Generalised Additive Models and Random Forest Approach as effective methods for predictive species density and functional species richness

Abstract

Species distribution modelling (SDM) is a family of statistical methods where species occurrence/density/richness are combined with environmental predictors to create predictive spatial models of species distribution. However, it often turns out that due to complex multi-level interactions between predictors and the response function, different types of models can detect different numbers of important predictors and also vary in their predictive ability. This is why we decided to explore differences in the predictive power of two most common methods, such as the Generalised Additive Model (GAM) and the Random Forest (RF) on the example of the Great Spotted Woodpecker Dendrocopos major and the Great Grey Shrike Lanius excubitor, as well as on the taxonomic and functional species richness. For each of the two bird species' densities and for two measurements of biodiversity, two sets of SDMs were generated: One based on the GAM, and the other on the RF. According to the out-of-bag, the Akaike Information Criterion (AIC) and an independent evaluation, we demonstrated that the GAM is the best method for predicting density of the Great Spotted Woodpecker and taxonomic species richness, whereas the RF has the lowest prediction error for the density of the Great Grey Shrike and functional species richness. It also becomes apparent that the GAM is responsive to taxonomic species richness and species with broad tolerance to environmental factors, i.e. the Great Spotted Woodpecker, while the RF detects more subtle relationships between density and environmental variables, rendering it more suitable for functional species richness and species with a narrow tolerance range to habitats factors, i.e. the Great Grey Shrike. Thus, effective predictive modelling of animal distribution requires considering several different analytical approaches to produce biologically realistic predictions.

Spatial Distribution and Abundance of Mesopelagic Fish Biomass in the Mediterranean Sea

“…Our results highlight a general different pattern among these tools, not only in the absolute biomass estimates but also in the temporal and spatial distribution. These two models are commonly used to assess species distributions in both marine and terrestrial realms (e.g., Moisen et al, 2006;Rooper et al, 2017;Kosicki, 2020), however RF models have been seen to have substantially more performance power and control over the model overfitting compared to GAM approaches (Elith et al, 2006;Rooper et al, 2017), as it is also shown in this study. This is likely because of the RF algorithm, which is an ensemble of models (regression trees), each built on a random selection of relatively few predictors (Breiman, 2001).…”

Species Distribution Modelling performance and its implication for Sentinel-2-based prediction of invasive Prosopis juliflora in lower Awash River basin, Ethiopia

“…The higher performance of RF is attributed to its ability to avoid overfitting as it combines and votes the most popular class from several individual trees (Breiman 2001). Due to its higher performance, it is the most flexible and widely applied machine learning algorithm for various field of studies such as land cover classification (Abdi 2020), forest monitoring (Ma et al 2020), species richness, and density (Kosicki 2020), and invasive species modelling (Ng et al 2018). In addition, its capability for remote sensing-based studies is also immense as it requires minimum time for satellite image classification (Sabat-Tomala et al 2020).…”

“…Random Forest (RF) RF (Breiman 2001), a combination of tree predictors, is the most commonly used machine learning algorithm (Abdi 2020). It is an effective method for predicting species richness and density (Kosicki 2020).…”

Intra‐season variations in distribution and abundance of humpback whales in the West Antarctic Peninsula using cruise vessels as opportunistic platforms

“…Model‐based estimates were achieved by fitting a detection function (the probability of animal detection as a function of perpendicular distance) followed by fitting generalized additive models (GAMs; Wood et al, 2016) to explore the relationship of animal presence and abundance to environmental covariates. This relationship was then used to predict animal density in the study area limited by the range of observed environmental values fitted in the GAM (Franchini et al, 2020; Kosicki, 2020; Mannocci et al, 2014). In addition, standard distance sampling estimates were also calculated, which can be found in the Appendix .…”