Introduction

Land-use change, including deforestation, forest degradation and intensification of farming, threatens terrestrial biodiversity (WWF 2020; Meyfroidt et al. 2022). Land use and land cover (LULC) change is projected to intensify, exacerbating pressure on biodiversity (Powers and Jetz 2019) and endemic species in particular (Kobayashi et al. 2019). It is therefore vital to anticipate what some plausible future land-use changes might look like and how they would impact specific aspects of biodiversity.

Biodiversity modelling based on scenarios of land-use change can be useful for understanding plausible future biodiversity loss (e.g., Leclère et al. 2020; Sala et al. 2000). However, most studies have modelled biodiversity impacts at global and regional scales, using climate and LULC data at coarse (e.g., 1 km) spatial resolution (but see Nelson et al. 2009). Such broad scale approaches are valuable but cannot capture local nuances, which limits their usefulness for local stakeholders and local conservation planning (Franklin 2010a). Here, we focus on developing fine-scale analysis to support local decision-making.

Analyzing land-use change at the landscape scale—that is, within fine-grain land cover mosaics ranging in extent from some square kilometers to hundreds of square kilometers (Fischer and Lindenmayer 2007)—can be particularly useful, because landscapes can be conceptualized as interlinked social-ecological systems (Wu 2013; Oteros-Rozas et al. 2015; Fischer et al. 2017). However, landscape-scale scenarios are often ‘narrative’ in nature, without explicitly modelled biodiversity outcomes. Better connecting landscape-scale narrative scenarios to specific biodiversity outcomes would be valuable to address potential trade-offs between alternative plausible futures.

In this paper, we modeled woody plant species diversity under four plausible future social-ecological scenarios in southwestern Ethiopia, a recognized biodiversity hotspot (Mittermmeier et al. 2011). The region supports moist evergreen Afromontane forests (Hylander et al. 2013) and is the origin of Arabica coffee Coffea arabica (Senbeta and Denich 2006). Woody plant species have important social and ecological functions in the region. They support many other species including globally significant populations of birds and mammals and thereby contribute significantly to local biodiversity (e.g., Gove et al. 2008; Hylander and Nemomissa 2008). They also contribute to multiple ecosystem services, such as house construction, farm implements, poles and timber, fuel wood, animal fodder, soil fertilization and honey production, that support the livelihoods of local people (e.g., Ango et al. 2014; Shumi et al. 2021).

In this study, we build on our earlier translations of four narrative land use scenarios into spatially explicit land use scenarios (Duguma et al 2022). The scenarios are entitled Gain over grain (emphasis on cash crops such as coffee, khat and eucalyptus), Mining green gold (emphasis on coffee plantation), Coffee and conservation and Food first (emphasis on food crops) that are summarized in more detail in Sect. “Social-ecological scenarios and their implications for land use” below. Drawing on the corresponding land use maps of the present and the future, we used statistical models to relate empirical data on woody plant species richness to indicators of human disturbance and environmental conditions in order to make spatial predictions at the landscape level. Our goal was to spatially model total woody plant species richness (hereafter, total species richness) and forest specialist woody plant species richness (hereafter, forest specialist species richness), both for the present-day situation and for the four future land use scenarios. Our findings can help to raise awareness of the sensitivity of woody plant biodiversity to different social-ecological development pathways and can support ecologically informed decision-making within the study region. In particular, such spatially explicit biodiversity mapping can be reintegrated into further rounds of social-ecological scenario planning that consider the tradeoffs between ecological and socioeconomic outcomes.

Methods

Study area

The study area encompassed the forest and farmland mosaic of three administrative districts of Jimma zone of Oromia region in southwestern Ethiopia—namely, Gera, Gumay and Setema. Each district includes several smaller administrative units called “kebeles”, that are an important level of local management and decision-making (Fig. 1). Dominant woody plant species in the study area include Croton macrostachyus, Vernonia auriculifera, Vepris dainellii, Bersama abyssinia, Galiniera saxifrage, Syzygium guineense, Pouteria adolfi-friederici, Chionanthus mildbraedii, Dracaena afromontana, Allophylus abyssinicus, Coffea arabica and Millettia ferruginea (Shumi et al. 2018, 2019b). C. arabica is native to the region and naturally grows under the shade of trees in moist evergreen Afromontane forest. The study area is undulating and falls within elevations of approximately 1200 to 3000 m above sea level (supplementary Fig. S1). The landscape is dominated by smallholder farmers, whose main economic activities are cereal crop production, livestock rearing and coffee production. We divided the study area into farmland and forest (patches of woody vegetation ≥ 1 hectare; Fig. 1). Because we expected different social-ecological drivers to operate in these two very different environments, we modeled farmland and forests separately and then merged results to summarize for the whole landscape and by kebele.

