Summer rain and wet soil rather than management affect the distribution of a toxic plant in production grasslands

In the northern forelands of the Alps, farmers report an increase of Jacobaea aquatica in production grasslands. Due to its toxicity, the species affects grassland productivity and calls for costly control measures. We are investigating the extent to which management practices or climatic factors are responsible for the increase of the species and how the situation will change due to climate change. We tested for effects of management intensity, fertilization, agri-environmental measures, and soil disturbance, and modeled the occurrence of the species under rcp4.5 and rcp8.5 scenarios. The main determinants of the occurrence of the species are soil type and summer rainfall. A high risk is associated with wet soils and > 400 mm of rain between June and August; an influence of the management-related factors could not be detected. Under the climate-change scenarios, the overall distribution decreases and shifts to the wetter alpine regions. Thus, the current increase is rather a shift in the occurrence of the species due to the altered precipitation situation. Under future climatic conditions, the species will decline and retreat to higher regions in the Alps. This will decrease the risk of forage contamination for production grassland in the lowlands.

• What are the climatic, soil, and management factors that determine the regional distribution and abundance of J. aquatica?• How is climate change over the past decades related to regionally declining or increasing occurrence of the poisonous species?• How will future climate change affect the distribution of the species?
To answer these questions, we applied habitat suitability and a random forest model that disentangled the role of soil, climate, and management in the spread of J. aquatica, and assessed the climate-related changes in the probability of occurrence over the past decades and the future.

