Bio-Inspired Hybridization of Artificial Neural Networks: An Application for Mapping the Spatial Distribution of Soil Texture Fractions

Round 1

Reviewer 1 Report

This is an interesting topic manuscript for publishing at this journal, also, it has written very well. I suggest accept it by a minor revise:

The introduction section can be updated using newly published papers about using bio-inspired in engineering, like below papers:

https://doi.org/10.3390/app11031286

https://doi.org/10.3390/w12113015

https://doi.org/10.3390/rs13010132

The innovation of the paper should be further characterized.

In the discussion part, it will be better if the results will be compared by other previous researches

Provide one graphical abstract for better understanding.

Author Response

Reviewer 1

Q4: This is an interesting topic manuscript for publishing at this journal, also, it has written very well. I suggest accept it by a minor revise:

Thank you for your very valuable and positive feedback.

Q5: The introduction section can be updated using newly published papers about using bio-inspired in engineering, like below papers: ttps://doi.org/10.3390/app11031286;https://doi.org/10.3390/w12113015;https://doi.org/10.3390/rs13010132

The recently published papers were added in the introduction section.

Q6: The innovation of the paper should be further characterized.

Research into the integration of these hybridized approaches and their associated benefits have yet to be reported in soil science and digital soil mapping. Furthermore, the ANN approach is particularly beneficial in handling compositional data due to the softmax function and whereby the predictions of sand, silt, and clay must sum to 100%. Lastly, there are few studies that have mapped soil PSFs at the provincial-scale using a hybridized ANN approach and a large number of environmental covariates. These points were reinforced throughout the paper.

Q7: In the discussion part, it will be better if the results will be compared by other previous researches

The result and discussion are presented in two separate sections. The comparative examples for prediction of PSFs at large-scales were added in Section 5.1

Q8: Provide one graphical abstract for better understanding.

Graphical abstract is added.

Reviewer 2 Report

This study achieved a spatial prediction of soil texture fractions using a series of hybridized artificial neural networks (ANN) models, 1595 soil samples, and 64 environmental covariates. The authors embraced a very important topic and the study was well-designed. However, I am not sure if the manuscript is within the scope of Remote Sensing because this study only used some spectral indices from remote sensing. I also concerned the novelty and applicability of the method.

1. As mentioned by the authors, one model is not universally better than others. How did you support the used method in this study that is optimal? 
2. Additional explanations and comparison with other study areas are required for the contribution of environmental covariates.
3. All the variables were converted into data with a spatial resolution of 30 m. But how did you consider the conversion from 1 km to 30 m?

Other comments read as below.

Abstract:

Line 24, most physical ….of soil；property?

Materials and Methods:

The quality of figure 1 is low.

Line 146: what are the digitized polygon maps?

More descriptions are necessary for the different data source such as the times of climatic datasets.

I think the description of different ANN algorithm could be simplified.

Author Response

Reviewer 2

Q9: This study achieved a spatial prediction of soil texture fractions using a series of hybridized artificial neural networks (ANN) models, 1595 soil samples, and 64 environmental covariates. The authors embraced a very important topic and the study was well-designed. However, I am not sure if the manuscript is within the scope of Remote Sensing because this study only used some spectral indices from remote sensing. I also concerned the novelty and applicability of the method.

In our study we used 46 satellite-derived datasets as covariates taken from three different sources of optical and radiometric satellite data (Landsat 8, Sentinel 2, and MODIS). They were chosen because it is well known that spectral signals from minerals (which represent soil texture indirectly by their typical weathering behaviour and source rock, for example more quartz in the sand fraction and more biotite, illite and chlorite in the clay fraction) are well reflected by these datasets. If we are able to extract high-quality information on soil texture and quantify the particle size distribution of a soil from space, this is a big step forward and can greatly advance our knowledge about soils and their geographical distribution, for example in remote and less accessible areas like desserts or high-mountain regions where we are still short in good-quality information about soil properties. In Figure 4 you can see that almost half of all environmental covariates to predict soil texture using the MBO-ANN algorithm origin from remote sensing data. We think that this is a great success for remote sensing and worth to be published to its audience. Even if looking at table 5, which shows that remote sensing-based covariates are outperformed by categorical and climate data in predicting the soil´s particle size fractions in this study, it is written that this is mainly due to the limited spatial resolution of MODIS, and the great importance of Sentinel-2 datasets (which are already moderately important) underpins the great potential of spatially high-resolution remote sensing data of soil science.

