Effects of meteorological reconstruction on hydrological and deep learning streamflow models in Lanyang Watershed, Taiwan

Abstract

Study region: The Lanyang River watershed in northeastern Taiwan, a steep mountainous basin characterized by strong climatic gradients and frequent discontinuities in meteorological observations. Study focus: To evaluate how different meteorological reconstruction methods influence watershed scale streamflow simulation using SWAT+ and deep learning models, and to quantify how variable specific reconstruction errors propagate through hydrological forecasts across multiple testing intervals. New hydrological insights for the region: Maximum and minimum temperature variables can be reconstructed reliably across stations, whereas precipitation remains the variable with the highest level of uncertainty even when supported by the high-resolution background fields provided by the Taiwan Climate Change Projection Information and Adaptation Knowledge Platform. Reconstruction quality directly affects hydrological behavior, as different machine learning reconstruction approaches modify streamflow timing, seasonal coherence, and event scale variability in distinct ways. Deep learning models, especially Long Short-Term Memory, outperform SWAT+ in temporal efficiency and variability reproduction, typically improving coefficient of efficiency (CE) by about 10-35% compared with SWAT+ model, demonstrating the value of combining improved meteorological reconstruction with advanced forecasting approaches to enhance hydrological reliability in this data scarce mountainous basin.