Modeling the spatial distribution and abundance of deep‐water red shrimps in the Mediterranean Sea: a machine learning approach

Abstract

Spatially-explicit models are invaluable tools for analyzing the species-environment interactions, even at scales beyond that of direct observations. In fisheries context, the observations on species usually consist of data derived from survey campaigns, such as the Mediterranean International Bottom Trawl Surveys (MEDITS) programme. MEDITS survey foresees the use of a standardized protocol for data acquisition on demersal species, such as the blue and red shrimp Aristeus antennatus and the giant red shrimp Aristaeomorpha foliacea. These two species are recognized as highly valuable marked resources accounting for about 5 % of the trawl fishing income in the Mediterranean basin. Here, we developed a modeling framework for the analysis of the MEDITS data on those species. Within our modeling framework we aimed at detecting the existence of a divergence in the spatial patterns that could guide the definition of targeted management actions for those two valuable fishing resources. A Random Forest (RF) machine learning approach has been used to model both the occurrence (i.e., presence/absence) and the biomass index (kg/km²) of both species in four Geographical SubAreas (GSAs) located in the central part of the Mediterranean and the Ionian Sea. The RF showed high level of accuracy (i.e., K=0.83 and K=0.88, for A. antennatus and A. foliacea, respectively) in modeling species occurrence, and good level of performance (i.e., R²=0.63 and R²=0.74, respectively) in modeling their biomass index (kg/km²). The niche overlap and statistical analyses we performed on the models outputs revealed the existence of a significant divergence in the spatial patterns between these species. This provides crucial ecological knowledge for the definition of targeted (i.e., species-related) management actions. Afterwards, the models have been extrapolated at the spatial scale of the Mediterranean Sea based on an approach we defined, called hyperspace. The hyperspace approach, while showing technical and ecological soundness, was meant to guarantee the reliability of model predictions in unknown areas. It reduces the need for a proper interpretation of “what is beyond a predicted value”, offering a straightforward method for model extrapolation. Our effort aims to provide insights for prioritizing key areas in conservation strategies and marine spatial planning. It also represents an important contribution towards adopting an ecosystem-based approach to fishery resource management in the Mediterranean basin.