Estimation of Organic Carbon in Anthropogenic Soil by VIS-NIR Spectroscopy: Effect of Variable Selection

Round 1

Reviewer 1 Report

The comment are in attached documment.

Comments for author File: Comments.pdf

Author Response

Response to Reviewer 1 Comments

 

We are greatly appreciated to you for your constructive suggestions and insightful comments, which helps us a lot to improve this manuscript. Based on your and other reviewers’ suggestions, we carefully revised the manuscript. Please find our point to point responses to your comments and the corresponding revisions in the revised manuscript.

 

Point 1: The manuscript Xu et al. search the effect of selection variables to estimate organic carbon in anthropogenic soil. Estimating the organic carbon content in the soil is important because of its contribution to the system. however, the use of spectroradiometers for this purpose is already known in many researches, as well as the use of pre-treatments in the spectrum.

Author writes that the differential of his research is the use of a soil with anthropogenic actions, however, there are several articles predicting SOC in agricultural land. Instead, the author could have assessed whether the quality of the SOC is interfered with by anthropogenicity, for example.

Other point is the parsimony of the models. The authors inform which the use of minor number of bands could be utilized to production of the low-cost sensor.

 

Response 1: Thank you very much for your comments and suggestions. We agree with you that the use of spectroradiometers for SOC content estimation in agricultural land is already known in many researches. This study focuses on utilizing spectral variable selection techniques to improve the accuracy and parsimony of SOC estimation models. And the study area is different from agricultural lands in most articles (Stenberg et al., 2009, Wetterlind et al., 2010, and Hong et al., 2020). Their patches of agricultural lands are large with similar human activities. Our study area has a high variability of human activities on a small scale.

 

Figure 1. Location of the study area and soil sampling.

 

The study area is located in the Jianghan Plain, China. It is known as ‘Country of Fish and Rice’. The land-use types include cropland, woodland, and meadows. Cropland patches are highly fragmented and some of them are close to settlements and various water bodies (breeding, ponds, irrigated canals, lakes, and rivers) (Wu et al., 2019). And diverse land management practices are carried out in our study area according to our field survey. Intensive human activities have led to the heterogeneity of the relationship between VIS-NIR spectra and SOC for a long time (Liu et al., 2014).

Figure 2. Photos showing the sampling sites of Dataset 1, 2, and 3 characterized with different land use and land cover.

 

A total of 108 soil samples were collected from three different sites in this study area, and the geographical distributions were shown in Figure 1. The total collected soil samples (Dataset 0, n = 108) were divided into three subsets according to sampling locations, land use and land cover types (Dataset 1, n = 49, Dataset 2, n = 16, and Dataset 3, n = 43). Samples of the three datasets were collected from three sites with different human activities on a small scale (Liu et al., 2014). Some photos of our field survey and sampling campaign were shown in Figure 2. Samples of Dataset 1 were collected from cropland that was adjacent to a breeding pond. Dataset 2 was sampled from cropland that was surrounded by cropland. Dataset 3 included samples of various land-use types (cropland, artificial forest, meadows and breeding ponds).

Figure 3. Correlation coefficient curves calculated between the raw visible and near-infrared (VIS-NIR) spectra and soil organic carbon (SOC) for four datasets. The blue line, green line, red line, and magenta line refer to correlation coefficient curves for Dataset 1, Dataset 2, Dataset 3, and Dataset 0, respectively. The blue ‘+’, green ‘+’, and magenta ‘+’ symbols refer to locations of VIS-NIR spectral variables having significant correlation for Dataset 1, Dataset 2, and Dataset 0, respectively (at a significance level of 0.05). The ‘s’ symbol refers to location of spectral variables having the lowest correlation coefficient.

 

The differences in the correlation coefficient curves reveal heterogeneous relationships between raw VIS-NIR spectra and SOC (Figure 3). In Figure 3, blue ‘+’, green ‘+’, and magenta ‘+’ symbols refer to locations of VIS-NIR spectral variables having significant correlations for Dataset 1, Dataset 2, and Dataset 0, respectively (at a significance level of 0.05). It was revealed that SOC had a significantly negative correlations with raw VIS-NIR spectra in the region of 400 - 2449 nm for Dataset 1 and Dataset 0. The spectral variables with significant negative correlations distributed in the region of 480 - 900 nm and 1170 - 1870 nm for Dataset 2, whereas no significant correlations were observed for Dataset 3.

Dataset 1 had the strongest correlations among these four datasets. The absolute correlation coefficients slowly decreased as the wavelength increases after 670 nm. The correlation coefficients of Dataset 2 had faint change in the spectral range of 870 - 2449 nm. Dataset 0 was the combination of Dataset 1, Dataset 2 and Dataset 3. Absolute correlation coefficients for Dataset 0 increased in the range of 400 - 570 nm, which was different trend compared to Dataset 1, Dataset 2 and Dataset 3. The highest absolute correlation coefficients of Dataset 1 (670 nm), Dataset 2 (570 nm), Dataset 3 (660 nm), and Dataset 0 (730 nm) were of different magnitude. This provides a vivid evidence that a heterogeneous relationship exists between VIS-NIR spectra and SOC for soils with intensive human activities.

Thank you very much for your suggestion about exploring the quality of the SOC interfered with by anthropogenicity. We will explore this topic in our future research.

 

Point 2: Lines 21 -“In homogeneous soils"? And heterogeneous, it's not possible. Have any limitation? No. The use of the spectral library can resolve this problem?

 

Response 2: Many thanks for your comments and suggestions. We agree with your opinions. This study focuses on the heterogeneous relationships between VIS-NIR spectra and SOC, not just heterogeneous soils. We also attempted to use three spectral libraries to resolve this problem.

The first spectral library is CSSL. The CSSL comprised 1581 samples from several provinces in China (Shi et al., 2014, and Shi et al. 2015). CSSL represents 16 soil types based on the genetic soil classification of China (GSCC). The GSCC is different from the United States (US) Soil Taxonomy System and World Reference Base for Soil Resources (WRB). The details of CSSL could be found in Shi et al. (Shi et al., 2014, and Shi et al. 2015).

The second spectral library is named CSSL-552. The type of our soil samples is similar to the type “paddy soil” in CSSL. Therefore, a total of 552 soil samples that belong to paddy soil in CSSL were selected as the second library.

