Cross-validation of species distribution models: removing spatial sorting bias and calibration with a null model

Species distribution models are usually evaluated with cross-validation. In this procedure evaluation statistics are computed from model predictions for sites of presence and absence that were not used to train (fit) the model. Using data for 226 species, from six regions, and two species distribution modeling algorithms (Bioclim and MaxEnt), I show that this procedure is highly sensitive to "spatial sorting bias": the difference between the geographic distance from testing-presence to training-presence sites and the geographic distance from testing-absence (or testing-background) to training-presence sites. I propose the use of pairwise distance sampling to remove this bias, and the use of a null model that only considers the geographic distance to training sites to calibrate cross-validation results for remaining bias. Model evaluation results (AUC) were strongly inflated: the null model performed better than MaxEnt for 45% and better than Bioclim for 67% of the species. Spatial sorting bias and area under the receiver-operator curve (AUC) values increased when using partitioned presence data and random-absence data instead of independently obtained presence-absence testing data from systematic surveys. Pairwise distance sampling removed spatial sorting bias, yielding null models with an AUC close to 0.5, such that AUC was the same as null model calibrated AUC (cAUC). This adjustment strongly decreased AUC values and changed the ranking among species. Cross-validation results for different species are only comparable after removal of spatial sorting bias and/or calibration with an appropriate null model.     

The effect of species response form on species distribution model prediction and inference

Ecological theory and current evidence support the validity of various species response curves according to a variety of environmental gradients. Various methods have been developed for building species distribution models but it is not well known how these methods perform under various assumptions about the form of the underlying species response. It is also not well known how spatial correlation in species occurrence affects model performance. These effects were investigated by applying an environmental envelope method (BIOCLIM) and three regression-based methods: logistic regression (LR), generalized additive modelling (GAM), and classification and regression tree (CART) to simulated species occurrence data. Each simulated species was constructed as a sum of responses with varying weights. Three basic species response curves were assumed: Gaussian (bell-shaped), Beta (skew) and linear. The two non-linear responses conform to standard ecological niche theory. All three responses were applied in turn to three simulated environmental variables, each with varying degrees of spatial autocorrelation. GAM produced the most consistent model performance over all forms of simulated species response. BIOCLIM and CART were inclined to underrate the performance of variables with a linear response. BIOCLIM was less sensitive to data density. LR was susceptible to model misspecification. The use of a linear function in LR underestimated the performance of variables with non-linear species response and contributed to increased spatial autocorrelation in model residuals. Omission of important environmental variables with non-linear species response also contributed to increased spatial autocorrelation in model residuals. Adding a spatial autocovariate term to the LR model (autologistic model) reduced the spatial autocorrelation and improved model performance, but did not correct the misidentification of the dominant environmental determinant. This is to be expected since the autologistic approach was designed primarily for prediction and not for inference. Given that various forms of species response to environmental determinants arise commonly in nature: (1) higher order functions should always be tested when applying LR in modelling species distribution; (2) spatial autocorrelation in species distribution model residuals can indicate that environmental determinants with non-linear response are missing from the model; and (3) deficiencies in LR model performance due to model misspecification can be addressed by adding a spatial autocovariate to the model, but care should be taken when interpreting the coefficients of the model parameters.     

The importance of spatial autocorrelation,extent and resolution in predicting forest bird occurrence

Concerns about declines in forest biodiversity underscore the need for accurate estimates of the distribution and abundance of organisms at large scales and at resolutions that are fine enough to be appropriate for management. This paper addresses three major objectives: (i) to determine whether the resolution of typical air photo-derived forest inventory is sufficient for the accurate prediction of site occupancy by forest birds. We compared prediction success of habitat models using air photo variables to models with variables derived from finer resolution, ground-sampled vegetation plots. (ii) To test whether incorporating spatial autocorrelation into habitat models via autologistic regression increases prediction success. (iii) To determine whether landscape structure is an important factor in predicting bird distribution in forest-dominated landscapes. Models were tested locally (Greater Fundy Ecosystem [GFE]) using cross-validation, and regionally using an independent data set from an area located ca. 250 km to the northwest (Riley Brook [RB]). We found significant positive spatial autocorrelation in the residuals of at least one habitat model for 76% (16/21) of species examined. In these cases, the logistic regression assumption of spatially independent errors was violated. Logistic models that ignored spatial autocorrelation tended to overestimate habitat effects. Though overall prediction success was higher for autologistic models than logistic models in the GFE, the difference was only significantly improved for one species. Further, the inclusion of spatial covariates did little to improve model performance in the geographically discrete study area. For 62% (13/21) of species examined, landscape variables were significant predictors of forest bird occurrence even after statistically controlling for stand-level variability. However, broad spatial extents explained less variation than local factors. In the GFE, 76% (16/21) of air photo and 81% (17/21) of ground plot models were accurate enough to be of practical utility (AUC > 0.7). When applied to RB, both model types performed effectively for 55% (11/20) of the species examined. We did not detect an overall difference in prediction success between air photo and ground plot models in either study area. We conclude that air photo data are as effective as fine resolution vegetation data for predicting site occupancy for the majority of species in this study. These models will be of use to forest managers who are interested in mapping species distributions under various timber harvest scenarios, and to protected areas planners attempting to optimize reserve function.     

