In the paper, authors suggest to apply a neural network regression model to study a potential impact of external parameters of a physical model on its simulation results. They test how the neural network regression model VESPER can help to assess the potential impact the updated lake cover (lake fraction), lake depth and newly introduced lake salinity flag fields, on the simulations of the skin temperature (Land+Lake Surface Temperature, LST) by the NWP model IFS. They use VESPER to correct the LST results of IFS, applying several additional predictors, including updated and newly introduced lake data. For that, they use MODIS observations of LST as a ground truth. They show that a network regression model can help to see potential benefits and highlight potential problems. This is a very important and interesting finding! It is still doubtful from the manuscript, however, that this approach allows to calculate the accuracy of the possible physical model updates: many quantitative estimates in the paper are questionable.

There are several important essential comments and some editorial comments. Since there are many of them, I would suggest major revision of the paper. For the detailed comments, see the Supplement file.

Reviewer comments on

Deep Learning for Verification of Earth-System Parametrisation of

Water Bodies by

Reviewer comments on

Deep Learning for Verification of Earth-System Parametrisation of

Water Bodies by

Tom Kimpson, Margarita Choulga, Matthew Chantry, Gianpaolo Balsamo,

Souhail Boussetta, Peter Dueben and Tim Palmer

This study presents, evidently for the first time, VESPER, kind of

"statistical twin" to the IFS dynamical earth-system model. The idea

is to apply deep learning method to test the impact of land surface

description to predicted land surface (skin) temperature. For this,

selected transient and static near-surface fields from ERA5 reanalysis

are used as input to neural network regression model. For validation

and comparison, surface temperature from Aqua MODIS is used. The study

focuses on lake surface temperature, that is treated by Freshwater

Lake parametrizations in the IFS model when producing ERA5 reanalysis.

The study presents new ideas, results, interesting

discussions. Especially I liked the nice combination of global view to

analysis of local details. Unfortunately the manuscript is written in

a very raw and messy way. It needs a major revision in order to focus

in essential and present the work in an organised way.

I would suggest to improve and possibly extend the presentation of

basics of VESPER. My feeling is that your ideas have emerged in

connection to the development of the lake data base and then quickly

grown towards other areas of the surface physiography

description. Evidently, VESPER is a powerful tool for impact studies

beyond the lakes, as indeed discussed in the paper. This is why it is

important to present it systematically, with sufficient basic facts

included. In the introduction you might even shortly describe how it

evolved, and to make clear that later in this paper you will focus on

evaluation of the impacts of lake physiography description updates. To

keep the paper size reasonable, you might consider if all results and

examples now included are essential. One possibility could be to keep

them minimal for illustration of the power and possibilities of VESPER

only, and write a second part already entering the details of lake

applications?

My expertise is not sufficient to judge the applied machine learning

method itself. This is why I will asume everything is fine with it and

instead make some questions and suggestions concerning data,

variables, application of the results. I have not systematically

checked for typos or language mistakes, will only comment unclear

formulations that prevent understanding.

Specific comments

-----------------

I would suggest modification of the manuscript strucure to

something like this:

1. Introduction

2. Constructing VESPER

   2.1.   Features and targets

   Table about the variables, their sources  (now in 3.1) 

   feature transient/static source                updated

                            surface description

(GlobCover2009 and GLDBv1 GSWE and GLDBv3 ...)

                ERA reanalysis

                MODIS

   2.2    Data sources

   2.1.1  ERA5

   

   2.1.2  Aqua MODIS

   2.3    Joining the data

   2.4    Constructing a regression model

   add a Table about realizations:  

   V15 V15x V20 V20x

   what they mean, how differ, what were the updates done

   (now some explanations are in 3.1, some jump only among the results much later)

3.  Results (or Evaluation of ...)

4.  Discussion

5.  Conclusion 

l.54:

I find the definition "when we refer in this work to “lake

parametrisation" we mean both the model and the parameters"

problematic. In the IFS model, FLake is a model that is used as

parametrization scheme in grid cells where is some fraction of

lake. You show that in ERA5, a significant percentage of the 31-km

cells indeed contain some amount of lake. If there is no lake in

gridpoint, then FLake does not work there and vice versa. When FLake

is run coupled to atmosphere and surface inside the IFS model, it

needs in addition the lake depth. Can you confirm that in the IFS

model, every surface type (lake, sea, ice/glacier, forest etc) is tied

to a specific parametrization scheme? If yes, then you might perhaps

indeed bundle each scheme to the physiography description it needs.