The third spectral library Honghu-272 is from our research group. A total of 272 soil samples were collected from Honghu City in 2013. Most of the sampling were distributed in cropland and the rest were distributed in unused land. The details of this spectral library could be found in Hong et al. (Hong et al., 2019).

CSSL, CSSL-552, and Honghu-272 are used as calibration datasets to build SOC estimation models by PLSR, respectively. The calibration dataset used in this study is also used for model calibration. The validation dataset that is the same as this study is used to evaluate PLSR models.

 

Table 1 Accuracies of full-spectrum PLSR models by the use of the spectral libraries.

Spectral Library	Spectral pretreatments	LVsa	Calibration dataset		Validation dataset		RPD	b
			Rc2	RMSEc	Rp2	RMSEp
This study	None	9	0.79	2.93	0.70	3.60	1.72	1.81
	FD	7	0.78	3.01	0.80	3.17	1.96
	Log(1/R)	11	0.86	2.44	0.76	3.37	1.84
	MC	10	0.86	2.36	0.75	3.24	1.92
	MSC	8	0.78	3.02	0.70	3.66	1.70
	SNV	8	0.78	3.02	0.70	3.66	1.69
CSSL	None	16	0.48	4.94	0.37	6.49	0.96	1.08
	FD	10	0.50	4.74	0.35	6.22	1.00
	Log(1/R)	13	0.57	4.39	0.55	5.34	1.16
	MC	11	0.51	4.70	0.49	5.71	1.09
	MSC	13	0.56	4.43	0.55	5.46	1.14
	SNV	13	0.58	4.34	0.60	5.33	1.16
CSSL-552	None	18	0.68	4.02	0.45	4.66	1.33	1.26
	FD	18	0.60	4.53	0.33	5.06	1.23
	Log(1/R)	18	0.77	3.37	0.57	5.30	1.17
	MC	17	0.74	3.54	0.57	4.81	1.29
	MSC	15	0.74	3.55	0.61	4.87	1.27
	SNV	15	0.74	3.54	0.62	4.90	1.27
Honghu-272	None	13	0.73	3.78	0.58	4.26	1.46	1.56
	FD	13	0.71	3.86	0.66	3.72	1.67
	Log(1/R)	11	0.74	3.81	0.66	3.71	1.67
	MC	10	0.70	3.94	0.56	4.50	1.38
	MSC	11	0.73	3.73	0.68	3.93	1.58
	SNV	11	0.73	3.73	0.68	3.90	1.59

^a Number of latent variables; ^b Mean of RPD

The results of full-spectrum PLSR models by three spectral libraries and our measured spectra are shown in Table 1. The performance of PLSR models by three spectral libraries is always poorer than that by our measured spectra. The  of PLSR models by CSSL, CSSL-552, and Honghu-272 are 1.08, 1.26, and 1.56, respectively. It may be that CSSL has sparse sampling density on a national scale, and cannot reflect spectral diversity on a small scale. PLSR models by Honghu-272 perform better than that by CSSL. It may be that Honghu-272 has a similar sampling scale and location with the data used in our study. Results show that the heterogeneity of the relationship between VIS-NIR spectra and SOC are complex, which is worthy of further study. This finding coincides with Guerrero et al. They proposed that SOC estimation models derived from very small-sized spectral libraries could provide accurate SOC estimation, and outperform those models derived from a large spectral library (Guerrero et al., 2016). Besides, they suggested that large spectral libraries might not be needed for local scale SOC assessment.

Although the results of using spectral libraries are not satisfactory, some potentials by spectral libraries could be found to improve the accuracy of SOC estimation models on a small scale. We will work on the role of the spectral library in SOC estimation in future research. Thank you again.

 

Point 3: Line 23 - What these anthropogenic soils? Anthroposol? SOC accumulated with the anthropogenic use? Culture, accumulation of the other cultures in passed recent? Define.

 

Response 3: Thanks very much for your comments. The term “Anthropogenic soils” refers to the soils that are influenced by high human activities and land management in the long-term tillage. According to the Canada system of soil classification (CSSS), an Anthroposols has one or more of their natural horizons removed, replaced, added, or significantly modified by human activities. Depth of the anthropogenic disturbance, modification or addition must be >10 cm above or below the surface of the soil. But tillage practices may have been disturbed horizon > 10 cm, while do not constitute changes sufficient for a soil to be called an Anthroposol (Anne Naeth et al., 2012). Therefore, soils in this study do not belong to an Anthroposol.

 

Point 4: Line 23 - Parsimony? What? Do you utilized the full spectrum. The use of the spectroradiometer was continuous. The use of the statistical model, algorithms and software was full. Where is the parsimony?

 

Response 4: Thank you for your comments. Spectral variable selection algorithms are used to improve the parsimony of PLSR models in this study. They are competitive adaptive reweighted sampling (CARS) and random frog (RF). CARS selected 16 - 31 spectral variables, while RF selected 21 - 106 spectral variables from the full spectra to build PLSR models. This strategy removes uninformative variables and reserves important variables. The parsimony of the CARS/RF models has been greatly improved when compared to the full-spectrum models.

 

Point 5: Lines 24 25 - Repetitive.

 

Response 5: Thanks very much for your comments. We have revised this in our manuscript (Lines 23 - 24).

Lines 23 - 24: This fact results in poor performance of SOC estimation models. To improve model accuracy and parsimony, we investigated the performance of two variable selection algorithms.

 

Point 6: Lines 48-50: There are many studies focused on SOC content with VIS-NIR spectroscopy, and showing the efficiency of this technique in heterogeneous soils.

 

Response 6: Many thanks for your comments. We agree that many studies focus on SOC content and show the efficiency of VIS-NIR spectroscopy in heterogeneous soils. This study focuses on the heterogeneity in the relationship between VIS-NIR spectra and SOC. The details can be found in Response 1.

 

 

Point 7: Lines 104-116: Author could detail the location and then reference. What kind of anthropogenic activity is there? The samples were collected in 2011, when was the spectral analysis done? Could the storage time of these samples have interfered with the result?