Moreover, 10-fold cross validation of each prediction model is calibrated ten times guaranteeing each data point was used as validation at least once. The uncertainty analysis, to some-extent, showed the applicability of proposed method as well.

Furthermore, research into the integration of these hybridized approaches and their associated benefits have yet to be reported in soil science and digital soil mapping. Furthermore, the ANN approach is particularly beneficial in handling compositional data due to the softmax function and whereby the predictions of sand, silt, and clay must sum to 100%. Lastly, there are few studies that have mapped soil PSFs at the provincial-scale using a hybridized ANN approach and a large number of environmental covariates. These points were reinforced throughout the paper.

Q10: As mentioned by the authors, one model is not universally better than others. How did you support the used method in this study that is optimal?

We acknowledge the inconsistency; however, the objective of this work was to examine the use of different types of optimization techniques and couple it with ANN. Here, our model comparison and selection was based on our comparison of a variety of hybridized models and we selected the optimal hybridization based on a repeated 10-fold cross validation procedure.

Future work could involve the use of these optimization techniques with other machine learners; however, that is beyond the scope of this study. Furthermore, we deemed to ANN model to be appropriate based on its intrinsic ability to incorporate a softmax transfer function, which is important when handling compositional data. That being said, other areas of research could involve the use of other types of log-ratio transformations on other machine-learners; however, that is also beyond the scope of this work.

Q11: Additional explanations and comparison with other study areas are required for the contribution of environmental covariates.

It was added in the text.

Q12: All the variables were converted into data with a spatial resolution of 30 m. But how did you consider the conversion from 1 km to 30 m?

Great point. All environmental covariates were aggregated (average resampling) or disaggregated (bilinear resampling) to a common grid of 30 × 30 m spatial resolution and rescaled using Z-score standardization.

Q13: Line 24, most physical ….of soil；property?

It was corrected

Q14: The quality of figure 1 is low.

Figure 1 is improved.

Q15: Line 146: what are the digitized polygon maps?

In table 1A the digitized polygon maps were addressed. In this study three polygon maps including the geology, soil type and physiographic unit maps.

Q16: More descriptions are necessary for the different data source such as the times of climatic datasets.

Great point. It was corrected. Data representing the mean annual rainfall (mm) and mean annual temperature (°C) data for the period 1970-2000 with spatial resolutions of 1000 meter were acquired from the WorldClim data repository to represent the climatic factors that control the soil forming process for the region

Q17: I think the description of different ANN algorithm could be simplified.

Many thanks for your comment. We revised accordingly. The proposed ANN-based algorithms which are among the most efficient methods are: genetic algorithm (GA-ANN), particle swarm optimization (PSO-ANN), bat (Bat-ANN), and monarch butterfly optimization (MBO-ANN) algorithms. The training for ANN had been performed using four different evolutionary algorithms. As you advised we have summarised the descriptions and presented them more briefly. We highlighted the revisions and hope you would agree with the updated version and it can meet your expectation. 

Reviewer 3 Report

Review summary

Taghizadeh-Mehrjardi et al., submitted the article "Bio-inspired hybridization of artificial neural networks to im-prove the spatial prediction of soil texture fractions" for review. The article approaches an important topic providing technical ANN solutions that are tested on a case study and which produce a gridded map of soil texture on the Mazandaran province, Iran. It has also practical impact potentially for agriculture practice but also to land surface modeling as the results could improve the accuracy of infiltration/runoff variables to a certain extent depending on the quality of data currently used in such models.

The article develops ANNs methods with sufficient illustration on the method. Furthemore, it proposes a deep investigation of 64 potential covariates to explain the particle size fractions (PSFs). By so, the authors are able to identify which variable contributes the most to PSF. The authors support their finding with convincing quality Figures and rigorous methodology.

