Major distribution shifts are projected for key rangeland grasses under a high-emission scenario in East Africa at the end of the 21st century

Grasslands and savannahs dominate large parts of East Africa and are critical for biodiversity, wildlife tourism, and the livelihoods of pastoralists and agro-pastoralists. Mobility enables pastoralists to cope with high climate variability, but increasing land fragmentation and access restrictions threaten this strategy. Grasses as foundation species underpin ecosystem structure and services such as forage provision, carbon sequestration, and hydrological regulation. This study focuses on seven native, perennial grasses important to East African rangelands—Cenchrus ciliaris, Cynodon dactylon, Cenchrus mezianus, Cynodon plectostachyus, Digitaria macroblephara, Digitaria milanjiana and Themeda triandra—chosen for their pastoral utility and ecological roles. Despite their importance, there is a lack of regional assessments of climate change impacts on these species, compounded by sparse historical mapping of grasslands and limitations of coarse-resolution global climate models for regional applications. The research questions address how climate change could alter suitable habitats for these key species, modify species composition and co-occurrence patterns, and what the implications may be for pastoralists, livestock, wildlife, and conservation planning. The study aims to provide fine-scale projections of current and future distributions under a high-emission scenario using dynamical downscaling coupled with machine-learning species distribution models, to inform strategic planning (e.g., corridors, conservation areas) in East Africa.

Previous work has shown the importance of grasses in ecosystem function, pastoral livelihoods, and conservation in East Africa. Only one prior climate impact study on a focal species (Cenchrus ciliaris) exists for southern North America, where it is invasive and projected to expand under warming and precipitation changes, highlighting the need for regional assessments in East Africa. Broader climate projections generally indicate warming and a tendency toward wetter conditions in East Africa, but downscaling is needed to resolve fine-scale climatic and topographic effects critical for SDMs. The literature also documents species-specific ecological traits of the seven focal grasses, including drought tolerance, grazing responses, and roles in fire ecology and soil protection. Additionally, land-use change, woody encroachment under elevated precipitation and CO2, overgrazing, land fragmentation, and invasive species are known to alter grassland dynamics, sometimes exceeding the direct influence of climate change in determining grassland trajectories. The scarcity of spatially explicit, long-term grassland records further complicates inference, motivating model-based projections.

Study region: East Africa with a focus on Kenya and bordering areas in Ethiopia, Somalia, Uganda, and Tanzania, subdivided into nine physiographic units representing arid lowlands, rift highlands/plateaus, and the coast. Environmental predictors: Initially 37 variables spanning topography (elevation, slope), hydrology (distance to rivers and water bodies), vegetation (tree cover from Hansen 2010; NDVI from Landsat-8 dry-season mosaic), soils (SoilGrids: sand, silt, clay, texture, CEC, pH, N, SOC, salinity, soil depth, soil class; Afrisoils: Al exchange capacity), and human pressure (Human Footprint Index). All rasters resampled to 1 km. Climate model chain: Global CESM1.04 provides historical (1850–2005) and RCP8.5 (2006–2100) forcing. Regional downscaling with WRF v3.8.1 using a parent domain at 27 km and a nested domain at 9 km over Kenya (49 vertical levels to 50 hPa). Two 30-year time slices: present (1981–2010) and future (2071–2100). Bias adjustment via a delta-change approach: WRF forced by ERA5 (1999–2018) forms the baseline; monthly differences between CESM-WRF future and present bioclimatic variables are added to the ERA5-based climatology to create adjusted future bioclim variables. Nineteen BioClim variables computed; five key bioclim variables used for change analysis (Bio1, Bio2, Bio4, Bio12, Bio14). All predictors resampled to 1 km. Species data: Presence records compiled for eight grasses from GBIF, sPlot, RAINBIO, VDEA, and SWEA-dataveg; sPlot and VDEA also provide real absences. Absences were thinned by a 5×5 km grid (one random absence per cell), removing absences within 5 km of any presence. Due to insufficient data and poor model performance, Pennisetum stramineum was excluded; seven species remained. SDM algorithms and evaluation: Four algorithms tested (MaxEnt, SVM, Random Forest [RF], Boosted Regression Trees [BRT]). Performance assessed via 10-fold cross-validation using AUC and TSS; only RF and BRT met accuracy thresholds (majority of species with AUC ≥ 0.8 and TSS ≥ 0.5). Collinearity reduction via Spearman’s |rho| > 0.6 yielded 11 final predictors: Bio1, Bio2, Bio4, Bio12, Bio14, slope, tree cover (TC), soil texture, CEC, HFI, distance to waterways (d2ww). Final RF (randomForest) and BRT (dismo, gbm.step, Bernoulli) models trained with tuned parameters (per species), then applied to future predictors (only climate variables changed; soils/landscape held constant). Model performance (RF TSS) ranged 0.52–0.75 across species. Binary maps and change categories: Presence probability thresholds set to each species’ TSS to derive binary suitable/unsuitable maps. Change classes per grid cell: absent (absent in both periods), no change (present in both), contraction (present→absent), expansion (absent→present). Co-occurrence maps computed as the count of species present per grid cell for present and future. Model verification and extrapolation check: CESM-WRF change signals compared to WorldClim CMIP5 downscaled ensemble (2061–2080, RCP8.5) showed similar magnitude but finer-scale, physically consistent patterns. Novel climate analysis identified limited areas, mainly in eastern domain, where future values exceed present ranges (notably mean annual temperature), warranting caution in interpretation there.

