Spatial Prediction of Soil Continuous and Categorical Properties Using Deep Learning Approaches for Tamil Nadu, India

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript evaluated the potential of deep learning in mapping soil continuous and categorical properties at a regional scale. Using more than 2,000 soil observations and covariates selected by recursive feature elimination, authors well predicted soil continuous (pH and OC) and categorical variables (Order and Suborder) using 1 Dimensional Convolutional Neural Net (1D CNN). The investigated topic is interesting, while several aspects should be greatly improved: (1) the soil formation used in this study have not been well described; (2) some methodology is confusing; (3) benchmark model should be compared; (4) results and discussion should be improved. Therefore, a major revision is suggested.

Specific comments:

Line 19: 27,098 Nos. of top soil profile observations. This is not clear, a mixture with topsoil and profile.

Keywords: Please use full names rather than abbreviations.

Lines 48-50: Please support your statements with recent publications. Two review articles are suggested below.

https://doi.org/10.1016/j.geoderma.2021.115567

https://doi.org/10.1016/j.earscirev.2020.103359

Line 59: CNN has not been defined before.

Line 68: radar not RADAR.

Line 76: Taghizadeh-Mehrjardi, et al. please delete the comma here.

Lines 84-87: The closing sentence to define the objectives of this study is missing. The relevant knowledge should be better described.

Line 93: km2 please use superscript for 2. Correct the similar typos.

Figure 1: Locational of the study area. Please also indicate what does figure 1a present? A satellite images? If so, for which date and relevant data sources (Landsat or Sentinel?)

Lines 115-117: What is time span of soil samples collection in this large database?

Line 119: It does not make sense to represent topsoil using 0-60 cm. Logically, 0-20 or 0-30 cm should be used.

Lines 124-125: What is the source of this 273 soil observations? What is the sampling design of this data?

Lines 249: R2 please use superscript for 2. Correct the similar typos. In addition, do not repeat the full names along with abbreviation once they have been defined (Line 250).

Table 3: Maybe also summarize the soil order and suborder here.

Lines 323-324: Why used the default parameters? Better to use cross-validation to optimized these parameters.

Figure 4: What is the soil property here? pH or OC? I expect two plots for two soil properties. For soil order and suborder, two plots are also expected.

Table 4: I suggest authors to add the performance Random forest as a benchmark model here. I seem that two model performed not well for pH and OC.

Lines 588: Limitations of this study and perspectives should be added.

Author Response

• Most of the specific comments demarcated by the reviewer such as the time span of the soil datasets, detailed sampling design, soil depth, and addition of limitations and perspectives have been addressed, among others.
• The summarization of the soil order and suborder such as area proportion statistics was not adhered as it would much complicate the content.
• The cross-validation procedure suggested by the reviewer was not implicated considering the larger number of training data implicated for the study and as an alternative the random holdback procedure was already implemented.
• A detailed explanation has been facilitated in the RFE part of the results and discussion on ranking based variable selection implemented and how all of the soil properties were integrated in the same figure no. 4
• Though the authors have suggested to implement the RF as the benchmark model for comparison, considering lengthening of the manuscript and to stay within the scope of the objectives, the author have added the ground truth images of the respective soil attributes as suggested by the other reviewers.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript introduced a deep learning method comparison study of soil classification. The network structure is quite common in machine learning fields and the comparison between methods are also limited. Reviewer appreciates authors’ hard working, this work is well organized and well written, can be accepted after making following major revision.

1.     CNN and MLP are often mixed in deep learning, there is limited meaning to compare the effect of CNN and MLP.

2.     If authors want to compare the effect of CNN and MLP, authors should list the accuracy, inference speed, model parameters. Like the accuracy change in line 533.

3.     For the figures 5-8, authors are suggested to add the ground truth image for better comparison.

Author Response

• The specified matter on regard of the mix up between the MLP and 1D CNN have been tried addressing the introduction part and since the model uses only the flattened or vector data for the classification and prediction, the numerical comparison of the inference speed will be negligible, but were added subjectively in the discussion part.
• The ground truth image as advised by the reviewer have been added.

