Why to choose Random Forest to predict rare species distribution with few samples in large undersampled areas? Three Asian crane species models provide supporting evidence

Abstract

Species distribution models (SDMs) have become an essential tool in ecology, biogeography, evolution, and more recently, in conservation biology. How to generalize species distributions in large undersampled areas, especially with few samples, is a fundamental issue of SDMs. In order to explore this issue, we used the best available presence records for the Hooded Crane (Grus monacha, n=33), White-naped Crane (Grus vipio, n=40), and Black-necked Crane (Grus nigricollis, n=75) in China as three case studies, employing four powerful and commonly used machine learning algorithms to map the breeding distributions of the three species: TreeNet (Stochastic Gradient Boosting, Boosted Regression Tree Model), Random Forest, CART (Classification and Regression Tree) and Maxent (Maximum Entropy Models) Besides, we developed an ensemble forecast by averaging predicted probability of above four models results. Commonly-used model performance metrics (Area under ROC (AUC) and true skill statistic (TSS)) were employed to evaluate model accuracy. Latest satellite tracking data and compiled literature data were used as two independent testing datasets to confront model predictions. We found Random Forest demonstrated the best performance for the most assessment method, provided a better model fit to the testing data, and achieved better species range maps for each crane species in undersampled areas. Random Forest has been generally available for more than 20 years, and by now, has been known to perform extremely well in ecological predictions. However, while increasingly on the rise its potential is still widely underused in conservation, (spatial) ecological applications and for inference. Our results show that it informs ecological and biogeographical theories as well as being suitable for conservation applications, specifically when the study area is undersampled. This method helps to save model-selection time and effort, and it allows robust and rapid assessments 35 developed an ensemble forecast by averaging predicted probability of above four models results.36 Commonly-used model performance metrics (Area under ROC (AUC) and true skill statistic 37 (TSS)) were employed to evaluate model accuracy. Latest satellite tracking data and compiled 38 literature data were used as two independent testing datasets to confront model predictions. We 39 found Random Forest demonstrated the best performance for the most assessment method, 40 provided a better model fit to the testing data, and achieved better species range maps for each 41 crane species in undersampled areas. Random Forest has been generally available for more than 42 20 years, and by now, has been known to perform extremely well in ecological predictions. 43 However, while increasingly on the rise its potential is still widely underused in conservation, (Heikkinen et al., 2012). In this study, 84 we used generality (transferability) as the concept of generalizing distribution from sampled areas 85 to unsampled areas (extrapolation beyond the data) in one study area....

Outstanding Challenges in the Transferability of Ecological Models

“…Random forests and boosted regression trees, two data-driven approaches that are relatively immune to overfitting and can handle predictor interactions, can demonstrate high performance in unsampled areas (e.g., [56] [8]. In recent years, dynamic models capable of tracking the temporal aspects of a species' behavior and distribution, and joint species distribution models designed to simultaneously account for the co-occurrence of multiple species, have also gained traction.…”

“…Random forests and boosted regression trees, two data-driven approaches that are relatively immune to overfitting and can handle predictor interactions, can demonstrate high performance in unsampled areas (e.g., [56]). MaxEnt, another machine learning method, has been ranked the most transferable in some studies (e.g., [57]).…”

Input matters matter: Bioclimatic consistency to map more reliable species distribution models

“…We performed correlative models for each species and database (i.e., three SDMs per species) using the Random Forest algorithm (RF: Breiman, ) which is considered to be among the best performing machine learning methods (Cutler et al., ; Mi, Huettmann, Guo, Han, & Wen, ). The chosen predictors were mean annual precipitation and mean annual temperature for terrestrial mammals, and bathymetry (from MARSPEC) and sea surface temperature for marine mammals.…”

Climate change and the future restructuring of Neotropical anuran biodiversity

“…Range maps have been previously used in amphibian biogeographical studies at large spatial scales (McKnight et al 2007, Buckley and Jetz 2008, Lawler et al 2009), and recent findings showed a strong agreement between SDMs trained with IUCN range maps and SDMs trained with true occurrences (Fourcade 2016). Random forests are machine learning algorithms based on classification of regression trees returning the class that is the mode of the classes output by individual trees (Breiman 2001); they have been widely used in SDMs, and proven to be effective for predictive modelling (Prasad et al 2006, Mi et al 2017. Random forest classifiers are robust to over-fitting but tend to better fit the observed data than other algorithms, and hence producing more accurate predictions of species' ranges (Lawler et al 2006).…”