Machine-Learning Modeling of Benthic Systems: Baseline Assessment and Monitoring

• Ano de defesa: 2025
• Autor(a) principal: Watanabe, Luciana Erika Yaginuma
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Bacia de Santos; Ecossistema marinho; Macrofauna; Marine ecosystem; Meiofauna; Microbiota; Nematoda; Santos Basin
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/21/21134/tde-11062025-162009/

In the face of global ecosystem collapse and our urgent need to respond effectively through conservation actions, statistical modeling can be a powerful tool as it provides valuable baseline and predictive information. This study aims to enhance baseline assessments and monitoring programs by proposing analytical frameworks suited to different community targets and machine-learning modeling approaches. The database from the Santos Basin benthic system was used as a case study. All data were obtained from 100 sampling stations distributed along 11 transects with eight points each, and an additional 12 stations over the São Paulo Plateau, during two campaigns between June-November 2019 and February-June 2021. Taxonomic composition data and ecological indicators of the microbiota (bacteria and archaea), meiofauna, and macrofauna were obtained in each station, along with 132 environmental variables from the bottom water and sediment. Over four chapters, different machine-learning techniques were applied to model different benthic targets. In Chapter 2, meiofauna univariate descriptors were modeled using environmental variables as predictors and random forest regression algorithms. Based on the spatial distribution of the predictions, six benthic zones were delineated in the Basin. In Chapter 3, multivariate nematode genera data were aggregated into associations by using a self-organizing map (SOM) and hierarchical clustering (HC) analysis. Those associations were then modeled using observed and simulated environmental predictors. Six nematode associations were obtained and their spatial distribution analytically confirmed the Basin organization in a mosaic of six benthic zones. Also, high accuracy was retrieved by the association model based on the simulated data of eight significant environmental variables suggesting that its spatial distribution can be monitored using only those variables. In Chapter 4, temporal variability was added to the meiofauna and macrofauna univariate data and the model\'s results highlighted the influence of temporal variation of oceanographic processes on those benthic communities mainly in the continental shelf. Also, the use of a structure modeling framework resulted in a loss of 3% accuracy compared to a non-structure model, with the advantage of making generalizations of the first one. Finally, in Chapter 5, more spatial complexity was added by integrating multivariate composition data from the microbiota, nematode, and macrofauna through multilayer SOM. Eight major benthic communities were identified and were highly correlated with each benthic group separately. The accuracy of the model was perfect (100%) and nine environmental variables were indicated as significant ones. All modeling approaches presented in the chapters provide valuable baseline information on the benthic system, with composition and relative densities references, and lists of the most important environmental variables for prediction. The frameworks allowed the accuracy evaluation of the different model structures, subsidizing monitoring decisions. Based on the proposed structured model it is also possible to make predictions for future scenarios. This is a major advantage of machine-learning approaches for adaptive monitoring programs, as modeling can be constantly improved and simulated as more data is acquired. In conclusion, the study has high potential to support management and monitoring programs, and ultimately the conservation of marine systems.