Effects of sample size on the performance of species distribution models

A wide range of modelling algorithms is used by ecologists, conservation practitioners, and others to predict species ranges from point locality data. Unfortunately, the amount of data available is limited for many taxa and regions, making it essential to quantify the sensitivity of these algorithms to sample size. This is the first study to address this need by rigorously evaluating a broad suite of algorithms with independent presence-absence data from multiple species and regions. We evaluated predictions from 12 algorithms for 46 species (from six different regions of the world) at three sample sizes (100, 30, and 10 records). We used data from natural history collections to run the models, and evaluated the quality of model predictions with area under the receiver operating characteristic curve (AUC). With decreasing sample size, model accuracy decreased and variability increased across species and between models. Novel modelling methods that incorporate both interactions between predictor variables and complex response shapes (i.e. GBM, MARS-INT, BRUTO) performed better than most methods at large sample sizes but not at the smallest sample sizes. Other algorithms were much less sensitive to sample size, including an algorithm based on maximum entropy (MAXENT) that had among the best predictive power across all sample sizes. Relative to other algorithms, a distance metric algorithm (DOMAIN) and a genetic algorithm (OM-GARP) had intermediate performance at the largest sample size and among the best performance at the lowest sample size. No algorithm predicted consistently well with small sample size (n < 30) and this should encourage highly conservative use of predictions based on small sample size and restrict their use to exploratory modelling.

Modelling plant species distribution in alpine grasslands using airborne imaging spectroscopy.

Remote sensing using airborne imaging spectroscopy (AIS) is known to retrieve fundamental optical properties of ecosystems. However, the value of these properties for predicting plant species distribution remains unclear. Here, we assess whether such data can add value to topographic variables for predicting plant distributions in French and Swiss alpine grasslands. We fitted statistical models with high spectral and spatial resolution reflectance data and tested four optical indices sensitive to leaf chlorophyll content, leaf water content and leaf area index. We found moderate added-value of AIS data for predicting alpine plant species distribution. Contrary to expectations, differences between species distribution models (SDMs) were not linked to their local abundance or phylogenetic/functional similarity. Moreover, spectral signatures of species were found to be partly site-specific. We discuss current limits of AIS-based SDMs, highlighting issues of scale and informational content of AIS data.

Species distribution models for crop pollination: a modelling framework applied to Great Britain

Insect pollination benefits over three quarters of the world's major crops. There is growing concern that observed declines in pollinators may impact on production and revenues from animal pollinated crops. Knowing the distribution of pollinators is therefore crucial for estimating their availability to pollinate crops; however, in general, we have an incomplete knowledge of where these pollinators occur. We propose a method to predict geographical patterns of pollination service to crops, novel in two elements: the use of pollinator records rather than expert knowledge to predict pollinator occurrence, and the inclusion of the managed pollinator supply. We integrated a maximum entropy species distribution model (SDM) with an existing pollination service model (PSM) to derive the availability of pollinators for crop pollination. We used nation-wide records of wild and managed pollinators (honey bees) as well as agricultural data from Great Britain. We first calibrated the SDM on a representative sample of bee and hoverfly crop pollinator species, evaluating the effects of different settings on model performance and on its capacity to identify the most important predictors. The importance of the different predictors was better resolved by SDM derived from simpler functions, with consistent results for bees and hoverflies. We then used the species distributions from the calibrated model to predict pollination service of wild and managed pollinators, using field beans as a test case. The PSM allowed us to spatially characterize the contribution of wild and managed pollinators and also identify areas potentially vulnerable to low pollination service provision, which can help direct local scale interventions. This approach can be extended to investigate geographical mismatches between crop pollination demand and the availability of pollinators, resulting from environmental change or policy scenarios.

Habitat predictive modelling of demersal fish species in the Azores

Dissertação de Mestrado, Estudos Integrados dos Oceanos, 25 de Março de 2013, Universidade dos Açores.

