Evaluating rainfall-driven LSTM for statistical analysis of low-flow regimes in the Nakagawa basin, Japan
Year: 2026
Type: article
Abstract: Nakagawa Basin in Japan. The suitability of a rainfall-driven long short-term memory (LSTM) rainfall-runoff model was evaluated for low-flow regime analysis. Hourly discharge for 1994–2022 was simulated from basin-mean rainfall and rainfall-derived state variables representing cumulative rainfall, antecedent wetness, drying persistence, and seasonality. The model was trained on 1994–2014 and validated on 2014–2022. Evaluation combined global performance metrics with a two-sample Kolmogorov-Smirnov test on daily mean flows, flow-duration-curve diagnostics, and threshold-based low-flow signatures derived from the observed daily-discharge flow-duration curve. Low-flow occupancy, event occurrence, duration, seasonality, and recession characteristics were assessed with bootstrap-based uncertainty estimates. The model reproduced the overall discharge series well, but substantial discrepancies emerged in the low-flow tail. Bias increased toward more extreme low-flow conditions, and threshold-based diagnostics showed progressive deterioration in low-flow occupancy and event detection from moderate to extreme low-flow regimes. At the most extreme threshold, only one valid simulated recession event was identified, limiting recession-scale comparison. These results show that strong global model skill can coexist with poor reproducibility of extreme low-flow behavior and highlight the need for regime-specific evaluation before applying rainfall-driven deep-learning runoff predictions to drought and environmental-flow decision support in the Nakagawa Basin. • Rainfall-driven LSTM simulated hourly discharge at Noguchi (Nakagawa, Japan) • High NSE/KGE coexisted with distribution shifts and positive FDC tail bias • Moving-block bootstrap quantified uncertainty in FDC and low-flow signatures • Low-flow tail (Q75–Q99) was overestimated (up to ∼+20% at Q99) • Diagnostics support drought and environmental-flow applications of AI models
Authors: Do-yup Kim, Naoki Shirakawa
Institution: University of Tsukuba
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,600
OpenAlex ID: W7144099064