Methods
Study region.Our study covers the state of Bavaria (Germany), which has an area of 70,550 km 2 (Fig. 1) and is a suitable model region.Bavaria includes seven climatic sub-regions, i.e., the Alps, Alpine Forelands, Southern Bavarian Hills, the Danube and Main regions, the Spessart-Rhön Highlands, and the Eastern Bavarian Hills and Mountains 30 .Mean annual temperatures vary from 8.5 °C in the Main region over 7.3 °C in the Alpine forelands to 5.7 °C in the Alps, with an overall average of 7.9 °C.The warmest months are July and August, with mean temperatures of 16.3 °C; the coldest months are January and February, with − 0.5 °C.Mean temperatures in spring (April-May) reach already 7.7 °C, and the vegetation period lasts from May to October.Precipitation increases with altitude and proximity to the Alps.While the Main region only receives about 700 mm, the alpine forelands get 1500 mm, locally > 2000 mm in the Alps.Precipitation is highest in summer, with 1.5-fold the amount of the other seasons.
In the study region, 10,573 km 2 is permanent grassland; most (98%) are meadows, mowing pastures, or pastures, while the remaining 2% are litter meadows or conservation grassland 31 .The proportion of grassland varies among regions, ranging from 10 to 30% in the hilly parts of southern Bavaria, the Danube and Main region, and in the Spessart-Rhön region, up to 80-100% in the Alps and their foothills with highest densities in the Oberallgäu and Garmisch-Partenkirchen districts.In the East Bavarian Hills and Mountains, there is a gradient from 30% in the northern to 80% in the southern part of the region.Higher percentages of grassland are usually associated with increased precipitation and a more pronounced relief.
Study species.Jacobaea aquatica (syn.Senecio aquaticus; Asteraceae) is native to Western and Central Europe, where it occurs in (semi)natural wet grasslands, grassy floodplains, and along the banks of watercourses 17,32 .The species has a high light demand and is usually biennial.Frequent cutting can prevent flowering and significantly extend its life span 17 .In highly infested grasslands, J. aquatica can locally exceed 100 plants per m 2 (up to 30-50% cover), and each plant may produce several hundred seeds with a pappus facilitating wind dispersal 25 .Seeds germinate quickly under favorable moisture and light conditions 19,22 , and the species forms a persistent soil seed bank of between 350 and 2000 seeds per m 218, 25,33 .There is no information on the dispersal distances of J. aquatica.However, in the closely related species J. vulgaris, 89% of the achenes could not overcome a 5-m distance, and none was found > 14 m 34 .As J. aquatica produces considerably larger achenes, long-distance dispersal has apparently even less importance for this species 10 .The high pyrrolizidine content of J. aquatica makes the species poisonous for livestock and cattle and can even be harmful to humans via milk 13,35 .Already at a density of one plant per 10 m 2 , grassland is deemed unsuitable for forage purposes for cattle and Data acquisition and preprocessing.In 2017, 259 fields of 126 farms in the study region were surveyed for the occurrence of J. aquatica during the main vegetation period (May-September), with at least 2 weeks between the last cut or grazing.The abundance of J. aquatica was determined for each field along two opposing 5-m wide transects, diagonal to the field.As each transect differed in length, this abundance was projected to a standardized plot area with a size of 100 m 2 , and assigned to frequency classes (Supplementary Table A1).Simultaneously, major gaps in the grassland sward, such as tire tracks or livestock tread marks, were recorded.
Additionally, the respective farmers answered a questionnaire including general practices and the specific management of the infested grasslands.It included the type of management (conventional, organic), the implementation of agri-environmental schemes, the frequency and type of grassland use (grazing, mowing, combinations), fertilization (mineral, liquid or stable manure), drainage systems as well as gaps in the sward caused by tires or livestock trampling (Supplementary Table A1).Farm size ranged from 4 to 266 ha, mean farm size was 41 ± 35 ha; 52% of the farms were conventional, and 48% were organic.For 68% of the farms, dairy farming was the main production sector, 12% produced suckler cows, and 20% operated different production types.The average stocking rate was 1.1 livestock units ha −1 .The main types of grassland used were meadows and hay-grazed meadows (80%), pastures (10%), and litter meadows (10%).50 reference plots (ca.20%) were unaffected by J. aquatica, the remaining 209 plots were infested with various densities.
To capture the impact of climatic factors and soil type on the general occurrence and changes of J. aquatica we used a generalized boosted regression model to predict the species occurrence risk in Bavaria for the periods 1988-1997 ('past'), 2008-2017 ('current'), and 2028-2037 ('future') under the rcp4.5 and rcp8.5 scenario.The current time period was used for training, and the model was then projected to the past period and the future.For the training of the model, the presence/absence data of J. aquatica were complemented by data from the Bavarian survey of habitats with high nature conservation value 36 sampled from 2008 to 2017.We only used records from quadrants of the habitat survey maps (scale 1:25,000) that were not occupied by our own findings.This survey comprises the species inventory of over 400,000 habitats recorded from 1984 to 2020; however, only 2012 of them were surveyed in the period 2008-2017.
Data collected from 2008 to 2017 were mainly sampled to complete former sampling and to replace outdated records.Therefore, only small areas were mapped in this time period so that the analyzable data are clumped and cover only a small percentage of the study region.Hence only 87 sites from the current period could be used for model training, while 1320 sites, recorded from 1988 to 1997, were used to validate the model prediction for this time period.Thus, 296 presence points for J. aquatica could be included.As the absences recorded in our own dataset were placed on the same farms as the presence points, these real absences were excluded from the climate model due to high spatial correlation.Instead, the complete presence data were complemented with ten sets of 296 pseudo-absences each.These pseudo-absences were randomly placed in such quadrants of the habitat survey maps where J. aquatica has not been reported until 2018, and with a minimum distance of 3000 m to other growth areas.Each set of pseudo-absences was then combined with our set of presence points to provide the data for ten models.
Considering the preference of J. aquatica for wet and water-influenced soils 37 , we further included soil type as a predictor.These data were taken from the digital vector map of Bavaria 38 .This map was compiled from geological and forestry maps (1:25,000), and was validated against soil profiles.For this map, the German soil classification system was used, and therefore, units do not exactly correspond to the World Reference Base for Soil Resources 39 .To improve comprehensibility, the main soil classes were adapted to the international system (Supplementary Table A2).However, as these systems only match at the level of soil classes, German terms were kept for the subclasses.The respective classes and their subclasses were numerically ordered along a gradient of descending soil moisture (Supplementary Table A2).We assumed soil classes to remain constant during the past and future model periods.
The following climate variables were used as predictors for our habitat suitability model: all-year minimum and maximum air temperature (2 m AGL), mean precipitation per year, in spring (March-May), summer (June-August) and autumn (September-October) as well as the number of ice, frost, and hot days per year.For Germany, these data were available as interpolated raster data since the 1980s with a resolution of 1 km × 1 km 40 .
To relate the recent changes in the distribution of J. aquatica in Bavaria to the development of the climate conditions, data from 1988 to 1997 were compared with current records from 2008 to 2017, each sampling period comprising 10 years.For each period, the respective raster data were averaged and confined to the study region.
The future distribution for the decade from 2028 to 2037 was calculated using the climate data predicted under the rcp4.5 and rcp8.5 scenarios by the Bavarian climate model.Data for each year were obtained from the Bavarian Climate Information System 30 with a resolution of 5 km × 5 km and the respective variables were averaged for the period.
All raster data were projected to ETRS89; EPSG25832 and resampled to a final resolution of 50 m × 50 m to allow for the detailed representation of soil classes.