Fig. 1
figure 1

A the study area in Jimma Zone, Oromia region (green grey) within Ethiopia (other regions are tan-colored); B the district boundaries (delimited by a thick black line and labelled in red) and lower administrative boundaries (kebeles; thin black lines) in the study area. The underlying land cover map illustrates the distribution of forest and farmland (from Duguma et al. 2022)

Full size image

Social-ecological scenarios and their implications for land use

Four qualitative narrative scenarios (i.e., ‘Gain over grain’, ‘Mining green gold’, ‘Coffee and conservation’ and ‘Food first’) describing potential social-ecological conditions in 2040 were developed through participatory scenario planning (Jiren et al. 2020). These considered plausible environmental, social and economic changes and were translated into spatially explicit, quantitative land use maps (Fig. 2, supplementary Fig. S2A–E) (Duguma et al. 2022).

Fig. 2
figure 2

A A cross-sectional illustration of the present landscape and the four scenarios (reproduced from Jiren et al. 2020) and B spatially explicit, quantitative LULC maps of the current and the scenario landscapes (reproduced from Duguma et al. 2022). High resolution versions of the maps are provided in Fig. S2A-E

Full size image

Briefly, the ‘Gain over grain’ scenario prioritizes farmers’ specialization and commercialization to boost development: traditional food cropping is abandoned in favor of cash crops such as coffee, the stimulant drug khat (Catha edulis) and fast-growing trees (e.g., Eucalyptus). The ‘Mining green gold’ scenario is characterized by the intensification of coffee production through large investors who use modernized, high-input production: farmland and forests conducive for coffee investment have been transferred to capital investors for the creation and expansion of coffee plantations. The ‘Coffee and conservation’ scenario is based on a more balanced land use approach: a diversified mosaic of forest and farmland consisting of a core zone with unused natural forest, a buffer zone for low-intensity production of local coffee, wild honey and other forest products and an outer zone with interspersed cropland, pastures and tree plantations. The ‘Food first’ scenario is characterized by the production of large amounts of cereal crops through intensive, large-scale agriculture, which involves extensive land consolidation, including the clearing of woody vegetation and expansion of cropland on flatlands and drained wetlands (Fig. 2).

Data

Woody plant diversity

Woody plant species of all ages (with a minimum height of 150 cm) were surveyed in 72 farmland sites (Shumi et al. 2018) and 108 forest sites (Shumi et al. 2019b). Sites were 20 m × 20 m quadrats, which were stratified across major landscape gradients. Species were classified into forest specialist, generalists and pioneer species. Here, we focused on two response variables: total species richness and forest specialist species richness.

Candidate predictors

We identified fifteen candidate predictors based on our ecological knowledge of the landscape, as well as data availability for the baseline and future scenarios. These included topographic variables and indicators of human disturbance. Coarse-scale climatic variables and latitude/longitude information were considered less important locally than topography and were therefore not included (Table 1, (see also Franklin 2010b)). Topographic variables were derived from the 30 m ASTER DEM (NASA/METI/AIST/Japan Spacesystems and U.S./Japan ASTER Science Team 2009). As measures of human disturbance, we included current and historical distance of sites from the forest edge, farmland type, forest type, percent woody vegetation at different scales (at 1 ha, 200 m, 500 m and 2 km) and land cover diversity (Table 1).

Table 1 Variables used in the model building process
Full size table

Variable preprocessing is detailed in the supplementary (S1) and summarized here. We examined predictor variables using histograms and transformed if this improved normality. We removed strong co-linear variables. We included quadratic terms if these were indicated by theory and supported by the data. We removed variables that were deemed unreliable (due to remaining skew resulting in poorly supported and unexpected bivariate relationships).