The article only presents minor corrections such as the consistency in article citation,Figure axis resition, equations and discussions of the limitation of the study It overall an interesting research study with on top practical and useful results.

Questions

#1

If I understand clearly you implemented the bio-inspired metaheuristic optimization algorithms for this specific application (It is not fully clear with abstract line 29 ” were built for predicting PSFswere built for predicting PSFs”). You may highlights in the title such as, just a compromise between the theoretical development and it application, i.e. something like

Bio-inspired hybridization of artificial neural networks: An application to soil texture fraction spatial distribution, Mazandaran province, Iran

Line 215, it is mentioned "were used to train"

-->

were implemented to train ??

Maybe clary a unique term of each of the sentence that reflect the best your work

#2

In your abstract you compare the current ANN with a backpropagation training algorithm (BP-ANN).

Is the backpropagation training algorithm a good reference comparaison ? compared to recent ANN development ? I am sure it is still the favorite flavor of ANN developer ...

It is again mentioned dLine 407 BP-ANN (i.e. traditional implementation of ANN). Would you still argue that ANN implementation remains BP-ANN even if its limitation is well known in several areas. However, it is still valid to use it as reference comparison...

#3

Looking at the unequal distribution of the soil sample (Figure 1) mainly in the North and North-East and the impact of the 2 highly sampled yellow zones. Would it be possible that the ANN model is overfitting with silt detection or let's say better at detecting silt. You may have to discuss the impact of the sparse sampling point in the Loam/Sandy Loam area , could that potentially impact/limit your ANN model in any case ?

#5

An interesting conclusion is that the highest relevant covariates are found in the categorical class which heavily depend on in-situ sampling and field work. This highlights the use of direct observations against satellite and climatic data ...It is further reinforced by Line 547 Overall, RS-derived covariates were not strongly relevant which may question this paper to be in remote sensing ... geology also uses rs .... You may present that point ... some climate data are also derived (cloud cover, rainfall, ground temperature )... that may open the discussion of the auxiliary data classification.

Abstract

Line 37-39.

This study demonstrates the effectiveness of bio-inspired algorithms for im-

proving ANN models ...

That a strong point in a remote sensing- NN point of view. It also confirms that the main parameters controlling the soil texture are the soil types, geology, maps, rainfall, temperature. That would be an interesting remark in soil physics.

Figures

Figure 1

Add a,b,c on your Figures

Adapt caption to it.

Figure 1.Spatial distribution of soil sampling sites for the Mazandaran province of northern Iran.

--> i.e.

Figure 1.a. Localization of Iran b. Localization of the Mazandaran province in Iran

c. Spatial distribution of soil sampling sites for the Mazandaran province of 129northern Iran.

Precise exactly the difference between yellow rectangular and yellow point.

--> ?? Line 138 croplands and orchards using gridded sample design with 2 km × 2 km

grid spacing ?? Is the significance of the yellow rectangular ?

Seems that yellow points are the distribution of soil sampling then what are the rectangular for

Please increase the font of the lat lon on the Iran and world map at least that can be read.

Figure 2

Figure is cut short on the top, it cutting the start box ...

Figure 4

Try to increase size/resolution of the legend (pixel effect), there still enough space,

Why do you call predictors family auxiliary data. To my understanding, it is more data that share common characteristics. Auxiliary meaning some extra data or data of 2nd order ... data class or data types may be more easy to understand.

Equations

Line 268

Add Equation numbers

Line 284-290

Add Equation numbers

Line 349-352

Add Equation numbers

Tables

Table 1

are Hyper-parameters non dimensional ? Please add their units if not.

Table 5

keep same parameter in caption

a

-->

W

b

-->

Sl

..

There is some difference between the covariables list and the auxiliary data of Figure 4. It seems the same to me, you may want to use the same classification and call covariables as class data or auxiliary data for consistency ....

You may also try to keep the same order as the legend Figure 4 or vis versa for easy reading.

Specific comments

Line 116-117

The predominant soil temperature regime class is classified as being thermic followed by mesic and cryic

What is the meaning of a temperature regime class ? Add a reference or a definition.

Could you please decide a citation format. Better to have all in [XX] and not mixed citations [XX] and authors name [XX].