Climate change signal (RCP8.5, 2071–2100 vs. 1981–2010): - Warming of 2–3.5 °C, strongest over east Kenyan plains and Ethiopian highlands; least over Turkana and northern Kenyan highlands. Mean diurnal range (Bio2) decreases by ~2–3 °C (≈15–20%) with disproportionate increases in daily minimum temperatures. Temperature seasonality (Bio4) increases in Turkana and parts of Kenyan highlands, decreases over Serengeti plains, Tanzanian highlands, and Kitui County. - Annual precipitation (Bio12) increases widely: +400–600 mm year−1 in Turkana, up to +800 mm year−1 in central Kenyan highlands; ~+200 mm year−1 over much of the east Kenyan lowlands; decreases in south-eastern areas including Serengeti plains and Tanzanian highlands. Driest-month precipitation (Bio14) strongly increases along the coast and high elevations in Kenyan highlands (doubling or tripling), but decreases up to 10 mm month−1 in parts of northern Serengeti (some grid cells with zero rainfall in the driest month). Predictor importance (RF and BRT): - Tree cover (TC) is the most important predictor across species. Temperature variables (Bio1, Bio2, Bio4) are generally second in importance; precipitation of the driest month (Bio14) is most critical for Cynodon plectostachyus. Human Footprint Index (HFI) and CEC show minor influence. RF variable importance (IncNode Purity) highlights TC as top for multiple species; RF TSS per species: C. ciliaris 0.63; C. dactylon 0.58; C. plectostachyus 0.75; D. macroblephara 0.66; D. milanjiana 0.52; C. mezianus 0.70; T. triandra 0.72. Range changes (RF model, areas in 100 km² units; positive change indicates expansion): - Cenchrus ciliaris: +296 (≈+29,600 km²), +4.6% overall. Expansion in parts of higher elevations in the west; absence over much of Kenyan highlands and around Mount Kilimanjaro. - Digitaria milanjiana: +122 (≈+12,200 km²), +2.7%. Expansion in north Ugandan plains, Kenyan highlands, Serengeti plains, and Tanzanian highlands. - Cynodon dactylon: −919 (≈−91,900 km²), −16.6%. Contraction over east Kenyan plains, Kitui County, northern Lake Turkana, and northern Kenyan highlands. - Cynodon plectostachyus: −2456 (≈−245,600 km²), −44.4%. Nearly disappears from Lake Turkana region; strong contraction overall (RF), with smaller decrease in BRT (~−10.8%). - Cenchrus mezianus: −1659 (≈−165,900 km²), −34.6%. Contraction around Lake Turkana, NE Kenyan highlands, and eastern east Kenyan plains; some expansion in southern east Kenyan plains. - Digitaria macroblephara: −1047 (≈−104,700 km²), −16.7%. Patchy changes; contraction on the boundary of east Kenyan plains and Ethiopian highlands. - Themeda triandra: −763 (≈−76,300 km²), −18.8%. Contraction around Lake Turkana; slight expansions in east Kenyan plains and Tanzanian highlands (larger in BRT, smaller in RF). Co-occurrence: - Co-occurrence (number of species per grid) declines markedly over the greater Turkana region and much of arid eastern Kenya; increases in the west (north Ugandan plains, western Kenyan highlands, Serengeti plains, Tanzanian highlands). RF and BRT show similar regional patterns. Sparse data coverage in the east implies higher uncertainty there.

