Assessing the Efficiency of Remote Sensing and Machine Learning Algorithms to Quantify Wheat Characteristics in the Nile Delta Region of Egypt
, Zaher Mundher Yaseen 6,7,8 and Salah Elsayed 9,*Round 1
Reviewer 1 Report
Title: Assessing the Efficiency of Remote Sensing and Machine Learning Algorithms to Quantify Wheat Characteristics in the Nile Delta Region of Egypt
Manuscript ID: agriculture-1585186
Overall, I thought the writing was clear. There were a few choppy sections, and I point out a few of the more obvious ones in my comments.
First, I would prefer that the results and the discussion be separated into 2 different sections.
Second, review the whole article. I have noticed a number of misspellings.
Lines 111 to 114: Sentence is too long and confusing. Rewrite this.
Line 118: Define ANN and RF. Then in the methods, justify the use of only these two? What about logistic regression or support vector machine that are both widely used? Or the parametric classification method Maximum Likelihood? Justify the choice of methods by comparing it with other known machine learning methods. You may have to read and cite this:
Song, X., Z. Duan, and X. Jiang. 2012. “Comparison of Artificial Neural Networks and Support Vector Machine Classifiers for Land Cover Classification in Northern China Using a SPOT-5 HRG Image.” International Journal of Remote Sensing 33 (10): 3301–3320. https://www.tandfonline.com/doi/abs/10.1080/01431161.2011.568531?journalCode=tres20
Or why not use an ensemble of models?
Salas EAL, Subburayalu SK, Slater B, Zhao K, Bhattacharya B, Tripathy R, Das A, Nigam R, Dave R, Parekh P (2019) Mapping crop types in fragmented arable landscapes using AVIRIS-NG imagery and limited field data. International Journal of Image and Data Fusion, 1-24. https://www.tandfonline.com/eprint/VQ672BMUPXDGU7V7PYIQ/full?target=10.1080/19479832.2019.1706646
Since parameterization is crucial, what are the parameterization of RF? Or the ANN? How did you adjust the parameters to optimize the classification performance?
Line 199: How was the spectra smoothed?
Line 218: What is meant by constructing manual validation?
Table 2: Again, explain why these Vis were selected. Most of these Vis are ratio-based. There are also VIs that are distance-based, but are not part of these list.
Where are the maps produced by each of the MLC method?
Table 4: What’s the reason that water has been misclassified?
Line 419: Change ‘was’ to ‘were’. Review of the rest of the paper for grammatical errors.
Explain the differences observed among values in Tables 6 and 7. For example, RDVI was the most sensitive for the in-situ, while NDVI for the Quickbird.
Also, r^2 values for QuickBird were comparably lower than the in-situ.
Conclusion: What are the real implications of these results? Knowing that QuickBird stopped taking images of the earth ~ 2015, could your results be applied using other satellite images? How about using medium-coarse resolution sats? All these need to be addressed and discussed to make your results useful.
Author Response
Reviewer 1
Response: We greatly appreciate your critical observations as well as your constructive and helpful comments. We hope that we could address your questions/comments by the explanations and revisions made in the manuscript.
Overall, I thought the writing was clear. There were a few choppy sections, and i point out a few of the more obvious ones in my comments.
Response: Thanks very much, we went through the whole manuscript to edit it and changed the correction you have recommended.
First, I would prefer that the results and the discussion be separated into 2 different sections.
Response: many thanks for comment. We write results and discussion together because we tested different objectives in this study so it was difficult to write them separately. These objectives are to (1) investigate the effects of better irrigation practices, water and salinity stress conditions on the four measured wheat characteristics (LAI, plant-h, AGB, and SPAD value); (2) having mapped different crops via remote sensing, quantify wheat characteristics through remotely sensed data across healthy, moisture and salinity stressful conditions; (3) evaluate the efficiency of different classification algorithms for mapping varying crops throughout the entire study site; (4) evaluate the efficiency of varying V-SRIs extracted from in situ spectroradiometry measurements and QuickBird imagery to estimate four measured wheat characteristics; and (5) assess the effectiveness of ANN and RF dependent on V-SRIs obtained from in situ spectroradiometry data and QuickBird images and their combination to detect different measured wheat characteristics of wheat. For example the aim of classifying images is totally different from the other objectives. Therefore, the results and discussion of each objective were combined and presented and discussed in details under different subtitles in the results and discussion section.