Modeling climate and soil effects on plant occurrence.
To identify potential habitats for J. aquatica based on climate and soil characteristics, we applied generalized boosted regressions using the R packages gbm 41 and biomod2 42 modeling presence/absence of the species with soil type and the averaged climate data for the current period (2008-2017) as predictors (Supplementary Table A3).For tuning purposes, we ran different models with one of the generated presence/absence data sets and an initial parameter set (Supplementary Table A4), and subsequently gradually modified the parameters shrinkage, interaction depth, minimum objects in node, and bag fraction.Each of the resulting models was evaluated using gbm.perf, and the respective root mean square error (rsme) was calculated.Finally, the parameter set resulting in the lowest rmse was chosen for our actual model.This model was applied to our ten generated sets of presence/pseudo absence data using a random 70% fraction of the data for training and 30% for model evaluation, assuming a prevalence of 0.5.
The resulting models were evaluated by determining the relative importance of the predictor variables, AUC, and TSS.Extrapolation risk due to projection to areas with non-analogous conditions 43 was assessed using the mop package 44 .ROC curves were generated using the performance function of the ROCR package 45 , and sensitivity, specificity, and cut-off values were determined using the roc function of the pROC package (Version 1.18.0;Robin et al. 46 ) with the method best.Partial dependence plots were generated for the two most influential variables using the gbm.plot function 41 .To assess the independence of the two most influential variables, multipredictor partial dependence plots were generated for each model and averaged.
Then, each model was used to predict the probability of the occurrence of J. aquatica in Bavaria based on the predictors for the current and past period and for the future period under scenarios rcp4.5 and rcp8.5.Model predictions for each period were averaged, the ROC curves were plotted, and the final AUC of the averaged prediction was determined.The importance of the respective predictors was averaged from their relative importance in the respective models.Finally, the prediction performance for the period from 1988 until 1997 was validated against the 1320 sites of the Bavarian habitat survey, including J. aquatica.To account for possible inaccuracies in the soil and climatic maps, the validation was repeated with a 250-m buffer around the respective biotope.
Finally, to visualize changes between the study periods and to delineate potential current and future risk areas, these areas were identified where the occurrence or the disappearance of J. aquatica was predicted by the model for the first time.Additionally, areas where species occurrence was predicted for both periods but the probability of occurrence decreased or increased by more than 20% were indicated.

Modeling the management effects on plant occurrence. To determine whether and to what extent
farm-specific factors and management practices influence the occurrence (presence/absence) or frequency of J. aquatica we used a random forest classification approach 47 .Random forest classifiers provide a high prediction accuracy with factorial data, they are less sensitive to over-parameterization and tolerate collinearity 48 .They further provide information about the importance of the single variables which have been recently applied in various fields of ecology.Using the randomForest package (Version 4.6-14 49 ) in R 50 , we trained and evaluated ten random forest classifiers for the presence/absence and frequency of J. aquatica, each constructed of 1000 decision trees based on land-use data (Supplementary Table A1).For each model, our data were randomly divided into a training set with 130 sites and 129 sites used for model evaluation.
Predictor variables were organic farming (true/false), land-use type (pasture, meadow, mowing pasture, litter meadow or unused), use intensity (number of uses per year), autumn grazing (yes/no), fertilization (yes/no), implementation of agri-environmental measures (yes/no), and apparent soil disturbance (vehicle tracks, animal tracks, other).To identify the influence of particular practices, each land-use type and the applied fertilizer types (fertilizer: mineral, manure, slurry), were additionally included in the model as Boolean type factor, and use intensity was converted into a factor as well (Supplementary Table A1).The performance of the resulting model for absence/presence data was assessed using its ROC and AUC value.The respective model for the frequency of J. aquatica was assessed by its confusion matrix and overall error.

