Adriano Fonseca
Ph.D. Thesis Proposal Defense
Ocean Engineering
Thursday, May 21, 2026, 9:00am
Chase 1305
Abstract
Modern multibeam echosounder (MBES) surveys produce massive, irregularly sampled sounding sets whose processing remains a bottleneck for hydrographic surface production. A dominant challenge is to convert these soundings into statistically validated surfaces while preserving small, safety-critical features and maintaining traceability and uncertainty awareness consistent with hydrographic standards.
In this proposal, I develop and validate a point-cloud-native pipeline that combines (i) conservative, uncertainty-aware subsampling and (ii) graph-based learning to support scalable bathymetric reconstruction. The first contribution is a subsampling stage (Bathy Subsample) that reduces redundancy without committing to raster gridding. The method partitions soundings into spatially coherent neighborhoods, preserves anomalous measurements by design, and consolidates redundant points using surface-aligned, TVU-informed vertical binning and Bayesian Gaussian mixture modeling. The output is a reduced point cloud with explicit strength weights that preserve density information and support downstream learning and surface construction.
The second contribution is a graph neural network (GNN) framework (GraBN) for learning on bathymetric soundings. Preliminary results demonstrate point-wise reconstruction scoring for denoising; the core proposal extends this proof of concept toward explicit local surface reconstruction by coupling a GNN encoder with surface-generating decoders (including FoldingNet-style folding layers).
Finally, I propose a comparative validation method that quantifies the impact of preprocessing choices by evaluating models trained on full-resolution soundings, GMM-subsampled soundings, and CUBE/CHRT-informed reductions, using both surface-level and hydrographically relevant risk/uncertainty metrics. Although outside of the scope of this proposal, this thesis represents a real-world challenge with a real need for robust implementation. With this in mind, the study will also address the subject of performance and optimization throughout the work.
Bio
Adriano Fonseca was born in Rio de Janeiro, Brazil. He received a B.Sc. in Electronic and Computer Engineering from the Universidade Federal do Rio de Janeiro (UFRJ) in Rio de Janeiro, Brazil. He went on to earn an M.Sc. in Integrated Circuits and System from the Université Paris-Saclay and a Specialization degree in Micro and Nano Electronics from CentraleSupélec both in Paris, France. His work focused mostly on integrated circuit design for harsh environments. He then earned an M.Sc. in Electrical Engineering from Coppe/UFRJ with a focus on computer vision and machine learning and an emphasis on Generative Adversarial Networks.
He is currently pursuing a Ph.D. in Ocean Engineering at UNH and researching crowdsourced bathymetry acquisition and processing.
