Soil Salinity Prediction in an Arid Area Based on Long Time-Series Multispectral Imaging

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

in my opinion the work is interesting and well written. The research described focuses on the environmentally important problem of excessive soil salinity, which significantly affects vegetation and thus agricultural production.

 

The article was prepared in accordance with IMRAD principles, it can be seen that the Authors had an idea for this work, met the purpose of the work and demonstrated the high effectiveness of LM in assessing soil salinity.

 

Below I send some comments to be taken into account when preparing an update of the manuscript.

Please comment on why the study area has high soil salinity?

"The average temperature of data collection is 33 degrees Celsius." Isn't this a mistake?

Why were such months of surveying chosen?

Line 220 - The % after 70 was missing.

Why weren't cases separated for the test set?

Please add 5-10 papers in the results discussion section. Cite these papers in the text.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

General comment:

·       Please also calculate the Mean percent error and/or the normalized RMSE, these metrics represent errors in percentage so that a reader who is not familiar with SSC quantities can have a better understanding of your model accuracy

·       It is not clear to me if the spectral indices were taken over bare soil or soil covered with vegetation. If the former is correct then it makes sense to correlate the spectral data to the soil salinity. Nevertheless, since the spectral data is only relevant to the surface, please explain how the soil salinity samples that were collected below the surface can be correlated to the surface spectral data. If the latter is true (i.e., spectral data collected over vegetation) then how can the spectral indices (that represent vegetation) indicate soil salinity?

·       Why did you apply the model for each year? Why not use the data from 2021 and 2022 as a training set and the data from 2023 as a test set?

·       Can you provide a map with points representing the locations where the samples were taken?

·       You should test all predictors (e.g., bands, indices, features) for co-linearity. The way to do that is to calculate the correlation (r) between all of them. This way you’ll find predictors that have a high correlation (>0.9) with each other, meaning that some of them are redundant.

·       You have a small dataset for using ML algorithms. Please discuss that in the discussion. Also, please discuss how to scale it up to satellites, because using UAVs is limited compared to satellites.

 

 

 

 

Please see the specific comments below (L=lines):

• L14:  please add Northwest China after Heihe River Basin
• L47:  what is the meaning of hm²?

·       In lines 51-53 you write that Traditional methods of soil salinity measurements are not suitable for large-scale monitoring of soil 52 salinity in cultivated land areas. I agree, but UAVs also are not 100% ideal because they pose technical/financial/spatial limitations compared to satellites.

• L55-57: Compared with satellite remote sensing, UAVs can provide high-resolution remote sensing images with less expenditure, ensuring large-scale and rapid monitoring of soil moisture and crop growth. I’m not sure how UAVs have less expenditure than freely available satellite images. With UAVs you need to have a UAV, then physically go to the field, wait for a suitable time, calibrate the UAV, fly it over the field, change batteries if needed, tackle technical and logistics issues, then process a huge image file, whereas in satellite images you can have the image in a second using a simple API code. Also, the scale of satellite images is many times higher than the UAV. Please rephrase this sentence or explain it better.
• L67 – how is the nitrogen study related to soil salinity literature review
• L79 – this sentence must be better explained. You argue that the accuracy of soil salt models needs to be improved, but you have not provided any accuracy metrics achieved by other studies. As a reader, how would I know that the studies you mentioned achieved inferior results that need to be improved? Please also try to think of other merits of your study compared to others.
• L95: what do you mean by highlighting good prediction accuracies?
• L108: Glcm should be GLCM
• L104-119: half of this text should be in the materials and methods section.
• Table1 and line 137 – please change the order of the spectral band from the shortest wavelength to the longest.
• Section 2.2.1- what was the vegetation stage./crop cover percentage when the images were taken? Was it taken over bare soil? Maybe report the min/max NDVI so that the reader can have a better understanding. This information is needed to understand if the images sensed the soil or the vegetation
• L163 – can you give an example of how you replaced the red and NIR with red edge? It seems that you used indices that include both (e.g., DVI-reg)
• Figure 3 is difficult to understand. What are the units? What is the scale? Maybe you can remove this image as it does not add relevant information to the text.
• L205: please explain the initials of CART and DART
• Section 2.4.1: you mention 3 models (BPNN, Rf, and XGB) but you only elaborate on one of them.
• Figure 6. It should be a bar chart and not lines. Also please add labels to the y-axis

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Referencing the author's responses:

·       Response 1: I read your answer but I still think you should calculate an additional metric with percentage units.

·       Response 2: please add this clarification and explanation to the text. The reader needs to understand what the soil samples represent (surface, subsurface, etc) and what the spectral measurements represent (bare soil, mixed pixel of soil and vegetation) with respect to the soil salinity.

·       Response 3: My suggestion here was to mimic a real-case scenario in which you do not have the current year’s data, so you cannot use it as part of the training set. In any case, following your response, please explain in the text in a coherent way which model and variable setting you find the best, meaning that from your perspective you can apply this model to a new dataset. Also, please add variable importance results for the best model.

·       Response 4: Thank you for the additional figure. Please refine it by adding a scale bar, and a north arrow as well as mentioning the image source (is it the drone/Google Maps?) and date. Also please remove the +- sign (upper right corner).

·       Response 5: I think you misinterpreted my Response. What I meant was to calculate collinearity between all input variables. In Figure 3 collinearity was calculated only for some of the input variables, and the text only refers to the correlation with the observed soil salt contents. For example, according to Figure 3, there is a high correlation between the Green and the Blue bands. This begs the question – why use both? The same goes for all other input variables. Not accounting for the collinearity between the input variables is a drawback of the methodology and thus the model.

·       Response 7: please add the word “China” to line 14

·       Response 8: a hectare is already a unit of area, so there is no need to add m2.

·       Response 12: I meant adding an accuracy metric (numbers) to studies that you cite and predicted soil salinity. For example, Zhu [13] combined UAV-based hyperspectral visible and near-infrared spectroscopy with two feature selection techniques to predict soil salinity, and achieved an RMSE of / R2 / MAE…

·       Response 18: The text reads: replace the red and near-infrared bands with a newly introduced red edge band. But you mean to replace the red or the near-infrared…. Because replacing them both in the same index does not make sense. Also, the red-edge band is not new, so I would remove any such reference.

·       Response 21: The general audience reading such papers is not necessarily soil salinity or ML experts. You need to make your text understandable to a general audience that is interested in the scope of this journal. As so, you need to provide even basic explanations for things that might be obvious to you. Please add 2-3 sentences describing each model that you use, and explain that BPNN and RF are already very common in current soil salinity prediction research whereas the XGBoost algorithm is currently not common in soil salinity prediction research.

Author Response

Please see the attachment

Author Response File: Author Response.pdf