Results
Current and potential distribution of Jacobaea aquatica.The gradient-boosted model based on climate parameters and soil type was well suited to predict the potential distribution of J. aquatica for both the current and historic climatic situation.The average AUC value for the ten models (mean ± SD) was 0.96 ± 0.01, and the final averaged model achieved an AUC value of 0.98.The best cut-off based on optimal sensitivity and specificity for the averaged model was determined as 0.35, with a specificity of 98% and a sensitivity of 91% (Supplementary Table A5).Taking all variables into account, the proportion of non-analog cells was ca.11% for the rcp45 scenario and ca.30% for the rcp85 scenario.The variables frost days, ice days, and hot days which exclusively caused the deviations, were only of minor importance in the model.The removal of these variables reduced the percentage of non-analogous cells to almost 0% (Supplementary Fig. A2).
Applying the averaged model and the cut-off of 0.35 to the past period and validating the prediction results against the occurrences of J. aquatica recorded in the Bavarian habitat survey of the respective time period resulted in a true positive rate of 75% when inaccuracies of the soil map were ignored and even 86% when a 250-m buffer was considered.Among the 1320 habitats with J. aquatica surveyed between 1988 and 1997, 331 (ca.25%) occurred in regions where the model did not predict this species.However, when the uncertainty reflecting potential inaccuracies of the soil map and during surveying were included, the rate of false negatives decreased to 14%, and 1135 of 1320 habitats with J. aquatica were correctly predicted.A check of the remaining 14% of the recorded presences, where the model excluded species occurrence, revealed that the habitat survey for all these sites indicates wet soil conditions and wet grassland vegetation.A cross-check with the soil map further showed that all but three of these sites were not reflected by the soil map, probably due to the low resolution and lack of precision of this map.
In all ten individual models, summer rainfall (rainJJA; Fig. 2a) was the most important factor for the occurrence of J. aquatica, with 49.2 ± 4.5% relative importance, followed by soil type with a relative importance of 15.5 ± 2.3%, and then by the other climatic factors (Supplementary Table A5).The partial dependence plots for summer rainfall reveal a distinctive threshold, i.e., above 406 ± 5 mm of rain in June and August, the probability of J. aquatica occurrence raised steeply.
A similar but negative threshold was evident for the soil classes (Fig. 2b).The probability of occurrence was restricted to wet soils such as peatlands and water-logged mineral soils, including histosols, gleysols, and stagnosols.In a combined multi-predictor partial dependence plot (Fig. 3) with summer rainfall and soil classes as predictors, however, it became clear that while on these wet soils an occurrence of the species is generally possible, the occurrence on drier soils such as luvisols, pelosols, and cambisols is linked to summer rainfalls > 406 mm.An occurrence on Ah/C soils and terrestrial raw soils can be excluded.
For both the past and current periods, the highest risk for infestation and spread of J. aquatica was in the climate regions Alpine Forelands and Alps (Fig. 4).There, this risk extended over large and contiguous areas.In the lower and drier regions, probability of occurrence was restricted to floodplains, along streams and rivers, and to peatlands.In the Alps with above-threshold rainfall, the species occurrence was controlled by temperature, namely by the average ice and frost days per year.Thus, the relatively dry Main and Spessart-Rhön regions are low-risk areas.A2).
A comparison of the past with the current period clearly shows that the area with an increased infestation risk decreased (Supplementary Table A6).However, this distribution pattern was not uniform across Bavaria.In the Alpine forelands and the Alps, the occurrence probability was still pronounced and the risk extended over a wide contiguous area with even more distinct boundaries.In the rest of the state, occurrences retreated to the wet lowlands along rivers and streams.While J. aquatica still occurred in many of these areas in 1988, current conditions made the survival of these populations unlikely, while other regions had shifted from low to highrisk areas (Fig. 5).In total, the risk area decreased by 12% from 15,262 km 2 in 1988-1997 to 13,452 km 2 today.An area of 3031 km 2 where the model predicted a declining risk and 1221 km 2 where there is an increased risk today contributed to the total change.Of all the areas which exceeded the threshold risk for the occurrence of J. aquatica in both time periods, the probability of occurrence had decreased by > 20% on 1372 km 2 , and increased by > 20% on 1438 km 2 (Fig. 5).www.nature.com/scientificreports/Predicted distribution of Jacobaea aquatica.The future scenarios both predicted a substantial decrease in risk areas outside the alpine regions of Bavaria (Fig. 6).There was a strong probability that occurrences in the lowlands, which had already decreased in wetlands and river floodplains since 1988, will largely disappear or retreat to larger wetlands and the largest floodplain areas of the Rivers Danube and Isar.This decrease is associated with reduced summer rainfall and mainly affects populations on the wet soil groups of histosols, stagnosols, and gleysols (rcp4.5:91%; rcp8.5:82%; Supplementary Table A6).Decreasing summer rainfall also narrowed the  www.nature.com/scientificreports/extent of the high-risk area in the alpine forelands to areas where summer rainfall remained above the 400 mm threshold or even increased.This risk became more pronounced with scenario rcp8.5, characterized by slightly higher summer precipitation.Both scenarios also predict substantial new risks for higher elevation areas (Supplementary Fig. A1).In addition to future summer rainfall, which will continue to exceed the 406-mm threshold, the decrease of ice and frost days, which limit the occurrence of J. aquatica in high precipitation areas, will make this environment accessible.
Influence of management practice on the occurrence of Jacobaea aquatica.In contrast to the clear influence of climate and soil, none of the grassland management variables tested in our models showed a significant impact on the occurrence or frequency of J. aquatica.The average AUC value of all ten random forest classifiers for the absence/presence of J. aquatica was only 0.469 ± 0.062.Similarly, the random forest model using the species frequency as a response variable had an overall error of 54%, and the confusion matrix shows only a reasonable prediction for the frequency class 'B' representing 0-1 individuals/m 2 (Fig. 7).Hence, as none of the classifiers held any explanatory strength, the tested predictors cannot explain the presence (or absence) or frequency of the species.Consequently, the relative influence of individual variables is not further indicated or discussed.