Second, review the whole article. I have noticed a number of misspellings.
Response: Thanks for your valuable comment. We went through the whole to correct mistakes.
Lines 111 to 114: Sentence is too long and confusing. Rewrite this.
Response: Thanks again. We have rewritten the sentence again to be shorter and clearer.
Line 118: Define ANN and RF. Then in the methods, justify the use of only these two?
Response: Thanks for your valuable comment. ANN and RF were defined in the introduction in line 125.
What about logistic regression or support vector machine that are both widely used? Or the parametric classification method Maximum Likelihood? Justify the choice of methods by comparing it with other known machine learning methods.
Response: Thanks for your valuable comment. The use of two models is clearly described in the materials section. The logistic regression and support vector machine are widely used but not the models proposed in this study. We chose the ANN and RF models because of their advantages in applying regression as described in lines 125-140 in the Introduction section. We have tested both support vector machine and partial least square regression for calibration and validation but both ANN and RF models performed better than them. For that we selected ANN and RF in this study
You may have to read and cite this:
Song, X., Z. Duan, and X. Jiang. 2012. “Comparison of Artificial Neural Networks and Support Vector Machine Classifiers for Land Cover Classification in Northern China Using a SPOT-5 HRG Image.” International Journal of Remote Sensing 33 (10): 3301–3320. https://www.tandfonline.com/doi/abs/10.1080/01431161.2011.568531?journalCode=tres20.
Salas EAL, Subburayalu SK, Slater B, Zhao K, Bhattacharya B, Tripathy R, Das A, Nigam R, Dave R, Parekh P (2019) Mapping crop types in fragmented arable landscapes using AVIRIS-NG imagery and limited field data. International Journal of Image and Data Fusion, 1-24. https://www.tandfonline.com/eprint/VQ672BMUPXDGU7V7PYIQ/full?target=10.1080/19479832.2019.1706646
Response: Thanks for your valuable comment. they was cited in the revised manuscript from line 565 to 568.
Or why not use an ensemble of models?
Response: Thanks for your valuable comment. In lines 149-152, there is limited evidence available to evaluate the ANN and RF approaches for predicting the LAI, plant-h, AGB, and SPAD value of wheat dependent upon a combined approach of vegetation-SRIs extracted from both in situ spectroradiometry dataset and QuickBird images.
Since parameterization is crucial, what are the parameterization of RF? Or the ANN? How did you adjust the parameters to optimize the classification performance?
Response: Thanks for your valuable comment. The ANN and RF models were optimized by selecting the best hyperparameters. The ANN parameters were number of neurons in two hidden layers, and activation function but the number of trees (ntree) and features (ntry) at every tree node for training RF model. Generally, the spectral indices were fed randomly to the model in the 1st loop, the low-level features were dropped during each loop, and the superb features were organized in an ascending order concerning the highest contribution. During looping, the best hyperparameters were picked up and the rest were excluded. Then, the ANN and RF outputs were compared to decide high-ranking variants and superior hyperparameters at minimum RMSEV that could improve the prediction.
Line 199: How was the spectra smoothed?
Response: Thanks for your valuable comment. The spectra was smoothed using the ASD software which was done for all datasets acquired by the instrument (spectroradiometer). It was written from line 204 to 205.
Line 218: What is meant by constructing manual validation?
Response: Thanks for your valuable comment. It means that during field work visits certain locations were identified as different classes (e.g. wheat, clover, bare soil and water bodies). These points were used in ENVI package using the QuickBird image as a validation dataset which was created manually in the software.
Table 2: Again, explain why these Vis were selected. Most of these Vis are ratio-based. There are also VIs that are distance-based, but are not part of these list.
Response: Thanks for your valuable comment. The explanation was written from line 255 to line 259.
Where are the maps produced by each of the MLC method?
The maps were produced in the school of Biological and Environmental Sciences, University of Stirling, United Kingdom.
Table 4: What’s the reason that water has been misclassified?
Response: Thanks for your valuable comment. The reason for that could be attributed to the reflected energy from some irrigated fields which have recently cultivated with other crops grown at early spring season such as squash which is sown at this time of year.
