N/A

The submitted Technical Note describes a potential cost-effective methodology for monitoring evapotranspiration (ET) based on Smartphones and portable sensor. 

While the idea is interesting and there is a need of dense ET measurements, I do not think the Technical Note is fit for publication in EGU journals. I understand this is only a preliminary study, but the results are still too limited in my opinion.

The main reason behind my decision is that the models proposed by the authors are developed and tested on a "well-instrumented site". What is the guarantee that the proposed model will work elsewhere when there is no lysimeter or EC data available for calibration? What parameters to use? To show the actual feasibility of this approach, the authors should have tested on other sites not used for calibration. What errors can we expect by applying parameters calibrated elsewhere? What are influential factors hindering generalization? Also, how are results affected by the individual mobile phone used?



Other issues of relevance include the inappropriate use of the terms "machine learning" to define multivariate regression done with very limited data (machine learning is data hungry), as well as the lack of information concerning the rationale behind the formula used. Other detailed comments follow below.

Lines 24 - 32: Can the authors refer to existing literature when highlighting these gaps?

Line 44: built-in

Line 48: not sure referring to the figure without a thorough explanation is suitable at this point in the introduction. Better in the methodological section?

Lines 49-50: "[...] the question is how these estimates can work in 

50 concert under field conditions to produce accurate ET." this part is not sufficiently clear.

Line 54: RQ1 seems to be very generic; is the focus beyond that of ET?

Line 55: RQ2 comes a bit out of the blue. It is the first time machine learning is mentioned. There is no story leading to it. 

Figure 1: Not sure why Figure 1 is made of two parts given that they look quite different. Would it make more sense to seprate them? Also, I don't find the caption particularly informative and sufficiently correlated with the images. 

Line 82: Add a space before "An overview..."

Line 88: "we use the following multivariate regression as a simple form of machine learning to estimate the evapotranspiration..." <- why calling this machine learning? It is a simple multivariate regression, and there is nothing wrong with it, per se. The authors should refrain from calling this machine learning and change the manuscript accordingly. On the other hand, what is the rationale behind the formula used? What are the parameters that need to be calibrated? There is no explanation.

Line 89: The equation has not been numbered.

Line 93: how much data is available in total?

Figure 2: the legend and captions are confusing, please amend.

Line 113: "training Eq. 1" you do not train equation and that equation is not machine learning.

Line 117-120: I think the authors overstate the results they obtained. They are fitting their model to the two target variables, EC or lysimeter using values of said variables for parameters calibration (in the "training" dataset). On the other hand, lysimeter and EC are obtained independently.

Line 130: I don't this is a great practice to add comments in parentheses? E.g., "it should be noted...".

Line 130-133: This paragraph is convoluted. Please rephrase and consider splitting it.

Figure 3: The caption is not clear and the figure as well. Why not showing with different markers calibration vs validation data?

Why the two vertical bars separating the two parts of the image?

Discussion and Outlook: I find that this section lacks a proper discussion on the limitation of the proposed approach. How many precision lysimeters do we need to calibrate a world-wide network of phone-based algorithms? Is a linear method sufficient to generalize to location with no calibration data? Perhaps a non-linear machine learning model would be more useful to improve model generalization by processing external data (i.e., rural/urban catchment characteristics, see [1] for instance.

[1] Kratzert, Frederik, et al. "Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets." Hydrology and Earth System Sciences 23.12 (2019): 5089-5110.