Predicting demersal fish distributions using presence-only algorithms
Monk, J and Ierodiaconou, D and Bellgrove, A and Harvey, E and Laurenson, L, Predicting demersal fish distributions using presence-only algorithms, AMSA 2009 Program, 05-09 July, Adelaide, SA, pp. 146. (2009) [Conference Extract]
PDF (Potential sources of error in the application of towed video data for benthic habitat characterisation) Pending copyright assessment - Request a copy 3Mb
Information on the spatial distribution and essential habitat requirements of fishes is important for developing management plans that balance resource extraction and conservation needs in marine environments. Spatially explicit modelling approaches provide the capacity to make predictions about where a particular species could be found based on existing seascape information. When compared to landscape applications, however, marine researchers have little guidance when choosing between competing methods because few comparative studies exist. This study compares six modelling methods for five commonly observed demersal fish species. We used presence-only data to fit models, and independent presence-pseudo-absence data to evaluate predictions. Species distribution data were based upon observations from towed underwater video footage. Thirteen seafloor-related variables were derived from multibeam sonar to define seafloor characteristics for model input. Bioclimatic envelope model (BIOCLIM; DIVA-GIS), DOMAIN (DIVA-GIS) and four habitat suitability algorithms (Median, Distance geometric mean, Distance harmonic mean and Minimum distance; Biomapper 4.06) were used for each species to build distribution models. Models were developed using 75/25% split of training and validation data. Area-Under-Curve (AUC) validation tests were used to assess predictive performance. Validation statistics for the test data found that across all species Median Distance, BIOCLIM, Harmonic and Geometric Mean models yielded the lowest predictive capabilities, although all differ from a purely random prediction (i.e. greater than 50%). The Minimum Distance and DOMAIN algorithms produced significantly better prediction performance, with Minimum Distance yielding highest predictive capabilities for four out of the five species investigated.