Source-linked AI summary
Learning the Shoreline: A Very High-Resolution Approach to Reef Island Dynamics
Tobias Fischer, B Stoll
TL;DR
Conventional shoreline proxies can miss localized atoll changes that matter for ecosystems, infrastructure, and local stakeholders. The study combines very high-resolution Pléiades imagery with XGBoost transfer learning to map emerged-land shorelines across French Polynesian atolls, achieving high accuracy and generalization without local retraining. The workflow reveals fine-scale shoreline shifts while retaining stated limitations related to clouds, tides, and ambiguous transitional zones.
Problem
Conventional imagery, planform metrics, and vegetation-line proxies can miss localized shoreline changes, while site-specific models require local training data that limits Pacific-wide scalability.
Method
The study uses Pléiades spectral and textural features with an XGBoost classifier to delineate the outer edge of emerged land and transfer one model across atoll scenes.
Results
Mean MAPE was 1.28 m overall and 1.44 m for two held-out scenes, while IoU consistently exceeded 0.97, demonstrating stable classification across diverse atoll settings and timeframes.
Takeaways & Limitations
Very high-resolution shoreline mapping reveals subtle shifts, sediment redistribution, and motu reshaping that can remain undetected when assessment relies on total land area.
Takeaways & Limitations
Cloud and shadow artifacts required manual correction, and tidal normalization was not applied because tide gauges and beach profiles were unavailable.
Abstract
from arXiv · showhide
Pacific atoll islets are often described as stable in global-scale studies, typically based on long-term shoreline proxies such as vegetation line or morphometrics like planform surface area. While informative, these approaches can obscure short-term, localized coastal dynamics -including changes in island shape and position -that are critical for ecosystem function, cultural practices, and coastal infrastructure resilience. This study presents a transferable, automated approach to shoreline monitoring using very high-resolution Pl{é}iades imagery and a XGBoost classifier. The method integrates spectral indices and textural features to delineate the outer limit of emerged land, including vegetated areas, beaches, man-made surfaces, and beach rock. This shoreline definition supports finescale, spatially explicit monitoring of reef island dynamics, even in morphologically complex environments. Developed and tested on multiple atolls in French Polynesia (Tetiaroa, Tikehau, Hao, and Puka Puka), the model achieves high accuracy (mean Intersection over Union $\approx$ 0.99; Mean Absolute Positional Error $\approx$ 1.28 m) and demonstrates strong performance on both training and held-out sites, validating its spatial transferability. The extracted shorelines reveal subtle but significant island-scale changes in extent, configuration, and spatial position that remain undetected by conventional shoreline proxies and surface metrics. By enabling highprecision, scalable shoreline monitoring, this method provides a more nuanced understanding of atoll change processes. It supports Pacific efforts to move beyond narratives of passive loss toward frameworks of resilience and adaptation, while providing spatial tools tailored to low-lying island realities.
1 Introduction
Atoll islets are highly exposed to climate threats, yet conventional shoreline assessments can miss localized changes. This study develops a transferable, high-precision approach for monitoring such dynamics across diverse atoll settings.
- Atoll islands are vulnerable to sea-level rise, coastal erosion, and extreme events because of their low elevation, narrow shape, and dynamic structure.Monitoring these changes supports assessment of habitability, infrastructure risks, and ecosystem health.
- Conventional imagery, planform metrics, and vegetation-line proxies can miss motu reshaping, sand movement, and artificial expansion.These localized changes matter to local stakeholders despite the usefulness of conventional approaches for long-term trends.
- Very high-resolution imagery and machine learning enable detailed shoreline monitoring, but most existing studies remain site-specific and require local training data.This limits scalability across the Pacific’s diverse island systems.
- The study evaluates a transferable shoreline classification approach using Pléiades imagery and an XGBoost model across diverse French Polynesian atolls.The approach applies transfer learning for cross-site classification without retraining.
2 Study Area and Data
The study uses four French Polynesian atolls selected for contrasting geomorphologies, reef structures, connectivity, and land use. Pléiades imagery from 2016–2023 provides coverage across these varied settings, with some cloud-affected scenes and targeted analysis areas.
- Four atolls—Tetiaroa, Tikehau, Puka Puka, and Hao—were selected for contrasting geomorphologies, reef structures, and land use.These variations provide a test of shoreline-monitoring transferability.
- The atolls differ in lagoon-ocean connectivity, shoreline configuration, and human modification, including tourism infrastructure, conservation, plantations, military development, and shoreline armoring.Tetiaroa and Puka Puka are closed atolls, whereas Tikehau and Hao are larger open atolls with major passes.