Lines 253-257

Line 275

304

374

397

508

511

524

555

561

563

564

568

Line 392 ??

variations

-->

variability

Line 596

for decision makers in order to inform agricultural practices

--> ??

for decision makers in order to improve agricultural practices

Line 505

The relationship between these datasets and PSFs was apparent in this

study; for example, the spatial distribution of alluvial plains and flood basins were two

physiographic units of the northeastern region of the study areas and which also coincides

with the high clay contents that was observed.

-->

most likely due to the river erosion/transport process ...

Line 619

It can also potentially improve land surface modelling.

Author Response

Reviewer 3

Q18: Taghizadeh-Mehrjardi et al., submitted the article "Bio-inspired hybridization of artificial neural networks to im-prove the spatial prediction of soil texture fractions" for review. The article approaches an important topic providing technical ANN solutions that are tested on a case study and which produce a gridded map of soil texture on the Mazandaran province, Iran. It has also practical impact potentially for agriculture practice but also to land surface modeling as the results could improve the accuracy of infiltration/runoff variables to a certain extent depending on the quality of data currently used in such models.

Thank you for your very valuable and positive feedback.

Q19: The article develops ANNs methods with sufficient illustration on the method. Furthermore, it proposes a deep investigation of 64 potential covariates to explain the particle size fractions (PSFs). By so, the authors are able to identify which variable contributes the most to PSF. The authors support their finding with convincing quality Figures and rigorous methodology.

Thank you for your very valuable and positive feedback.

Q20: The article only presents minor corrections such as the consistency in article citation, Figure axis resition, equations and discussions of the limitation of the study It overall an interesting research study with on top practical and useful results.

Figures, equations, and discussion are improved.

Q21: If I understand clearly you implemented the bio-inspired metaheuristic optimization algorithms for this specific application (It is not fully clear with abstract line 29 ” were built for predicting PSFs were built for predicting PSFs”). You may highlights in the title such as, just a compromise between the theoretical development and it application, i.e. something like Bio-inspired hybridization of artificial neural networks: An application to soil texture fraction spatial distribution, Mazandaran province, Iran Line 215, it is mentioned "were used to train" were implemented to train ?? Maybe clary a unique term of each of the sentence that reflect the best your work

That's correct. The sentence and title are corrected.

Q22: In your abstract you compare the current ANN with a backpropagation training algorithm (BP-ANN).

Is the backpropagation training algorithm a good reference comparaison ? compared to recent ANN development ? I am sure it is still the favorite flavor of ANN developer ...

It is again mentioned dLine 407 BP-ANN (i.e. traditional implementation of ANN). Would you still argue that ANN implementation remains BP-ANN even if its limitation is well known in several areas. However, it is still valid to use it as reference comparison...

It is a good reference for comparison due to its common usage in most literature/research. As you mentioned the backpropagation algorithm is the most used training algorithm but not the best. Due to the shortcomings of this algorithm including local minima, convergence rate, and tendency to overfit the data we proposed alternative training algorithms, which were compared. Given the effectiveness of the alternative approaches from our results, it would be worthwhile for users of ANN models to give greater consideration the hybridizations that we tested.

Q23: Looking at the unequal distribution of the soil sample (Figure 1) mainly in the North and North-East and the impact of the 2 highly sampled yellow zones. Would it be possible that the ANN model is overfitting with silt detection or let's say better at detecting silt. You may have to discuss the impact of the sparse sampling point in the Loam/Sandy Loam area, could that potentially impact/limit your ANN model in any case ?

The soil data in this study came from different sources, that why two highly sampled yellow zones were out-looked. Due to the repeated nested 10-fold-cross validation we do not expect our accuracy metrics to be influenced by the overfitting effect. By the way, the Loam/Sandy Loam area (Figure 5), did not coincided with the area with highly dense sampled. The silt content is highly related to the loess-derived soil that is discussed in the paper.

Q24: An interesting conclusion is that the highest relevant covariates are found in the categorical class which heavily depend on in-situ sampling and field work. This highlights the use of direct observations against satellite and climatic data ...It is further reinforced by Line 547 Overall, RS-derived covariates were not strongly relevant which may question this paper to be in remote sensing ... geology also uses rs .... You may present that point ... some climate data are also derived (cloud cover, rainfall, ground temperature )... that may open the discussion of the auxiliary data classification.

