Species occurrence records for Lantana camara, Chromolaena odorata, and Pontederia crassipes were compiled from field observations, global biodiversity databases (including GBIF: https://doi.org/10.15468/dl.hyx9xn and iNaturalist) and records provided by a collaborating researcher. We used R and the readxl, dplyr, sf, raster and sp packages for data preprocessing. We kept only records with valid geographic coordinates, converted coordinate fields to numeric values, removed missing or incorrect records and projected all data to WGS 84 (EPSG:4326). Finally, we cropped the cleaned records to the Mara and Simiyu Regions in northern Tanzania (33.0°E–36.0°E; −3.0° to 0.0° latitude).

To reduce spatial sampling bias and prevent excessive clustering of records, we filtered the data using a 1-km grid-based thinning method in R using the raster and sf packages. We created a raster grid with a resolution of 0.008333° to approximate 1 km and kept only one record per grid cell, using cell identifiers from the environmental grid. After this filtering, we had 234 records for L. camara, 104 for C. odorata and 21 for P. crassipes for modelling. (Table 1).

In the third step, we addressed spatial sampling bias in presence-only data by using bias-weighted background sampling, following the target-group approach from Phillips et al. (2009). We created a kernel density surface from GBIF vascular plant records and used it as a bias file in MaxEnt to guide background-point selection. This ensured that background sampling matched the spatial distribution of survey effort, reducing the chance that model estimates were affected by sampling intensity rather than species–environment relationships. By combining spatial thinning, grid-based filtering and bias-weighted background sampling, we built a strong framework to minimise spatial bias and improve model reliability at both local and regional levels.

To model the potential distribution of the three invasive alien plant species, we selected environmental predictors representing both climatic and land-use drivers. We selected predictors based on their ecological relevance to plant establishment, growth, survival, dispersal, and habitat suitability, as well as on previous species distribution modelling studies of invasive plants (Guisan and Thuiller 2005, Elith et al. 2011,Booth et al. 2014). Bioclimatic variables capture environmental conditions that directly influence plant physiological performance and geographic distributions, whereas land-use and land-cover variables represent habitat disturbance, landscape modification and anthropogenic pressures that facilitate biological invasions. Integrating both predictor groups enabled the assessment of the combined influence of climate and land-use change on future invasion risk.

We modelled the current and future potential distributions of Lantana camara, Chromolaena odorata and Pontederia crassipes using MaxEnt v.3.4.3 (Phillips et al. 2006), implemented through the ENMeval v.2.0.5.2 framework (Muscarella et al. 2014, Kass et al. 2021). Each species was modelled independently using its own occurrence dataset and tuning procedure. Models were fitted using presence-only occurrence data and environmentally heterogeneous predictors.

For all species, predictor sets initially included seven bioclimatic variables selected after filtering for multicollinearity. For the two terrestrial invaders (L. camara and C. odorata), we additionally incorporated land-use information using ESA land-cover data for the Mara–Simiyu Region. To retain landscape heterogeneity, we aggregated categorical land-cover layers to derive continuous fractional-cover predictors representing the proportion of each land-use class within each 1-km cell. This approach reduces information loss associated with categorical resampling and aligns with best practice for incorporating land-use data into species distribution models (Franklin 2010, Guisan et al. 2017). All predictors were resampled to a standard 30-arc-second grid and converted to the Terra SpatRaster format prior to analysis.

Due to the limited number of spatially filtered occurrences available for P. crassipes (21 records after 1-km filtering), model complexity was deliberately constrained to reduce the risk of overfitting. Overfitting is particularly problematic in species distribution models with small sample sizes and high predictor dimensionality (Breiner et al. 2015). Previous studies have demonstrated that reliable MaxEnt models can be developed with limited occurrence records when model complexity is appropriately controlled (Hernandez et al. 2006, Galante et al. 2018). Accordingly, we restricted predictors for P. crassipes to four ecologically interpretable, weakly correlated climate variables and limited feature classes to linear and quadratic forms. Additionally, because P. crassipes is an aquatic species whose distribution is primarily constrained by freshwater habitats, we did not include land-use variables in the final model projections and used climate-only models for subsequent analyses.

