Novel mapping approach for coastal boulders using deep learning

Abstract

Coastal boulders are often formed by tsunamis and storm surges and thus provide valuable insights into the dynamics of past inundation events. However, the mapping of these boulders has been constrained by inherent limitations in the speed and precision of manual processes. In this study, we introduced a novel approach for boulder mapping by integrating an unmanned aerial vehicle with a mask region-based convolutional neural network that enables the rapid and precise detection and volume calculation of boulders distributed along coastlines. The approach was validated on Ishigaki Island, Okinawa, Japan, and showed high precision in the detection and measurement of boulders, achieving a high F1-score of 0.863 for the target boulders. The digital surface model yielded more realistic and precise volume calculations than the traditional approximation approach, providing more accurate information for understanding the transport processes of coastal boulders during inundation events. This approach enhances our understanding of past inundation events and provides a practical tool for ongoing coastal monitoring and disaster response. Furthermore, it serves as a baseline model for future research in automated boulder mapping.