldram, 2009). On the one hand, the graminoids and other herbs on the ground that are made inaccessible by dense shrubs and thickets for cattle grazing can serve as fuel loads to initiate a fire during the dry season. On the other hand, increasing browsing pressure by goats and camels can thin the woody plant layer and suppress the growth of shrubs and trees, which can indirectly facilitate the growth of herbs on the ground. In the drier lowlands where the transition pathway from grassland to bushland dominates, the chance to apply prescribed fire to control woody encroachment will be less feasible due to the lack of ground fuels (Archibald, Roy, Wilgen, & Scholes, 2009). In such locations, adding more goats and camels while reducing the number of cattle in the herds could be crucial to make better use of the woody forage resources that are well suited as food source for browsing (Liao, Ruelle, & Kassam, 2016). Rather than simply living with bush encroachment, pastoralists can actively contribute to its mitigation by changing their livestock portfolios.

CO N FLI C T O F I NTE R E S T
None declared.

AUTH O R CO NTR I B UTI O N S
C.L., P.E.C, and S.D.D designed the study. C.L. conducted fieldwork, collected data, and performed data analysis. C.L., P.E.C, and S.D.D wrote the article.

DATA ACCE SS I B I LIT Y
This publication is supported by MODIS NDVI data, which is publicly available at https://modis.gsfc.nasa.gov/data/dataprod/mod13. php. Field vegetation assessment data will be available for access on figshare.