Model complexity was optimised using ENMeval by evaluating combinations of feature classes (L, Q, H, LQ, LQH) and regularisation multipliers ranging from 0.5 to 3.0. For each species–predictor set (climate-only and climate plus land use), the optimal model was selected using the minimum Akaike Information Criterion corrected for small sample sizes (AICc), thereby balancing goodness-of-fit and model parsimony (Warren and Seifert 2011, Kass et al. 2021).

Model performance was evaluated using multiple complementary metrics, including the Area Under the Receiver Operating Characteristic Curve (AUC), Continuous Boyce Index (CBI), omission rates (OR10 and ORMTP) and the Akaike Information Criterion corrected for small sample sizes (AICc), all generated using ENMeval. Validation AUC, calculated from spatially independent block cross-validation, was used to assess the model's ability to discriminate between suitable and unsuitable habitats (Roberts et al. 2017, Valavi et al. 2022). Detailed model performance statistics for all candidate models are provided in Suppl. material 2. We used the Continuous Boyce Index (CBI) to evaluate agreement between predicted habitat suitability and observed occurrence records; it is particularly appropriate for presence-background models such as MaxEnt (Boyce et al. 2002, Hirzel et al. 2006). Omission rates (OR10 and ORMTP) were used to quantify prediction errors under commonly applied threshold criteria, while AICc was used to identify the optimal model by balancing model fit and complexity (Warren and Seifert 2011, Muscarella et al. 2014). Together, these complementary metrics provided a robust assessment of model discrimination, calibration, prediction accuracy and model complexity, reducing reliance on a single evaluation metric.

MaxEnt continuous suitability outputs (cloglog transformation) were retained for interpretation and for projecting habitat suitability under current and future climate scenarios (Phillips et al. 2006, Elith et al. 2011). For map-based interpretation, suitable–unsuitable maps were derived using three widely applied threshold criteria: the 10^th percentile training presence (P10), equal sensitivity and specificity (ESS) and maximum sensitivity plus specificity (maxSSS) (Liu et al. 2005, Liu et al. 2016, Hellegers et al. 2025). We calculated thresholds using presence records and background points within the MaxEnt modelling framework. Employing multiple threshold criteria enabled comparison amongst species, while explicitly acknowledging uncertainty associated with threshold selection.

We projected future habitat suitability under four CMIP6 Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, SSP3-7.0 and SSP5-8.5), representing low- to high-emission futures (Mölg and Pickler 2022). We used three CMIP6 General Circulation Models (GCMs) for future climate projections: ACCESS-CM2, MIROC6 and MRI-ESM2-0. We chose these models because they provide complete coverage across all SSPs and future periods (2030, 2050, 2070 and 2090), offer bias-corrected and downscaled bioclimatic variables at approximately 1-km spatial resolution through WorldClim v.2.1, and represent different climate modelling approaches. This helps capture a wider range of projection uncertainty. In addition, CMIP6 models have demonstrated good performance for climate simulations in East Africa (Ayugi et al. 2021, Mölg and Pickler 2022, Makula and Zhou 2022). Using multiple GCMs rather than relying on a single climate model improves the robustness of future habitat suitability projections by accounting for uncertainties associated with climate model structure and emission scenarios.

We calibrated MaxEnt models with current data and projected them using the cloglog output, which provides a continuous, interpretable index of relative occurrence probability suitable for scenario and time-period comparisons (Phillips et al. 2009). We combined climate projections with land-use layers corresponding to each SSP for 2030, 2050, 2070 and 2090 and matched them to the relevant bioclimatic periods (2021–2040, 2041–2060, 2061–2080 and 2081–2100). We resampled and aligned all predictors to a common 1-km grid before projection.

For each species, SSP and time period, we projected suitability separately for each GCM. We then generated ensemble predictions by averaging across the GCMs with equal weights. To assess model uncertainty, we quantified the projected suitable area across global climate models (ACCESS-CM2, MIROC6 and MRI-ESM2-0). For each species, year and SSP, suitable habitat was defined as grid cells with predicted suitability ≥ the 10^th percentile training presence threshold. The total suitable area (km²) was calculated separately for each GCM within the AOI, and inter-model uncertainty was summarised using the mean, standard deviation and coefficient of variation of suitable area. For each species, year and SSP, suitable habitat was defined as grid cells with predicted suitability at or above the 10^th percentile training presence threshold. Total suitable area (km²) was calculated separately for each GCM within the AOI, and inter-model uncertainty was summarised using the mean, standard deviation and coefficient of variation of suitable area. We produced ensemble suitability maps, uncertainty layers and area-based summaries. We retained continuous cloglog predictions for all modelling, evaluation and uncertainty analyses.

