From Beach to Cliff: Adapting CoastSnap Citizen Science for Coastal Cliff Change Detection Using ML-Assisted Image Registration and Prompted Segmentation

Abstract

CoastSnap is a citizen science tool that uses repeat smartphone photographs from fixed stations to monitor coastal change, yet it has rarely been applied to coastal cliffs. We test a workflow for cliff change screening using a controlled pilot CoastSnap station at Ōnaero, New Zealand (iPhone 12), integrating (i) Machine Learningassisted ground control point (GCP) transfer for image registration, (ii) pinhole camera calibration and reprojection-based geo-rectification onto a curved cliff-surface model, and (iii) prompted segmentation for isolating cliff-related features. Auto-GCP benchmarking shows a tradeoff between precision and robustness: the Scale-Invariant Feature Transform (SIFT) model achieves low localisation error when successful, but is sensitive to changes in illumination. In contrast, the Local Feature ransformer (LoFTR) model provides a higher detection yield and more stable performance, but is sensitive to the threshold used. Nevertheless, it is well suited to operational use with human-inthe- loop verification. Calibration produced consistent parameters across repeat images, with focal length estimates matching device specifications. Visual segmentation based only in Regions of Interest (RoI) often merged adjacent objects, while text-guided Segment Anything Model (LangSAM) delineated cliff and debris more reliably than fracture-like features. Preliminary detection of the supply and removal of debris highlights the potential for quantifying the frequency and magnitude of mass movements at each cliff.