Combining food web and species distribution models for improved community projections

The ability to model biodiversity patterns is of prime importance in this era of severe environmental crisis. Species assemblage along environmental gradient is subject to the interplay of biotic interactions in complement to abiotic environmental filtering. Accounting for complex biotic interactions for a wide array of species remains so far challenging. Here, we propose to use food web models that can infer the potential interaction links between species as a constraint in species distribution models. Using a plant-herbivore (butterfly) interaction dataset, we demonstrate that this combined approach is able to improve both species distribution and community forecasts. Most importantly, this combined approach is very useful in rendering models of more generalist species that have multiple potential interaction links, where gap in the literature may be recurrent. Our combined approach points a promising direction forward to model the spatial variation of entire species interaction networks. Our work has implications for studies of range shifting species and invasive species biology where it may be unknown how a given biota might interact with a potential invader or in future climate.

Overcoming the rare species modelling paradox: test of a novel hierarchical framework with an Iberian endemic plant

Rare species have restricted geographic ranges, habitat specialization, and/or small population sizes. Datasets on rare species distribution usually have few observations, limited spatial accuracy and lack of valid absences; conversely they provide comprehensive views of species distributions allowing to realistically capture most of their realized environmental niche. Rare species are the most in need of predictive distribution modelling but also the most difficult to model. We refer to this contrast as the "rare species modelling paradox" and propose as a solution developing modelling approaches that deal with a sufficiently large set of predictors, ensuring that statistical models aren't overfitted. Our novel approach fulfils this condition by fitting a large number of bivariate models and averaging them with a weighted ensemble approach. We further propose that this ensemble forecasting is conducted within a hierarchic multi-scale framework. We present two ensemble models for a test species, one at regional and one at local scale, each based on the combination of 630 models. In both cases, we obtained excellent spatial projections, unusual when modelling rare species. Model results highlight, from a statistically sound approach, the effects of multiple drivers in a same modelling framework and at two distinct scales. From this added information, regional models can support accurate forecasts of range dynamics under climate change scenarios, whereas local models allow the assessment of isolated or synergistic impacts of changes in multiple predictors. This novel framework provides a baseline for adaptive conservation, management and monitoring of rare species at distinct spatial and temporal scales.

Improving the prediction of plant species distribution and community composition by adding edaphic to topo-climatic variables

Questions Soil properties have been widely shown to influence plant growth and distribution. However, the degree to which edaphic variables can improve models based on topo-climatic variables is still unclear. In this study, we tested the roles of seven edaphic variables, namely (1) pH; (2) the content of nitrogen and of (3) phosphorus; (4) silt; (5) sand; (6) clay and (7) carbon-to-nitrogen ratio, as predictors of species distribution models in an edaphically heterogeneous landscape. We also tested how the respective influence of these variables in the models is linked to different ecological and functional species characteristics. Location The Western Alps, Switzerland. Methods With four different modelling techniques, we built models for 115 plant species using topo-climatic variables alone and then topo-climatic variables plus each of the seven edaphic variables, one at a time. We evaluated the contribution of each edaphic variable by assessing the change in predictive power of the model. In a second step, we evaluated the importance of the two edaphic variables that yielded the largest increase in predictive power in one final set of models for each species. Third, we explored the change in predictive power and the importance of variables across plant functional groups. Finally, we assessed the influence of the edaphic predictors on the prediction of community composition by stacking the models for all species and comparing the predicted communities with the observed community. Results Among the set of edaphic variables studied, pH and nitrogen content showed the highest contributions to improvement of the predictive power of the models, as well as the predictions of community composition. When considering all topo-climatic and edaphic variables together, pH was the second most important variable after degree-days. The changes in model results caused by edaphic predictors were dependent on species characteristics. The predictions for the species that have a low specific leaf area, and acidophilic preferences, tolerating low soil pH and high humus content, showed the largest improvement by the addition of pH and nitrogen in the model. Conclusions pH was an important predictor variable for explaining species distribution and community composition of the mountain plants considered in our study. pH allowed more precise predictions for acidophilic species. This variable should not be neglected in the construction of species distribution models in areas with contrasting edaphic conditions.