For visualisation and area-based summaries, we converted continuous MaxEnt cloglog outputs into discrete suitability classes using statistically derived thresholds rather than arbitrary value ranges. Specifically, the 10^th percentile training presence (P10) threshold was used to distinguish between unsuitable and suitable conditions. The P10 threshold is widely used in presence-only species distribution modelling because it accommodates potential geolocation error and environmental uncertainty by allowing a small proportion (10%) of training occurrence records to fall below the suitability threshold (Liu et al. 2005, Pearson et al. 2007, Radosavljevic and Anderson 2014). The methodological steps used in the modelling process are illustrated in Fig. 2

Model performance was evaluated using multiple complementary metrics, including the Area Under the Receiver Operating Characteristic Curve (AUC), Continuous Boyce Index (CBI), omission rates (OR10 and ORMTP) and Akaike Information Criterion corrected for small sample sizes (AICc). We selected the optimal models based on low AICc values, while maintaining high validation AUC and CBI scores and low omission rates. For climate-only models, validation AUC values were 0.983 for L. camara, 0.978 for C. odorata and 0.776 for P. crassipes. Corresponding CBI values were 0.854, 0.782 and 0.590, respectively, indicating generally good predictive performance, although model performance for P. crassipes was comparatively lower due to the limited number of occurrence records available. When climate and land-use predictors were combined, predictive performance remained high, with validation AUC values of 0.980 for L. camara and 0.989 for C. odorata, accompanied by high CBI values (> 0.75) and low omission rates (< 0.13). Overall, the combination of evaluation metrics indicated that the selected models provided reliable estimates of habitat suitability across Tanzania's Lake Zone (Suppl. material 1).

In the climate-only models, temperature and precipitation variability were the primary drivers, but their influence varied across species. Annual temperature range (Bio7) was the most important for L. camara (47.29%) and C. odorata (47.29%). Temperature seasonality (Bio4) was also important, especially for P. crassipes (34.6%), L. camara (21.34%) and C. odorata (17.28%). Precipitation seasonality (Bio15) had a strong effect on C. odorata (39.92%) and L. camara (15.19%), while precipitation of the warmest quarter (Bio18) and mean temperature of the wettest quarter (Bio8) were key factors for P. crassipes (34.8%) and (30.6%), respectively Table 3.

In the combined climate–land-use models, Bio7 and Bio4 remained the dominant predictors for both L. camara (33.71% and 23.04%, respectively) and C. odorata (29.52% and 20.12%). Precipitation seasonality (Bio15) also contributed substantially, and land-use/land-cover under the water category emerged as an important predictor for C. odorata (14.84%) and L. camara (14.14%). Precipitation in the wettest quarter (Bio16) had a moderate relative contribution. However, its higher permutation importance indicates a strong influence on model predictions, despite a lower fitted contribution Table S2.

Jackknife analysis, based on regularised training gain, revealed that a subset of bioclimatic variables contributed disproportionately to the model structure across species (Fig. 3). For all three species, variables related to precipitation seasonality and temperature variability yielded relatively high gains when used in isolation, whereas omitting them led to notable reductions in model gain. Response curves (Fig. 4) further illustrate how predicted habitat suitability varies along individual environmental gradients while holding other predictors constant.

Across the Mara–Simiyu landscape, current habitat suitability was spatially heterogeneous, with suitable areas concentrated in the central–western and northern portions of the study area, particularly along riparian corridors and adjacent floodplain systems (Suppl. material 3, Fig. 5). Under climate-only models, C. odorata and L. camara exhibited broadly similar patterns, with approximately 13–17% of the region classified as marginal to highly suitable, while P. Crassipes showed slightly lower overall suitability (≈ 13%). Highly suitable areas for terrestrial species were largely restricted to zones with higher moisture availability and productive land cover, whereas the eastern and southern parts of the landscape were dominated by unsuitable conditions. Incorporating land-use information resulted in modest reductions in suitable habitat for C. odorata and L. camara, primarily through contraction of moderately and highly suitable areas in fragmented agricultural and urbanised zones, while suitability along riparian corridors remained comparatively stable. For C. odorata, the proportion of highly suitable area declined from 4.58% under climate-only conditions to 4.26% under the combined model, and for L. camara from 5.61% to 5.46%, with no substantial shifts in the overall spatial pattern.

