Abstract
The rapid development of deep learning (DL) is the most significant contemporary evolution of hydrological science, yet its limitations remain underexplored. This study aims to evaluate the effectiveness of DL-based, traditional, and hybrid hydrological models in streamflow simulation especially in data-deficient regions. Daily discharge flow of the four sub-catchments in Samar, Philippines were modeled using HEC-HMS, Univariate Long-Short Term Memory (LSTM) Network, Classical GR4J and DL-assisted, parametrically- optimized GR4J. Results show that the classical GR4J underestimated the discharge, while the LSTM overestimated peaks. The DL-assisted, parametrically-optimized GR4J achieved the highest consistency and balance between realistic peak and low flow estimation, with NSE = 0.63-0.84, IA = 0.85-0.93, LMI = 0.76-0.82, and low MAPE (≤ 0.03) and RSR (≤ 0.02). Although the Univariate LSTM captured general trends well, it underestimated some peaks despite reaching high values in performance metrics. Many studies highlight DL's hydrological modeling power but ignore its limits and hybrid model benefits. The results highlight that neither DL nor classical models alone are sufficient as DL-assisted approaches yield more reliable and realistic hydrological simulations, particularly in climate-sensitive, data-deficient regions like Samar.