Discussion
Our study revealed that the main factors determining the present distribution of J. aquatica in Central Europe are abundant summer rainfall and wet soils.With a mean average summer precipitation (June-August) of over 400 mm, J. aquatica grows on a broad range of soil types, including histosols, gleysols, stagnosols, luvisols, pelosols, and cambisols.When the average summer precipitation drops below this threshold, the occurrence of the species becomes increasingly restricted to wet and moist soil types, mainly occurring in peatlands, floodplains, or other wetlands.When summer precipitation falls below 200 mm, the occurrence of J. aquatica becomes unlikely.This preference for moist soils agrees with Forbes 10 , who described severe J. aquatica infestation problems on the Orkney Islands and found that "soil surface wetness" was the most significant predictor for the occurrence of the target species.The minor influence of precipitation there may be due to climatic differences.Mean annual temperatures 9 K below those in the Alpine foreland and the wide distribution of peat soils 51 suggest that summer precipitation does not limit the growth of J. aquatica on the Orkney islands due to low evapotranspiration.In our study region, temperatures had little influence on the occurrence of J. aquatica.Only in the higher elevated areas of the Alps, the number of frost and ice days limit the distribution despite the high summer precipitation.
Modeling the development of species in dependence on environmental trends has become an important tool to understand ecosystem changes and to devise sustainable management strategies 52,53 .Many of these models have been designed to predict the spread of harmful organisms and to limit actual and future damages 54 .Most of these target organisms were invasive species with a negative impact on native ecosystems.Jacobaea aquatica can be considered a 'native invader' 55 , which also shows a considerable spread at least in parts of its distribution range 16,56,57 ).For the territory of Bavaria, however, our study could not confirm a significant spread of this problematic species.In contrast, the overall area at risk considerably decreased since 1988-1997, and the probability of occurrence in the affected areas declined.The reason for this is a decrease in summer precipitation, especially in the lower Alpine forelands.A soil type particularly affected by this development is gleysol.Today, the risk areas are mainly restricted to two habitat types: (1) a narrow, sharply delineated area along the northern edge of the Alps where summer rainfall decreased only slightly or even increased, and (2) wet soils of wetlands and riverine lowlands where the previous water regime remained unchanged.Our models show a clear decrease in the species frequency since the 1990s accompanied by a distinct shift in the regional occurrence mainly due to a change in rainfall distribution.Here, our findings confirm an earlier study by Suttner et al. 16 , who also observed a decline at most of the monitoring points and a shift in the distribution patterns.However, our risk shifts modeled are not fully congruent to the changes observed by Suttner et al. 16 as these authors only used data from the 'Bavarian Biotope Mapping Database' , which mainly includes protected areas and not agricultural land.Nevertheless, both analyses come to the conclusion that the 'increase' reported by practitioners is more of a local phenomenon and does not represent an overall direction in the spatial distribution of J. aquatica.
The trends detected for the recent past will continue in the future: both climate scenarios rcp4.5 and rcp8.5 predict decreasing summer rainfalls and increasing until 2037.Most likely, decreasing summer rainfall will move the potential risk areas to higher altitudes with sufficient rainfall, where low temperatures actually limit growth.In the lowlands, the risk areas along streams and rivers and the general probability of occurrence will decrease.This effect is even more pronounced with the rcp8.5 scenario, where the expected summer rainfall decrease is higher.As a result of climate change, such shifts and the retreat to higher altitudes are observed or expected for many species [58][59][60] .
Due to higher rainfalls and temperatures and the considerable decrease of frost and ice days predicted by both climate scenarios, an expansion of J. aquatica to higher altitudes in the Alps can be expected creating new potential risk areas.In the rcp8.5 scenario, however, this is mitigated by a stronger decrease in precipitation also in higher regions, while the decrease of risk in the Alpine forelands is somewhat less due to slightly higher precipitation there.However, as these new risk areas are comparatively small, they will not compensate for the areas with reduced risk.For agricultural practice, problems due to the spread of J. aquatica to higher altitudes can be estimated as low due to the minor importance of agriculture there.For the riverine grasslands on wet histosols, stagnosols, and gleysols in the extra-Alpine lowlands, our models also predict a decreasing infestation risk which indicates a substantial reduction of future management problems by J. aquatica there.
A climate change phenomenon which could not be reflected by our model is the expected increase in the torrentiality of precipitation events in summer 30 .The reason for this is the lack of high-resolution data.However, it can be assumed that more torrential precipitation will impair water infiltration by soils and therefore reinforce soil dryness, further reducing the infestation risk.
Management practice did not show significant effects on the occurrence of J. aquatica in our study.Neither organic vs. conventional farming or the implication of conservation schemes, nor land-use intensity in the form of stocking rates, the frequency of mowing and grazing, or the type and amount of fertilizationi or the showed an impact on J. aquatica occurrence.
While these results agree well with Forbes 10 who also found little or no impact of fertilization, cutting frequencies, or stocking rates on the occurrence of J. aquatica on 96 farms in Scotland, it contradicts various studies that report significant effects of different management practices from field experiments (e.g. 16,17,21,23,25).A major reason for this contradiction may be that the management methods applied in the experiments were specifically targeted at reducing J. aquatica populations.In contrast, our study and the analysis of Forbes 10 reflect the actual farming practice where the cost-effective achievement of fodder yields rather determines management decisions than on targeted weed control.Furthermore, J. aquatica does not always respond linearly to control measures and complex interactions of different measures also play a role 21,23 .In the study area, J. aquatica mainly occurs in wet meadows with a low to intermediate management intensity 21 .At this level of land-use intensity, conventional and organic grassland farming have much of their management practices in common, and the most noteworthy difference between the systems is in the application of organic or mineral fertilizers.As both types of fertilizers similarly stimulate the growth of J. aquatica, no significant difference between the effects of the two systems could be detected.Also, interactions between the schedule of management measures and population development can play a decisive role in establishing the species.Although J. aquatica may find favorable site conditions within a risk area, it needs dispersal and suitable germination conditions to colonize and infest potential areas.
Generally, the production of large numbers of wind-dispersed seeds enables J. aquatica to rapidly occupy areas in the vicinity of existing populations.Due to the light requirement of establishing seedlings, gaps in the grassland sward essentially facilitate the colonization of so far unoccupied areas 22 .Therefore, maintaining a close and vigorous plant canopy is an important tool to prevent infestation 19 .In our study, however, no such correlation between J. aquatica infestation and the occurrence of vegetation gaps was observed.Several reasons could have caused this result.Hence, when phases of seed dispersal and the availability of suitable gaps do not overlap, the risk of seed predation, mortality or false germination is substantially increased.Considering the short dispersal distances of Jacobaea sp.seeds 10,19,34 and the impact of wind direction, gaps suitable for germination may not have been close enough to seed-producing J. aquatica plants in our study.
Due to the relevance of the spread of the poisonous plant, J. aquatica's spatial and temporal interactions between seed production, seed dispersal, and the accessibility of sites with suitable germination conditions should become an important issue of further experiments and models.For practical farmers, prevention of seed production and maintenance of a close grassland sward should become essential precautionary measures to avoid future infestation problems.