Line 419: Change ‘was’ to ‘were’. Review of the rest of the paper for grammatical errors.
Response: Too much appreciated, you are right as the word refers to plural not singular was has been changed to were in the revised manuscript.
Explain the differences observed among values in Tables 6 and 7. For example, RDVI was the most sensitive for the in-situ, while NDVI for the Quickbird. Also, r^2 values for QuickBird were comparably lower than the in-situ.
Response: We should consider the time difference between collecting spectra and satellite acquision time. When targeting a region to have an image, it is easy when using airborne but in case of satellite you do not have a guarantee to have images. We failed many times to have a targeted satellite image on time and alternatively we searched for the closest to in situ spectra. Another reason is that in situ spectra are collected at nadir position while satellite images are not. Lastly we can rely that difference to the accuracy of both instruments. Spectroradiometers have an accuracy of 1 nm while QuickBird satellite has a resolution of 2m which can affect the derivation of different indices. The explanation was written in revised manuscript from line 512 to 516.
Conclusion: What are the real implications of these results? Knowing that QuickBird stopped taking images of the earth ~ 2015, could your results be applied using other satellite images? How about using medium-coarse resolution sats? All these need to be addressed and discussed to make your results useful.
Response: Yes you are right QuickBird satellite stopped taking images of the Earth in 2015 but it is similar to today's satellites in terms of spatial resolution (2 m resolution). Like many other newly working satellites such as worldview 2 and 3 (1.24 and 2.4 multispectral resolution) there is no big difference between QuickBird satellite and newly launched ones in terms of resolution so QuickBird still useful in comparison to up to date satellites. Regarding the use of medium and coarser resolution satellites, it will be restricted by the small-size field system in the Nile Delta and Valley. As a result of cultivated land fragmentation, most fields in these regions have a width of less than 10 m which will limit medium and coarser resolution satellite (Sentinel 1 and 2). Therefore the detection of stress in these regions needs high resolution satellites which are very costly to have time series satellite imagery.
We hope that our revisions meet with your approval.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The first problem of this manuscript is its poor English writing. Deep English editing is required for the possible publication of the manuscript.
The second issue: both In situ spectroradiometry measurements and remote sensing imagery acquisition are related to “2007”. Do you think that these results are applicable for the current situation that we faced with? There is day to day environmental changes. How can you generalize your models and your obtained results with the present conditions?
Other comments:
Abstract:
Line 22: …crop growth performance, by …
Line 27: wheat crop: Which kind of wheat?
Line 28: SPAD value: Please write the full spell of SPAD and then abbreviation
Line 28: healthy: What is your mean from healthy? Is it normal (non-stress) conditions?
Line 28: water: Your mean is “Drought” stress?
Line 32: Was assessed to ..
Lines 35-36: results demonstrated…were found to be…. Please rewrite this sentence.
Line 46: … QuickBird, and in situ spectroradiometry measurements …
Line 47: health: Non-stress and stress (drought and salinity) conditions …
Line 47: of them: of who?
Keywords: Keywords are not in alphabet order. In addition, “ANN and RF” is not a one keyword.
Introduction
Line 56: …supplementing supply??
Line 60: is and?
Line 61: Please delete “agricultural”
Line 68: “stressors”. “Stress” is enough
Line 69: are considered “as” major …
Lone 73: to maximise to maximize (Repeated).
Line 97: ….field, which …
Line 97: What is your mean from “site-specific management”?
Line 100: … the prediction of yield …
Line 118: ANN, and RF: Please first spell full-name then abbreviation.
Line 120: … ANNs have… as a…?
Line 123: As a result of their …
M&M:
Line 234: What is “4.” Before the formula?
Results:
Classifying Wheat and Other Crops: Please specify the other crops. Which crops?I only see the one crop “Clover”
Table 8: What is the “F” parameter?
Author Response
Reviewer 2
Response: We greatly appreciate your critical observations as well as your constructive and helpful comments. We hope that we could address your questions/comments by the explanations and revisions made in the manuscript.
The first problem of this manuscript is its poor English writing. Deep English editing is required for the possible publication of the manuscript.
Response: Many thanks for this comment. English writing was improved as you suggested in whole manuscript.