Climate-only model

Future climate projections indicated substantial changes in habitat suitability across species, scenarios and time periods (Fig. 6).

Niche overlap between Lantana camara and Chromolaena odorata remained moderate to high across all climate scenarios (Schoener’s D = 0.45–0.81; Hellinger’s I = 0.78–0.97). Under low-emission conditions (SSP126), overlap remained stable over time, reflecting persistent climatic niche similarity. In contrast, intermediate- and high-emission scenarios showed greater variability amongst general circulation models. Some projections indicated increased climatic convergence by the end of the century (e.g. SSP585–2090; D = 0.81), while others suggested partial niche divergence (e.g. SSP370–2090; D = 0.45). These findings demonstrate that, although the two terrestrial invaders currently share broadly similar environmental requirements, the extent of future climatic overlap will depend strongly on both the emission trajectory and the climate model structure (Suppl. material 8).

Spatial overlap in projected suitable areas, measured using the Jaccard index of binary P10-thresholded maps, showed greater variability amongst models than environmental niche overlap. Under SSP126 in 2030, approximately 47–49% of the suitable habitat was shared between the two species. In subsequent periods, spatial overlap varied widely amongst general circulation models, ranging from a complete absence of co-suitable area (0%) in scenarios where one species lost suitability, to values exceeding 50% in projections with broad co-suitability. When both species retained suitable habitat, a large proportion of the smaller species' suitable range often overlapped with that of the other species, frequently exceeding 70%, indicating substantial geographic convergence. Comprehensive spatial overlap metrics, including co-suitable area (km²), Jaccard overlap and overlap of the smaller suitable range, are presented in Tables Suppl. material 9 and Suppl. material 10.

This study provides a subnational, spatially explicit assessment of the current and future distribution of three invasive plant species across contrasting ecosystems in the Mara–Simiyu Region using a threshold-based MaxEnt modelling framework (Suárez-Mota et al. 2016, Shilky Kishore et al. 2023). The study focuses on two terrestrial species, Chromolaena odorata and Lantana camara and one water-based species, Pontederia crassipes (formerly Eichhornia crassipes). These species were modelled under scenarios considering only climate and both climate and land use, using cloglog outputs and a set 10^th-percentile training presence threshold (Pearson et al. 2007).

The models show that, under current climate and land use, the main hotspots for the terrestrial invasive species are in the central, western and northern parts of the region. These areas include farmland, grazing areas, settlements and roads, which often disturb the land, increase the spread of invasive seeds and reduce competition from native plants. This makes it easier for invasive species to take hold (Pyšek and Richardson 2010). Similar links between C. odorata/L. camara and disturbed areas have been found in Tanzania and other parts of East Africa, where changes in land use, livestock movement and transport routes are major factors in the spread of these plants (Mgumia and Oba 2003, Mwangi and Swallow 2008). The clustering of suitable habitats along these human-altered landscapes shows how both the environment and human activity shape invasion patterns. In contrast, P. crassipes is mainly found along Lake Victoria, nearby wetlands and connected rivers. This matches the species’ need for permanent freshwater, slow-moving water and nutrient-rich conditions that help it grow and spread quickly. Runoff from farms adds nutrients and connected waterways help floating mats of the plant move across different aquatic areas. Similar patterns have been seen throughout the Lake Victoria Basin and other tropical freshwater systems in East Africa (Albright et al. 2004, Njiru et al. 2012).

Analysis of variable contributions indicates that temperature variability and seasonal precipitation are key determinants of species distributions, although their relative importance differs among species (Phillips et al. 2006, Elith et al. 2011).

