Summary

Aqel et al. present a neural network (NN) based approach to predict matric potential from soil water content observations. Using an autoencoder, they extract the most relevant features of the soil water retention dynamics. They input their results into a deep neural network (DNN), which increases the transferability of the DNN.

Assessment

The approach presented in this paper is convincing. The manuscript is well-written. Prediction of hysteresis in soil water retention is of interest to the soil hydrology and soil physics community.

I don't have major comments. Thus, I recommend accepting the manuscript after minor revisions. I have some minor comments below.

Nash-Sutcliffe efficiency

The Nash-Sutcliffe coefficient tends to emphasise maxima in a time series, which might bias the results. An additional interesting metric would be the Kling-Gupta efficiency (Knoben et al., 2019; doi: 10.5194/hess-23-4323-2019).

Local nature of the model

The results obtained are from sites that share similar climate and topography. I wonder if this workflow would work as well in different regions of the world, or if further adjustments must be made.

Physically-based modelling

I partially agree that the reductionist mechanistic models might be unable to account for the full complexity inherent in the soil water retention process. Input-agnostic approaches such as neural networks surely have an advantage when it comes to predicting matric potential. However, physically-based modelling is also a tool for process understanding that could potentially help us disentangling the effects of all the interacting processes that control soil water retention. I know that there are efforts to make machine learning a tool for process understanding as well. Perhaps the authors could comment briefly on this and place their work in this discussion?

General Comments

This manuscript presents an approach for predicting soil water potential and its hysteresis under natural field conditions by combining deep neural networks (DNN) with autoencoder neural networks. This integration leverages the strengths of both methods, with the autoencoder effectively compressing and capturing site-specific features of soil moisture dynamics, and the DNN utilizing these features to enhance prediction accuracy.

Overall, the method is promising and convincing, and the manuscript is well-organized and clearly written. I have only a few concerns and suggestions, primarily regarding the model's generalization capability to clay soils and regions with significantly different climatic conditions, and the model's interpretability.

Specific Comments

1. Lines 106-109: Generalization Capability: Autoencoders are highly dependent on the quality and diversity of the training data. As shown in Figure 1, the selected region has relatively similar climatic conditions and soil types, mainly loams with a clay fraction less than 50%. I am curious about the model's generalization capability to different regions with varying climatic conditions and soil types, especially for clayey soils. Suggest expanding Section 4.1 to discuss this point and potential approaches to address this issue. Additionally, suggest discussing the possibility of using other autoencoders, such as variational autoencoder (VAE).
2. Lines 218: Model Interpretability: The interpretability of the autoencoder's hidden layer representations is typically challenging. Suggest Including a discussion in the results analysis or discussion section on potential techniques to visualize the features learned by the autoencoder's hidden layers, which can help readers understand the model's internal workings.
3. Lines 289-290: Why adopt an NSE value > 0.80 as the criterion for an optimal model? Please provide the rationale for selecting this value.
4. Figure 2: The common unit for matric potential is -kPa. Please explain the relationship between the -kPa and -cm used in this manuscript.
5. Equation 1: Please ensure that all parameters are clearly defined after the equation, and that their mathematical notation (bold, italic) is consistent throughout the manuscript.