Comparing wildlife habitat suitability models based on expert opinion with camera trap detections

Expert knowledge is used in the development of wildlife habitat suitability models (HSMs) for management and conservation decisions. However, the consistency of such models has been questioned. Focusing on 1 method for elicitation, the analytic hierarchy process, we generated expert-based HSMs for 4 felid species: 2 forest specialists (ocelot [Leopardus pardalis] and margay [Leopardus wiedii]) and 2 habitat generalist species (Pampas cat [Leopardus colocola] and puma [Puma concolor]). Using these HSMs, species detections from camera-trap surveys, and generalized linear models, we assessed the effect of study species and expert attributes on the correspondence between expert models and camera-trap detections. We also examined whether aggregation of participant responses and iterative feedback improved model performance. We ran 160 HSMs and found that models for specialist species showed higher correspondence with camera-trap detections (AUC [area under the receiver operating characteristic curve] >0.7) than those for generalists (AUC < 0.7). Model correspondence increased as participant years of experience in the study area increased, but only for the understudied generalist species, Pampas cat (β = 0.024 [SE 0.007]). No other participant attribute was associated with model correspondence. Feedback and revision of models improved model correspondence, and aggregating judgments across multiple participants improved correspondence only for specialist species. The average correspondence of aggregated judgments increased as group size increased but leveled off after 5 experts for all species. Our results suggest that correspondence between expert models and empirical surveys increases as habitat specialization increases. We encourage inclusion of participants knowledgeable of the study area and model validation for expert-based modeling of understudied and generalist species.     

Predicting tree distributions in an East African biodiversity hotspot: model selection, data bias and envelope uncertainty

The Eastern Arc Mountains (EAMs) of Tanzania and Kenya support some of the most ancient tropical rainforest on Earth. The forests are a global priority for biodiversity conservation and provide vital resources to the Tanzanian population. Here, we make a first attempt to predict the spatial distribution of 40 EAM tree species, using generalised additive models, plot data and environmental predictor maps at sub 1 km resolution. The results of three modelling experiments are presented, investigating predictions obtained by (1) two different procedures for the stepwise selection of predictors, (2) down-weighting absence data, and (3) incorporating an autocovariate term to describe fine-scale spatial aggregation. In response to recent concerns regarding the extrapolation of model predictions beyond the restricted environmental range of training data, we also demonstrate a novel graphical tool for quantifying envelope uncertainty in restricted range niche-based models (envelope uncertainty maps). We find that even for species with very few documented occurrences useful estimates of distribution can be achieved. Initiating selection with a null model is found to be useful for explanatory purposes, while beginning with a full predictor set can over-fit the data. We show that a simple multimodel average of these two best-model predictions yields a superior compromise between generality and precision (parsimony). Down-weighting absences shifts the balance of errors in favour of higher sensitivity, reducing the number of serious mistakes (i.e., falsely predicted absences); however, response functions are more complex, exacerbating uncertainty in larger models. Spatial autocovariates help describe fine-scale patterns of occurrence and significantly improve explained deviance, though if important environmental constraints are omitted then model stability and explanatory power can be compromised. We conclude that the best modelling practice is contingent both on the intentions of the analyst (explanation or prediction) and on the quality of distribution data; generalised additive models have potential to provide valuable information for conservation in the EAMs, but methods must be carefully considered, particularly if occurrence data are scarce. Full results and details of all species models are supplied in an online Appendix.     

Modeling species-habitat relationships with spatially autocorrelated observation data.

