Abstract

Summary After the dramatic eutrophication‐induced decline of intertidal seagrasses in the 1970s, the Wadden Sea has shown diverging developments. In the northern Wadden Sea, seagrass beds have expanded and become denser, while in the southern Wadden Sea, only small beds with low shoot densities are found. A lack of documentation of historical distributions hampers conservation management. Yet, the recovery in the northern Wadden Sea provides opportunity to construct robust habitat suitability models to support management. We tuned habitat distribution models based on 17 years of seagrass surveys in the northern Wadden Sea and high‐resolution hydrodynamics and geomorphology for the entire Wadden Sea using five machine learning approaches. To obtain geographically transferable models, hyperparameters were tuned on the basis of prediction accuracy assessed by non‐random, spatial cross‐validation. The spatial cross‐validation methodology was combined with a consensus modelling approach. The predicted suitability scores correlated amongst each other and with the hold‐out observations in the training area indicating that the models converged and were transferable across space. Prediction accuracy was improved by averaging the predictions of the best models. We graphically examined the relationship between the consensus suitability score and independent presence‐only data from outside the training area using the area‐adjusted seagrass frequency per suitability class (continuous Boyce index). The Boyce index was positively correlated with the suitability score indicating the adequacy of the prediction methodology. We used the plot of the continuous Boyce index against habitat suitability score to demarcate three habitat classes – unsuitable, marginal and suitable – for the entire international Wadden Sea. This information is valuable for habitat conservation and restoration management. Divergence between predicted suitability and actual distributions from the recent past indicates that unaccounted factors limit seagrass development in the southern Wadden Sea. Synthesis and applications. Our methodology and data enabled us to produce a robust and validated consensus habitat suitability model. We identified highly suitable areas where intertidal seagrass meadows may establish and persist. Our work provides scientific underpinning for effective conservation planning in a dynamic landscape and sets monitoring priorities.

Highlights

  • Seagrass beds are highly valued for their ecosystem services (Costanza et al 1997; Orth et al 2006) related to nutrient cycling and provisioning resources and habitat for other species (Beck et al 2001)

  • We graphically examined the relationship between the consensus suitability score and independent presence-only data from outside the training area using the area-adjusted seagrass frequency per suitability class

  • Before further detailing the modelling approach, we present a conceptual model describing how desiccation, hydrodynamics and geomorphology control the distributions of intertidal Z. noltii and Z. marina (Fig. 1) in the Wadden Sea (Den Hartog & Polderman 1975)

Read more

Summary

Introduction

Seagrass beds are highly valued for their ecosystem services (Costanza et al 1997; Orth et al 2006) related to nutrient cycling and provisioning resources and habitat for other species (Beck et al 2001). The effectiveness of seagrass conservation management such as safeguarding suitable habitats is often hampered by the lack of systematic documentation of historical seagrass distributions and by changes in environmental conditions in space or time. Another complication is that seagrass distributions in energetic environments like intertidal zones show local extinction– colonization dynamics which necessitates identification and protection of vegetated and suitable unvegetated habitat (Valle et al 2013; Suykerbuyk et al 2016). In the case of the spatial variant of this cross-validation method, data are split into geographical subsets that are more distinct than random subsets The idea behind this method is that the validation sets differ from training sets as the area for which predictions are required. This method yields more transferable models and more accurate predictions under new conditions than models that are tuned by random cross-validation (Wenger & Olden 2012)

Objectives
Methods
Results
Discussion
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call