Model selection

We developed four models for woody plant species across farmland and forest areas (two in each environment), using total species or forest specialist species richness as outcome variables. As our aim was to produce spatial predictions, our model selection process focused on finding models that would both fit the data well and perform well in the prediction processes. We used generalized linear modeling due to the interpretability of these models with respect to the predictor variables. The model evaluation and cross-validation process is detailed in the supplementary material (S1) and summarized here. To be able to test the predictive power of the models and avoid overfitting, we first split the data into training and testing sets based on key environmental PCA dimensions (S1). Models were fit and selected on the training set and evaluated in terms of root mean square error (RMSE) of the (1) training set predictions, (2) test set predictions, (3) k-fold cross-validation across the whole dataset, with sevenfold based on environmental strata and (4) k-fold cross-validation across the whole dataset, with 10 random folds. We conducted modelling in R v 4.1.2 (R Core Team 2019), including the packages identified below.

The model selection process (detailed in S1 and summarized here) started by selecting the appropriate regression family. We first fit a Poisson GLM (log link) to the full models and modified this to a negative binomial (R package: MASS; Venables and Ripley 2002) where overdispersion was indicated (i.e., for the total species richness models in forest and farmland). For the model of total woody species richness in farmland we fit truncated (> 0) Poisson and negative binomial GLMs (log link; R package: glmmTMB; (Brooks et al. 2017)), as the sampling only took place in plots where woody vegetation was present (i.e., the data had no zero values). We then reduced these full models by identifying the top subsets of each of these that performed well either in terms of AICc or mean RMSE from the random-strata cross-validation (identified with the R package MuMIn; Barton 2020). We selected 5–10 models per task which showed high accuracy and indicated a good model fit (R package DHARMa; Hartig 2021). We re-fit these on the whole sample data and repeated the examination of the models and model residuals. Where these indicated poor fits, we repeated the selection process with modified or additional variables, including interaction terms. If spatial autocorrelation was indicated, we accounted for this by including the spatial covariance matrix as an exponential spatial random effect (R package: glmmTMB; Brooks et al. 2017).

Final models were then chosen based on the spatial predictions conforming with theoretical expectations and expert opinion across all scenarios. We projected out selected models (2–3 per modelling task) to the study region, including for the current landscape and future scenarios. Models with spatial components were projected without known random (spatial) effects to better reflect underlying patterns driven by the predictors. To ensure that our models were not extrapolated into substantially novel parameter spaces, we constrained the input raster predictor variables to a floor and ceiling observed in our field-based data. This constraint produced predictions consistent with current observed parameter ranges for most models. We further constrained the maximum value output for the model on total species richness in forest, as the prediction ranged substantially higher in a small subset of cells for this model, due to novel parameter combinations. We checked models for logical consistency, specifically that total species richness > forest specialist species richness scores for both the forest and farmland.

As our models did not explicitly distinguish coffee plantations from other types of forest cover—because intensive plantations are currently rare in the landscape—in the ‘Mining green gold’ scenario we applied a ceiling of four total species and two forest specialist species in plantation areas, based on highest values of species richness in the forest plots sampled that most closely resembled intensified coffee plantations.

Spatial predictions and model summaries at landscape and kebele level

The main purpose of the modelling was to produce spatial predictions of woody plants species richness for the current landscape as a baseline and the four scenarios. In each case, the results for farmland and forest were merged to produce a single predicted map for the entire study area (i.e. the landscape). Finally, we summarized our results at the kebele level to support local decision-makers using ArcGIS 10.6.1. Specifically, we subtracted the mean predicted values in each kebele for each scenario from the mean predicted values of the current landscape to obtain average changes in species richness at kebele level.

Results

Selected models

In farmland, total species richness increased with a decrease in historical distance to the forest edge, with an increase in land cover diversity and with an increase in slope; and decreased with topographic wetness index (Table 2). Forest specialist richness peaked at intermediate elevation, while it decreased with both current and historical distance from the forest edge and increased with greater land cover diversity (Table 2).