Spatial autocorrelation in wildlife observation data arises when extrinsic environmental processes and patterns that influence the spatial distribution of wildlife are themselves spatially structured, or when species are subject to intrinsic population processes, causing contagion or dispersion effects. Territoriality, Allee effects, dispersal limitations, and social clustering are examples of intrinsic processes. Both forms of autocorrelation can violate the assumptions of generalized linear regression models, resulting in biased estimation of model coefficients and diminished predictive performance. Such consequences may be avoided for extrinsic autocorrelation when autocorrelated environmental variables are available for use as model covariates, whereas intrinsic spatial autocorrelation requires an alternative modeling approach. The autologistic model provides an approach suited to the binary observations often obtained in wildlife surveys, but its performance has not been tested across widely varying sampling intensities or strengths of intrinsic spatial structure. Here we use simulated data to test the autologistic model under a range of sampling conditions. The autologistic model obtains better fits and substantially better predictive performance than the standard logistic regression model over the full range of sampling designs and intensities tested. We provide a simple Bayesian implementation of the autologistic model, which until now has not been achieved with standard statistical software alone. A step-by-step procedure is given for characterizing and modeling spatial autocorrelation in binary observation data, along with computer code for fitting autologistic models in WinBUGS, a freeware Bayesian analysis package. This approach avoids normal approximations to the pseudo-likelihood, in contrast to previous Bayesian applications of the autologistic model. We provide a sample application of the autologistic model, fitted to survey data for a gliding marsupial in southeastern Australia.     

The performance of state-of-the-art modelling techniques depends on geographical distribution of species

We explored the effects of prevalence, latitudinal range and clumping (spatial autocorrelation) of species distribution patterns on the predictive accuracy of eight state-of-the-art modelling techniques: Generalized Linear Models (GLMs), Generalized Boosting Method (GBM), Generalized Additive Models (GAMs), Classification Tree Analysis (CTA), Artificial Neural Network (ANN), Multivariate Adaptive Regression Splines (MARS), Mixture Discriminant Analysis (MDA) and Random Forest (RF). One hundred species of Lepidoptera, selected from the Distribution Atlas of European Butterflies, and three climate variables were used to determine the bioclimatic envelope for each butterfly species. The data set consisting of 2620 grid squares 30′ × 60′ in size all over Europe was randomly split into the calibration and the evaluation data sets. The performance of different models was assessed using the area under the curve (AUC) of a receiver operating characteristic (ROC) plot. Observed differences in modelling accuracy among species were then related to the geographical attributes of the species using GAM. The modelling performance was negatively related to the latitudinal range and prevalence, whereas the effect of spatial autocorrelation on prediction accuracy depended on the modelling technique. These three geographical attributes accounted for 19–61% of the variation in the modelling accuracy. Predictive accuracy of GAM, GLM and MDA was highly influenced by the three geographical attributes, whereas RF, ANN and GBM were moderately, and MARS and CTA only slightly affected. The contrasting effects of geographical distribution of species on predictive performance of different modelling techniques represent one source of uncertainty in species spatial distribution models. This should be taken into account in biogeographical modelling studies and assessments of climate change impacts.     

Analyzing and modelling spatial distribution of summering lesser kestrel: The role of spatial autocorrelation

In modelling spatial distribution of species, ignoring spatial autocorrelation (SA) and multicollinearity may lead to false ecological conclusions. Here we take into account both issues for examining and modelling the spatial pattern of abundance of the globally threatened lesser kestrel (Falco naumanni) during summer in a 38,400 ha area of northwestern Spain where large premigratory aggregations of the species occur. Spatial pattern was examined using Moran's correlogram, and models were built including geographical coordinates and autocovariate terms (which account for SA) in generalized linear models (GLM) and hierarchical partitioning (HP) models. HP models allow to alleviate multicollinearity. A grid-based approach was used by dividing the study area in 24 contiguous 4 km × 4 km squares where birds were counted in 2–3 visits per square (response variable). Environmental coarse-grained variables were extracted from a geographic information system (GIS) at three spatial extents. Moran's correlogram showed that lesser kestrel mean abundance per square was spatially autocorrelated up to 4–8 km. The results from both GLM and HP analyses were roughly compatible. The GLM models explained 80.0% of the variation in kestrel abundance and were the same at the three spatial extents. Lesser Kestrel abundance was not significantly explained by landscape variables, but was negatively related to both the distance to the nearest communal roost and distance to the nearest breeding colony with more of 10 breeding pairs of lesser kestrel. An autocovariate term added later in the GLM models improved both their explanatory power (from 74.5 to 80.0%) and model residuals, which were not longer spatially autocorrelated, fulfilling thus the statistical assumption of independent errors. Findings suggest that the spatial distribution of abundance of summering lesser kestrel is, at least, partially driven by endogenous causes, such as conspecific attraction. Exogenous causes such as finer-scale variables (e.g. type of crops and food available) are yet likely needed for lesser kestrel-environment relationships.     