Reviewer 3 Report

Comments and Suggestions for Authors

1. Please add 1-2 sentences at the end of the abstract and explain your results and conclusions briefly, and include quantifiable metrics and numbers if possible.
2. I am a bit confused by the phrase "an attempt is made ....". It is not clear what is the novelty and contribution of the paper, and how the outputs are significant and compliment the scientific literature. Please clarify this in both Abstract (briefly) and Introduction (comprehensively).
3. I don't think the metho used in Figure 2 is a deep learning method. Can you use the words Artificial Neural Networks instead? Please revise the text as well accordingly.
4. It seems that the testing data quantity is almost only one percent of training data samples, which is extremely low. I recommend that you keep at least 30% of the training data out as testing. Otherwise, the accuracy measurements could be misleading. You can do a a sensitivity analysis and see how much accuracy drops if you use 70% of training data, and test the model on the remaining 30%.
5. Please justify the choice of Recursive Feature Elimination and Permutation Feature Importance for covariate selection. Please also ;provide justification for selecting specific covariates and the potential impact of excluding others.
6.  The models used are standard in the field of machine/deep learning, and the paper does not introduce any significant advancements in model architecture, data handling, or interpretability techniques. If the novelty of the paper is in application or case study, the paper still does not explore how the proposed models might overcome the limitations of existing methods in a novel way. No compaartive analysis is performed to prove superiority of the proposed approach.
7. The R2 values are very low, and could indicate very low predictive accuracy. Can you explain it?
8. The prediction accuracy for soil orders and suborders is lower, but you do not provide deep insights or solutions to improve this performance.
8. I highly recommend showng pair maps of ground-truth data versus prediction for visual observation of accuracy.
9. The discussion section attempts to contextualize the results but falls short in providing a thorough critical analysis of the findings. There is a lack of comparison with existing studies or alternative approaches that could have been taken or suggested for future efforts.
10. Please make sure you paraphrase your text with similarity to your other papers published by MDPI. It seems that some text are similar to the other paper.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

• The revision of the abstract and the introduction part specified by the reviewer have been addressed.
• The DL-MLP model was used in the context as specified by the paper: https://link.springer.com/chapter/10.1007/978-3-030-31332-6_19 and a substantiating statement have also been added in this regard.
• The revisions as specified have been added and the dataset partitioning that was original proposed was training, validation, and test datasets. The validation dataset was initially removed as the authors felt it may hinder the readability of the paper. As advised by the reviewer the results of the validation datasets were also added.
• The justification on the utilization of RFE and PFI was added in the introduction part and the reason of excluding some covariates have also been addressed.
• The potential of the proposed model was then compared through the visual interpretation with the legacy soil information maps.
• The substantiations on the prediction accuracies of the continuous and categorical variables were added and the ground truth images for visual comparison was also facilitated.

Reviewer 4 Report

Comments and Suggestions for Authors

1. It is recommended to add the main focus and innovations of the study at the end of the first chapter to help readers better understand the uniqueness of the research.

2. In line 99, the latitude and longitude labels in Figure 1 are blurry. It is suggested to increase the font size and adjust all images with latitude and longitude grids to ensure clarity.

3. In section 2.3, the meteorological data is from 1970-2000, while the soil profile data is from 2023. It is recommended to explain the reason for choosing these datasets to help readers understand the rationale behind the time difference.

4. Lines 199-202 and 231-233 describe the hyperparameter settings of the two model platforms. It is suggested to combine these descriptions and provide a separate section for the hyperparameter settings of the experimental platforms. This will make it easier for readers to compare the hyperparameter settings of the two models. Additionally, please introduce the experimental platform and GPU model.

5. Lines 204 and 222 each contain an image with similar descriptions. It is suggested to provide appropriate captions for these images.

6. The table in line 237 describes the network design of the CNN model. Noticing the varying kernel sizes, it is recommended to explain whether this design is inspired by other papers or if there are other reasons. Additionally, it is suggested to conduct ablation experiments to verify the impact of this clever design of varying kernel sizes on model accuracy compared to conventional kernel sizes.

7. It is recommended to combine lines 269-272 into a single paragraph, similar to the description in lines 258-260.

8. In Figures 5, 6, 7, and 8, there are color discontinuities, such as horizontal or vertical lines. It is suggested to explain the reasons for these experimental results.

9. In section 4.2, lines 585-588 state: "In the categorical predictions provided by the DL-MLP, almost every covariate parameter contributed substantially. Still, in the case of 1D-CNN, most of the contributions were facilitated by the parent material covariates and the relief parameters, explaining the inclusion often origin and the topography for the taxonomical predictions [75]. The finer delineations provided by the DL-MLP can be explained by their equal contributions facilitated by all the covariate layers. From the overall contributions, irrespective of the soil attributes, Geomorphology, Carbonate Difference Ratio, Rainfall, MRRTF, MRVBF, LULC, Clay Difference Ratio, Landsat - Band 5, and Landsat - Band 4 are some of the covariates, which considered influential in predicting the soil attributes." It is recommended to provide complete experimental results to fully support this conclusion.

