Premium
Bayesian networks for habitat suitability modeling: a potential tool for conservation planning with scarce resources
Author(s) -
Tantipisanuh Naruemon,
Gale George A.,
Pollino Carmel
Publication year - 2014
Publication title -
ecological applications
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.864
H-Index - 213
eISSN - 1939-5582
pISSN - 1051-0761
DOI - 10.1890/13-1882.1
Subject(s) - threatened species , habitat , biodiversity , ecology , range (aeronautics) , bayesian network , environmental resource management , geography , biology , machine learning , computer science , environmental science , materials science , composite material
Bayesian networks (BN) have been increasingly used for habitat suitability modeling of threatened species due to their potential to construct robust models with limited survey data. However, previous applications of this approach have only occurred in countries where human and budget resources are highly available, but the highest concentrations of threatened vertebrates globally are located in the tropics where resources are much more limited. We assessed the effectiveness of Bayesian networks in generating habitat suitability models in Thailand, a biodiversity‐rich country where the knowledge base is typically sparse for a wide range of threatened species. The Bayesian network approach was used to generate habitat suitability maps for 52 threatened vertebrate species in Thailand, using a range of evidence types, from relatively well‐documented species with good local knowledge to poorly documented species, with few local experts. Published information and expert knowledge were used to define habitat requirements. Focal species were categorized into 22 groups based on known habitat preferences, and then habitat suitability models were constructed with outcomes represented spatially. Models had a consistent structure with three major components: potential habitat, known range, and threat level. Model classification sensitivity was tested using presence‐only field data for 21 species. Habitat models for 12 species were relatively sensitive (>70% congruency between observed and predicted locations), three were moderately congruent, and six were poor. Classification sensitivity tended to be high for bird models and moderate for mammals, whereas sensitivity for reptiles was low, presumably reflecting the relatively poor knowledge base for reptiles in the region. Bayesian network models show significant potential for biodiversity‐rich regions with scarce resources, although they require further refinement and testing. It is possible that one detailed ecological study is sufficient to develop a model with reasonable sensitivity, but BN models for species groups with no quantitative data continue to be problematic.