Modelling the horizontal spatial structure of planktonic community in Lake Trasimeno (Umbria,Italy) using multivariate geostatistical methods

Spatial heterogeneity of ecological systems has been recognised in recent years as an important ecological feature of an ecosystem, rather than a mere statistical nuisance. However, although considerable interest has been paid to the development of statistical methods for the analysis of spatial environmental data, when in presence of more species or environmental variables common analyses still fail to recognise the necessity of a joint modelling of the whole correlation structure. In this paper, we propose to study the multivariate spatial autocorrelation of a plankton community by making explicit reference to a spatial linear factor model entailing a set of constraints for the spatial structure of the planktonic species. The data set examined come from an intensive 2-day sampling survey performed in July 1991 on Lake Trasimeno (Italy) to investigate the horizontal spatial heterogeneity and distribution of the planktonic community, from small (50 m) to large (1000–10,000 m) scale. The analysis revealed that zooplankton and phytoplankton essentially have different degrees of heterogeneity and different spatial structures which required separate modelling. On the other hand, the similarity of the spatial autocorrelation found within zooplankton and phytoplankton communities, indicates that at the investigated scales of observation the horizontal organisation of both components is not appreciably affected by species-specific behaviours. The analysis of the multivariate spatial patterns emerging from the mapping of the extracted factors suggested an interpretation of the distribution of macrozooplankton and phytoplankton assemblages in terms of planktonic responses to environmental factors of a lake-size scale.     

Assessing area-specific relative risks from large forest fire size in Canada

The evaluation of area-specific risks for large fires is of great policy relevance to fire management and prevention. When analyzing data for the burned areas of large fires in Canada, we found that there are dramatic patterns that cannot be adequately modelled by traditional hierarchical modelling assuming spatial autocorrelation. In this paper, we use the robust locally weighted scatterplot smoothing (LOESS) technique to remove spatial and temporal trends; and we account for periodical cycles by employing the relevant periodic functions as covariates in a hierarchical Gamma mixed effects model. Based on the results of this generalized multilevel analysis of large fire size, we provide an area-specific relative risks ranking system for Canada and confirm that lightning tends to cause more severe damage in terms of fire size than human factor. A diagnostic check on the modelling shows that large fires data are reasonably modelled using this combination of semiparametric and mixed effects modelling approaches.     

Generalized site occupancy models allowing for false positive and false negative errors

Site occupancy models have been developed that allow for imperfect species detection or "false negative" observations. Such models have become widely adopted in surveys of many taxa. The most fundamental assumption underlying these models is that "false positive" errors are not possible. That is, one cannot detect a species where it does not occur. However, such errors are possible in many sampling situations for a number of reasons, and even low false positive error rates can induce extreme bias in estimates of site occupancy when they are not accounted for. In this paper, we develop a model for site occupancy that allows for both false negative and false positive error rates. This model can be represented as a two-component finite mixture model and can be easily fitted using freely available software. We provide an analysis of avian survey data using the proposed model and present results of a brief simulation study evaluating the performance of the maximum-likelihood estimator and the naive estimator in the presence of false positive errors.     

Using Niche-Based Models to Improve the Sampling of Rare Species

Abstract:  Because data on rare species usually are sparse, it is important to have efficient ways to sample additional data. Traditional sampling approaches are of limited value for rare species because a very large proportion of randomly chosen sampling sites are unlikely to shelter the species. For these species, spatial predictions from niche-based distribution models can be used to stratify the sampling and increase sampling efficiency. New data sampled are then used to improve the initial model. Applying this approach repeatedly is an adaptive process that may allow increasing the number of new occurrences found. We illustrate the approach with a case study of a rare and endangered plant species in Switzerland and a simulation experiment. Our field survey confirmed that the method helps in the discovery of new populations of the target species in remote areas where the predicted habitat suitability is high. In our simulations the model-based approach provided a significant improvement (by a factor of 1.8 to 4 times, depending on the measure) over simple random sampling. In terms of cost this approach may save up to 70% of the time spent in the field.     

Sensitivity of ecological models to their climate drivers: statistical ensembles for forcing.

