Machine Learning for Marine Dock Detection Using LiDAR Intensity and Detectron2

Abstract. The Province of British Columbia is undertaking a multi-year LiDARmapping program to deliver high-quality elevation data and support open-access geospatial products. Although dock detection was not an initial priority, this study demonstrates the value of leveraging LiDAR acquisition programs to support scalable, automated coastal infrastructure mapping. We evaluate the impact of LiDAR intensity normalization on CNN-based marine dock detection using Detectron2. Through a controlled ablation, we compare three strategies—raw intensity, scan-angle correction, and range-based correction—to isolate radiometric effects on detection performance. Detection metrics (precision, recall, F-score, IoU) and shape-fidelity measures are reported separately to clarify trade-offs. Results show scan-angle correction delivers balanced precision and recall, while range-based correction improves precision at the cost of recall. Beyond technical findings, this study offers comparative evidence to guide preprocessing choices and provides evidence-based guidance for AI-derived coastal mapping products in British Columbia.