Improving data quality to build a robust distribution model for Architeuthis dux

Highlights

• We estimate the potential distribution map of the giant squid Architeuthis dux.

• We detect important environmental parameters and produce absences from AquaMaps.

• We combine information using feed-forward neural networks.

• Our model is the most similar to an expert drawn map respect to other models.

• Our approach is also applicable to other rare species.

Abstract

The giant squid (Architeuthis) has been reported since even before the 16th century, and has recently been observed live in its habitat for the first time. Among the species belonging to this genus, Architeuthis dux has received special attention from biologists. The distribution of this species is poorly understood, as most of our information stems from stranded animals or stomach remains. Predicting the habitat and distribution of this species, and more in general of difficult to observe species, is important from a biological conservation perspective. In this paper, we present an approach to estimate the potential distribution of A. dux at global scale, with relative high resolution (1-degree). Our approach relies on a complex preparation phase, which improves the reliability of presence, absence and environmental data correlated to the species habitat. We compare our distribution with those produced by state-of-the-art approaches (MaxEnt and AquaMaps), and use an expert-drawn map as reference. We demonstrate that our model projection is in agreement with the expert's map and is also compliant with several biological assessments of the species habitat and with recent observations. Furthermore, we show that our approach can be generalized as a paradigm that is applicable to other rare species.

Introduction

In recent years, niche models that estimate species distribution have become widely used in conservation biology (Guisan and Zimmermann, 2000). Rare species are examples where the prediction of suitable habitats is paramount to support fisheries management policies and conservation strategies (Pearce and Boyce, 2006, Márcia Barbosa et al., 2003). Defined by Cao et al. (1998) as species that occur at lower frequency or in low number in a sample of certain size, rare species have a key role in affecting biodiversity richness and by consequence they are indicators of degradation for aquatic ecosystems (Lyons et al., 1995, Cao et al., 1998). In this context, predictive models can considerably support the qualitative and quantitative criteria used to assign a “status” to a species (IUCN Species, 2001), by providing accurate, applicable and reliable spatial predictions to species population monitoring and sampling (Guisan et al., 2006). As discussed in many studies, the methodological progresses of Species Distribution Models (SDMs) allow nowadays to apply robust techniques to rare and endangered species (Guisan and Thuiller, 2005, Ferrier, 2002, Gibson et al., 2007, Razgour et al., 2011, Ovaskainen and Soininen, 2011, Rebelo and Jones, 2010, Wisz et al., 2008, Lomba et al., 2010).

Here, we propose a procedure to generate a niche model for a species of the giant squid family (Architeuthis dux), based on both presence and estimated absence locations. Our aim is to produce a map that is more accurate with respect to the ones that can be produced by commonly used models. Although giant squids have recently received special attention, little has been published regarding the population demographics and the ecology of these rare species. Most of the records refer to dead stranded animals, individuals captured alive by nets or from the remains found in the stomach of marine mammals (Clarke, 2006). When modelling the distribution of these species, high quality data are crucial but very scarce. This problem is especially important for rare species prediction, where models training is highly dependent on data quality.

Given this context, our study investigates a combination of presence only and presence/absences techniques to identify potentially suitable areas for A. dux subsistence. We also expect the results to help defining guidelines for use of SDMs for rare species.

We illustrate our approach using data from authoritative sources of observation records. Furthermore, we use an expert system to produce absence locations. In order to ensure high quality for the environmental variables associated to presence information, we use the maximum entropy (MaxEnt) model (Phillips et al., 2006, Berger, 1996) as a filter to select the variables that are important to define the potential habitat of the species. These are the variables that are mostly correlated to the species observations, among those we selected from reference studies. When possible, we make environmental variables values range from 450 to 1000 m, encompassing the deep ocean waters usually inhabited by A. dux (Guerra et al., 2010). Finally, we train an Artificial Neural Network on these datasets and compare the results with (i) a presence-only method, (ii) an expert system and (iii) an expert drawn map.

The paper is organized as follows: Section 2 reports the effort made to model or understand the potential habitat of rare species, and in particular of A. dux. Section 3 reports the details of our method and its expandability as a general approach to rare species modelling. Section 4 reports the results of both a qualitative and a quantitative comparison with other distribution maps for A. dux. Section 5 discusses the results and Section 6 draws the conclusions.