The study demonstrates that under a high-emission scenario, fine-scale, dynamically downscaled climate projections coupled with SDMs project substantial redistribution of key East African rangeland grasses by late century. Tree cover emerged as the dominant predictor of habitat suitability, consistent with competitive light limitation and the likelihood of woody encroachment under wetter conditions and elevated CO2. Temperature and precipitation patterns set broad limits and elevational bounds, while soils refine local distributions. Species-specific responses reflect nonlinear interactions among predictors: for example, range expansion of drought-tolerant Cenchrus ciliaris into warmer, drier, and some higher-elevation areas; contraction of Themeda triandra around Lake Turkana associated with increased temperature seasonality; declines in Cynodon dactylon potentially reducing dry-season forage and erosion control; strong contraction of Cynodon plectostachyus linked to increased precipitation and temperature seasonality in Turkana; and modest gains for Digitaria milanjiana in higher elevations. The aggregate decrease in co-occurrence over much of eastern Kenya and around Turkana signals a risk of reduced grass species diversity and potential loss of ecosystem function and grazing value, with implications for pastoral and wildlife mobility and increased risks of floods, erosion, and landslides in regions seeing both bioclimatic shifts and species declines. While climate is a key driver explored here, land-use pressures, overgrazing, woody encroachment, and invasive species can amplify or even dominate outcomes, potentially leading to further biodiversity loss and altered pasture availability. The results provide spatially explicit foresight to inform management, conservation planning, and the design of livestock and wildlife corridors under climate change.

Under present climate, arid lowlands of eastern and northern Kenya are broadly suitable for the studied grasses. By the end of the 21st century under RCP8.5, notable distribution shifts are projected: slight overall expansions for Cenchrus ciliaris and Digitaria milanjiana, but substantial contractions for Cynodon dactylon, Cynodon plectostachyus, and Cenchrus mezianus; Themeda triandra and Digitaria macroblephara generally contract. The Turkana region stands out with a projected near-absence of most studied species, driven by increased precipitation and temperature seasonality. Some higher-elevation areas may become newly suitable for certain species as temperatures rise. These shifts imply altered pasture composition and availability and potential changes in wildlife and pastoral mobility, elevating risks of resource conflict. The projections can guide long-term, climate-aware planning of corridors and protected areas. Future work should integrate vegetation transitions (e.g., woody encroachment), land-use change, human activities, and interspecific competition, and consider species beyond those included here. Results represent potential habitat suitability and do not account for human-led cultivation or reseeding of rangelands.

Key limitations include: (1) Data sparsity and spatial bias in presence/absence records, particularly in arid eastern Kenya, increasing uncertainty there. (2) Climate forcing based on a single GCM-RCM realization and a single high-emission scenario (RCP8.5); more robust precipitation change assessments would benefit from an ensemble of high-resolution simulations. (3) Use of a delta-change bias-adjustment approach may affect perfect physical consistency, though it is widely used for impact studies. (4) Soil and landscape predictors are held constant between periods; actual future changes (e.g., land degradation, woody encroachment) are not represented. (5) SDMs identify correlative, nonlinear relationships but do not establish causality; some projected changes arise from complex predictor interactions. (6) Novel climate conditions occur in parts of the domain (notably the east), implying some extrapolation beyond the training range. (7) Binary thresholding using TSS may simplify continuous suitability gradients. (8) Human management, grazing regimes, fragmentation, and invasive species dynamics are not explicitly modeled.