Trying to estimate soil fungi with vis-NIR spectroscopy is venturing down a very slippery slope. Soil fungal abundance and diversity have no spectrally active components. This is recognized by the authors in the Introduction. At best, estimation of soil fungal abundance and diversity is based on indirect correlation with properties that are spectrally active, such as clay minerals functional groups. However, it is not even clear from the literature whether soil fungal abundance and diversity have much relationship with these spectrally active soil components. With the use of complex machine learning models, there is a serious risk of finding accurate results simply because the model finds fortuitous relationships between spectra/variables and the property of interest. Without any surprise, in this study the model performing best is the most complex one: the deep learning model. The simplest model, PLSR, performed poorly in nearly all cases. PLSR is yet the benchmark method for spectroscopic modelling, and is not per se a simple model, it uses the whole spectrum and makes statistical decomposition of the spectral data (principal component analysis). This clearly suggests that the results of this study cannot be trusted because the good results obtained by the 1D-CNN models are not based on direct nor known indirect relationship with spectrally active soil components. At best, there is some kind of indirect relationship with spectrally active soil components which serve as proxy for the soil fungal abundance and diversity, but it is difficult to prove that this relationship indeed exists based on the literature. The reliability of this relationship would also then depend strongly on the relationship found among the data by the deep learning model and on the specific study case and calibration data. For this reason alone, I do not endorse publication of this manuscript. There is no reason from the existing literature to justify the use of vis-NIR spectroscopy to estimate soil fungal abundance and diversity. It is misleading to claim that spectroscopy can “serve to supplement the more expensive and laborious molecular approaches”.

In addiction to this major conceptual problem, I have also several important comments that need to be addressed, please see below.

General comments

This paper is very similar to the Soil Biology and Biochemistry paper: same methodology, same data (fungi instead of bacteria, but still on the Australian BASE dataset), same covariates, same concept.

Combining spectral data and terrain into the modelling is not recommended and not common is spectroscopic research. The authors provide as input to the model a vector comprising both spectral data and environmental data found at site. The rational for doing this is unclear, and providing too much data to complex models will only aggravate the problem of finding spurious relationship among the data. This way of modelling brings additional problem with the covariates used. For example, the covariates called “mineralogy”, i.e. kaolinite, illite and smectite (Appendix), are an interpolation of vis-NIR spectra band depth. So these covariates add redundant information. Combining spectral data and environmental variables into a single vector used for prediction is going too far into “unconscious” soil modelling. 

To my understanding (because the writing needs to be improved and made more precise) the modelling is made on the absorbance spectra whereas the interpretation is made on a different model based on the continuum removed reflectance spectra. This is not correct. Interpretation should be made on the same model that the authors choose to be the best, not on a model fitted on different data. The authors are interpreting a model that was not used during modelling.

Data

The BASE dataset contains more than 577 samples. Also, in the SBB paper 681 samples are used for bacteria. Why are not all data used? What is the reason to discard some observations? This should be clearly described in the Methods section of the manuscript. 

The environmental variables used as covariates are outdated (Appendix). For example, the prescott index map for Australia is made at 5 km, but indeed downscaled at 90 m. The mineral maps (kaolinite, illite and smectite) are an interpolated product based on the band depths of vis-NIR spectra -should not be used here to avoid redundancy. The soil texture maps of 2015 are outdated, there are new ones since several months (see the paper https://www.publish.csiro.au/sr/pdf/SR20284). The vegetation maps are also very outdated. The authors used the old ones based on 250 m and 1 km resolution products, but there are new ones since several months based on Landsat 30 m data, see for example: http://data.auscover.org.au/xwiki/bin/view/Product+pages/Landsat+Seasonal+Fractional+Cover. Also, all the topographic covariates are now available at 30 m resolution and easy to download through the TERN repository: https://shiny.esoil.io/Covariates/.

Methods

The use of various machine learning methods is not of interest and there is the risk that this study simply becomes a comparison of validation statistics. Not clear what the added value is of applying seven methods. In this way, the manuscript runs the risk of being about the models and their comparison, rather than about understanding whether vis-NIR spectral data can effectively predict soil fungi. I think SOIL’ readership is rather interested in the latter. Any model that here predicts better than another is case-study specific, and is unlikely to interest soil scientists who are the readers of this journal. Further, a model is usually chosen carefully with the problem in hand. Several of the models used by the authors are in fact very similar, they are all non-linear. The problem is here that none of the models described in the Appendix are thoroughly explained and it is not clear how the authors actually implemented all this. I could make many questions on each of the models, but my best advice here would be to reduce the model count to the minimum (PLSR and DL), otherwise most readers will probably see this manuscript as a programming exercise.