Conclusions
The occurrence of Jacobaea aquatica, an indigenous species of wet grassland in Central Europe, is mainly determined by climatic factors.Although individual, targeted management measures in controlled experiments showed to reduce or avoid infestations by J. aquatica, our study shows that in the real, complex agricultural www.nature.com/scientificreports/environment different management does not influence the occurrence of the species.Simple solutions such as extensification or organic farming as a solution to avoid J. aquatica infestation are therefore out of the question.
Our study reveals how decreasing precipitation can deteriorate habitat quality mainly on terrestrial soils and thus lead to a large-scale decline of this species.However, this study also displays that the habitat conditions of J. aquatica have not deteriorated across its whole distribution range.In some regions where climate change does not negatively affect the living conditions, the species even show an increase.
Our results agree with climate models, which generally indicate that future decline in precipitation patterns will considerably reinforce such losses of species richness in wet grassland.Considering the high nature conservation value and the strong decline of such habitats, measures to preserve these ecosystems by rewetting and extensive management seem appropriate.On the other hand, this decline of highly poisonous J. aquatica will also facilitate the production of uncontaminated food for humans and animals.Therefore, a major future challenge for the management of J. aquatica habitats is to find a way to meet both objectives, the conservation of the species threatened by climate and land-use change and the production of healthy food.