Accuracy of plant assemblage predictions from species distribution models varies along environmental gradients

Aim: Climatic niche modelling of species and community distributions implicitly assumes strong and constant climatic determinism across geographic space. This assumption had however never been tested so far. We tested it by assessing how stacked-species distribution models (S-SDMs) perform for predicting plant species assemblages along elevation. Location: Western Swiss Alps. Methods: Using robust presence-absence data, we first assessed the ability of topo-climatic S-SDMs to predict plant assemblages in a study area encompassing a 2800 m wide elevation gradient. We then assessed the relationships among several evaluation metrics and trait-based tests of community assembly rules. Results: The standard errors of individual SDMs decreased significantly towards higher elevations. Overall, the S-SDM overpredicted far more than they underpredicted richness and could not reproduce the humpback curve along elevation. Overprediction was greater at low and mid-range elevations in absolute values but greater at high elevations when standardised by the actual richness. Looking at species composition, the evaluation metrics accounting for both the presence and absence of species (overall prediction success and kappa) or focusing on correctly predicted absences (specificity) increased with increasing elevation, while the metrics focusing on correctly predicted presences (Jaccard index and sensitivity) decreased. The best overall evaluation - as driven by specificity - occurred at high elevation where species assemblages were shown to be under significant environmental filtering of small plants. In contrast, the decreased overall accuracy in the lowlands was associated with functional patterns representing any type of assembly rule (environmental filtering, limiting similarity or null assembly). Main Conclusions: Our study reveals interesting patterns of change in S-SDM errors with changes in assembly rules along elevation. Yet, significant levels of assemblage prediction errors occurred throughout the gradient, calling for further improvement of SDMs, e.g., by adding key environmental filters that act at fine scales and developing approaches to account for variations in the influence of predictors along environmental gradients.

Very high-resolution environmental predictors in species distribution models: moving beyond topography?

Recent advances in remote sensing technologies have facilitated the generation of very high resolution (VHR) environmental data. Exploratory studies suggested that, if used in species distribution models (SDMs), these data should enable modelling species' micro-habitats and allow improving predictions for fine-scale biodiversity management. In the present study, we tested the influence, in SDMs, of predictors derived from a VHR digital elevation model (DEM) by comparing the predictive power of models for 239 plant species and their assemblages fitted at six different resolutions in the Swiss Alps. We also tested whether changes of the model quality for a species is related to its functional and ecological characteristics. Refining the resolution only contributed to slight improvement of the models for more than half of the examined species, with the best results obtained at 5 m, but no significant improvement was observed, on average, across all species. Contrary to our expectations, we could not consistently correlate the changes in model performance with species characteristics such as vegetation height. Temperature, the most important variable in the SDMs across the different resolutions, did not contribute any substantial improvement. Our results suggest that improving resolution of topographic data only is not sufficient to improve SDM predictions - and therefore local management - compared to previously used resolutions (here 25 and 100 m). More effort should be dedicated now to conduct finer-scale in-situ environmental measurements (e.g. for temperature, moisture, snow) to obtain improved environmental measurements for fine-scale species mapping and management.

Effects of alternative sets of climatic predictors on species distribution models and associated estimates of extinction risk: a test with plants in an arid environment

Aim To evaluate the effects of using distinct alternative sets of climatic predictor variables on the performance, spatial predictions and future projections of species distribution models (SDMs) for rare plants in an arid environment. . Location Atacama and Peruvian Deserts, South America (18º30'S - 31º30'S, 0 - 3 000 m) Methods We modelled the present and future potential distributions of 13 species of Heliotropium sect. Cochranea, a plant group with a centre of diversity in the Atacama Desert. We developed and applied a sequential procedure, starting from climate monthly variables, to derive six alternative sets of climatic predictor variables. We used them to fit models with eight modelling techniques within an ensemble forecasting framework, and derived climate change projections for each of them. We evaluated the effects of using these alternative sets of predictor variables on performance, spatial predictions and projections of SDMs using Generalised Linear Mixed Models (GLMM). Results The use of distinct sets of climatic predictor variables did not have a significant effect on overall metrics of model performance, but had significant effects on present and future spatial predictions. Main conclusion Using different sets of climatic predictors can yield the same model fits but different spatial predictions of current and future species distributions. This represents a new form of uncertainty in model-based estimates of extinction risk that may need to be better acknowledged and quantified in future SDM studies.