Global and regional numerical models for terrestrial ecosystem dynamics require fine spatial resolution and temporally complete historical climate fields as input variables. However, because climate observations are unevenly spaced and have incomplete records, such fields need to be estimated. In addition, uncertainty in these fields associated with their estimation are rarely assessed. Ecological models are usually driven with a geostatistical model's mean estimate (kriging) of these fields without accounting for this uncertainty, much less evaluating such errors in terms of their propagation in ecological simulations. We introduce a Bayesian statistical framework to model climate observations to create spatially uniform and temporally complete fields, taking into account correlation in time and space, spatial heterogeneity, lack of normality, and uncertainty about all these factors. A key benefit of the Bayesian model is that it generates uncertainty measures for the generated fields. To demonstrate this method, we reconstruct historical monthly precipitation fields (a driver for ecological models) on a fine resolution grid for a climatically heterogeneous region in the western United States. The main goal of this work is to evaluate the sensitivity of ecological models to the uncertainty associated with prediction of their climate drivers. To assess their numerical sensitivity to predicted input variables, we generate a set of ecological model simulations run using an ensemble of different versions of the reconstructed fields. We construct such an ensemble by sampling from the posterior predictive distribution of the climate field. We demonstrate that the estimated prediction error of the climate field can be very high. We evaluate the importance of such errors in ecological model experiments using an ensemble of historical precipitation time series in simulations of grassland biogeochemical dynamics with an ecological numerical model, Century. We show how uncertainty in predicted precipitation fields is propagated into ecological model results and that this propagation had different modes. Depending on output variable, the response of model dynamics to uncertainty in inputs ranged from uncertainty in outputs that matched that of inputs to those that were muted or that were biased, as well as uncertainty that was persistent in time after input errors dropped.     

Error propagation in spatial modeling of public health data: a simulation approach using pediatric blood lead level data for Syracuse,New York

Lead poisoning produces serious health problems, which are worse when a victim is younger. The US government and society have tried to prevent lead poisoning, especially since the 1970s; however, lead exposure remains prevalent. Lead poisoning analyses frequently use georeferenced blood lead level data. Like other types of data, these spatial data may contain uncertainties, such as location and attribute measurement errors, which can propagate to analysis results. For this paper, simulation experiments are employed to investigate how selected uncertainties impact regression analyses of blood lead level data in Syracuse, New York. In these simulations, location error and attribute measurement error, as well as a combination of these two errors, are embedded into the original data, and then these data are aggregated into census block group and census tract polygons. These aggregated data are analyzed with regression techniques, and comparisons are reported between the regression coefficients and their standard errors for the error added simulation results and the original results. To account for spatial autocorrelation, the eigenvector spatial filtering method and spatial autoregressive specifications are utilized with linear and generalized linear models. Our findings confirm that location error has more of an impact on the differences than does attribute measurement error, and show that the combined error leads to the greatest deviations. Location error simulation results show that smaller administrative units experience more of a location error impact, and, interestingly, coefficients and standard errors deviate more from their true values for a variable with a low level of spatial autocorrelation. These results imply that uncertainty, especially location error, has a considerable impact on the reliability of spatial analysis results for public health data, and that the level of spatial autocorrelation in a variable also has an impact on modeling results.     

Hierarchical analysis of species distributions and abundance across environmental gradients

Abiotic and biotic processes operate at multiple spatial and temporal scales to shape many ecological processes, including species distributions and demography. Current debate about the relative roles of niche-based and stochastic processes in shaping species distributions and community composition reflects, in part, the challenge of understanding how these processes interact across scales. Traditional statistical models that ignore autocorrelation and spatial hierarchies can result in misidentification of important ecological covariates. Here, we demonstrate the utility of a hierarchical modeling framework for testing hypotheses about the importance of abiotic factors at different spatial scales and local spatial autocorrelation for shaping species distributions and abundances. For the two orchid species studied, understory light availability and soil moisture helped to explain patterns of presence and abundance at a microsite scale (<4 m2), while soil organic content was important at a population scale (<400 m2). The inclusion of spatial autocorrelation is shown to alter the magnitude and certainty of estimated relationships between abundance and abiotic variables, and we suggest that such analysis be used more often to explore the relationships between species life histories and distributions. The hierarchical modeling framework is shown to have great potential for elucidating ecological relationships involving abiotic and biotic processes simultaneously at multiple scales.     