Table 2 Modeling results for the farmland for total and for forest specialist richness of woody plants
Full size table

Total species richness increased in primary forest and decreased with distance from the forest edge. In addition, total species richness peaked at intermediate levels of both elevation and percent woody vegetation cover (Table 3). Forest specialist richness in forest showed similar relationships with forest type, elevation and percent woody vegetation cover. Even though current distance was not included in the selected model for forest specialist, forest specialist richness did appear to increase with increasing distance from forest edge (Fig. 3).

Table 3 Modeling results for the forest for total and for forest specialist richness of woody plants
Full size table
Fig. 3
figure 3

Spatial predictions under the baseline and future scenarios for total species richness (upper panels, A–E) and forest specialist species richness (lower panels, F–J)

Full size image

Spatial predictions across the landscape

Higher total species richness was found in forest compared to farmland, with scenarios differing largely due to the distinct distribution of woody vegetation in these main two habitats (Fig. 3). Under the ‘Gain over grain’, ‘Coffee and conservation’ and ‘Food first’ scenarios the distribution curves showed two peaks of total species richness near five and 19, depicting their relative distribution in farmland and forest, respectively (Fig. 4). However, while the ‘Gain over grain’ and ‘Coffee and conservation’ scenarios saw a slight general increase in species richness compared to the current baseline, the ‘Food first’ scenario experienced a general decline in species richness (Fig. 4). Under ‘Mining green gold’ where large parts of the forest were lost, the distribution curve for total species richness was concentrated around four.

Fig. 4
figure 4

Density distributions showing changes in woody plants species richness under the scenarios (GoG ‘Gain over grain’, MGG ‘Mining green gold’, CC ‘Coffee and conservation’, FF ‘Food first’) against the current landscape (CL). The left panels A–D illustrate changes in total species richness, while the right panels E–H present changes in forest specialist species richness

Full size image

Patterns for forest specialist were similar to total richness. Forest specialist richness was lower than total species richness overall, and forest specialist diversity was higher in forest than in farmland, especially towards the interior of the forest (Fig. 3). Peaks in their distributions were around one in farmland and 14 in forest (Fig. 4).

Changes in woody plant species richness at the kebele level

For total species richness, most kebeles showed little to no change in the ‘Gain over grain’ scenario, relative to the baseline, with only two kebeles (corresponding to two district towns) showing a loss in total species richness (Fig. 5). Both the ‘Mining green gold’ and ‘Food first’ scenarios showed losses in total species richness across a majority of kebeles. Changes were more heterogeneous for ‘Mining green gold’ compared to the more uniform responses under other scenarios, but also included the strongest losses where some kebeles lost up to 14 species of 18 total species. In the ‘Food first’ scenario, most kebeles lost one to four species of woody plants. In contrast, no kebele experienced losses in woody species richness in the ‘Coffee and conservation’ and more kebeles increased in species richness under this than in any other scenario (Fig. 5). Broadly speaking, similar patterns were found for forest specialist species (Fig. 5).

Fig. 5
figure 5

Changes in mean species richness at the kebele level. The mean changes for each scenario were calculated against the current landscape as a reference point. The upper panels A–D show total species richness, while the lower panels E–H depict forest specialist richness

Full size image

Discussion

Models of biodiversity at the landscape scale are powerful tools for natural resource management (Mateo et al. 2018). Here, we combined landscape-scale biodiversity models, including locally important variables that influence biodiversity, with future land use maps generated from participatory scenario planning to spatially predict plausible future changes in woody plant diversity. Our approach revealed clear differences in biodiversity responses among the scenarios. Biodiversity declined strongly in two scenarios, the ‘Mining green gold’ and ‘Food first’ scenarios. Slightly positive responses were observed for the ‘Gain over grain’ scenario, while richness was most positively affected in the ‘Coffee and conservation’ scenario. Below, for each scenario, we provide a more detailed discussion of possible trends at the landscape and the kebele levels, before considering general implications for conservation.

‘Gain over grain’

At the landscape scale, richness values changed only slightly from the current baseline in this scenario. In farmland, slight increases resulted from degraded farmland on steep slopes being covered by eucalyptus plantations (Duguma et al. 2022). Forest specialist richness showed no change mainly because it was associated with primary forest cover (Shumi et al. 2018), which remained unchanged in this scenario. At the kebele scale, most kebeles showed low to moderate changes in woody plant richness. Only two kebeles, namely those that were assumed to experience a strong future growth in human settlements (Duguma et al. 2022) lost species under this scenario.

