Estimation of Soil Heavy Metal Content Using Hyperspectral Data

Round 1

Reviewer 1 Report

In general, the manuscript contain a number of inaccuracies ("absolute reflectance "in L128; "field soil" preparation in laboratory in 2.3), conceptual errors ("field spectra"under halogen lamp, L124), methodological errors (spare treatment of spectra of dry powdered soil with destroyed texture  and spectra of moist soils with natural texture - probably because the authors did not provide a clear description of the procedure; no proper validation of models); important elements of the procedures have not been described (sampling protocol for Conghua area), is inconsistent (?? L221 and L226-228) and contain numerous errors in graphs descriptions ( Fig. 11 - Cd from a to d) and equations (ie Y instead of EXP Y) in eq. 5, 6, 8 and 9).

Regression models for estimating As, Cd and Hg in soil based on soil spectra were developed without proper validation. On a small collection of 65 soil samples, spectral measurements were carried out in laboratory conditions, measurements conditions were not described - surface formation, sensor height, diameter of field of view. Dry soil spectral measurements (DSSR) were made on samples grinded to the size of 0.84mm, however, the method of preparing samples designated as "field soil" (FSSR, and should be remarked as MSSR) was not described. DSSR of all soil samples were used for selection of variables and model calibration. There is no validation stage. In section 3.3, only R² for model calibration is shown (0.60, 0.91 and 0.65), without error measures (RMSE, RRMSE, etc). Models calibrated for DSSR (eq. 4-6) were used for the MSSR data of the same full data set. I can't agree that this is correct validation. The results presented in Table 3 suggest that substitution of MSSR data in place of DSSR in practice does not change anything, because R² in all cases is the same as for calibration (0.60, 0.91 and 0.65 - amazing?!). However, quantitative estimation error measures (RMSE and RRMSE) have been added, with no corresponding values for calibration, this is not the case. I am sorry, but this approach to modeling and evaluating results is pseudoscientific, the results of such investigations do not explain anything. Reading such investigations can't help anyone.

In Chapter 3.4, moist laboratory spectra (MSSR) of 65 samples were used for selecting spectral bands and model construction / calibration (equations 7-9). These models  were used (validated) for HJ-1a satellite data. However, in L322, the authors describe that to solve eq. 7-9 the dry soil spectra from HJ-1a were recalculated using eq. 3 with coefficients from table 4. The procedure described in the paper is not logical and does not explain anything.

The number of samples used to calibrate the models is insufficient. The authors did not assess the effectiveness of modeling (RRMSE) in relation to criterion used to assess the performance of predictive models (eg Conte et al. 1986, RRMSE <25% as reasonable accuracy)

SD Conte, HE Dunsmore, and VY Shen (1986), Software Engineering Metrics and Models, Benjamin / Cummings, Menlo Park CA.

The work is chaotic, the presentation is sloppy, and the obtained results do not give the recipient any new knowledge. In its present form, the work does not deserve publication in such a highly ranked scientific journal. After thorough rethinking, methodological corrections and elimination of errors, I propose to submit a new version of the manuscript in a magazine of lesser prestige.

Additional remarks

Graphics are insufficient - the drawings are illegible.

L94-97 - field soil sampling protocol is not described, which is important when the analytical results are related to a pixel with a size of 100x100m,

L125 - so called "field soil" preparation is not described, it is not field spectra measurement, but only the measurement of moist soil spectra under halogen lamp light,

L125 - no description of the method of scanned soil surface preparation and the height of the sensor, the diameter of the field of view,

L128 - no white reference panel description, own or certified, no description or measurements were converted to absolute white by CF of white reference,

L131 - window size for SG calculations,

L173 - introduced the abbreviation MZSF, probably incorrectly (MZSA?)

L202, 207, 215, L286-287 - The authors mistakenly use the term "field soil spectra", measurement under artificial lighting can't be treated as such, it is only "moist soil spectral reflectance" (MSSR) (grounded or not ?)

L207 - "discovered" is used, but the term "identified" should be used,

L221 - it ends with a colon, after which something is missing.

Fig. 8 - the drawing is illegible, descriptions in the legend should be in line with the explanations in the text L141-142, descriptions of the x axis above 2500nm are superfluous

Fig. 10 - the form of the Z-score presentation suggests that at> 0.75 spectral variables significant, earlier in the text (L165) it was suggested that the significance is determined by the Z-score of original features above the corresponding value from random forest.

eq. 7-9 - it makes no sense to present a determination coefficient (R2) with an accuracy of more than two decimal places

Author Response

Thank you very much for your comment. We improved the manuscript by addressing all your comments and suggestions.

Please see attachment for details, thank you very much.

Author Response File: Author Response.docx

Reviewer 2 Report

see attached document

Comments for author File: Comments.pdf

Author Response

Thank you very much for your comment. We improved the manuscript by addressing all your comments and suggestions.

Please see attachment for details, thank you very much.

Author Response File: Author Response.docx

Reviewer 3 Report

Summary:

The objective of the study was to acquire the optimal hyperspectral inversion model for the determination of heavy metals from dry soil spectral indexes and field measured soil heavy metal content. The dry soil spectral indexes were acquired form soil spectral data by building the relationship model among soil moisture contents, field soil spectral data and dry soil spectral data to eliminate the effect of soil moisture. The optimal hyperspectral inversion model proposed by the manuscript was verified by comparing the measured and estimated values of soil heavy metal content in the selected region of China using HuanJing-1A (HJ-1A) HyperSpectral Imager (HSI) data.The relationship models were built between the optimal relevant spectral variables and observation data on Cd, Hg and As contents. More over, it was also concluded that the proposed approach can be efficiently applied to detect soil heavy metal Cd content at the regional scale.

Specific comments:

1st and 4th keywords should be replaced with new why they just repeat phrases and words from the manuscript title.

in line 88, you provide the number of sample points 65. With the total are under observation it represents the one sample point on 2,765 square kilometres. Are you convinced it is sufficient?

in line 94, in regional scale you provide the number 33 samples. Do you think, in conjunction with previous comment that these resolutions are sufficient for modelling?

in lines 98-101 you are stating the needs of used data corrections. You need to justify this needs.

in lines 108-114 you provide good methodology. It is standardised procedures or recommended? Standard/Reference?

From the Table 1 it is apparent that the variation in the data is massive which is returning the issue to the 2nd and 3rd comments.

Devices used in any analyses needs to be defined in closer details. Device (Model, Version, Producer, Country ...).

in line 132,  procedure is recommended? Support with reference if applicable.

in line 139-146, Same as above comment. Recommended/standard/reference?

in equations 4 and 5 there are obvious error in subscript representing the spectral variables (wave lengths).

Not all of the observed issues and results were discussed properly. More importantly, in lines 378-391 you do hit the same problem with low sampling resolution.

Moreover, there are no single reference from Remote Sensing journal, which I believe is the reason why introduction and the discussion is at lower level of quality. Single skimming through the journal paper allow me to get a lot of data and ideas how this manuscript content can be improved. Authors may focus on it as well.

Final judgement:

Manuscript is of higher quality however its improvement is possible as well. Depending on the specific comments minor revision can be recommended. On the other hand I would suggest authors to invest even more time into manuscript texts quality improvement and come up with manuscript after major revision.

Author Response

Thank you very much for your comment. We improved the manuscript by addressing all your comments and suggestions.

Please see attachment for details, thank you very much.

Author Response File: Author Response.docx