Peer review on manuscript

"Predicting environmental gradients with ...................."

Introduction
Revision Recommendations

This study tests the proposal that ..... species can Question: Minor
indicate cation concentration and the proportions of Data: Accept
sand, silt and clay in .... It finds strong correlations Methods: Major
between observed and predicted cation concentration
Inference: Minor
(good QUALITATIVE predictions: correlation
coefficients [calculated from the r2 values given] Writing: Minor
varying from 0.81 to 0.87), though how
QUANTITATIVELY accurate the predictions were is not very clear (see Critique). The
predictions were much poorer for particle size classes, both qualitatively (r=0.37 to 0.69) and
quantitatively (RMSE values around half the typical values of the variables). Interestingly,
using ...... data for ... ... did not improve the predictions over ... ... ... data. When the
analysis was done the other way round, ... ... species composition was strongly related to soil
cation concentration (the flip-side of the above), and most species were typically found in
quite a narrow range of soil cation concentrations - preferences that appear to align quite well
with those in ... ... ... ... .... Therefore, many of the species can be used as indicators of soil
cation levels.

Merits

The study uses a large dataset, with reasonable coverage across ... ... .... The main analyses
are appropriate in my opinion, though can be improved in their application and reporting (see
Critique). The conclusions are mostly well supported by the results. The predictions are done
using two quite different approaches, which yield very similar results, reassuringly. The
manuscript builds on a good body of research on this topic in the region and is written with
some authority. It is also quite well written, apart from lack of clarity in some places, and
some specifics. The authors make clear the intended practical application of the research, as
well as its novelty. I see no major flaws.

Critique

I could not discover whether the RMSE values for cation concentration in Table 2 are in units
of log10(sum of bases) or untransformed sum of bases. (Line 136 says that these values were
logtransformed before analysis.) This means the reader cannot tell how big the typical
prediction error was for this variable. This is important because it is the main result, and the
second conclusion stated in the abstract is "Species composition in a plot can provide good
quantitative estimation of nutrient concentration in the soil." It may be that although the
correlation between predicted and measured cation concentration was strong, the exact
predictions were not very accurate, in which case the quantitative estimation was not as good
as claimed. The predicted values may also have systematic errors, such as underprediction at
high cation levels and overprediction at low levels. Or a non-linear relationship with the
measured values. If an ultimate aim is to produce a map of important soil properties (final
sentence of Discussion) then these things matter. The predictions were generated using a
leave-one-out approach, at the plot level. But the 305 plots analysed were aggregated within
about 38 'locations' (I count 38 triangles on Fig.1), with 5-30 plots per location (line 91). Each
plot was at least 1km from all others (line 88), but I suspect many were in quite uniform
terrain. Particularly for the K-NN analysis, which used the 4 nearest neighbours (in terms of
species composition) to predict values (line 548), this could lead to predictions that are unduly
good. That is, sites close together geographically may tend to be very similar in both species
composition and soil properties. While that would not invalidate the results, it would suggest
that to get the sorts of prediction strengths reported in the manuscript, you need known values
from sites within a few km of the focal one - so the usefulness for low-cost, broadscale
mapping is diminished. In fairness, the WA analysis is probably much less affected by this
concern (but it could be quite strongly affected if the key indicator species have small
geographic ranges). It is not clear how large the geographic ranges of the 54 .. .... species
are. Overall, for a study explicitly aimed at an eventual mapping outcome, I was surprised at
the lack of analysis of spatial structure in the data and errors.

All four soil variables were positively skewed. Cation concentration was log-transformed but
there is no mention of transformation of any of the other variables. This could affect many of
the analyses, particularly the r2s of the regressions of predictions vs measured values (Table
2; the main results).

Finally, it seems possible that ... ... alter the soil properties, rather than only responding to
them. Was there any consideration of this in the soil sampling? Perhaps worth discussing.