Notably, smallholder-driven land use intensification as simulated in this scenario could have a positive impact on income and employment in rural households (Appelt et al. 2022). However, although our models showed slightly positive effects on biodiversity in the short term (20 years), in the long-term, especially the expansion of eucalyptus plantations may negatively impact biodiversity and thus ultimately also may be negative for local ecosystem services and human wellbeing (Rasmussen et al. 2018; Beckmann et al. 2019). Furthermore, our analysis focused on biodiversity impacts that follow spatial changes in land use and cover. An increased use of agrochemicals under the ‘Gain over grain’ scenario, which is already common practice in more intensified cash crop landscapes elsewhere in Ethiopia (e.g., Gessesse Dessie 2013; Cochrane and O’Regan 2016), may contribute to future negative impacts on biodiversity that our current analysis was not able to capture.

‘Mining green gold’

This scenario showed strong losses in total and forest specialist species richness compared to other scenarios, because this scenario was mainly driven by the clearance of small patches of woody vegetation in farmland within future coffee elevation and by conversion of forest to large scale coffee plantation. Our results are consistent with evidence of a rapid decline in woody plant biodiversity with increasing yields in coffee agroforests, especially under intensive management (Geeraert et al. 2019; Zewdie et al. 2022). Similarly, investments intended for the production of export-oriented commodities often result in strong biodiversity losses in tropical forest (Davis et al. 2020).

Notwithstanding the overall negative effects of this scenario on biodiversity, effects among kebeles were more heterogeneous compared with other scenarios. High and low elevation kebeles (which were mainly outside the elevation range suitable for coffee) that were dominated by arable land and with little woody vegetation showed no change in both total and forest specialist species richness. Other kebeles showed particularly strong losses in species richness (up to 14 total species and 10 forest specialist species; Table S1); these kebeles had high proportions of woody vegetation (from 73 to 95%) and were located at intermediate elevations, that are most suitable for growing coffee (Hylander et al. 2013; Moat et al. 2017). Future differences in land use and landscape suitability for specific crops thus drove diverging responses in this scenario.

‘Coffee and conservation’

Positive outcomes for the diversity of woody vegetation were observed for this scenario, which was largely based on the concept of sustainable land management (Jiren et al. 2020). The resulting heterogeneous landscape allowed for an increase in both total and forest specialist species richness in previously homogeneous farmland areas. Indeed, this scenario reflected key characteristics for biodiversity-friendly farming such as the maintenance of native vegetation, the improvement of heterogeneity and structural complexity and an avoidance of the use of agrochemicals (Perfecto and Vandermeer 2010; Fischer et al. 2013). Moreover, the scenario would help to protect the gene pool of wild Arabica coffee through conservation of the last remaining coffee forests, which is considered a high conservation priority (Aerts et al. 2015). The positive effects on woody vegetation at the landscape level resulted in relatively consistent, positive outcomes at the kebele level (Fig. 5). Our results thus suggest that the establishment of a new biosphere reserve, or similar approach that would result in a sustainably managed, diversified landscape, could make major contributions to biodiversity conservation in the region. Especially if published guidelines for the establishment and management of biosphere reserves are followed (Stoll-Kleemann and Welp 2008; Van Cuong et al. 2017), it is highly plausible that the ‘Coffee and conservation’ scenario would indeed contribute to the urgently required conservation of Ethiopian coffee forest. The creation of a new biosphere reserve would lead to a spatially connected cluster of several biosphere reserves within southwestern Ethiopia (Urban and Beswick 2018)—that together, would generate a currently unreached level of regional connectivity for the moist Afromontane forest biome.