In contrary with many studies, this study demonstrated that the RS-derived covariates were not strongly relevant to the PSFs prediction. As you may know, the PSFs is a soil characteristic with static behavior. In the other word, the soil texture cannot change in a short time and it cannot change easily during a year that the spectral data was acquired.  It is highly influenced by the parent material, geology and soil units, that is why the weights of these covariates for PSFs prediction are higher than other DEM and RS derived covariates. 

Q25: Line 37-39. This study demonstrates the effectiveness of bio-inspired algorithms for improving ANN models ...That a strong point in a remote sensing- NN point of view. It also confirms that the main parameters controlling the soil texture are the soil types, geology, maps, rainfall, temperature. That would be an interesting remark in soil physics.

Thank you for your very valuable and positive feedback.

Q26: Figure 1: Add a,b,c on your Figures Adapt caption to it. Figure 1.Spatial distribution of soil sampling sites for the Mazandaran province of northern Iran. Figure 1.a. Localization of Iran b. Localization of the Mazandaran province in Iran c. Spatial distribution of soil sampling sites for the Mazandaran province of 129 northern Iran. Precise exactly the difference between yellow rectangular and yellow point.

Thank you. The figure 1 is improved.

Q27: Line 138 croplands and orchards using gridded sample design with 2 km × 2 km. grid spacing ?? Is the significance of the yellow rectangular ? Seems that yellow points are the distribution of soil sampling then what are the rectangular for

No, the data for yellow rectangular belong to two individual projects with 250 m × 250 m grid spacing, that's why the rectangular soil sampling box appears in Figure 1.

Q28: Please increase the font of the lat lon on the Iran and world map at least that can be read.

Thank you. The figure 1 is improved.

Q29: Figure2 is cut short on the top, it cutting the start box ...

Thank you. The figure 2 is improved.

Q30: Figure 4: Try to increase size/resolution of the legend (pixel effect), there still enough space, Why do you call predictors family auxiliary data. To my understanding, it is more data that share common characteristics. Auxiliary meaning some extra data or data of 2nd order ... data class or data types may be more easy to understand.

Thank you. The figure 4 is improved.

Q27: Equations: Line 268;Add Equation numbers: Line 284-290;Add Equation numbers: Line 349-352;Add Equation numbers

Equation numbers were added for all equations.

Q31: Table 1: are Hyper-parameters non dimensional ? Please add their units if not.

Units were added.

Q32: Table 5: keep same parameter in caption a --> W  b-->Sl

It was corrected 

Q33: There is some difference between the covariables list and the auxiliary data of Figure 4. It seems the same to me, you may want to use the same classification and call covariables as class data or auxiliary data for consistency .... You may also try to keep the same order as the legend Figure 4 or vis versa for easy reading.

Great point. It was corrected in the table.

Specific comments

Q34: Line 116-117: The predominant soil temperature regime class is classified as being thermic followed by mesic and cryic. What is the meaning of a temperature regime class ? Add a reference or a definition.

I add the reference to this sentence.

Q35: Could you please decide a citation format. Better to have all in [XX] and not mixed citations [XX] and authors name [XX]. Lines 253-257, Line 275, 304, 374, 397, …

We changed the authors name [XX] to the [XX] citation format throughout the manuscript.

Q36: Line 392 ?? Variations to variability

It was corrected.

Q37: Line 596 for decision makers in order to inform agricultural practices to for decision makers in order to improve agricultural practices

It was corrected.

Q38: Line 505 The relationship between these datasets and PSFs was apparent in this study; for example, the spatial distribution of alluvial plains and flood basins were two physiographic units of the northeastern region of the study areas and which also coincides with the high clay contents that was observed. most likely due to the river erosion/transport process ...

It was corrected.

Q39: Line 619

It can also potentially improve land surface modelling.

It was added.

Round 2

Reviewer 2 Report

All the comments are addressed and no additional comments.

Author Response

Point 1: All the comments are addressed and no additional comments.
Response 1: Thank you.