Abstract Deep learning models (DLMs) have gained attention for estimating daily streamflow, often outperforming traditional process‐based models (PBMs). However, only recently have studies begun comparing the performance of DLMs and PBMs under uncertain data conditions, and the existing work overlooks streamflow measurement errors, consider only a subset of uncertainties, and does not examine how different training schemes affect DLM sensitivity. This study evaluates the sensitivity of PBM (Sacramento, SAC), DLM (long short‐term memory, LSTM), and a hybrid model (LSTM incorporating SAC outputs) performance to data uncertainties: streamflow measurement error, precipitation input error, and out‐of‐sample (OOS) conditions. Precipitation and synthetic streamflow, estimated by a population PBM, are considered error‐free and used to construct controlled uncertainty scenarios, enabling a systematic comparison of model sensitivities under known‐error conditions. Results show that DLMs are more sensitive than PBMs to measurement errors, making PBMs preferable when streamflow data quality is low. However, DLMs are less sensitive to input errors than PBMs in gauged and ungauged basins. All models exhibit substantially higher sensitivity to input error during the testing period than training. DLMs struggle to capture changes in extreme flood under OOS conditions, whereas the PBM effectively captures these events, suggesting their suitability for applications under OOS conditions. Moreover, multi‐basin training can effectively reduce sensitivity to measurement errors in DLMs. All models demonstrate increased sensitivity in basins characterized by high elevation, low temperature, and dry conditions. This study provides timely and practical insights for selecting appropriate modeling approaches, especially given the growing global use of DLMs.
Kim et al. (Fri,) studied this question.