‘Food first’

This scenario was associated with major losses in total and forest specialist species at both the landscape and kebele levels, largely due to decreased farmland heterogeneity and significant deforestation in the wake of large-scale agricultural land expansion and intensification. Especially forest specialist declined in response to agricultural expansion and intensification that homogenized the landscape (Fig. 4). This scenario would represent a continuation of current trends of forest change in our study region, where deforestation has increased since 1973, especially in areas not suitable for growing (forest) coffee (Hylander et al. 2013). While forest loss already is widely acknowledged as a primary driver of biodiversity loss (Le Roux et al. 2019; Caro et al. 2022), our results further support the important role of forests for future biodiversity conservation. Beside its effect on biodiversity loss, several studies in Ethiopia and Rwanda have also revealed negative social-ecological impacts of large-scale intensive farming on smallholder farmers, including losses in the quantity and diversity of food and income, as well as reductions in forest-based ecosystem services (e.g., Moreda 2017; Kim et al. 2022).

General conservation implications

For forest specialist, which have a high conservation priority (Shumi et al. 2019b, a), differences between primary and secondary forests and percent woody vegetation cover were the main driving factors in forest, whereas current and historical distance from the forest edge and land cover diversity were the main driving factors in farmland. Our results also indicate that woody species in general are sensitive to local differences and changes in future landscapes within our study area. Two scenarios, ‘Mining green gold’ and ‘Food first’, resulted in relatively strong losses of biodiversity, while two other scenarios, ‘Gain over grain’ and ‘Coffee and conservation’, resulted in slight and relatively strong gains in biodiversity, respectively.

The ‘Mining green gold’ and ‘Food first’ scenarios are relatively well-aligned with Ethiopia’s current large-scale land investment policy that intends to improve food security, earn foreign currency, generate income from land rent fees and create employment (Rahmato 2011; Shete and Rutten 2015). However, existing government policy that favors agricultural intensification through highly hierarchical biodiversity and food security governance (Jiren et al. 2018) largely ignores the critical importance of woody plant species diversity and associated ecosystem services for local communities (e.g., Kassa et al. 2016). Future regional development strategies and land use policies therefore need to more fully recognize the needs and practices of local communities with respect to managing woody vegetation.

Furthermore, especially in highly forested areas, large-scale agricultural intensification for monoculture crops typically leads to losses of biodiversity and ultimately has negative impacts on many important ecosystem services and ultimately also on human-wellbeing (Rasmussen et al. 2018). Thus, these two scenarios are unlikely to be sustainable, or at the very least would necessitate exceptionally proactive measures to avoid losses in biodiversity and important ecosystem services (Tadesse et al. 2014; Shumi et al. 2021). While the ‘Gain over grain’ scenario indicates slightly positive biodiversity responses, this scenario involved land use and management changes that could negatively affect biodiversity in the long run and which are not captured by our spatial analysis. Proactive management thus would also be needed under this scenario to avoid unsustainable outcomes.

The ‘Coffee and conservation’ scenario showed the most clearly positive changes for biodiversity and also could result in longer-term sustainable outcomes. In addition to conservation benefits, the scenario may lead to positive socioeconomic outcomes for smallholder households, who closely depend on the ecosystem services generated from the landscape (Tadesse et al. 2014; Ango 2018; Shumi et al. 2021). While the short-term socioeconomic gains may be more moderate compared to the benefits that a limited group of households may experience under the ‘Gain over grain’ scenario, whatever gains there are would possibly be distributed more equitably across the study region and among households. Our work thus supports the idea of “working lands conservation” (Kremen and Merenlender 2018) being able to support both biodiversity and goods and services for humanity, while maintaining the abiotic conditions necessary for long-term sustainability and resilience. The creation of an additional biosphere reserve, to form a cluster with existing other reserves, could therefore represent a sustainable trajectory for southwestern Ethiopia.

Future research and limitations

We have shown in our research how spatially explicit land use and land cover scenarios that were co-generated with stakeholders, can be used to map and predict woody plant biodiversity under different social-ecological conditions. We acknowledge that there are limitations to our approach—for example it can be challenging to translate narrative scenarios into spatially explicit maps (e.g., Duguma et al. 2022) and maps cannot fully capture all aspects of social change that may take place in the future. Moreover, a realistic outcome for the study area may well be a combination of aspects of all different scenarios, or different aspects of the scenarios may manifest in different parts of the landscape. Future research should further explore these questions, for example by working with stakeholders to identify which aspects of different scenarios are most desirable in different sections of the study area.