 

Response 7: Thank you very much for your suggestions. We have detailed these contents in our manuscript (Lines 118 - 123, and Lines 127 - 139). The details of anthropogenic activity are also shown in Response 1. The samples were collected from 20 December 2011 to 21 December 2011, and were transported to the laboratory on 22 December 2011. It took about one weeks to pretreat soil samples (e.g., air drying). Both SOC content and the VIS-NIR spectra used in this study were measured after the sample pretreatments. Therefore, we assume that the storage time for about one weeks has minor or little interference with the result.

 

Point 8: Line 115:1 do not understand. Soil is heterogeneous or not?

 

Response 8: Thanks for your comments. Soil is heterogeneous. Soils in our study area are influenced by relatively homogenous natural environment factors (due to samples taken on a small scale) and heterogeneous human activities.

 

Point 9: Lines 123- 128: Where did this methodology come from? The methodology used for this reading has no reference.

 

Response 9: Thank you for your comments. We have added the reference to this methodology in our manuscript (Lines 142 - 159).

Lines 142 - 159: In the laboratory, soil samples were air-dried at 20 - 30 ℃ for one weeks, then ground, and passed through a 2-mm sieve [54]. An ASD FiledSpec 3 portable spectro-radiometer with a spectral range of 350 - 2500 nm, and a spectral resolution of 1 nm was used to scan soil samples in a dark room to avoid stray light interference. All samples were put separately in dishes with a 20-cm diameter. A halogen lamp placed at 30 cm distance and an angle of 45 ° was used to illuminate soil samples. The detection fiber probe was placed vertically to soil samples at 12 cm distance. A white Spectralon panel was used to calibrate the spectrometer before measuring spectrum of the first soil sample and repeated every six soil samples [53]. A total of ten scans were recorded for each soil sample which were averaged in one sample spectrum [55]. Through these procedures, the reflectance spectra of the 108 samples were obtained. The spectra in the range of 350 - 399 nm and 2450 - 2500 nm were removed due to serious noises. The remained spectra (400 - 2449 nm) were further resampled to 10 nm to extract 205 wavebands.

The SOC content was measured by wet oxidation at 180 °C with a mixture of potassium dichromate and sulfuric acid [18]. It should be noted that the oxidation of active organic carbon by this approach is incomplete, which underestimates the SOC content. A “standardized” corrective factor ranging from 1.10 to 1.40 could be used in practice [56]. In our study, we used the “raw” SOC content without using a “standardized” corrective factor. This allowed comparing the results of our study with other studies that also use the “raw” SOC content.

 

Point 10: Lines 429- -443: One of the aims of the article is to evaluate PLSR models in highly heterogeneous soil, however this heterogeneity is not clear in the text. The CV is low for all soil samples clusters.

 

Response 10: Many thanks for your comments and suggestions. Wilding (1985) categorized the CV (coefficient of variation) values into three classifications: CV > 35%, high variability; 15% < CV < 35%, moderate variability; CV < 15%, low variability. The CV of the total, calibration, and validation dataset are 0.40, 0.40, and 0.39, respectively. Therefore, the SOC contents of the datasets are of high variability. High variability of SOC contents provides evidence that soils in our study could be heterogeneous.

 

In addition, this study mainly focuses on the heterogeneity in the relationship between VIS-NIR spectra and SOC. The details about the heterogeneity could be found in Response 1.

 

Point 11: Do you believe that the computational effort undertaken is compensatory for the results you obtained? R - 0.80 to 0.81 for CARS; improvement of 4% compared for RMSE.

Why do you did not use Stepwise? Is practical, security, simple etc.

 

Response 11: Thanks very much for your comments and suggestions. This study focuses on the use of spectral variable selection techniques rather than the use of spectral pretreatments. We simply test the performance of spectral pretreatments with spectral variable selection techniques. We found that the best R² increases from 0.80 (by full spectra) to 0.81 (by CARS) with different pretreatments. We admit that the accuracy improvement is minor. Nevertheless, the model parsimony has improved significantly. The number of selected spectral variables decreases from 205 (by full spectra) to 31 (by CARS). From a practical perspective, one can simply use the spectral variable selection algorithm such as CARS without any spectral pretreatment. This simplified strategy can also benefit the outcomes, in which the best R_p² increases from 0.70 (by full spectra) to 0.78 (by CARS), while the number of selected spectral variables decreases from 205 (by full spectra) to 21 (by CARS).

       Thank you for your suggestion of using Stepwise. Stepwise selection is well supported in a variety of statistical packages. Nevertheless, it is known that Stepwise selection does not consider all possible combinations of explanatory variables. In addition, the selection of variables using a stepwise regression could be unstable in the case that we have a small sample size compared to the number of potential explanatory variables.

       CARS algorithm consists of two steps to spectral variable selection, including exponentially decreasing function (EDF) and adaptive reweighted sampling (ARS). EDF eliminates spectral variables rapidly in the first stage, and slowly in the second stage. ARS is employed to further eliminate spectral variables based on the ‘survival of the fittest’ principle. The combination of EDF and ARS realizes objective and stable selection. Therefore, CARS was used in this study.

 

Point 12: This is a manuscript that tries to expose the data to exhaustive tests and statistical models so that its results are satisfactory for that data set. There is no cause and effect relationship in the discussions. Why is one algorithm better than the other? Can it detect more carbon, lower moisture content? What is the relationship between energy-matter? Is the remote sensing article or statistical article?

 

Response 12: Thank you very much for your comments and suggestions. We further explained why one algorithm is better than the other, and the potential cause and effect relationship of what we found. These discussions could be found in Section 4 (Lines 420 - 446 and Lines 470 - 483).

Lines 420 - 446: For the CARS based models, Log (1/R) and MC enhanced the accuracy of PLSR models, while FD, MSC and SNV led to accuracy deterioration. FD has not improved the estimation performance of the PLSR model. This might be due to the lack of variables in the spectral region of 1200 - 1800 nm, a spectral range that contains relevant overtones and combinations of fundamental bonds of O–H and C=O groups that are associated with SOC. The amplification of noise may also hinder useful signals [62]. MSC has failed to improve model performance as only one variable was selected in the visible region of 400 - 780 nm, which is an important region associated with soil color variation in the blue band around 450 nm. The darker the soil color, the higher the SOC content [63]. This single selected variable may not accounted for sufficient information to capture variation in soil darkness that could be linked with variation in SOC. The number of selected spectral variables with SNV was large than that with MSC, which explains why the PLSR model with SNV outperforms that with MSC. For the RF-based models, all pretreatments enhanced the accuracy of PLSR models. It could be attributed to the fact that RF could select important spectral variables as comprehensively as possible while reducing collinearity between spectral variables. Among them, Log (1/R) + CARS/RF + PLSR models provided the best estimation accuracy. Since SOC absorbs energy at specific frequencies in VIS-NIR, Log (1/R) enhanced its absorption feature [64].