Are niche-based species distribution models transferable in space?

Aim To assess the geographical transferability of niche-based species distribution models fitted with two modelling techniques. Location Two distinct geographical study areas in Switzerland and Austria, in the subalpine and alpine belts. Methods Generalized linear and generalized additive models (GLM and GAM) with a binomial probability distribution and a logit link were fitted for 54 plant species, based on topoclimatic predictor variables. These models were then evaluated quantitatively and used for spatially explicit predictions within (internal evaluation and prediction) and between (external evaluation and prediction) the two regions. Comparisons of evaluations and spatial predictions between regions and models were conducted in order to test if species and methods meet the criteria of full transferability. By full transferability, we mean that: (1) the internal evaluation of models fitted in region A and B must be similar; (2) a model fitted in region A must at least retain a comparable external evaluation when projected into region B, and vice-versa; and (3) internal and external spatial predictions have to match within both regions. Results The measures of model fit are, on average, 24% higher for GAMs than for GLMs in both regions. However, the differences between internal and external evaluations (AUC coefficient) are also higher for GAMs than for GLMs (a difference of 30% for models fitted in Switzerland and 54% for models fitted in Austria). Transferability, as measured with the AUC evaluation, fails for 68% of the species in Switzerland and 55% in Austria for GLMs (respectively for 67% and 53% of the species for GAMs). For both GAMs and GLMs, the agreement between internal and external predictions is rather weak on average (Kulczynski's coefficient in the range 0.3-0.4), but varies widely among individual species. The dominant pattern is an asymmetrical transferability between the two study regions (a mean decrease of 20% for the AUC coefficient when the models are transferred from Switzerland and 13% when they are transferred from Austria). Main conclusions The large inter-specific variability observed among the 54 study species underlines the need to consider more than a few species to test properly the transferability of species distribution models. The pronounced asymmetry in transferability between the two study regions may be due to peculiarities of these regions, such as differences in the ranges of environmental predictors or the varied impact of land-use history, or to species-specific reasons like differential phenotypic plasticity, existence of ecotypes or varied dependence on biotic interactions that are not properly incorporated into niche-based models. The lower variation between internal and external evaluation of GLMs compared to GAMs further suggests that overfitting may reduce transferability. Overall, a limited geographical transferability calls for caution when projecting niche-based models for assessing the fate of species in future environments.

Do stacked species distribution models reflect altitudinal diversity patterns?

The objective of this study was to evaluate the performance of stacked species distribution models in predicting the alpha and gamma species diversity patterns of two important plant clades along elevation in the Andes. We modelled the distribution of the species in the Anthurium genus (53 species) and the Bromeliaceae family (89 species) using six modelling techniques. We combined all of the predictions for the same species in ensemble models based on two different criteria: the average of the rescaled predictions by all techniques and the average of the best techniques. The rescaled predictions were then reclassified into binary predictions (presence/absence). By stacking either the original predictions or binary predictions for both ensemble procedures, we obtained four different species richness models per taxa. The gamma and alpha diversity per elevation band (500 m) was also computed. To evaluate the prediction abilities for the four predictions of species richness and gamma diversity, the models were compared with the real data along an elevation gradient that was independently compiled by specialists. Finally, we also tested whether our richness models performed better than a null model of altitudinal changes of diversity based on the literature. Stacking of the ensemble prediction of the individual species models generated richness models that proved to be well correlated with the observed alpha diversity richness patterns along elevation and with the gamma diversity derived from the literature. Overall, these models tend to overpredict species richness. The use of the ensemble predictions from the species models built with different techniques seems very promising for modelling of species assemblages. Stacking of the binary models reduced the over-prediction, although more research is needed. The randomisation test proved to be a promising method for testing the performance of the stacked models, but other implementations may still be developed.

