AbstractPeatlands play a critical role in the global carbon cycle, with water level dynamics strongly controlling their function as carbon sinks or sources. While process-based models are commonly used to simulate peatland hydrology, the potential of data-driven approaches remains largely unexplored at large spatial scales.Here, we assess the capability of a Long Short-Term Memory (LSTM) model to simulate daily water level in natural northern peatlands (40°N–75°N), trained on a diverse set of in situ water level observations. Model performance is evaluated against the same in situ water level observations using a strict block-wise cross-testing scheme. Furthermore, model performance is benchmarked against simulations from NASA’s Catchment Land Surface Model with peatland modules (PEATCLSM).The LSTM model demonstrates improved agreement with in situ water level observations compared to PEATCLSM in terms of root mean square difference and bias, while the PEATCLSM exhibits higher spatial and temporal correlation with the in situ observations. Feature importance analysis indicates that the LSTM model captures key hydrological controls on water level dynamics, with precipitation and reference evapotranspiration emerging as dominant drivers, followed by leaf area index and snow water equivalent.The lack of sufficient in situ water level observations for model training, both in terms of record length and spatial coverage across peatland sites, restricts the development of a model with additional input variables that could enhance performance. Despite these limitations, the LSTM model shows spatial patterns consistent with the process-based model, supporting its reliability. These findings highlight the potential of deep learning approaches such as LSTM-based modeling to complement traditional process-based modeling of peatland hydrology. Future improvements will depend on collaborative data sharing to enhance training datasets and support informed climate and environmental decisions.