It should be noted that the effects of spectral pretreatments are determined by the quality of raw spectra. In other words, spectral pretreatments would be necessary in the case that raw spectra are influenced by variable soil physical conditions and by the surrounding environment during measurement (e.g., temperature, ambient light). In this study, however, all the samples were air-dried, ground and passed through a 2-mm sieve, and VIS-NIR spectra were measured in a well-controlled environment. Therefore, both CARS and RF could develop satisfactory SOC estimation models without spectral pretreatments. Meanwhile, this study provided that Log (1/R) should be adopted to further improve the accuracy of SOC estimation models.

 

Lines 470 - 483: In summary, the selected spectral variables in the visible region by both methods are mainly concentrated in the locations of 400 nm, 530 nm and 610 nm. These wavelengths are associated with variation in chromophores and the darkness of humic acid (e.g. due to blue absorption band at 450 nm and perhaps red color absorption band at 680 nm) [65]. Spectral variables in the near-infrared region are chosen because some fundamental vibrational bonds are associated with SOC. These bonds mainly include C–H, N–H, C=O, C–O, O–H and Al–OH [54, 66]. Table 3 shows the possible assignments of fundamental bonds, absorption wavelength, and related soil constituents for main selected spectral variables in the near-infrared region by RF and CARS. Some selected spectral wavelengths in this study do not coincide with possible wavelengths of fundamental vibrational bonds. It could be attributed to slight shifts in wavelength locations due to inharmonic molecules vibrations [54].

Table 3. Possible assignments of fundamental bonds, absorption wavelength, and related soil constituent for the selected spectral variables in the near - infrared region by competitive adaptive reweighted sampling (CARS) and random frog (RF) [54, 66].

Locations of selected spectral variables (nm)	Possible fundamental bonds	Possible wavelength (nm)	Possible related soil constituents
800	C–H	825	Organics (aromatics)
1000	N–H	1000	Organics (amine)
1100	C–H	1100	Organics (aromatics)
1200	C–H	1170	Organics (Alkyl asymmetric-symmetric doublet)
1420	O–H	1380	Water
1500	C=O	1524	Organics (amides)
1800	C–H	1754	Organics (Alkyl asymmetric-symmetric doublet)
1920	O–H	1915	Water
2000	C=O	2033	Organics (amides)
2100	N–H	2060	Organics (amine)
2200	Al–OH	2230	Clay minerals
2350	C–O	2381	Organics (Carbohydrates)

 

 

Point 13: Line 48-50. I suggest a more advanced search in articles on the topic presented.

 

Response 13: Many thanks for your suggestion. We have followed your suggestion and improved the introduction (Lines 50 - 71).

Lines 50 - 71: Intensive human activities increase the heterogeneity in the relationship between SOC and VIS-NIR spectra, which leads to a new challenge to apply VIS-NIR spectroscopy for SOC estimation [18]. Several soil spectral libraries (SSLs) have been established to cover the heterogeneous spectral characteristics of different soil types as comprehensive as possible, thereby improving the accuracy of the SOC estimation. Existing SSLs include ICRAF-ISRIC world soil spectral library [19], European spectral library [20], Brazilian soil spectral library [21], Australian soil spectroscopic database [22], China soil spectral library [23], etc. These soil samples in the spectral libraries are collected on a large scale with the low sampling density. For example, the mean soil sampling density of the European spectral library is 77 samples per 10000 km² [24]. It is not enough to reveal the heterogeneous relationship between the VIS-NIR spectra and SOC on a small scale [25]. Therefore, many researchers made efforts on small-scale studies with soil samples collected from farmland, when VIS-NIR spectroscopy has shown good performance [26-30]. These farmlands are continuous, and have large area with similar human activities. However, the heterogeneity in the relationship between SOC and VIS-NIR spectra is more complex in the highly fragmented farmland with various human activities, which can weaken the performance of the SOC estimation model by VIS-NIR spectra. To improve model performance, previous studies adopted the strategy of using representative calibration samples [26-29]. Nevertheless, the efficiency of this strategy is susceptible to sample size [30]. Therefore, it is essential to investigate other approaches that may improve model accuracy. Besides, VIS-NIR spectra featured by high spectral resolution may contain abundant spectral information, which may complicate the SOC estimation models [31, 32]. Thus, it is necessary to establish new approaches to improve model parsimony.

 

Point 14: 362 - 372. “Due to the difference of selected spectral variables with different spectral pretreatments, the PLSR models have different accuracy.'

Ok. I have a soil analyses laboratory and I want adapted your methodology? I have one methodology and I don't have time to works with the spectrum. I want the continuous process. What do you do? Do you keep testing the algorithms until you find the right one?

 

Response 14: Thank you very much for your comments. This study aims to use spectral variable selection techniques to improve the accuracy of PLSR models. According to our results, both CARS and RF can achieve more satisfactory performance of PLSR models than the full-spectrum models (For raw spectra, full-spectrum models: R² = 0.70 and RPD = 1.72, CARS models: R² = 0.78 and RPD = 2.03, RF models: R² = 0.72 and RPD = 1.86).

From a practical perspective, we would suggest a combination of Log (1/R) and CARS (or RF), which hopefully guarantee a fairly good improvement of VIS-NIR estimation of SOC.

 

 

Point 15: 366 - of lack of variables in the spectral region of 1200 - 1800 nm and 2200 - 2400 nm . and the fewer number of selected spectral variables.

If you algorithm don't choose the variable in the spectral region (1200 - 1800 and 2200 -2400) is because the bands in the region is not important.

 This is a clear example of the problem of using many algorithms to model soil data using spectra with big number of data with high collinearity. Spectral bands resulting from the presence of water in the SOC (1400 and 1900nm) and, therefore, differentiable since each type of SOC can absorb more or less water have not entered the model. The same case for 2000 - 2400 nm (OH group in minerals silicate associate with SOC). That is, where we could have an explanation of cause and effect are discarded from the model for statistical reasons and not of the researcher's tacit knowledge.