Discussion
Some environmental variables are quite well mapped, even for areas with little field sampling
(eg from satellite imagery). Soil variables are much more difficult because they usually
require intensive, quite costly field sampling (and laboratory work). At least for ecologically
important soil variables, a good proxy is potentially very useful, for example for conservation
planning (as the authors state), or other purposes. The manuscript can make useful advances
in this respect. The results reported show that cation concentration is important for ... ... ,
where the climate is mostly 'benign' and not strongly limiting. Whether other soil variables are
needed for other taxonomic groups remains to be seen, as the authors note; thus it is not clear
whether the usefulness is restricted to ... ... , or is much wider. Some suggestions for
improving the research within this manuscript:
*Report and interpret the slope and intercept of the fit line for measured vs predicted soil
values.
*Use the leave-one-out approach at the 'location' level. That is, leave out all the plots from the
focal location when generating the predictions for each plot.
*Divide ... ... geographically into a few 'regions' - Fig.1 and Appendix S1 suggest some
quite 'natural' clusters of sampling points - and try predicting from one region to others. This
may really help in establishing the limits to prediction (especially for mapping purposes), and
thus inform data-collection needs for the future.
*Examine the distributions of the errors in geographic and environmental space, try non-
linear fits, etc.

Finally, a possible strength of this approach is not mentioned. Species' presence reflects
conditions over a much longer period than the moment of the field sampling, and over a
broader area than the exact soil-sample location. So, a bit like using invertebrate species to
indicate water quality, using ... ... to indicate soil properties may have advantages over
direct soil measurements.

References

[1] Anonymous authors (2013) Predicting environmental gradients with ... ... ....
(unpublished manuscript) - Peerage of Science

Additional comments for authors

Will you publish the ... ... plot data? I hope so because they could be useful for other
purposes.

The equivalents of Fig. 3 for the particle size variables would be nice to publish in
supplementary material.
There are various typos and very minor errors, which I have not listed here (e.g. in line 22 it
should be 'Main conclusions').

The last sentence of the abstract (lines 26-27) is incomplete.

Lines 41-42 -- this sentence needs rewriting because at the moment it says that biotic
heterogeneity is an important determinant of biotic heterogeneity!

Lines 81-82 -- the word 'if' should be used as a conditional. Here the correct word is 'whether'.
Also, data are plural; this sentence has both singular and plural. So the sentence should read
'Finally, we assess whether species abundance data are needed to obtain useful predictions, or
whether the more easily obtainable presence-absence data are adequate.'

Lines 93-94 -- if the idea was NOT to sample any environmental gradient then why were
transects used?

'Data collection' section (and rest of manuscript): I could find no mention of WHEN the
floristic data were collected.

Line 224 -- latitude and longitude were tested, apparently. Why is there no mention of these in
the methods or Table 1, or anywhere else in the manuscript? Was anything else tested but not
reported?

Line 229 -- I do not follow "All twenty indicator species of the richer soils". In Fig.2, I count
18 indicators (and 21 at the second level for the richer soils). I have no idea what is being
referred to here.

Line 230-231 - I do not follow the end of the same sentence! It does not seem to relate to
results reported in Fig.2. Is it referring to a possible alternative first-level split?

Line 252 -- "a minimum of four neighbouring plots". I do not understand. From lines 172-174
I thought that the final predictions would use a fixed number of neighbouring plots, once the
'best' k had been established, NOT a variable number with a fixed minimum. Also Table 2
legend suggests it is k=4 not k>=4.

Line 270 -- better to join this paragraph to the previous one so the first paragraph of the
discussion summarises the key findings.

Lines 312-313 -- Surely you can investigate this by sub-sampling your data. I suggest you
either do this section properly or remove it. From what you say on page 15 the better option
seems to be to keep the section and do the extra analysis.
References -- various formatting inconsistencies.

Table 1 -- how appropriate is it to report standard deviations for the variables that are highly
rightskewed?

Table 2 -- as well as saying whether the RMSEs are on logged values or not, the formatting
needs fixing.

Figure 1 -- the legend says "crosses" but they look like triangles to me.

Figure 2 -- why is the figure repeated?