The second issue: both In situ spectroradiometry measurements and remote sensing imagery acquisition are related to “2007”. Do you think that these results are applicable for the current situation that we faced with? There is day to day environmental changes. How can you generalize your models and your obtained results with the present conditions?
Spectra measurements in 2007 were collected using a spectrometer of 1 nm resolution which still advanced. Yes I think we can still benefit from these results when using new statistical techniques such as multivariate models and machine learning algorithms. Regarding the environmental changes, these changes do not affect the accuracy of in situ measurements when choosing the optimum time (especially free-cloud days) for collecting spectra. Satellite images can also be acquired over times of clear conditions which are available most of the year in regions like the Nile Delta and Valley. Again QuickBird is still comparable to many satellites working today.
Other comments:
Abstract:
Line 22: …crop growth performance, by …
Thanks for your valuable comment. It has been changed as recommended in line 22.
Line 27: wheat crop: Which kind of wheat?
Thanks for your valuable comment. The kind of wheat is Sakha 6 and it was added in line 27.
Line 28: SPAD value: Please write the full spell of SPAD and then abbreviation
Response: Many thanks for this comment. We mentioned the full term Soil Plant Analysis Development (line 28) for the abbreviation SPAD
Line 28: healthy: What is your mean from healthy? Is it normal (non-stress) conditions?
Response: Thanks for your comment. Yes, healthy refers to non-stress condition or under well irrigated. It was modified in line 29.
Line 28: water: Your mean is “Drought” stress?
Response: Thanks for your comment. It has been changed to drought stress instead of water stress in the revised manuscript.
Line 32: Was assessed to ..
Response: Thanks for your comment. We changed were to was in line 32.
Lines 35-36: results demonstrated…were found to be…. Please rewrite this sentence.
Response: Thanks for your comment. As recommended, the sentence has been rewritten in the revised manuscript in lines 37-38.
Line 46: … QuickBird, and in situ spectroradiometry measurements …
Response: Thanks for your valuable comment. It was modified in line 48.
Line 47: health: Non-stress and stress (drought and salinity) conditions …
Response: Thanks for your valuable comment. It was modified in lines 49-50.
Line 47: of them: of who?
Response: Thanks for your revision. The sentence has been changed in the revised manuscript in line 50.
Keywords: Keywords are not in alphabet order. In addition, “ANN and RF” is not a one keyword.
Response: Thanks for your valuable comment. It was modified in lines 52-53.
Introduction
Line 56: …supplementing supply??
Response: Thanks for your valuable comment. It means it can be a resource of low quality water (e.g. agricultural drainage water) which sometimes be used to avoid sever stress on crops until having the chance to irrigate using normal canal water.
Line 60: is and?
Response: Thanks very much for your comment. It was modified in line 63.
Line 61: Please delete “agricultural”
Response: Thanks much. The word agriculture has been deleted as recommended in line 65.
Line 68: “stressors”. “Stress” is enough
Response: Thanks very much for your comment. It was modified in line 73.
Line 69: are considered “as” major …
Response: Thanks very much for your comment. It has been added in line 74.
LIne 73: to maximise to maximize (Repeated).
Response: Thanks very much for your comment. To maximize has been deleted in line 78.
Line 97: ….field, which …
Response: Thanks very much for your comment. A comma has been added to the text in line 103.
Line 97: What is your mean from “site-specific management”?
Response: Thanks very much for your comment. The sentence was modified to be clear for the reader in line 103.
Line 100: … the prediction of yield …
Response: Thanks very much for your comment. It was modified in line 106.
Line 118: ANN, and RF: Please first spell full-name then abbreviation.
Response: Your comment is too much appreciated. The full terms of both abbreviations have been added in lines 127-128.
Line 120: … ANNs have… as a…?
Response: Thanks very much for your comment. The sentence was modified to be clear for the reader in lines 128-130.
Line 123: As a result of their …
Response: Thanks; it has been changed as recommended to as a result of their in line 133.
M&M:
Line 234: What is “4.” Before the formula?
Response: Thanks very much for your comment. The number 4 is not included in the equation. It was deleted from line 245.
Results:
Classifying Wheat and Other Crops: Please specify the other crops. Which crops?I only see the one crop “Clover”
Response: Yes you are right it is just one more crop (clover). The sentence has been changed to classifying wheat and other classes. In Egypt there are two main crops grown in the winter mainly wheat and clover. I did mean different classes included in the obtained image (QuickBird).