For Lantana camara and Chromolaena odorata, temperature seasonality (Bio4) and annual temperature range (Bio7) were the main factors influencing where these plants can thrive. Temperature changes affect key processes such as photosynthesis, respiration, water use, and seed germination, which in turn affect plant growth and survival. Both species can tolerate a wide range of conditions and adapt easily, allowing them to keep growing even when temperatures change and to take advantage of disturbed areas where native species may be less competitive. Thermal variability may also alter competitive interactions by increasing stress on native vegetation while favouring invasive species capable of rapid establishment and resource acquisition (Bradll et al. 2010). Climate projections for Tanzania indicate increasing mean temperatures and greater temperature variability during the 21^st century (Luhunga et al. 2018, Luhunga and Songoro 2020, Magesa 2024). Such changes may alter seasonal temperature regimes, potentially expanding suitable conditions in some regions while increasing climatic stress in others.

For Pontederia crassipes, temperature seasonality (Bio4) and precipitation during the warmest quarter (Bio18) were the most important climate factors, showing that this species depends on warm, moist conditions for rapid growth (Villamagna and Murphy 2010). Regional projections indicate increasing temperatures, changes in rainfall seasonality and more frequent extreme precipitation events (Luhunga and Songoro 2020, Magesa 2024).

While climate variables were the strongest predictors in our models, the spread of P. crassipes is also strongly affected by aquatic environmental conditions that our models did not include. High levels of nutrients, especially nitrogen and phosphorus from farm runoff and wastewater, help the plant grow quickly and form thick floating mats Wilson et al. 2005, Villamagna and Murphy 2010) Connections between water bodies, such as lakes, rivers, wetlands, and irrigation systems, make it easier for the plant to spread, and water flow patterns affect how populations start, survive, and break apart (Albright et al. 2004, Villamagna and Murphy 2010). As a result, local invasion patterns depend on both climate and aquatic ecosystem features. These factors may explain local distribution patterns that climate-based models do not fully capture and should be included in future habitat suitability models.

Precipitation of the wettest quarter (Bio16) showed moderate contribution, but high permutation importance, indicating that model performance declined when this variable was randomised. This suggests that Bio16 captures seasonal water availability not fully represented by other rainfall variables. Seasonal rainfall patterns strongly influence plant establishment and survival in tropical ecosystems (Baltzer and Davies 2012), while water availability regulates biomass production across wet–dry gradients (Muller‐Landau et al. 2021). Disturbance events interacting with seasonal moisture conditions may further enhance invasion success (Setterfield et al. 2005).

When land-use variables were incorporated, the relative contribution of climatic predictors decreased, but remained dominant. Temperature-related variables declined by approximately 30–40% in relative importance compared with climate-only models, but they remained the primary drivers of suitability. This indicates that invasion risk is shaped by both broad-scale climatic constraints and local habitat availability. Consequently, these findings support the view that invasion dynamics result from interactions between environmental conditions and anthropogenic landscape modification rather than climate alone (Roura-Pascual et al. 2011, Early et al. 2016, Ren et al. 2025).

When considering only climate projections, the two invasive species maintained mostly stable core areas under the low-emission scenario (SSP1–2.6). In contrast, under the higher-emission scenarios (SSP3–7.0 and SSP5–8.5), particularly towards the end of the century, the models projected increasing fragmentation and contraction of the most suitable areas. This pattern is consistent with projections of rising mean temperatures and greater temperature variability across Tanzania under high-emission scenarios (Luhunga et al. 2018, Luhunga and Songoro 2020). Since annual temperature range (Bio7) and temperature seasonality (Bio4) were the most influential predictors in our models, the projected decline in suitability may indicate that future climatic conditions could move beyond the thermal ranges currently associated with these species, rather than simply shifting their distributions. Increased warming and greater temperature variability may impose physiological stress, reduce competitive performance and limit establishment in already warm tropical environments.

Accordingly, some parts of the study area may become less suitable for L. camara and C. odorata as the climate warms, while other areas could remain suitable or change only slightly. This matches findings from other studies, which show that tropical invasive plants can lose suitable habitat if temperatures exceed their tolerance or if the climate becomes much more variable (Kriticos et al. 2005, Bellard et al. 2018). Still, these predictions do not mean that future ranges will definitely shrink, since actual distributions will also depend on factors like dispersal, propagule pressure, land-use changes, disturbance and interactions with other species.