 

Response 15: Many thanks for your comments and suggestions which help us greatly improve our manuscript. The explanation of cause and effect can be found in Response 12 as well as in Lines 420 - 446 and Lines 470 - 483.

 

Point 16: 375 - 376 Model parsimony is an important issue that has been less investigated in previous studies. With the increasing demand of heterogeneous soil information and the design of low-cost spectrometer, less spectral variables are desired for SOC estimation.

Low-cost spectrometer design? How? The authors guarantee that exactly these bands will be used to evaluate SOC in any region? Regardless of the algorithm used, you need to start the process using the spectroradiometer. You need a complete curve. Each different type of organic matter will have different spectral bands than the ones you used. In 369-390 the authors show this fact.“Hong et al. reported that spectral variables at 400 - 800 nm and 2000 - 2400 nm were related to SOC. It should be noted that sensitive spectral variables for SOC may differ from one dataset to another.

 Did you a metanalyses to verify recurrent SOC bands in other papers?

The problem of the high cost of the hyperspectral image sensor goes beyond the number of bands. There are sensors with a greater or lesser signal/noise ratio. The diffracting grid would have to have specific windows, which can be more expensive than a standard grid. Another problem is to think that an orbital sensor will have the same spectral behavior as the laboratory. I would first do tests with sensors with maximum resolution and then choose sensors with specific windows, and not the other way around.

Computationally, the problem of having models with a large number of bands is irrelevant. The important thing is to know if we are not losing information that was thrown out of the model or the model is overloaded with unnecessary bands. Therefore, knowledge of the energy-matter relationship is more important to explain why band 1 enters the model and band N does not.

 

Response 16: Thank you very much for your explanations for the spectrometer and sensors design. We agree that the sensitive spectral variables for SOC may differ from one dataset to another. Therefore, both case study of a specific area and metanalyses of recurrent SOC bands in other papers could be beneficial. We have revised our statement as “With the demand for advanced regression algorithms, fewer spectral variables are desired for SOC estimation and mapping with hyperspectral remote sensing data.”. Nonetheless, we attempt to argue that the use of spectral variable selection can reduce the number of spectral variables, thereby improving computing efficiency and reducing computing costs. In addition, the correlation between VIS-NIR spectra and SOC may be nonlinear. The nonlinear regression algorithms are applied to estimate SOC content (e.g. support vector machine, artificial neural network, random forest, etc.). But these algorithms have poor performance for a small sample size with a large number of spectral variables. Spectral variable selection techniques can help to solve this problem.

In addition, we agree with you that knowledge of the energy-matter relationship is more important to explain why band 1 enters the model and band N does not. However, near-infrared spectra are dominated by weak overtones and combinations of fundamental vibrational bands (Chang et al., 2001). Because of the broad bands and overlapping absorption of soil constituents, VIS-NIR spectra are nonspecific and difficult to interpret (Stenberg et al., 2010). The locations of overtones and combinations of fundamental groups also often slightly shift from the exact expected location because real molecules do not behave totally harmonically (Viscarra Rossel et al., 2010). Therefore, multivariate calibrations are used to analyze the correlation between spectra and soil properties from mathematics (e.g. multiple linear regression, partial least squares regression, support vector machine, and artificial neural networks, etc.) (Wetterlind et al., 2013). This study utilizes CARS/RF with PLSR to interrogate the correlation of spectral variables and SOC, which belongs to multivariate calibrations. This strategy mainly utilizes VIS-NIR techniques as an aid in precision farming and both assessment and management of soil quality. This study also provides a possible explanation for selected spectral variables (Lines 414 - 438 and Lines 462 - 474). We need more collaborative research with other domains to understand the physical basis for the reflection of light from soils (Stenberg et al., 2010). To our knowledge, there seems to be no applicable law to specifically explain the energy-matter between VIS-NIR spectra and SOC due to the complexity of soil properties and spectra. In-depth knowledge of the energy-matter relationship will be future research direction in this domain.

 

Point 17: Line 430 - CARS + PLSR and RF + PLSR with five different spectral pretreatments were used to estimate SOC in Jianghan Plain of China.

Write only CARS and RF algorithms since PLSR was used in all stages. You are not comparing PLS with another mathematical model.

 

Response 17: Many thanks for your suggestion. We have revised this in our manuscript (Lines 508 - 511).

Lines 508 - 511: In this study, competitive adaptive reweighted sampling (CARS) and random frog (RF) algorithms in combination with five different spectral pretreatments were used to select spectral variables, which were used as input in partial least squares regression (PLSR) to estimate soil organic carbon (SOC) in the Jianghan Plain of China.

 

We look forward to hearing from you. Thank you very much for your time and kind consideration.

 

 

Reference:

Anne Naeth, M., et al. (2012). "Proposed classification for human modified soils in Canada: Anthroposolic order." Canadian Journal of Soil Science 92(1): 7-18.

Chang, C.-W., et al. (2001). "Near‐infrared reflectance spectroscopy–principal components regression analyses of soil properties."  65(2): 480-490.

Guerrero, C., et al. (2016). "Do we really need large spectral libraries for local scale SOC assessment with NIR spectroscopy?" Soil and Tillage Research 155: 501-509.

Wu, Z., et al. (2019). "Estimating soil organic carbon density in plains using landscape metric-based regression Kriging model." Soil and Tillage Research 195.

Liu, Y., et al. (2014). "Transferability of a Visible and Near-Infrared Model for Soil Organic Matter Estimation in Riparian Landscapes." Remote Sensing 6(5): 4305-4322.

Shi, Z., et al. (2014). "Development of a national VNIR soil-spectral library for soil classification and prediction of organic matter concentrations." Science China Earth Sciences 57(7): 1671-1680.

Shi, Z., et al. (2015). "Prediction of soil organic matter using a spatially constrained local partial least squares regression and the Chinese vis-NIR spectral library." European Journal of Soil Science 66(4): 679-687.

Liu, Y., et al. (2015). "Comparing geospatial techniques to predict SOC stocks." Soil and Tillage Research 148: 46-58.

Hong, Y., et al. (2019). "Application of fractional-order derivative in the quantitative estimation of soil organic matter content through visible and near-infrared spectroscopy." Geoderma 337: 758-769.

