The authors have presented an application of machine learning techniques to create a global model of groundwater levels in Germany. They compared two different input model settings: one with static features and one with dynamic features. Additionally, they compared these settings with two reference cases, one with random entity variables and a second without entity variables. Their results indicate that both model settings perform well under in-sample conditions, but their performance diminishes in out-of-sample conditions. The work aligns with the growing trend in this field, introducing entity-aware methods to hydrogeology and yielding promising results. I believe the paper has the potential for publication, but there are some points that need to be addressed before publication.

• Limited Machine Learning Methods Tested: The authors only tested two machine learning methods, LSTM and CNN. There are powerful alternatives like Transformers, which have outperformed LSTM in other studies (https://doi.org/10.1016/j.apr.2023.101833). LSTM models are designed for handling transient conditions, whereas CNNs are adapted to do so. There are various other methods like extreme gradient boosting that have been applied in hydrogeology that can be a powerful alternative to CNN (cite: https://doi.org/10.1016/j.watres.2023.119745, https://doi.org/10.1007/s10661-020-08695-3, https://doi.org/10.1016/j.scitotenv.2021.151065). The authors should explain why they chose not to explore more advanced models before attributing out-of-sampling prediction issues to dataset limitations.

• Data Fusion Method Comparison: In Line 164, the authors propose a new data fusion method. Did they compare this method to existing methods to demonstrate its benefits and limitations? Comparing it to cited methods would provide valuable insights.

• Introduction: The introduction lacks a description of the study area, which should be addressed since the model applies to a single case study. Adding a figure depicting the study area and well distribution would enhance the paper's context.

• Well data selection (Line 104): Quantitatively explain what "spatial coverage as representative as possible" means. Clarify if there is a minimum distance between wells, data density per area, or any specific criteria used for well selection. Provide the original dataset size from which data was picked.

• Upscaling (Line 136): Elaborate on the importance of not having too fine-grained categorical data. Describe the upscaling process and how the authors ensured that each category is correctly represented. Provide references or explore the effect of upscaling on training and prediction.

• MLP Classifier (Line 168): Explain the advantages of adding an MLP classifier rather than providing static features directly. Address concerns about uncertainty propagation due to MLP output in the concatenation.

• MLP Output Nodes (Line 170): Specify the number of output nodes in the MLP.

• Validation MSE (Figure 2): Explain the phenomenon where the validation MSE is smaller than the training MSE, especially in the initial epochs. This could indicate a bias in the validation dataset, and clarification is needed.

• Feature Importance (Line 245): Suggest using the SHAP method (Lundeberg and Lee, 2017, https://doi.org/10.48550/arXiv.1705.07874) for more stable feature ranking, as it has been used effectively in similar studies (Ransom et al., 2020, https://doi.org/10.1016/j.scitotenv.2021.151065).

Overview

The authors develop an entity-aware deep learning model for spatially and temporally continuous groundwater level prediction using a combined Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP) network. They rely on ground observations from 108 wells in Germany and other dynamic and static predictor data obtained from multiple sources to train, validate, and test the model. The authors also perform some interesting comparisons of four model variants, namely, the time series feature-driven model (TSFeat), environmental feature-driven model (ENVfeat), random static features (RNDfeat), and dynamic inputs only features (DYNonlyfeat). While there are some issues with spatial generalizability in the out-of-sample setting, the model shows satisfactory performance with the Nash-Sutcliffe Efficiency (NSE) > 0.8 in an in-sample setting.

Overall, the manuscript is generally well-structured, with detailed explanations of the methodologies and data. However, the following comments should be addressed before this manuscript is published.



Please see the attached PDF for my comments.