Daniel Lew

Bio: Daniel Lew is an academic researcher from La Salle University. The author has contributed to research in topics: Model selection. The author has an hindex of 1, co-authored 1 publications receiving 1064 citations.

AUC: a misleading measure of the performance of predictive distribution models

Abstract: The area under the receiver operating characteristic (ROC) curve, known as the AUC, is currently considered to be the standard method to assess the accuracy of predictive distribution models. It avoids the supposed subjectivity in the threshold selection process, when continuous probability derived scores are converted to a binary presence‐absence variable, by summarizing overall model performance over all possible thresholds. In this manuscript we review some of the features of this measure and bring into question its reliability as a comparative measure of accuracy between model results. We do not recommend using AUC for five reasons: (1) it ignores the predicted probability values and the goodness-of-fit of the model; (2) it summarises the test performance over regions of the ROC space in which one would rarely operate; (3) it weights omission and commission errors equally; (4) it does not give information about the spatial distribution of model errors; and, most importantly, (5) the total extent to which models are carried out highly influences the rate of well-predicted absences and the AUC scores.

Predicting species distributions from small numbers of occurrence records: A test case using cryptic geckos in Madagascar

Abstract: Aim: Techniques that predict species potential distributions by combining observed occurrence records with environmental variables show much potential for application across a range of biogeographical analyses. Some of the most promising applications relate to species for which occurrence records are scarce, due to cryptic habits, locally restricted distributions or low sampling effort. However, the minimum sample sizes required to yield useful predictions remain difficult to determine. Here we developed and tested a novel jackknife validation approach to assess the ability to predict species occurrence when fewer than 25 occurrence records are available. Location: Madagascar. Methods: Models were developed and evaluated for 13 species of secretive leaf-tailed geckos (Uroplatus spp.) that are endemic to Madagascar, for which available sample sizes range from 4 to 23 occurrence localities (at 1 km2 grid resolution). Predictions were based on 20 environmental data layers and were generated using two modelling approaches: a method based on the principle of maximum entropy (Maxent) and a genetic algorithm (GARP). Results: We found high success rates and statistical significance in jackknife tests with sample sizes as low as five when the Maxent model was applied. Results for GARP at very low sample sizes (less than c. 10) were less good. When sample sizes were experimentally reduced for those species with the most records, variability among predictions using different combinations of localities demonstrated that models were greatly influenced by exactly which observations were included. Main conclusions: We emphasize that models developed using this approach with small sample sizes should be interpreted as identifying regions that have similar environmental conditions to where the species is known to occur, and not as predicting actual limits to the range of a species. The jackknife validation approach proposed here enables assessment of the predictive ability of models built using very small sample sizes, although use of this test with larger sample sizes may lead to overoptimistic estimates of predictive power. Our analyses demonstrate that geographical predictions developed from small numbers of occurrence records may be of great value, for example in targeting field surveys to accelerate the discovery of unknown populations and species. © 2007 The Authors.

The effect of sample size and species characteristics on performance of different species distribution modeling methods

Abstract: Species distribution models should provide conservation practioners with estimates of the spatial distributions of species requiring attention. These species are often rare and have limited known occurrences, posing challenges for creating accurate species distribution models. We tested four modeling methods (Bioclim, Domain, GARP, and Maxent) across 18 species with different levels of ecological specialization using six different sample size treatments and three different evaluation measures. Our assessment revealed that Maxent was the most capable of the four modeling methods in producing useful results with sample sizes as small as 5, 10 and 25 occurrences. The other methods compensated reasonably well (Domain and GARP) to poorly (Bioclim) when presented with datasets of small sample sizes. We show that multiple evaluation measures are necessary to determine accuracy of models produced with presence-only data. Further, we found that accuracy of models is greater for species with small geographic ranges and limited environmental tolerance, ecological characteristics of many rare species. Our results indicate that reasonable models can be made for some rare species, a result that should encourage conservationists to add distribution modeling to their toolbox.

A maximum entropy approach to species distribution modeling

Abstract: We study the problem of modeling species geographic distributions, a critical problem in conservation biology. We propose the use of maximum-entropy techniques for this problem, specifically, sequential-update algorithms that can handle a very large number of features. We describe experiments comparing maxent with a standard distribution-modeling tool, called GARP, on a dataset containing observation data for North American breeding birds. We also study how well maxent performs as a function of the number of training examples and training time, analyze the use of regularization to avoid overfitting when the number of examples is small, and explore the interpretability of models constructed using maxent.