Author Response

• The comments of the reviewers on the introduction and the abstract were made.
• The latitude and longitude labels were changed as accordingly for all the images and if the issue persists, the authors will upload the high-quality images with legible labels at the production stage. As the comparison maps has to be clubbed together (Figure 5 to 8).
• The authors have revised the dynamic nature of the soil database maintained and the climate variable implemented will coincide the time span of the soil database. Considering the time component in the soil formation process, such inclusion of the time variable datasets was advocated for the study.
• The image descriptions were changed as pointed out and the nature of varying kernel size implementation and their reasons have been facilitated.
• The information of the experimental platform was added and though the reviewer has advised to change the hyperparameter description statements to a separate subheading the authors believe that it would affect the flow of the manuscript. If the reviewer still suggests them, the authors will make the necessary changes.
• The discontinuities in the final prediction maps and their reasons for occurrence was addressed.
• The required substantiation to support the influential parameters have been included in the supplementary materials.

Reviewer 5 Report

Comments and Suggestions for Authors

I have carefully reviewed the article with the ID "agronomy-3145787", which was submitted for publication in the special issue titled "Geospatial Artificial Intelligence (GeoAI) Applications in Agriculture for Smart Farming Solutions" in the "Precision and Digital Agriculture" section of the "Agronomy" journal.

This article focuses on the classification of categorical and continuous soil data with digital soil mapping (DSM) methodology and the prediction of categorical and continuous soil data with regression-based machine learning algorithms.

This approach is significant in terms of soil science. For example, EC values ​​in soil are obtained by laboratory analysis with numerical-continuous data. While this is modeled spatially with the regression problem, the classification of EC values ​​at certain thresholds and salinity phenomena (levels) are included in the modeling process with the classification problems.

The article applies this methodology well. However, it does not make any progress in terms of scientific innovation. In addition, "uncertainty" maps should be provided for each soil map produced as a result of current DSM studies. Because these maps can provide information to the end user and provide deep insights into the use of spatially open maps.

In line with these comments, the article, which does not contain any innovation and does not clearly state its contribution to scientific progress, should be restructured by taking into account the current developments of DSM and conducting uncertainty analyses. I have to recommend its rejection in its current form.

Best regards,

Author Response

• The reviewer advised for the implementation of a novel uncertainty assessment analysis for the implemented models. Most of the uncertainty assessment approaches incorporates the utilization of entire images/maps, besides generation of probabilistic intervals for the predictions.
• The current study did not include such implications, though a workaround method through deriving the probabilistic intervals from the residual maps (predicted and ground truth maps) can be possibly implemented. But considering the scope of the objectives proposed the methodology was not implicated.
• The current provided uncertainty analysis provides the appropriates insights on the prediction and were utilized in several of the studies for studying the efficiency of the DSM models.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised manuscript from the authors well addressed my concerns 

Author Response

• We express our sincere gratitude for your valuable time in improving the quality of the manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper's result is not great as accuracy is still low around 60%. Its network is also quite old. But for encouraging scholars to apply more neural network to computer vision task.

Author Response

• Considering the study as a benchmark work, the model architecture implemented can be further evolved with the ablation experiments. Though considered widely used, some of the amendments that were still made in the model implementation can be considered for evaluation in the future works. The same as been stated in the limitations and perspective section of the manuscript.
• Further, the study was concluded with the reasons and justifications for the lower measures of the evaluation metrics.

Reviewer 4 Report

Comments and Suggestions for Authors

This article has been improved and the amount of work done on this article is sufficient and innovative. The only thing that puzzles me at the moment is that deep learning algorithms have the advantage of dealing with raw data and work well when the amount of data is very large, so why did the authors use recursive feature elimination to filter the parameters? Why not input them all into the model? Please explain in detail and compare the accuracy before and after recursive feature elimination.

Author Response

• The reason for implementing RFE for the modeling process have been justified in the limitations and perspective section of the manuscript. Further, the inclusion of RFE have become a standardised workflow in many of the DSM studies, which is why the current study also investigated and reported RFE. Further, comparing the with and without nature of the RFE in the modeling process sounds promising, considering the scope of the manuscript and already lengthy portions, the authors have not added suggestion in the current manuscript.

Reviewer 5 Report

Comments and Suggestions for Authors

I reviewed the revision of the article.

It was found noteworthy to add section 4.2. Limitations and Persepectives in the discussion section.

However, there should be statements regarding the uncertainty findings, which are important deficiencies of the study.

In addition, current literature, especially studies comparing categorical and continuous data, studies including soil data such as salinity and soil texture should be discussed in the introduction and discussion sections.

If suggested corrections and additions are taken into account, it may be a timely contribution to the current literature and may be acceptable.

Author Response

• The additions stated by the reviewers have been added in the introduction part and the limitations and perspective section of the manuscript.