Mapeamento de sedimentos da margem equatorial brasileira usando algoritmos de aprendizado de máquina

Abstract

Large-scale mapping of seafloor sediments is essential for environmental management, offshore engineering, and the sustainable exploitation of resources, especially in optically

complex coastal regions such as the Brazilian Equatorial Margin (BEM). Although tradi- tional acoustic surveys provide high accuracy, their application is limited by operational

costs and low acquisition frequency. This thesis presents a reproducible machine learning

pipeline for regional sediment classification using Sentinel-2 imagery, systematically cover- ing the shallow shelf between 0 and 30 m depth. The methodology integrates Principal

Component Analysis (PCA), Multilayer Perceptron (MLP), and Synthetic Minority Over- sampling Technique (SMOTE), combined with a conservative curation protocol based on

historical granulometric consistency, local spatial agreement, and independence between image patches. The final predictive scope was defined as the binary Sand vs Mud scenario, using 2,199 spatially independent samples under 25 km block-based spatial validation with five folds. Among the evaluated model families, including convolutional neural networks, MLPs, SVMs, and EfficientNet models, the MLP+PCA+SMOTE pipeline was selected as the recommended configuration because it combined competitive performance, stability, and low computational cost. The optimized configuration achieved an accuracy of 92.64% ± 1.95, macro F1 of 89.04% ± 2.40, and computational cost 6.6 times lower than that of CNN-Base. Although no statistically significant differences were observed in relation to the evaluated convolutional architectures, the cartographic agreement of 84.9% with an independent sedimentological map indicates regional cartographic consistency, without constituting contemporary in situ validation. The results demonstrate that lightweight spectral models represent a viable and scalable alternative for regional sediment mapping in complex coastal environments, provided that they are accompanied by rigorous curation, adequate spatial validation, and cautious interpretation of the temporal limitations of the data.