It is unclear why the authors describe the accuracy statistics this way and what was done. With the information provided, the interpretation of these statistics is also unprecise. For example, the coefficient of determination of the linear model can be interpreted as amount of variance explained only in the case of a linear model with intercept. So the R2 that the authors report in the manuscript is similar to a squared correlation coefficient, and should not be interpreted as percent of variance explained. See also the Wikipedia page, last paragraph of the intro: https://en.wikipedia.org/wiki/Coefficient_of_determination. Further, decomposition of the RMSE into a bias and variance component is very well known in the literature and adds nothing to the paper. How the authors call it with “inaccuracy” and “imprecision” is not well accepted in the statistical literature. The decomposition is so well known that it is described in Wikipedia: https://en.wikipedia.org/wiki/Mean_squared_error or in any introductory statistic book, see, for example, page 298 in the book "Elementary Statistics for Geographer", by Burt et al. (2009), Third Edition.  https://www.routledge.com/Elementary-Statistics-for-Geographers/Burt-Barber-Rigby-Robeson-Horner/p/book/9781572304840. Note that in both cases, they do not use terms such as “inaccuracy” and “imprecision” and I would agree not to use them.

Variable Importance

The methods section of the manuscript does not clearly describe what was done. For example, there are many methods available to compute the variable importance. At lines 137-140, the only description is “VarImp for caret” and permutation importance for the 1D-CNN models. This is not a description that allows one to reproduce these results. What is done in caret VarImp? For the permutation importance of the 1D-CNN model, how is correlation among wavelengths accounted for (permutation is very sensitive to correlation)? How many permutations? What is the metric used (RMSE difference, ratio?)? What is the unit?

In Fig. 5, the importance should have a unit. What is the unit? How comes that covariates and spectra, which have very different units, can all have importance values between 0 and 1?  Usually a standardization can be made to the covariates and spectral data prior to use permutation on them. If this was done, please report it in the methods section and also use the same standardized input data for modelling. There is no reason to use different data and model for modelling and interpretation.

The manuscript by Yang and colleagues describes a study whose aim is to test the predictive capabilities of learning models based on different types of predictors such as near-infrared spectroscopy (point data) and large-scale environmental variables with respect to the abundance and diversity of soil fungi.

The study is based on a rather large Australian dataset, comprising several hundred soil samples from different regions of Australia under different land-use types, for which fungal abundance and diversity have already been measured and published in a previous study, and for which near infrared (NIR) spectra (and related environmental data) have also been acquired and published in previous papers.

The authors test different strategies to build their learning models, varying the type of predictors used (spectroscopic data alone, environmental data alone, or a combination of both types of information) and the learning algorithms used (7 different algorithms). The performances of the different learning models to predict the abundance and diversity of soil fungi are tested by cross-validation.

The objective of this manuscript is of great interest to SOIL journal readers: are we able to predict the abundance and diversity of soil fungi by learning models based on simpler-to-obtain predictors in unknown Australian soils?

The article is overall well written and is very well structured. The figures and tables are clear. The learning models seem to have been well constructed by a team that is very well versed in data science techniques and the application of spectroscopy technique to soils.

Major points

However, in reading this article, several major shortcomings appeared to me.

• Seasonality of the abundance and diversity of soil organisms

First, an important point: it seems to me that the seasonal variability of the abundance and diversity of soil organisms in general and soil fungi in particular is a major "unthought" of this manuscript. In this work, no reference to this major determinant of fungal abundance and diversity and no information on the date of sampling and climatic conditions in the month prior to sampling of each soil sample was collected and exploited as a potential predictor by the authors in their learning models. It seems to me that through this "unthought" of seasonallity, the authors give up, without ever informing the reader, an important part of the determinism of their soil fungus-related variables of interest (cf. e.g. reference Kivlin and Hawkes, 2016 which is cited by the authors in their manuscript). In this paper we can read: "The strong seasonal pattern in fungal richness and abundance suggests that fungal studies in tropical forests require temporal sampling to capture the full community. Indeed, the inconsistent correlations of fungi with climate and plant hosts across tropical ecosystems […] may be due to seasonal variation and spatial heterogeneity across single time point studies" (Kivlin and Hawkes, 2016)

In general, how can a soil biodiversity measurement at a time t of a soil can reliably represent the diversity and relative abundance of different species/groups of soil fungi for that soil in its different seasons? Applied to the authors' objective in this manuscript: how can one hope to reliably predict very dynamic soil variables (fungal organism abundance and diversity) with predictor variables that are much less dynamic in the soil or in the environment? At best, one could hope to predict large differences between very contrasting soil pedoclimates: something that Biogeography of soil organisms works have often already identified. However, such Biogeographic works do not constitute predictive models of soil biodiversity at a given time t.

