Reviewer comments to

HORAYZON v1.1: An efficient and flexible ray-tracing algorithm to

compute horizon and sky view factor 

by Christian R. Steger, Benjamin Steger, and Christoph Schär

The authors propose a new method for calculation of topographic

horizon and sky view factor based on ray tracing library and using a

high-resolution digital elevation model. For calculation of the

orographic radiation effects it is necessary to consider the geometry

of non-local terrain. The main parameter to consider is the local

horizon, from which the sky view factor can be derived. It is

important that the horizon is calculated with the highest possible

resolution of the surface elevation data. Such calculations,

especially if done using less optimal conventional algorithms, require

large computational resources in terms of memory usage and processing

time.

The authors propose, test and document a new and more efficient method

that is based on a high-performance ray tracing library. It is

demonstrated that the calculations could perform up to two orders of

magnitude faster than conventional ones. In addition to the

application the ray tracing method that stores terrain terrain

information in an efficient way, optimizations are related to

limitation of the calculation domain to only what is strictly

necessary at each point (terrain simplification in the boundary zone,

masking of ocean points). The suggested method is surely valuable for

the applications, like the numerical weather and climate prediction.

The manuscript is of applied, technical character which is fine in

this case when new software is described. It is well written, contains

detailed documentation and discussion of the suggested method, gives

sufficient background and demonstrates the authors' good understanding

the previous methods and applications. The paper can be used as basic

documentation of the method. The underlying data and the HORAYZON

source code are of public domain and available via github, even

together with user support, that makes the application especially

valuable.

The manuscript seems ready for publication with minor corrections. I

do not have sufficient expertise to verify the derivation of the

equations and technical details of the ray tracing method but rely on

the authors that these have been done and presented correctly. I would

however like to use the opportunity to raise for discussion some

general questions, suggestions, concerning application of HORAYZON in

numerical weather predition models (General comments). This is not to

suggest modifications to the manuscript but perhaps to take into

account for further developments and application. Some minor comments

concerning the manuscript text follow (Minor comments).

General comments

I would like to shortly describe our experience on preparing basic

terrain data for orographic radiation parametrizations within the NWP

models of HIRLAM and ACCORD NWP consortia. Here, methods described

first by Senkova et al. (2007) have been applied. In the latest

experiments, we took SRTM of 3" resolution over a limited domain

(e.g. over Caucasian mountains, Rontu et al., 2016, see also

https://www.ecmwf.int/sites/default/files/elibrary/2018/18234-radiation-and-orography-weather-models.pdf).

In each SRTM (lat,lon) point we calculated local horizon angles (LHA)

for (8) directional sectors. First we estimated the horizon for 360

sectors, resulting in one value per each one-degree sector, then

averaged these for 8 sectors. This was done using a simple home-made

fortran programme, searching maximum elevation within an assumed

radius around each point (for SRTM 3", we only took a radius of 5

km). In addition to the original (1) SRTM surface elevation field we

now got 8 extra LHA fields in the the same grid. Separately, we

calculated at each SRTM point the maximum slope angle and its azimuth

angle using 8 neighbours. This added two more SRTM-grid fields. We

used the standard tools (by gdal) for slope calculations. All these

calculations in SRTM grid were first done by using external programs

within a workstation, later more approximately within the physiography

generation phase of the NWP model, before aggregation of the data to

the model grid for derivation of slope, shadow and sky view factors.

The resulting 8 + 2 extra fine-resolution fields (could be e.g. 16 + 2

as well, with LHA sectors of 22.5 degrees instead of 45) are all we

need for the second step, (statistical) aggregation to the NWP model

grid. We also tried to calculate the sky view factor at each SRTM

point, possibly using the slope angle of the point in the sector LHA

was facing. In hindcast, it seems that SVF could rather be estimated

in the aggregation phase for the model grid, building on the

precalculated sectorial LHA and slope angles in the source grid.

After this long introduction comes the question/suggestion: would it

be possible to apply HORAYZON to the (almost global) NASADEM (or even

to the more local higher-resolution DEMs), in order to provide the

users of the DEMs with pre-calculated sectorial LHA and slope angles?