Hong, Y., et al. (2020). "Comparing laboratory and airborne hyperspectral data for the estimation and mapping of topsoil organic carbon: Feature selection coupled with random forest." Soil and Tillage Research 199: 104589.

Stenberg, B., et al. (2010). Visible and Near Infrared Spectroscopy in Soil Science: 163-215.

Stenberg, B. and J. Wetterlind (2009). "Small sized local vs. large sized national calibration sets and their combination for farm scale predictions by NIR."

Wetterlind, J., et al. (2013). "Soil analysis using visible and near infrared spectroscopy." Methods Mol Biol 953: 95-107.

Wetterlind, J., et al. (2010). "Increased sample point density in farm soil mapping by local calibration of visible and near infrared prediction models." Geoderma 156(3-4): 152-160.

Viscarra Rossel, A. V. and T. Behrens (2010). "Using data mining to model and interpret soil diffuse reflectance spectra." Geoderma 158(1-2): 46-54.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper is interesting from a methodological point of view but I’m wondering if it has its place in Remote Sensing. There is no RS at all in this paper (neither proximal sensing). I think that it should be submitted to a soil science or a chemometrics journal, this is why my opinion is to reject it and encourage resubmission elsewhere, but I let the editors decide.

There are, however, some flaws, which I detail hereafter

• The authors claim that VIS-NIR SOC relationships are more difficult to establish in anthropogenic soils. They do not elaborate why. It seems that this is not because of a high SOC variability because authors say that “the datasets were of medium variability” (lines 206-207).
• The method for determining SOC content is not appropriate. It is well known that Potassium dichromate does not extract total SOC but only a more labile part. (line 132). This is not even discussed. This is a major flaw of the paper.
• The classification of RPD in poor, fair, etc, models, has been recognized as arbitrary for a long time. (lines 185-187).
• There are too many self-cites (e.g. Hong et al)

Author Response

Response to Reviewer 2 Comments

 

Many thanks for your valuable comments and suggestions, which have led to significant improvement of this paper. Based on your and other reviewers’ suggestions, we carefully revised the manuscript. Please find our point to point responses to your comments and the corresponding revisions in the revised manuscript.

 

Point 1: This paper is interesting from a methodological point of view but I’m wondering if it has its place in Remote Sensing. There is no RS at all in this paper (neither proximal sensing). I think that it should be submitted to a soil science or a chemometrics journal, this is why my opinion is to reject it and encourage resubmission elsewhere, but I let the editors decide.

 

Response 1: Thank you very much for your comments and suggestions. Proximal soil sensing technology has been widely applied to estimate SOC content. Visible and near-infrared reflectance (VIS-NIR) spectroscopy, known as a typical type of proximal soil sensing technology, has outstanding performance on SOC content estimation (Stenberg et al., 2010, and Shi et al., 2018). VIS-NIR spectroscopy can be used both in situ and the lab. But VIS-NIR spectra measured in situ easily suffer from the influences of soil surface roughness, soil moisture, water vapor, light intensity, and other external environmental interference. The measurement of VIS-NIR spectra performed in the lab could effectively avoid these influences and increase the signal-to-noise ratio of VIS-NIR spectra.

This paper focuses on improving the accuracy and parsimony of SOC estimation models by VIS-NIR spectra. Spectral variable selection techniques are adopted to select important spectral variables to achieve this goal. And the selected spectral variables have a relatively high correlation with SOC content, which could provide a reference to design channels of proximal sensors. Therefore, we think that this study has its place in Remote Sensing (or proximal sensing). Thank you again for your suggestions.

 

Point 2: The authors claim that VIS-NIR SOC relationships are more difficult to establish in anthropogenic soils. They do not elaborate why. It seems that this is not because of a high SOC variability because authors say that “the datasets were of medium variability” (lines 206-207).

 

Response 2: Many thanks for your comments. In the revised manuscript, we explain why the VIS-NIR and SOC relationships are more difficult to establish in anthropogenic soils. It should be noted that the major difficulties lie in the heterogeneity in the relationship between VIS-NIR spectra and SOC content, not just the high variability of SOC content.

Figure 1. Location of the study area and soil sampling.

 

The study area is located in the Jianghan Plain, China. It is known as ‘Country of Fish and Rice’. The land-use types include cropland, woodland, and meadows. Cropland patches are highly fragmented and some of them are close to settlements and various water bodies (breeding, ponds, irrigated canals, lakes, and rivers) (Wu et al., 2019). And diverse land management practices are carried out in our study area according to our field survey. Intensive human activities have led to the heterogeneity of the relationship between VIS-NIR spectra and SOC for a long time (Liu et al., 2014).

Figure 2. Photos showing the sampling sites of Dataset 1, 2, and 3 characterized with different land use and land cover.

 

A total of 108 soil samples were collected from three different sites in this study area, and the geographical distributions were shown in Figure 1. The total collected soil samples (Dataset 0, n = 108) were divided into three subsets according to sampling locations, land use and land cover types (Dataset 1, n = 49, Dataset 2, n = 16, and Dataset 3, n = 43). Samples of the three datasets were collected from three sites with different human activities on a small scale (Liu et al., 2014). Some photos of our field survey and sampling campaign were shown in Figure 2. Samples of Dataset 1 were collected from cropland that was adjacent to a breeding pond. Dataset 2 was sampled from cropland that was surrounded by cropland. Dataset 3 included samples of various land-use types (cropland, artificial forest, meadows and breeding ponds).

Figure 3. Correlation coefficient curves calculated between the raw visible and near-infrared (VIS-NIR) spectra and soil organic carbon (SOC) for four datasets. The blue line, green line, red line, and magenta line refer to correlation coefficient curves for Dataset 1, Dataset 2, Dataset 3, and Dataset 0, respectively. The blue ‘+’, green ‘+’, and magenta ‘+’ symbols refer to locations of VIS-NIR spectral variables having significant correlation for Dataset 1, Dataset 2, and Dataset 0, respectively (at a significance level of 0.05). The ‘s’ symbol refers to location of spectral variables having the lowest correlation coefficient.