- Pléiades scenes acquired between 2016 and 2023 included four Tetiaroa images and one 2022 image for each of Hao, Puka Puka, and Tikehau.All images were acquired between May and July and delivered as atmospherically corrected, co-registered, pan-sharpened Level 2A products.
- Scenes with minimal cloud and haze were prioritized, although partial cloud cover affected parts of Tetiaroa and Tikehau.For Hao and Tikehau, analysis focused on representative sections near main villages and adjacent motu to balance coverage and computational cost.
3 Methods
The workflow defines shoreline as the outer edge of emerged land visible in Pléiades imagery, classifies land and water using spectral and textural features, and extracts continuous shorelines. Transferability is tested with one model trained across multiple scenes and evaluated on held-out temporal and spatial targets.
- The operational shoreline is the outer edge of emerged land visible in Pléiades imagery, including vegetated zones, beaches, reef platforms, and artificial structures.Depending on land cover, it corresponds to the beach toe, vegetation line, or stability line, while excluding submerged reef flats.
- Spectral indices and 3×3-window GLCM texture attributes from pan-sharpened Pléiades imagery support land–water classification.The texture features include mean, contrast, and entropy calculated from the NIR band.
- The XGBoost workflow produces a binary land–water raster, removes artifacts and internal gaps, and extracts shorelines with marching squares and linear interpolation.Extracted shorelines are compared with ground-truth references for accuracy assessment.
- A single transfer-learning XGBoost model was trained on five scenes and tested on Tetiaroa 2022 and Tikehau 2022 as temporal and spatial hold-outs.Evaluation used IoU, the proportion within 2 m, P95 shoreline-distance error, MAPE, and RMSE.
4 Results
Transfer learning matched or outperformed individually trained models across most test settings, including held-out scenes. Accuracy remained high across varied atoll environments, although performance varied by scene and cloud-affected data required correction.
- The transfer-learning model consistently matched or outperformed per-image trained models, including on temporally and spatially held-out scenes.The Original models were trained and tested separately per image, whereas TF was trained once on five scenes and tested on two held-out targets.
- IoU generally exceeded 0.97, and more than 87% of predictions were within 2 m of the reference shoreline in five of seven scenes.The 95th percentile error remained below 3.9 m in all but one case.
- Tetiaroa 2019 showed transfer-learning improvements across all metrics, with P95 decreasing from 5.36 m to 2.99 m and predictions within 2 m increasing from 88.8% to 90.6%.Similar improvements were observed for the Tetiaroa 2016 training image.
- Tikehau 2022 showed strong spatial generalization, with transfer learning outperforming the Original model across all indicators and reducing P95 from 4.19 m to 3.84 m.
- Cloud correction changed Tetiaroa 2016 MAPE from 2.55 m to 1.28 m and RMSE from 8.96 m to 3.84 m.MAPE and RMSE were supplemented with more robust spatial indicators because they are sensitive to outliers and classification ambiguity.
5 Discussion and Implications
Transfer learning and very high-resolution imagery support accurate, spatially transferable shoreline monitoring across diverse atoll settings. The resulting shorelines reveal localized changes and operational limitations that conventional area metrics and uncorrected imagery can miss.
- >0.97 IoU and 1.28 m mean MAPE demonstrate stable classification across diverse atolls and timeframes.Held-out scenes averaged 1.44 m MAPE, supporting spatial and temporal generalization.
- Very high-resolution imagery captured subtle shoreline shifts, sediment redistribution, and motu reshaping even when total land area remained unchanged.These observations demonstrate why direct shoreline mapping adds information beyond planform surface metrics.
- Tahuna Iti experienced 2.21 ha of accretion, 1.29 ha of erosion, and a 38.8 m northeastward shoreline shift between 2016 and 2023.The mapped changes illustrate island-scale dynamics that global area metrics can overlook.
- Manual correction was required for cloud and shadow artifacts, while missing tide gauges and beach profiles prevented tidal normalization.Ambiguity also persisted in transitional zones such as tide pools and overhanging vegetation.
- The workflow supports spatially nuanced resilience strategies beyond binary loss-or-gain narratives.Its operational framing is aligned with regional adaptation goals.
6 Conclusion
The study demonstrates that transfer learning with very high-resolution satellite imagery can provide automated, accurate shoreline detection in atoll environments. The workflow generalizes to held-out sites without local retraining while capturing changes invisible to conventional land-area metrics.
- A mean positional error of ≈ 1.28 m indicates high spatial accuracy for automated shoreline detection.
- The workflow achieved robust held-out-site accuracy without local data or retraining, supporting field-ready generalization.
- Integrating spectral and textural information captured subtle shoreline changes that conventional land-area metrics leave invisible.
- Despite cloud, tidal-variability, and shoreline-ambiguity limitations, the workflow supports scalable, spatially explicit monitoring for Pacific islands.