Prevalence-adjusted optimisation of fuzzy models for species distribution

Most performance criteria which have been applied to train ecological models focus on the accuracy of the model predictions. However, these criteria depend on the prevalence of the training set and often do not take into account ecological issues such as the distinction between omission and commission errors. Moreover, a previous study indicated that model training based on different performance criteria results in different optimised models. Therefore, model developers should train models based on different performance criteria and select the most appropriate model depending on the modelling objective. This paper presents a new approach to train fuzzy models based on an adjustable performance criterion, called the adjusted average deviation (aAD). This criterion was applied to develop a species distribution model for spawning grayling in the Aare River near Thun, Switzerland. To analyse the strengths and weaknesses of this approach, it was compared to model training based on other performance criteria. The results suggest that model training based on accuracy-based performance criteria may produce unrealistic models at extreme prevalences of the training set, whereas the aAD allows for the identification of more accurate and more reliable models. Moreover, the adjustable parameter in this criterion enables modellers to situate the optimised models in the search space and thus provides an indication of the ecological model relevance. Consequently, it may support modellers and river managers in the decision making process by improving model reliability and insight into the modelling process. Due to the universality and the flexibility of the approach, it could be applied to any other ecosystem or species, and may therefore be valuable to ecological modelling and ecosystem management in general.     

Rarefaction and nonrandom spatial dispersion patterns

Rarefaction estimates how many species are expected in a random sample of individuals from a larger collection and allows meaningful comparisons among collections of different sizes. It assumes random spatial dispersion. However, two common dispersion patterns, within-species clumping and segregation among species, can cause rarefaction to overestimate the species richness of a smaller continuous area. We use field studies and computer simulations to determine (1) how robust rarefaction is to nonrandom spatial dispersion and (2) whether simple measures of spatial autocorrelation can predict the bias in rarefaction estimates. Rarefaction does not estimate species richness accurately for many communities, especially at small sample sizes. Measures of spatial autocorrelation of the more abundant species do not reliably predict amount of bias. Survey sites should be standardized to equal-sized areas before sampling. When sites are of equal area but differ in number of individuals sampled, rarefaction can standardize collections. When communities are sampled from different-sized areas, the mean and confidence intervals of species accumulation curves allow more meaningful comparisons among sites. Electronic supplementary material  The online version of this article (doi:) contains supplementary material, which is available to authorized users.

Building statistical models to analyze species distributions.

Models of the geographic distributions of species have wide application in ecology. But the nonspatial, single-level, regression models that ecologists have often employed do not deal with problems of irregular sampling intensity or spatial dependence, and do not adequately quantify uncertainty. We show here how to build statistical models that can handle these features of spatial prediction and provide richer, more powerful inference about species niche relations, distributions, and the effects of human disturbance. We begin with a familiar generalized linear model and build in additional features, including spatial random effects and hierarchical levels. Since these models are fully specified statistical models, we show that it is possible to add complexity without sacrificing interpretability. This step-by-step approach, together with attached code that implements a simple, spatially explicit, regression model, is structured to facilitate self-teaching. All models are developed in a Bayesian framework. We assess the performance of the models by using them to predict the distributions of two plant species (Proteaceae) from South Africa's Cape Floristic Region. We demonstrate that making distribution models spatially explicit can be essential for accurately characterizing the environmental response of species, predicting their probability of occurrence, and assessing uncertainty in the model results. Adding hierarchical levels to the models has further advantages in allowing human transformation of the landscape to be taken into account, as well as additional features of the sampling process.     

Modelling plant species richness using functional groups

Conservation biologists increasingly rely on spatial predictive models of biodiversity to support decision-making. Therefore, highly accurate and ecologically meaningful models are required at relatively broad spatial scales. While statistical techniques have been optimized to improve model accuracy, less focus has been given to the question: How does the autecology of a single species affect model quality? We compare a direct modelling approach versus a cumulative modelling approach for predicting plant species richness, where the latter gives more weight to the ecology of functional species groups. In the direct modelling approach, species richness is predicted by a single model calibrated for all species. In the cumulative modelling approach, the species were partitioned into functional groups, with each group calibrated separately and species richness of each group was cumulated to predict total species richness. We hypothesized that model accuracy depends on the ecology of individual species and that the cumulative modelling approach would predict species richness more accurately. The predictors explained plant species richness by ca. 25%. However, depending on the functional group the deviance explained varied from 3 to 67%. While both modelling approaches performed equally well, the models of the different functional groups highly varied in their quality and their spatial richness pattern. This variability helps to improve our understanding on how plant functional groups respond to ecological gradients.