 

 

The differences in the correlation coefficient curves reveal heterogeneous relationships between raw VIS-NIR spectra and SOC (Figure 3). In Figure 3, blue ‘+’, green ‘+’, and magenta ‘+’ symbols refer to locations of VIS-NIR spectral variables having significant correlations for Dataset 1, Dataset 2, and Dataset 0, respectively (at a significance level of 0.05). It was revealed that SOC had a significantly negative correlations with raw VIS-NIR spectra in the region of 400 - 2400 nm for Dataset 1 and Dataset 0. The spectral variables with significant negative correlations distributed in the region of 480 - 900 nm and 1170 - 1870 nm for Dataset 2, whereas no significant correlations were observed for Dataset 3.

Dataset 1 had the strongest correlations among these four datasets. The absolute correlation coefficients slowly decreased as the wavelength increases after 670 nm. The correlation coefficients of Dataset 2 had faint change in the spectral range of 870 - 2449 nm. Dataset 0 was the combination of Dataset 1, Dataset 2 and Dataset 3. Absolute correlation coefficients for Dataset 0 increased in the range of 400 - 570 nm, which was different trend compared to Dataset 1, Dataset 2 and Dataset 3. The highest absolute correlation coefficients of Dataset 1 (670 nm), Dataset 2 (570 nm), Dataset 3 (660 nm), and Dataset 0 (730 nm) were of different magnitude. This provides a vivid evidence that a heterogeneous relationship exists between VIS-NIR spectra and SOC for soils with intensive human activities.

 

 

Point 3: The method for determining SOC content is not appropriate. It is well known that Potassium dichromate does not extract total SOC but only a more labile part. (line 132). This is not even discussed. This is a major flaw of the paper.

 

Response 3: Thanks very much for your comments and suggestions. The description of the SOC measurement has been revised and the limitation has been added (Lines 154 - 159).

Lines 154 - 159: The SOC content was measured by wet oxidation at 180 °C with a mixture of potassium dichromate and sulfuric acid [18]. It should be noted that the oxidation of active organic carbon by this approach is incomplete, which underestimates the SOC content. A “standardized” corrective factor ranging from 1.10 to 1.40 could be used in practice [56]. In our study, we used the “raw” SOC content without using a “standardized” corrective factor. This allowed comparing the results of our study with other studies that also use the “raw” SOC content.

 

 

Point 4: The classification of RPD in poor, fair, etc, models, has been recognized as arbitrary for a long time. (lines 185-187).

 

Response 4: Many thank you for your comments and suggestions. We have removed the classification of RPD in this paper. And this study does not classify the performance of PLSR models due to the arbitrariness of RPD classification. The values of RPD are used to evaluate the performance of PLSR models in Section 3.4 and Section 4.

 

 

Point 5: There are too many self-cites (e.g. Hong et al)

 

Response 5: Thank you very much for your comments and suggestions. We have removed some self-cites and retained part of self-cites which are highly related to this study.

 

We look forward to hearing from you. Thank you again for your time and consideration.

 

Reference:

Liu, Y., et al. (2014). "Transferability of a Visible and Near-Infrared Model for Soil Organic Matter Estimation in Riparian Landscapes." Remote Sensing 6(5): 4305-4322.

Stenberg, B., et al. (2010). Visible and Near Infrared Spectroscopy in Soil Science: 163-215.

Shi, T., et al. (2018). "Proximal and remote sensing techniques for mapping of soil contamination with heavy metals." Applied Spectroscopy Reviews 53(10): 783-805.

Wu, Z., et al. (2019). "Estimating soil organic carbon density in plains using landscape metric-based regression Kriging model." Soil and Tillage Research 195.

 

Author Response File: Author Response.pdf

Reviewer 3 Report

This manuscript examines the effect of variable selection in partial least squares regression (PLSR) for estimating soil organic carbon (SOC) from laboratory-measured VNIR reflectance. Two variable selection approaches are considered: competitive adaptive reweighted sampling (CARS) and random frog (RF). Both variable selection approaches were found to outperform PLSR alone, albeit only slightly.

 

The work appears to have been conducted in a thoughtful manner. I appreciate the authors’ thorough description of their methods and their informative figures. In my opinion, the most significant weaknesses of the study appear to be: 1) the relatively small impact of the different approaches of variable selection (e.g. R2 of 0.80 vs 0.81 vs 0.83) should be made more explicit, and 2) more focus should be given to the physical processes generating the observed features.

 

That being said, I am happy to report that the manuscript rises to the standard of Remote Sensing and I suggest it be published with only minor revisions to be made at the authors’ discretion.

 

Line-specific comments below:

 

Abstract and Introduction: Given that this is a remote sensing journal, it should be made clearer that this paper uses laboratory-only (and not airborne or satellite) measurements.

 

Lines 111-131: More details of sample storage, transport, and preparation should be given. Were the plastic bags airtight? How long were they in vehicles before reaching the laboratory, and what were the environmental conditions of transport? How long were they allowed to air dry? How frequently was calibration to Spectralon performed?

 

Figures 1 and 2. These figures are very informative.

 

Table 2. These differences are quite minor (R2p of 0.80 vs 0.81 vs 0.83). The tone of the text should be tempered to reflect this.

 

Lines 371-372. This sentence is not informative and should either be clarified or removed.

 

Lines 394-400. This portion of the manuscript – the specifics of the molecular bonds causing the differences in reflectance – is very interesting and should be expanded.

Author Response

Response to Reviewer 3 Comments

Point 1: Abstract and Introduction: Given that this is a remote sensing journal, it should be made clearer that this paper uses laboratory-only (and not airborne or satellite) measurements.

 

Response 1: Thank you very much for your comments and suggestions. We have made it clearer that this paper use laboratory-only measurements (Line 26 - 27, Lines 45 - 49 and Lines 142 - 143).

Line 26 - 27: A total of 108 samples were collected from Jianghan Plain, China, with the SOC content and VIS-NIR spectra measured in the laboratory.

Lines 45 - 49: VIS-NIR spectra can be measured in situ or the lab [16]. But VIS-NIR spectra measured in situ easily suffer from the influences of soil surface roughness, soil moisture, water vapor, light intensity, and other external environmental interference [17]. The measurement of VIS-NIR spectra performed in the lab could effectively avoid these influences.

Lines 142 - 143: In the laboratory, soil samples were air-dried at 20 - 30 ℃ for one weeks, then ground, and passed through a 2-mm sieve [54].