In contrast, Pontederia crassipes retained broad climatic suitability across most scenarios. Late-century projections frequently showed relatively uniform climatic suitability, with few areas classified as unsuitable. This pattern indicates that climatic conditions may remain broadly permissive, but it should not be interpreted as evidence that the species will occupy all climatically suitable areas. Although P. crassipes retained broad climatic suitability, its realised distribution remains constrained by the availability and connectivity of freshwater habitats, water quality, nutrient conditions, hydrological processes and dispersal opportunities. The continued climatic suitability is nevertheless consistent with evidence that floating aquatic plants may retain or expand their climatically suitable ranges as tropical regions warm (Villamagna and Murphy 2010, Kriticos et al. 2015). This also agrees with recent habitat suitability assessments that identified extensive invasion risk for P. crassipes across freshwater ecosystems, emphasising that climatically suitable regions remain vulnerable where suitable aquatic habitats and dispersal pathways exist (Adhikari et al. 2026).

If freshwater habitats are present and invasion pathways remain open, ongoing suitable climate conditions could put more pressure on aquatic ecosystems. This can happen through the growth of thick surface mats, less light reaching the water, lower oxygen levels and changes in nutrient cycles. These effects may lower native biodiversity and disrupt ecosystem function. They could also pose long-term risks for fisheries, irrigation, hydropower, navigation and people who rely on open-water resources. However, the magnitude of these impacts will depend on local environmental conditions, nutrient availability, management interventions and human-mediated dispersal. Therefore, identifying climatically suitable freshwater habitats remains essential to support early detection, targeted monitoring, and timely management interventions to prevent further spread and reduce invasion risk (Adhikari et al. 2026).

For the terrestrial species, divergence amongst GCMs increased under several high-emission scenarios, resulting in greater variability in projected highly suitable areas. This reflects increasing uncertainty in future temperature and rainfall conditions and their ecological consequences, which may exceed their tolerance (Buisson et al. 2010, Araújo and Peterson 2012). The projections, therefore, represent a range of plausible habitat-suitability outcomes rather than a single expected future distribution in East Africa.

However, these projections represent potential climatic suitability rather than realised invasion risk, as the modelling framework did not explicitly incorporate dispersal limitation, propagule pressure, or anthropogenic spread.

Including land-use change significantly altered patterns of climate-driven suitability. For C. odorata and L. camara, suitable habitats became even more limited under high-emission scenarios, and model certainty declined over time. These results highlight the importance of considering land-use dynamics in invasion modelling (Franklin 2010, Mainali et al. 2015, Thuiller et al. 2019). As uncertainty increases with greater warming, interpret late-century projections as risk estimates rather than precise predictions. Regions where models converge are more reliable for forecasting invasion risk, whereas areas with high variability require close monitoring and adaptive management.

Both species were often found to be highly suitable for croplands, showing their preference for disturbed areas. Activities like expanding agriculture, grazing, and building roads often remove native plants, increase light, expose bare soil, and cause repeated disturbances that support invasions. Roads and livestock can also spread seeds and help new plants take hold, while the edges of croplands and fallow fields offer good spots for colonisation. Similar links between L. camara, C. odorata and human-changed landscapes have been seen in tropical regions and East Africa (Václavík and Meentemeyer 2009, González‐Moreno et al. 2014, Shackleton et al. 2017, Pyšek et al. 2020). Forest areas had moderate variability that increased with stronger warming, suggesting that climate change could change habitat suitability there.

Selection ratios (SR > 1 for forest and shrubland; SR < 1 for water) indicate that land-use associations are not random, highlighting the importance of habitat context for invasion risk (González‐Moreno et al. 2014). Climate determines the overall suitability, but land-use patterns can either increase or limit invasion at the local level.

Moderate to high environmental niche overlap between L. camara and C. odorata, quantified using Schoener’s D and Hellinger’s I (Broennimann et al. 2012), indicates substantial similarity in climatic tolerances. However, spatial overlap in projected suitable areas provides a more direct estimate of potential co-occurrence. Under low-emission scenarios in 2030, approximately 47–49% of climatically suitable area was shared between the two species, indicating considerable potential for co-invasion.

Under mid- and late-century projections, spatial overlap became increasingly variable amongst GCMs. In some projections, Jaccard overlap exceeded 50%, suggesting strong spatial convergence, whereas in others, overlap declined sharply or approached zero due to the contraction of suitable habitat for one species. This divergence indicates that, while climatic niches remain broadly similar, geographic co-suitability is highly sensitive to model-specific climate trajectories.