Using species richness and functional traits predictions to constrain assemblage predictions from stacked species distribution models

Aim: Modelling species at the assemblage level is required to make effective forecast of global change impacts on diversity and ecosystem functioning. Community predictions may be achieved using macroecological properties of communities (MEM), or by stacking of individual species distribution models (S-SDMs). To obtain more realistic predictions of species assemblages, the SESAM framework suggests applying successive filters to the initial species source pool, by combining different modelling approaches and rules. Here we provide a first test of this framework in mountain grassland communities. Location: The western Swiss Alps. Methods: Two implementations of the SESAM framework were tested: a "Probability ranking" rule based on species richness predictions and rough probabilities from SDMs, and a "Trait range" rule that uses the predicted upper and lower bound of community-level distribution of three different functional traits (vegetative height, specific leaf area and seed mass) to constraint a pool of environmentally filtered species from binary SDMs predictions. Results: We showed that all independent constraints expectedly contributed to reduce species richness overprediction. Only the "Probability ranking" rule allowed slightly but significantly improving predictions of community composition. Main conclusion: We tested various ways to implement the SESAM framework by integrating macroecological constraints into S-SDM predictions, and report one that is able to improve compositional predictions. We discuss possible improvements, such as further improving the causality and precision of environmental predictors, using other assembly rules and testing other types of ecological or functional constraints.

Impact of model complexity on cross-temporal transferability in Maxent species distribution models: An assessment using paleobotanical data

Maximum entropy modeling (Maxent) is a widely used algorithm for predicting species distributions across space and time. Properly assessing the uncertainty in such predictions is non-trivial and requires validation with independent datasets. Notably, model complexity (number of model parameters) remains a major concern in relation to overfitting and, hence, transferability of Maxent models. An emerging approach is to validate the cross-temporal transferability of model predictions using paleoecological data. In this study, we assess the effect of model complexity on the performance of Maxent projections across time using two European plant species (Alnus giutinosa (L.) Gaertn. and Corylus avellana L) with an extensive late Quaternary fossil record in Spain as a study case. We fit 110 models with different levels of complexity under present time and tested model performance using AUC (area under the receiver operating characteristic curve) and AlCc (corrected Akaike Information Criterion) through the standard procedure of randomly partitioning current occurrence data. We then compared these results to an independent validation by projecting the models to mid-Holocene (6000 years before present) climatic conditions in Spain to assess their ability to predict fossil pollen presence-absence and abundance. We find that calibrating Maxent models with default settings result in the generation of overly complex models. While model performance increased with model complexity when predicting current distributions, it was higher with intermediate complexity when predicting mid-Holocene distributions. Hence, models of intermediate complexity resulted in the best trade-off to predict species distributions across time. Reliable temporal model transferability is especially relevant for forecasting species distributions under future climate change. Consequently, species-specific model tuning should be used to find the best modeling settings to control for complexity, notably with paleoecological data to independently validate model projections. For cross-temporal projections of species distributions for which paleoecological data is not available, models of intermediate complexity should be selected.

Overcoming limitations of modelling rare species by using ensembles of small models

1. Species distribution models (SDMs) have become a standard tool in ecology and applied conservation biology. Modelling rare and threatened species is particularly important for conservation purposes. However, modelling rare species is difficult because the combination of few occurrences and many predictor variables easily leads to model overfitting. A new strategy using ensembles of small models was recently developed in an attempt to overcome this limitation of rare species modelling and has been tested successfully for only a single species so far. Here, we aim to test the approach more comprehensively on a large number of species including a transferability assessment. 2. For each species numerous small (here bivariate) models were calibrated, evaluated and averaged to an ensemble weighted by AUC scores. These 'ensembles of small models' (ESMs) were compared to standard Species Distribution Models (SDMs) using three commonly used modelling techniques (GLM, GBM, Maxent) and their ensemble prediction. We tested 107 rare and under-sampled plant species of conservation concern in Switzerland. 3. We show that ESMs performed significantly better than standard SDMs. The rarer the species, the more pronounced the effects were. ESMs were also superior to standard SDMs and their ensemble when they were independently evaluated using a transferability assessment. 4. By averaging simple small models to an ensemble, ESMs avoid overfitting without losing explanatory power through reducing the number of predictor variables. They further improve the reliability of species distribution models, especially for rare species, and thus help to overcome limitations of modelling rare species.