 

 

Point 2: Lines 111-131: More details of sample storage, transport, and preparation should be given. Were the plastic bags airtight? How long were they in vehicles before reaching the laboratory, and what were the environmental conditions of transport? How long were they allowed to air dry? How frequently was calibration to Spectralon performed?

 

Response 2: Many thanks for your comments. We have added more details of sample storage, transport, and preparation (Lines 129 - 139 and Lines 142 - 153). The soil samples were put in sealed plastic bags. It took two days to collect soil samples in the field (20 December 2011, and 21 December 2011) and one day to send soil samples to the laboratory at room temperature (22 December 2011). In the laboratory, soil samples were air-dried at 20 - 30 ℃ for one weeks. A white Spectralon panel was used to calibrate the spectrometer before measuring the spectra of the first soil sample and after every six soil samples (Shi et al., 2014).

Lines 129 - 139: A total of 108 topsoil samples (0 - 15cm) were collected from 20 December 2011 to 21 December 2011. Each soil sample was a mix of five soil subsamples, which were collected from the center and four corners in a square of 1 m² [53]. The geographical coordinates of these samples were recorded by a hand-held global positioning system, and the geographical distributions were shown in Figure 1. The total collected soil samples (Dataset 0) were divided into three datasets according to sampling locations, land use and land cover types (Dataset 1, Dataset 2, and Dataset 3, respectively). Samples of the three datasets were collected from three sites with different human activities on a small scale [18]. Samples of Dataset 1 was collected from cropland that was adjacent to a breeding pond. Dataset 2 were sampled from cropland that was surrounded by cropland. Dataset 3 included samples of various land-use types (cropland, artificial forest, meadows, and breeding ponds). These samples were put in sealed plastic bags with sampling sequence labels, and then were sent to the laboratory at room temperature on 22 December 2011.

Lines 142 - 153: In the laboratory, soil samples were air-dried at 20 - 30 ℃ for one weeks, then ground, and passed through a 2-mm sieve [54]. An ASD FiledSpec 3 portable spectro-radiometer with a spectral range of 350 - 2500 nm, and a spectral resolution of 1 nm was used to scan soil samples in a dark room to avoid stray light interference. All samples were put separately in dishes with a 20-cm diameter. A halogen lamp placed at 30 cm distance and an angle of 45 ° was used to illuminate soil samples. The detection fiber probe was placed vertically to soil samples at 12 cm distance. A white Spectralon panel was used to calibrate the spectrometer before measuring spectrum of the first soil sample and repeated every six soil samples [53]. A total of ten scans were recorded for each soil sample which were averaged in one sample spectrum [55]. Through these procedures, the reflectance spectra of the 108 samples were obtained. The spectra in the range of 350 - 399 nm and 2450 - 2500 nm were removed due to serious noises. The remained spectra (400 - 2449 nm) were further resampled to 10 nm to extract 205 wavebands.

 

Point 3: Figures 1 and 2. These figures are very informative.

 

Response 3: Thanks very much for your comments.

 

Point 4: Table 2. These differences are quite minor (R2p of 0.80 vs 0.81 vs 0.83). The tone of the text should be tempered to reflect this.

 

Response 4: Many thank you for your comments and suggestions which help us greatly improve our manuscript. The tone of the text has been tempered to reflect the minor improvement (Lines 350 - 352).

Lines 350 - 352: The best Rp² in each spectral variable selection category slightly increased from 0.80 (Full spectra) to 0.81 (CARS) and 0.83 (RF), and RPD increased from 1.96 (Full spectra) to 2.05 (CARS) and 2.11 (RF).

 

Point 5: Lines 371-372. This sentence is not informative and should either be clarified or removed.

 

Response 5: Thank you very much for your comments and suggestions. We have removed this sentence.

 

Point 6: Lines 394-400. This portion of the manuscript – the specifics of the molecular bonds causing the differences in reflectance – is very interesting and should be expanded.

 

Response 6: Thank you very much for your suggestions. We have expanded this in Lines 470 - 483.

Lines 470 - 483: In summary, the selected spectral variables in the visible region by both methods are mainly concentrated in the locations of 400 nm, 530 nm and 610 nm. These wavelengths are associated with variation in chromophores and the darkness of humic acid (e.g. due to blue absorption band at 450 nm and perhaps red color absorption band at 680 nm) [65]. Spectral variables in the near-infrared region are chosen because some fundamental vibrational bonds are associated with SOC. These bonds mainly include C–H, N–H, C=O, C–O, O–H and Al–OH [54, 66]. Table 3 shows the possible assignments of fundamental bonds, absorption wavelength, and related soil constituents for main selected spectral variables in the near-infrared region by RF and CARS. Some selected spectral wavelengths in this study do not coincide with possible wavelengths of fundamental vibrational bonds. It could be attributed to slight shifts in wavelength locations due to inharmonic molecules vibrations [54].

Table 3. Possible assignments of fundamental bonds, absorption wavelength, and related soil constituent for the selected spectral variables in the near - infrared region by competitive adaptive reweighted sampling (CARS) and random frog (RF) [54, 66].

Locations of selected spectral variables (nm)	Possible fundamental bonds	Possible wavelength (nm)	Possible related soil constituents
800	C–H	825	Organics (aromatics)
1000	N–H	1000	Organics (amine)
1100	C–H	1100	Organics (aromatics)
1200	C–H	1170	Organics (Alkyl asymmetric-symmetric doublet)
1420	O–H	1380	Water
1500	C=O	1524	Organics (amides)
1800	C–H	1754	Organics (Alkyl asymmetric-symmetric doublet)
1920	O–H	1915	Water
2000	C=O	2033	Organics (amides)
2100	N–H	2060	Organics (amine)
2200	Al–OH	2230	Clay minerals
2350	C–O	2381	Organics (Carbohydrates)

 

Special thanks again to you for your good comments and suggestions which greatly improve our manuscript.

Reference:

Shi, T., et al. (2014). "Soil organic carbon content estimation with laboratory-based visible-near-infrared reflectance spectroscopy: feature selection." Appl Spectrosc 68(8): 831-837.

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors made the changes requested by this reviewer, in addition to clarifying the points in conflict.

Reviewer 2 Report

This paper has been improved. I still think that it would better fit to another journal but I let the editors decide.