Comparing discriminant analysis, neural networks and logistic regression for predicting species distributions: a case study with a Himalayan river bird

摘要: We assessed the occurrence of a common river bird, Plumbeous Redstart Rhyacornis fuliginosus, along 180 independent streams in Indian and Nepali Himalaya. then compared performance multiple discrimant analysis (MDA), logistic regression (LR) artificial neural networks (ANN) predicting this species’ presence or absence from 32 variables describing stream altitude, slope, habitat structure, chemistry invertebrate abundance. Using entire data (training set) threshold for accepting ANN LR set to P] 0.5, correctly classified marginally more cases (88%) than either (83%) MDA (84%). Model was two methods partitioning. In ‘leave-one-out’ approach, predicted (82%) (73%) (69%). However, holdout procedure, all performed similarly (73‐75%). All true (i.e. specificity holdout: 81‐85%) better sensitivity: 57‐60%). These effects reflect prevalence ( frequency occurrence), but are seldom considered distribution modelling. Despite occurring at only 36% sites, Redstarts one most Himalayan birds, problems will be greater with less species. Both require an arbitrary probability (often P0.5) which accept species model prediction. Simulations involving varied revealed that particularly sensitive effects. ROC plots (received operating characteristic) were therefore used compare on test range thresholds; always outperformed ANN. This case study supports need models data, use criteria assessing performance. do not yet have major advantages over conventional multivariate bird distributions. both efficient computer time ANN, also straightforward providing testable hypotheses about environmental occurrence. apparently subject chance significant explanatory variables, emphasising well-known risks based purely correlative data. © 1999 Elsevier Science B.V. rights reserved.