|
Title |
Transfer Learning for High-Concentration Algal Forecasting: Bridging Water Quality Characteristics B etween U p stream a nd D ownstream
|
|
Authors |
강덕준(Dejun Jiang) ; 권혁구(Hyuk-Ku Kwon) |
|
DOI |
https://doi.org/10.15681/KSWE.2026.42.4.325 |
|
Keywords |
Bidirectional GRU; Data scarcity; Harmful algal blooms; Spatial domain adaptation; Transfer learning |
|
Abstract |
Accurate prediction of chlorophyll-a (Chl-a) in spatially heterogeneous river networks is consistently hindered by data scarcity at newly established monitoring sites. Limited historical records prevent deep learning models from capturing the dynamics preceding acute bloom events. This study proposes a spatial transfer learning framework to address these limitations in the Geum River basin, South Korea. A cost-sensitive Bidirectional GRU (Bi-GRU) source model, which captures bidirectional temporal dependencies in hydrological data, was pre-trained on a decade-long dataset from the Gapcheon River. A sigmoid-weighted loss function was incorporated to prioritize rare high-concentration bloom events. The generalized hydro-chemical representations encoded in the pre-trained Bi-GRU layers were then transferred to two data-scarce target sites, Daecheong (upstream) and Buyeo (downstream), by adjusting the final model output to fit local conditions. Compared to conventional models trained solely on limited local data, the transfer framework demonstrated superior recovery in the critical high-concentration range (top 25%). The locally trained model achieved an overall R² of 0.83 at Daecheong but collapsed entirely under peak bloom conditions (HC R² = -0.01). In contrast, the Transfer model maintained meaningful high-concentration accuracy, achieving HC R² of 0.87 at the downstream Buyeo site. Feature importance analysis revealed a spatial drift in dominant drivers, with predictive reliance shifting from broad thermal controls at upstream sites to localized biogeochemical factors, including pH, TOC, and EC, at the downstream site. This framework offers a scalable early warning pathway for monitoring-limited rivers, presenting a compelling alternative for data-scarce modeling.
|