Structured machine learning modeling to support conservation of deep-sea benthic biodiversity.

Area of Science:

• Marine Biology
• Ecological Modeling
• Conservation Science

Background:

• Biodiversity monitoring programs require accurate, timely, and actionable predictions for effective management.
• Deep-sea ecosystems, like the Santos Basin benthos, face increasing anthropogenic pressures, necessitating robust monitoring strategies.

Purpose of the Study:

• To develop and evaluate a predictive monitoring program for deep-sea benthos in the Santos Basin, Brazil.
• To compare biodiversity predictions from a novel two-stage structured model using simulated (2M-Sim) and real (2M) environmental data against unstructured models (1M).

Main Methods:

• A two-stage structured modeling approach was developed using random forest algorithms.
• Benthic macro- and meiofaunal data were integrated with environmental (sediment, water column) and spatio-temporal variables.
• Model performance was assessed by comparing predictions from structured (2M, 2M-Sim) and unstructured (1M) models using a 20% validation dataset.

Main Results:

• Average model accuracies were 72% (1M), 69% (2M), and 68% (2M-Sim), with no significant accuracy loss between models.
• The 2M model identified 30 significant environmental variables, with bottom water parameters and sedimentary phytopigment/carbonate concentrations being key predictors.
• Model accuracy varied, generally higher for macrofauna (38%-84% for 2M).

Conclusions:

• The developed structured modeling approach is reliable for predicting deep-sea benthos biodiversity.
• This framework optimizes data requirements and sampling strategies, supporting data-driven management decisions for benthic biodiversity conservation.
• The study provides a valuable tool for enhancing deep-sea ecosystem monitoring and conservation in the Santos Basin.