At the very least, it seems to me that this point, which is an important intrinsic limitation of this study (and also of a similar previous study focusing on soil bacteria by Yang et al. 2019 in Soil Biology and Biochemistry; https://doi.org/10.1016/j.soilbio.2018.11.005); should be discussed in detail by the authors.

• Overly optimistic conclusions compared to reported results

A second major point concerns the conclusions that the authors draw from their results regarding the predictive capabilities of the learning models they have constructed for the abundance and diversity of soil fungal groups in Australia. These conclusions seem to me to be overly optimistic compared to the results the authors show in their manuscript. After reading this paper, if I had to answer the question: can we, in unknown Australian soil samples, robustly estimate the abundance and diversity of major soil fungal groups with the learning models constructed in this work, my answer would be: no, and we're a long way off. At most we can identify some major soil and environmental determinants of the abundance of these groups and their diversity, making this a work of Biogeography of soil organisms, much more than a paper that aims to predict these properties in unknown soils to study soil health.

Indeed, in the introduction to their article, the authors mention that "a paucity in the availability of soil microbial data is thought to be one of the main contributors to the uncertainty of soil health assessment and ecosystem management" while in concluding their paper, the authors state that "here, we show that spectro-transfer functions with readily accessible vis–NIR spectra and publicly available soil and environmental data can be developed to estimate soil fungal abundance and diversity" I disagree with the conclusions of the authors: using their models for soil health assessment of a particular Australian soil (for a criteria of soil health linked to soil fungal abundance or diversity for instance) would certainly not help improving the acuracy of this assessment given the predictive performance showed in this manuscript.

Specifically, the models with the best predictive capabilities use a combination of point spectroscopic data and continuous environmental data, and they most often use the deep learning algorithm 1D-CNNs. But despite these (still somewhat obscure) performance differences between the different algorithms, even the "best" results shown by Yang and colleagues are not very encouraging in terms of the actual predictive capabilities of the learning models built on unknown Australian soil samples:

- For the two (very) dominant soil fungus groups in Australia (Ascomycota; Basidiomycota) representing on average 80% of the total fungal abundance; Table 1), the "best" learning models show R²s below 0.6. I infer that the learning models constructed in this paper do not robustly quantify the abundance of groups representing 80% of the fungi in Australian soils.

- For the diversity of fungi estimated by the "ACE" indicator, the R² is 0.45, I deduce that the learning models built in this article do not allow to quantify in a robust way the diversity of fungi in Australian soils.

- A few models with slightly better predictive abilities are shown for the abundance of two groups of very low abundance fungi in soils (although it is difficult to quickly judge their performance in the absence of synthetic performance indicators such as the ratio of performance to deviation or the ratio of performance to interquartile distance).

Other points

Section 2.1: ecosystem types, please clarify the difference between woodland and forest.

Section 3.1: « In total, more than 60 million quality filtered sequences in the whole dataset were obtained, with an average of 107 310 sequences per sample. When we clustered the sequences at 97% similarity level 202 200 OTUs were detected. Each sample had an average of 666 OTUs » : this was already presented in the section 2.2, please remove it from the results section.

Figure 1b: please add some information to remind reader that this graph shows mean abundances (this graph does not represent errors on the mean, which is huged as shown in Table 1).

I agree with R1 that interpreting NIR spectra with continuum removed reflectance signals and using savitsky-golay first derivatives of absorbance signals in the modelling work can be misleading. Please interpret the NIR spectra with the signal used for modelling.

I wonder how does the sum of the model predictions for the relative abundance of the 5 main groups of soil fungi behave for the soil sample set used in this study: does this sum get close to 1 for all soil samples?

Reference

Kivlin, S. N. and Hawkes, C. V., 2016: Tree species, spatial heterogeneity, and seasonality drive soil fungal abundance, richness, and composition in Neotropical rainforests, Environmental Microbiology, 18, 4662–4673, 201