Figure 1 .
Figure 1.Location of the study sites in Bavaria with seven climate regions separated by dashed black lines (see main text).Black dots indicate sampling sites with Jacobaea aquatica and a set of randomly selected pseudoabsences.

Figure 2 .
Figure 2. Partial dependence plot of all individual models (in different colors) for the most influential predictors of the occurrence of Jacobaea aquatica in Bavarian grasslands: (a) summer rainfall (rainJJA), and (b) soil class ordered along a gradient of decreasing soil moisture (see Supplementary TableA2).

Figure 3 .
Figure 3. Multi-predictor partial dependence plot with the probability of occurrence of Jacobaea aquatica in Bavaria for summer rain (rainJJA) and soil type (see Supplementary TableA2: Soil classes).The orange line indicates the cut-off value at 0.35.
Figure 3. Multi-predictor partial dependence plot with the probability of occurrence of Jacobaea aquatica in Bavaria for summer rain (rainJJA) and soil type (see Supplementary TableA2: Soil classes).The orange line indicates the cut-off value at 0.35.
Figure 3. Multi-predictor partial dependence plot with the probability of occurrence of Jacobaea aquatica in Bavaria for summer rain (rainJJA) and soil type (see Supplementary TableA2: Soil classes).The orange line indicates the cut-off value at 0.35.

Figure 4 .
Figure 4. Predicted probability of the potential occurrence of Jacobaea aquatica in Bavaria for the time periods 1988-1997 and 2008-2017 based on an averaged gradient-boosted model.Colors represent percentages of risk areas from green (< 35% of all 50 m × 50 m grid cells show > 60% likelihood of occurrence) to red (> 80% of the cells have a high contamination risk).

Figure 5 .
Figure 5. Changes of the occurrence risk of grasslands by Jacobaea aquatica in Bavaria from 1988 to 1997 to 2008-2017.Dark red indicates areas where the model predicts a new occurrence of the species in 2007-2017, in the light red areas J. aquatica already occurred but the appearance became 20% more likely.Similarly, dark blue indicates areas where the model predicts the extinction of former findings and light blue represents areas where J. aquatica already occurred in 1988-1997 but survival became at least 20% more unlikely.

Figure 6 .
Figure 6.Predicted risk of Jacobaea aquatica occurrence in Bavaria for the time period 2028-2037 based on the rcp4.5 and rcp8.5 scenarios, respectively.

Figure 7 .
Figure 7. ROC curve of averaged random forest classifiers for the presence of Jacobaea aquatica in Bavaria with management parameters as predictors; and confusion matrix of the random forest classifiers for the frequency of the species in dependency of management factors. https://doi.org/10.1038/s41598-023-40646-z