Deep Learning Approach for Classification of Water Bottom and Surface from Bathymetric Lidar Point Clouds

This study investigates the application of PointNet, a deep learning architecture, to classify bathymetric LiDAR point clouds in shallow waters. Using a dataset from Marco Island's southern coast, Florida, the research categorized water levels into noise, surface, column, and bottom classes through Gaussian curve fitting and novel rule-based approaches. PointNet was trained considering critical parameters such as batch size, epochs, learning rate, and optimizer. Results indicated that a batch size of 8 yielded higher validation accuracy (0.7001) compared to 16 (0.6926). Evaluation showcased an approximate 70% accuracy, distinguishing noise, surface, bottom, and column points. While some ambiguity existed between surface and column points, differentiation between bottom, surface, and column was evident. This study demonstrates PointNet's feasibility for bathymetric LiDAR classification in shallow waters and emphasizes optimizing parameters for enhanced accuracy and performance.