Fourier-Transform Infrared Spectral Inversion of Soil Available Potassium Content Based on Different Dimensionality Reduction Algorithms
Round 1
Reviewer 1 Report
I believe this is an excellent manuscript with sound scientific merit and contributes greatly to the discipline. The authors of the manuscript have provided excellent background, adequate description of methods, results, and discussion. Here, the authors present the use of FTIR for estimating soil potassium content and compare data reduction methods for the best model. The authors found that the SPA algorithm was the most suitable, and in particular the SPA-PLSR model performed the best with an R2 of 0.49. Overall, the manuscript is very well written and provides significant results for the field. These results are presented well in figures and words, and I believe the manuscript is nearly ready for acceptance.
Figure 1: the subset map of china- what coordinate system/ projection was used? The map looks a bit flat.
Lines 147, 163: a bit awkward to say “as follows: .” perhaps reword to more direct language.
Line 200- “is” to “was” – please watch for tense here. Should be consistent.
Figure 3- really enjoy this figure
Really, a nice job.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
General comments:
The manuscript “Fourier-Transform Infrared Spectral inversion of soil available potassium content based on different dimensionality reduction algorithms” investigated the use of FTIR and dimensional reduction techniques for predicting soil available K at the site level. In general, it is a well-written manuscript and would be of general interest to the audience of Agronomy. I have some small suggestions/ minor revision requests to perhaps help the authors further improve their manuscript. In particular, I would like to see a bit more in-depth discussion about the implications of this work.
Specific comments:
L37-39. Awkward sentence.
L67-70. Expand a bit more either here or in the discussion. The model performance from previous studies was shown for SOC/ TN, but why not K? How much lower is the performance for K models? Some studies did include IR-based predictions for soil K contents (e.g. doi:10.2136/sssaj2014.09.0390; doi:10.2136/sssaj2018.05.0175; doi:10.2136/sssaj2011.0307), and they might be useful for comparison against results from this study. Also, what are the implications on IR models built for K, considering that the prediction likely relied on the indirect correlation between K and other soil components?
L95-97. Be clear about how frequently previous studies have utilized dimension reduction techniques to identify significant bands and/ or build IR models for K in soils. As far as I know such work has been done so ‘a lack of’ could be misleading if not explained well. The clarification is also key to bringing out the novelty of this study.
L198-199. Slightly awkward sentence.
L277. Add units for RMSE values.
L277-278. Fix the grammar error in this sentence.
L336-338. If possible, compare these results with previous studies and provide insights into model transferability. This study is carried out at a site level, so it’s important to discuss whether the selected bands for K could be useful for building models in a different region.
L370. “FTIR”.
L387-394. Discuss the implications of this finding carefully. If the performance of dimension reduction techniques differs among studies, then would the conclusion of SPA being superior of this study need to be interpreted with more caution? Discussion of model transferability/ domain would be helpful.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The manuscript regards an important issue focusing on forecasting the macroelements in the soil. The main problem of the study is to solve the problem of spectral reflectance selection to predict soil potassium content. Finally, the Authors used the PLSR method to establish the forecasting model, which relates the soil reflectance at selected spectra with the potassium content. Nevertheless, several elements require an explanation for better readability of the manuscript.
Section - Dimension reduction algorithms. The Authors used the PLSR method which decomposes both X and Y variables by iteratively finding latent variables (LVs), describing as much as possible of the covariance between X and Y. The results of this procedure itself enable the reduction of the dimension of full X variables. Please refer to this issue in the context of dimension reduction.
Lines 197-214 data partitioning section is hard to follow. I think it can be improved due to the model validation issue. Therefore this section should be clearly described. The division of the data set using the probability density function seems to be interesting but it could be better described by including the reasons for that.
Figure 3b presents only the average soil spectral reflectance in different soil available potassium grades. Why only average by groups?
Figure 6 is the context of the PLSR methods is not relevant because depending on the selection issue Authors used different numbers of the explanatory variables. The figure has only an informative issue.
Please refer to how many PCAs were taken for establishing the prediction models.
It can be interesting to make a prediction of potassium (using PLSR) for the full spectral range.
Name of section 2.3 please add s in the word algorithm(s)
Please add space 3.2 The results of different prediction models
Please change the letter in word limitation 4.3 limitation and uncertainty
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