Table 8: What is the “F” parameter?
F indicates the number of spectral indices with ground and aerial images, and the fusion of all features. It was modified to group in the tables.
We hope that our revisions meet with your approval.
Author Response File: Author Response.docx
Reviewer 3 Report
Dear Authors,
The presented article presents an interesting approach to monitoring strategic agricultural crops in Egypt in terms of crop growth performance using innovative digital precision agriculture tools. Unfortunately, I believe that the presentation of one-year data in the form of an article is insufficient for the final publication of the manuscript. You are probably aware that in decent agronomic research we present a minimum of 3 agronomic seasons. In modelling agricultural data it is best to use 4 seasons - 3 to build models and 1 to verify their correctness.
I suggest that the work be retracted and submitted as a communication. Besides, the paper contains described research results, but discussion of the results was carried out in a very truncated way. You refer to a very small number of scientific publications on similar topics. The paper should be improved - additionally DOI references should be added to the references section, a detailed description of the modeling methodology should be provided, which variables were taken into account while building the models, etc.
Author Response
Reviewer 3
Response: We greatly appreciate your critical observations as well as your constructive and helpful comments. We hope that we could address your questions/comments by the explanations and revisions made in the manuscript.
The presented article presents an interesting approach to monitoring strategic agricultural crops in Egypt in terms of crop growth performance using innovative digital precision agriculture tools. Unfortunately, I believe that the presentation of one-year data in the form of an article is insufficient for the final publication of the manuscript. You are probably aware that in decent agronomic research we present a minimum of 3 agronomic seasons. In modelling agricultural data it is best to use 4 seasons - 3 to build models and 1 to verify their correctness.
Response: Many thanks for this comment. We reconsider other two year’s datasets (in situ spectra datasets) to make sure that the modeling of the data can extrapolate the results of our research. We added the datasets of two seasons in 2014 and 2015 for in situ spectra datasets. The new data of measured parameters of two other year’s was added in Table 3. Also coefficient of determination for the association between various vegetation-SRIs obtained from in situ spectroradiometry and Egyptian wheat characteristics in two other year’s collected from the study site was added in table 6. As well as, the modified outcomes of calibration and validation models of ANN and RF after added the data of two years for the association between V-SRIs extracted from in situ spectrometry and satellite imagery and leaf area index, plant height, above ground biomass and SPAD value prior selecting the best features were added in Tables 8, 9, 10 and 11. Unfortunately, no other high resolution satellite images are accessible at the moment since there is a lack of fund to search images for the study area concurrent with all spectra datasets. You know it is too expensive to buy one image with less than 5m resolution. Another limitation to use other satellite images (medium and coarser resolution images;15-30m resolution) are not reliable to be used in Egypt's field system (field width is usually less than 10m) and therefore high interference from different crops which will be misleading. In Egypt, you can find three different crops grown in less than a hectare so there will mixed pixels from these crops that have different spectral signature.
It is now more reliable after using three-year datasets to extrapolate the results. The both models of ANN and RF of in situ spectroradiometry was tested using the two years datasets as calibration datasets (n = 72) and validation datasets (n = 36). The both models of ANN and RF of high resolution satellite images was tested using the two years datasets as calibration datasets (n = 24) and validation datasets (n = 12). As well as, the both models of ANN and RF of combining data from ANN and RF of in situ spectroradiometry and high resolution satellite images was tested using the two years datasets as calibration datasets (n = 96) and validation datasets (n = 48). These details were added in revised manuscript in lines 548 – 554.
You refer to a very small number of scientific publications on similar topics. The paper should be improved - additionally DOI references should be added to the references section,
Response: Many thanks for this comment. We went through the whole manuscript to amend this. The whole manuscript was improved and addition scientific publications were added. The DOI references are not a requirement of the MDPI according to guide for authors.
A detailed description of the modeling methodology should be provided, which variables were taken into account while building the models, etc.
Response: Many thanks for this comment. A detailed description of the SRIs, RF and ANN methodology is given in more details in the Materials and methods section.
We hope that our revisions meet with your approval.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Please start all keywords with Capital word.
Reviewer 3 Report
I accept.