I mean, applying high-performance computing facilities with graphical

processors and utilizing an available data base of some suitable

programme (like the COPERNICUS services, ECMWF computers) would allow

for doing the common basic work effectively and once for all, letting

the NWP consortia or other users to focus instead to the task of

(statistical) data aggregation for the specific parameters in their

specific grids?  There, plenty of different applications and

approaches, variable and changing grids, surely wait for development

of their specific solutions. The amount of resulting global

pre-calculated horizon data would be large but not much more than one

order of magnitude larger than that of the source DEM data. Most users

would only need to transfer data for specific domains anyway, and as

this is the question about orography fields, there are no worries of

their time evolution (as opposite to the output fields of NWP or

climate models).

Would you see principal or practical problems in such an approach,

e.g. in view of the SVF discussions within your sections 3.3, 4.1? In

several places you refer to Pillot et al. (2016), mentioning at l.306

that their algorithm was designed for point locations which makes its

run time substantially larger. Yours, according to Figure

5. calculates the horizon for a predefined small area (blue shaded

domain)? What would happen if you applied HORAYZON to calculate

horizon in the (transformed) DEM source grid points and returned the

resulting LHA fields back there? What about the additional inaccuracy

due to coordinate transforms? If you feel it appropriate, perhaps you

could discuss these questions in the concluding discussions. From the

practical point of view, their matlab code is most probably not

applicable in parallelised high performance computing environment

while yours might be?

Minor comments

l.10 (abstract)

Could you please add in the abstract one sentence, one number perhaps,

that would characterize the efficiency of the proposed method compared

to something conventional, already existing? On the line 320 you write

"In summary, the performance analysis revealed that the ray-casting

method is much faster for all considered terrain sizes (by about two

orders of magnitude)"

l.30 (introduction)

There are applications, like road weather models, that downscale the

radiation fluxes from NWP models and apply terrain corrections in

point scale for calculation of the road surface energy balance, for

discussion see e.g. Karsisto, 2019

(https://helda.helsinki.fi/handle/10138/305417).

Section 3.1.

Would it be possible to discuss the impact, loss of accuracy due to

the coordinate transformations? Are transformations kind of

reversible, i.e. would it be possible to return the calculated

variables back to the original source grid?

l.160

Indeed, the suggested masking approach might be useful for other

applications, too, e.g. in the surface data assimilation of NWP

models where practical coastline problems are met.

l.167

"High tesselation level" sounds a bit specific terminology for a

reviewer not familiar with computer graphics world.

Figure 6.

Would it be possible to indicate the horizontal (vertical) scales of

the valley shown?

Eq. (10) and (11)

To make sure I understood it correctly: here you allow that the point

you calculate SVF for is inclined, like in Manners et al. wanted to

assume (but made a mistake as you suggest)? In this case, your

Eq. (11) should have been applied also instead of Eq. (1) in Rontu et

al., 2016.

l.395

typo? "DEM data with high spatial resolution has to be processes*ed*,

which can be done..."

Section 5.1

I did not understand from the text how you did the (reference) spatial

aggregation of SVF to coarser grids? After l.420 in the next

section you do discuss simplifications, sampling density etc.

Section 5.2

Sub-grid SVF calculation is expensive, true, but somewhere you might

mention that such calculations are not done on daily basis but

only when new experiment (or operational NWP) domains are

defined. Perhaps not here but in introduction or discussions. (We also

applied sub-grid SVF, horizon and slopes in Senkova et al. and Rontu

et al., although using relatively coarse source DEMs.)

l.450

The approach by Helbig and Löwe could be characterized as a terrain

parametrization, whose results are later applied at another level of

parametrization within the radiative transfer calculations in a

NWP/climate model or in their postprocessing. In my opinion, it

represents a significant simplification.

l.484

"minor performance dependency on horizon search distance" is encouraging.

Appendices

I have not gone through the appendices.

see attachement