2025-09-18 京都大学
波崎海岸の観測桟橋
＜関連情報＞
- https://www.kyoto-u.ac.jp/ja/research-news/2025-09-18
- https://www.kyoto-u.ac.jp/sites/default/files/2025-09/web_2509_Mori-3970043afa218d07873908a509b351da.pdf
- https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025GL117764
少ないほど良い：短期ウィンドウ校正がモデリングにおける季節的海岸線予測を改善する Less Is More: Short-Term Window Calibration Improves Seasonal Shoreline Prediction in Modeling
Xinyu Chen, Masayuki Banno, Nobuhito Mori
Geophysical Research Letters Published: 28 August 2025
DOI:https://doi.org/10.1029/2025GL117764
Abstract
Coastal communities worldwide rely on shoreline models for risk assessment and management, yet these models often struggle to capture observed variability across different temporal scales. We analyzed 30 years of shoreline observations at Hasaki Beach, Japan, using Discrete Wavelet Transform to separate variation by timescale. Spectral analysis revealed wave-driven annual and semi-annual cycles, while long-term trends contributed significantly to total variance. The ShoreFor model, when calibrated using the full 30-year data set, severely underestimated seasonal variability. In contrast, 2-year calibration windows successfully reproduced seasonal variations both within calibration periods and, after DWT-based detrending, across the entire 30-year validation period. Our findings demonstrate that short-window calibration substantially enhances model capability for capturing wave-driven seasonal shoreline changes, offering a practical solution for coastal risk assessment using limited observational data. This approach is particularly valuable given increasing availability of satellite-derived shoreline data and the need for accurate seasonal predictions under changing climate conditions.
Plain Language Summary
Beaches are continuously being reshaped by the joint effects of coastal hydrodynamics. Beach erosion and accretion occur across timescales ranging from single storm events to year-to-year patterns, making it challenging for numerical models to capture all of these changes. We studied 30 years of high-frequency measurements from Hasaki Beach in Japan and discovered that when models are trained on decades of data, they attempt to explain everything simultaneously, from seasonal cycles to long-term trends, resulting in poor predictions at any specific timescale. Here, we propose a new approach: train the model using only short 2-year periods to make it focus exclusively on seasonal changes. Our results show that models trained with much less data actually reproduce seasonal shoreline cycles better than those trained on 30 years of data. After separating long-term trends using mathematical techniques, these short-window models maintained their predictive skill across the entire 30-year period. This finding offers a new calibration strategy for coastal modelers: let models focus on specific timescales rather than forcing them to explain everything at once. This approach is particularly valuable for locations with limited data, where just a few years of observations can provide robust predictions of wave-induced seasonal variations.
Key Points
- Short-window calibration (2 years) captures seasonal shoreline variations with sustained prediction skill across 30-year validation
- Long-term calibration (30 years) underestimates seasonal variability due to parameter bias from long-term trends and model drift
- Discrete Wavelet Transform effectively separates shoreline variability by timescale, enabling robust isolation of seasonal components