When their habitats overlap, these two species can co-exist in disturbed areas, which may increase propagule numbers and together reduce native vegetation. While this study did not examine direct competition between L. camara and C. odorata, their similar environmental needs suggest that co-invasion could worsen ecological impacts on invaded ecosystems. This variability means the risk of co-invasion could increase under some future climate scenarios but decrease under others, so it is important to interpret projected species interactions using a range of models. (Alexander et al. 2016).

Using the P10 threshold enabled consistent comparisons of suitability patterns across species and scenarios (Jiménez-Valverde and Lobo 2007, Elith et al. 2011). However, the final projections did not systematically test other threshold criteria. Choosing different thresholds could change the size of the predicted suitable areas.

There are several limitations to note. First, while MaxEnt performs better than other modelling methods with small sample sizes (Breiner et al. 2015, Moudrý et al. 2024), the limited number of occurrence records filtered for P. crassipes may still increase uncertainty and make it harder to apply the results to future climate scenarios. The model was kept simple to avoid overfitting, so predictions for this species should be viewed with caution. In addition, this study relied on a single modelling algorithm (MaxEnt), and future studies could compare multiple algorithms or use ensemble species distribution modelling approaches to reduce algorithm-specific uncertainty and improve the robustness of future projections.

Second, using a presence-only modelling approach makes it hard to distinguish between truly unsuitable environments and areas that have not been sampled, but could be suitable. This is a common issue in correlative species distribution models because results depend on the type and quality of the available occurrence data (Brotons et al. 2004, Yackulic et al. 2013, Jarnevich et al. 2015). Even with spatial bias correction and background weighting, predicting into new climate conditions can still involve extrapolation and uncertainty. In addition, the models estimated environmental suitability, but did not explicitly incorporate dispersal processes, propagule pressure or human-mediated spread. Consequently, climatically suitable areas may not necessarily be colonised if dispersal is limited, while anthropogenic transport through roads, trade and livestock movement may facilitate establishment beyond naturally accessible areas. The projected distributions should, therefore, be interpreted as potential habitat suitability rather than realised future invasion patterns.

Third, model performance was mainly measured using spatially structured test AUC. While AUC is commonly used and does not depend on thresholds, it does not fully reflect model calibration or ecological realism (Lobo et al. 2008). For this reason, AUC was considered along with omission rates, response curves and jackknife diagnostics. Although model transferability was evaluated using spatially structured validation, formal extrapolation analyses (e.g. multivariate environmental similarity surfaces or mobility-orientated parity analyses) were not conducted (Elith et al. 2010, Owens et al. 2013). Consequently, predictions in areas with novel future climatic conditions should be interpreted with caution because they may involve environmental conditions outside the range represented by the calibration data.

Fourth, the chosen general circulation models (GCMs) provided consistent data across all Shared Socioeconomic Pathways (SSPs) and time periods, but they were not the best-performing CMIP6 models for predicting East African rainfall. This trade-off between having complete data and obtaining the best regional results may lower the accuracy of predicted suitability patterns, especially for species sensitive to rainfall.

Finally, the predictor set was limited to climatic and land-use variables that are consistently available for both current and future CMIP6 scenarios. Other environmental factors, like soil characteristics, elevation, topographic indices, NDVI and distance to water, could improve model performance and ecological realism. These variables were not included because some lack reliable future projections, while others would reduce comparability between current and future models. Future research could include these predictors in more comprehensive modelling frameworks to improve invasion risk assessments.

To manage terrestrial invasive species, focus on agricultural areas, transport routes, grazing lands, and other disturbed habitats that are often hotspots for these species. Detecting and responding to invasions early in these places can be more cost-effective than removing them after they are established. Since cropland areas often show higher suitability for invasives, including invasive species management in agricultural extension programmes and land-use planning could help limit their future spread.

In aquatic ecosystems, the ongoing suitable climate for Pontederia crassipes means it is important to keep monitoring rivers, wetlands and the Lake Victoria catchment area. Managing this species should include both physical control methods and steps to reduce nutrient runoff from farms and settlements, since extra nutrients help water hyacinth grow and persist.