Using AUV’s and AI to Automate Data Collection and Imagery Data Analysis to Better Inform Offshore Energy Planning and Resource Management

Abstract

Benthic imagery collected by Autonomous Underwater Vehicles (AUVs) provide crucial ground-truth data needed for seabed habitat characterization, yet the manual annotation bottleneck required to convert these images into useful datasets remains a major challenge. A single AUV survey track can provide upwards of 8,000 images, requiring an estimated 13 workdays on manual annotation. This makes large-scale habitat characterization impractical without the use of automated systems. This study evaluates the viability of using a YOLO-based convolutional neural network (CNN), implemented within the software CoralNet-Toolbox to assist in automating the classification of benthic imagery collected in the NOAA R/V Seahawk mission in the Carolina Long Bay Wind Energy Areas. The secondary objective of this study was to assess how image patch size influences classification performance, testing three patch sizes of 164 x 164, 224 x 224, and 264 x 264 pixels. Results showed that classification performance is sensitive to patch size, with the 224 x 224 model achieving the highest overall accuracy (80.8%), balanced accuracy (76.5%), and macro-averaged F1 score (0.41) across four substrate classes of Rock, Sand, Shell, and Biological. The smallest patch size (164 x 164) exhibited increased confusion between Rock and Sand due to insufficient spatial context, while the largest patch size (264 x 264) maintained high overall accuracy but showed reduced minority-class recall, indicating a bias towards the Sand class. Minority class performance (Shell and Biological) remained unstable across all models due to there being limited training representation for those substrate types. These findings establish that patch size is a crucial design variable in patch-based benthic classification workflows. The framework developed in this study provides a reproducible methodology for integrating AI-assisted image analysis into large-scale offshore habitat mapping.