This was the case when ERA5 was prepared or is when the weather

forecast is run. Now, you possibly use the word model to denote the

statistical twin, VESPER, not the IFS model itself. (You might check

the consistency of your text in this sense). When building VESPER, you

use near-surface atmospheric fields from ERA5, and the same or updated

physiography fields that IFS model uses but you do not apply the

dynamical parametrizations like FLake here nor do you essentially

interact with the atmosphere.

The variables are listed in your Table 2 (to which I suggested

modifications above). Could you please add in the VESPER description

explanation how these variables were chosen? In that context, you

might clarify the roles of VESPER v.s. IFS model. More generally, you

might discuss, after presenting the results, why you think that the

the physiography fields will have similar type of impact in the earth

system model as they have in VESPER, why you expect that you can

generalise the result.

l. 123 ERA5

Please explain shortly how ERA5 skin temperature is obtained as this

is your central variable in this study. 

When explaining ERA5 please mention if FLake was a part of the ECLand.

About the ERA grid cells, please make clear that reduced Gaussian grid

cells are approximately of the same size all over the globe.

l. 133 MODIS

Please explain how MODIS surface (skin) temperature is derived. What

is the role of the surface emissivity, what are the related data and

assumptions when converting the outgoing LW radiation to Ts? To see

the surface, optical wavelenght instrument needs to clear the clouds,

perhaps you could shortly mention how this is done for MODIS.

l. 146 and Figure 3

What do you mean with "a 4 km resolution on a regular

latitude-longitude grid"? Either 4 km or regular lat/lon, not both?

Could you please explain if you indeed need the intermediate lat/lon

grid between the raw satellite pixels (of about equal size 1 km all

over the globe?) and ERA5 grid cells? Wouldn't it be possible to pick

boxes of e.g. 4 x 4 km size along the satellite track, with ca 16

pixels and use the box average, central coordinate and time when

collecting data into needed ERA5 cells? Is it possible that the

problems shown in Figure 3 - a lot of data at polar latitudes, lack of

data around Equator - are partly due to the artificial transformations

to lat/lon grid and back? The satellite does not care about the

coordinate convergence when flying over the rotating earth, simply

looks down over some view angle. In this sense, the Figure 3. caption

might need reformulation, too.

l. 184

Please move to this section the V15 etc model realizations, a new

table and descriptions.

Figure 4 and related text

Please define the prediction error (VESPER or ERA5 against MODIS,

which measure?) Please check the consistency in other Figures and

Tables.

I think Figure4 is important. Please discuss why VESPER overperforms

ERA5. This means also discussion of the grid-scale skin temperature

Ts. Aspects here:

- If you train VESPER Ts towards MODIS, does this mean that the

  improved result is closer to reality or only to MODIS, including

  also its problems over Himalayas etc?

- You showed a map of MODIS uncertainties, there are also specific

  studies like

  (https://www.tandfonline.com/doi/full/10.3402/tellusa.v66.21534

  Kheyrollah Pour et al., 2014) evaluating MODIS against in-situ lake

  surface water temperature. So how close MODIS is to reality? Any

  comments here in this respect? There are comments about uncertainty

  in the general discussion, though.

- If VESPER would indeed overperform ERA5 Ts, then why not to use it

  for forecasting instead the dynamical model? What could be the use

  of Ts beoynd the earth system model where it is an internal variable

  related to surface energy balance, interaction atmosphere via

  radiation and atmospheric turbulence etc?

- What about comparing outgoing LW radiation in clear-sky pixels by

  models to the corresponding MODIS variable? This would treat the

  surface emissivity in different way than the derived Tskin?

Section 3 Results

Please move the general VESPER realization materials with Table

earlier into a basic description, leave here the presentation of

lake-related results.

Please explain what happens in the physiography fields when e.g. Lake

Aral disappears - what comes instead of it? You discuss this in

diffent cases but perhaps a general explanation would be good in the

beginning.

(Somewhere later in discussions:) Would VESPER realizations and ECLand

parametrizations, including FLake and snow schemes etc. within IFS

model, react to the changes in similar way? This is a question of

applicability of VESPER results to the full earth system model

development.

In general, the Section of results is interesting and detailed, but

perhaps consider condensing it and/or making another part of the paper

out of it. There might be more things to study related snow and ice on

lakes and surroundings etc. that deserve attention later.

Section 4 Discussion

Above in the comments I have presented some possible questions to be

discussed here.