Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

A Method to Evaluate Spectral Analysis by Spectroscopy

A Method to Evaluate Spectral Analysis by Spectroscopy sensors Article 1 , 2 1 , 2 , 1 , 2 1 , 2 1 , 2 Yan Liu , Pingping Fan * , Huimin Qiu , Xueying Li and Guangli Hou Institute of Oceanographic Instrumentation, Qilu University of Technology, Qingdao 266061, China; [email protected] (Y.L.); [email protected] (H.Q.); [email protected] (X.L.); [email protected] (G.H.) Shandong Key Lab of Marine Environment Monitoring Technology, Qingdao 266061, China * Correspondence: [email protected]; Tel.: +86-186-6982-3823 Abstract: Visible and near infrared spectroscopy has been widely used to develop a method for rapidly determining organic carbon in soils or sediments (SOC). Most of these studies concentrated on how to establish a good spectral model but ignored how to evaluate the method, such as the use of detection range (max and min), resolution and error for SOC spectral analysis. Here, we proposed a method to evaluate the spectral analysis of SOC. Using 96 sediments sampled in the Yellow Sea and Bohai Sea, China, we established three spectral models of SOC after collecting their spectral reflectance by Agilent Cary 5000, ASD FieldSpec 4 and Ocean Optics QEPro, respectively. For both the calibration set and validation set in each spectrometer, the predicted SOC concentrations followed a distribution curve (function), in which the x-axis was the SOC concentrations. Using these curves, we developed these four technical parameters. The detection ranges were the SOC concentrations where the curve was near to or crossing with the lateral axis, while the detection resolution was the average difference between the two neighboring SOC concentrations. The detection errors were the differences between the predicted SOC and the measured SOC. The results show that these technical parameters were better in the bench-top spectrometer (Cary 5000) than those in the portable spectrometers when analyzing the same samples. For the portable spectrometers, QEPro had a broader detection range and more consistent detection error than FieldSpec 4, suggesting that the low-cost QEPro performed as well as the high-cost FieldSpec 4. This study provides a good example for evaluating spectral analysis by spectroscopy, which can support the development of the spectral method. Citation: Liu, Y.; Fan, P.; Qiu, H.; Li, X.; Hou, G. A Method to Evaluate Keywords: detection resolution; organic carbon; sediment; spectrometer; spectral analysis Spectral Analysis by Spectroscopy. Sensors 2022, 22, 5638. https:// doi.org/10.3390/s22155638 1. Introduction Academic Editor: Marc Brecht Visible and near infrared reflectance (Vis-NIR) spectroscopy has been widely used Received: 19 April 2022 to rapidly analyze many properties in various types of soils or sediments [1–12]. As the Accepted: 25 July 2022 primary chemical attribute, organic carbon in soils or sediments (SOC) is directly related Published: 28 July 2022 with Vis-NIR reflectance and has been mainly investigated for 20 years [13]. Past studies Publisher’s Note: MDPI stays neutral show that more effort was concentrated on how to establish an accurate spectral model of with regard to jurisdictional claims in SOC [1,3,13]. Therefore, many studies paid attention to the use of varied algorithms in the published maps and institutional affil- whole process of spectral analysis to accurately predict SOC using Vis-NIR spectroscopy. iations. For example, Liu et al. (2019) studied the spectral analysis of SOC in Chinese forests at the large (country) scale and found that the use of cubist after clustering could greatly improve the accuracy of SOC prediction compared with partial least squares regression (PLSR) [3]. At the local scale, PLSR was usually used to establish spectral models but Copyright: © 2022 by the authors. showed distinctive results in different studies [13]. Licensee MDPI, Basel, Switzerland. Obviously, spectrometers play an important role in spectral analysis [10,14–19]. During This article is an open access article the past three years, many studies have investigated the influences of spectrometers on distributed under the terms and the spectral analysis of SOC (Table 1). Results showed that different spectrometers could conditions of the Creative Commons produce varying spectral models, with distinctive R , RMSE and RPD or RPIQ, even for the Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ same samples (Table 1). For example, SOC in the Rothamsted long-term field experiments 4.0/). was analyzed by ASD FieldSpec 3 and Ocean Optics STS, respectively [15]. The spectral Sensors 2022, 22, 5638. https://doi.org/10.3390/s22155638 https://www.mdpi.com/journal/sensors Sensors 2022, 22, 5638 2 of 12 model derived from FieldSpec 3 had higher R and RPD, and lower RMSE than that from STS, and the difference became smaller outdoors compared with the laboratory (Table 1) [15]. The results also showed that laboratory spectrometers had a better performance than the field ones due to their technical advantages in terms of signal-to-noise ratio, stability of the measurement conditions, spectral resolution and wide spectral range [15,17,19]. For example, as mid-IR spectrometers, Vertex 70 (Bruker Optics, Ettlingen, Germany) and Agilent 4200 (Agilent Technologies, Santa Clara, CA, USA) are bench-top and portable, respectively [20]. SOC in an Australian farm was analyzed by Vertex 70 and Agilent 4200, respectively. The results showed that the spectral model derived from Vertex 70 was better than that from Agilent 4200 (Table 1) [20]. Table 1. Recent studies on spectral analysis of SOC by different spectrometers. Unit Spectrometers Spectral Models in Validation Set References ASD FieldSpec 4 R = 0.89, RMSE = 2.57, RPD = 3.04 gkg [14] Agilent 4300 R = 0.98, RMSE = 1.12, RPD = 7.01 ASD FieldSpec 3 R = 0.96, RMSE = 2.1, RPD = 5.4 gkg [15] Ocean Optics STS R = 0.94, RMSE = 2.4, RPD = 3.9 ASD FieldSpec 3 R = 0.89, RMSE = 3.9, RPD = 2.9 gkg [15] Ocean Optics STS R = 0.85, RMSE = 4.2, RPD = 2.6 Spectral Evolution R = 0.91, RMSE = 0.32, RPIQ = 1.31 FSR + 3500 % [16] Spectral Engines OY R = 0.80, RMSE = 0.46, RPIQ = 1.47 MEMS S2.2 Silver Spring RMSE = 0.23, RPIQ = 9.94 Foss XDS Ettlingen RMSE = 0.29, RPIQ = 8.01 gkg Bruker-TENSOR [17] ASD FieldSpec 3 RMSE = 0.83, RPIQ = 2.87 Agilent 4300 RMSE = 1.02, RPIQ = 2.28 East Norwalk R = 0.54, RMSE = 4.1 Ocean Optics USB2000 + gkg [10] Hamamatsu Photonics R = 0.49, RMSE = 4.5 C9914GB ASD AgriSpec R = 0.89, RMSE = 0.12 [18] NeoSpectra R = 0.78, RMSE = 0.16 Bruker Optics R = 0.96, RMSE = 0.17, RPIQ = 3.70 Vertex 70 % [19] Agilent 4200 R = 0.91, RMSE = 0.26, RPIQ = 2.46 ASD Labspec R = 0.88, RMSE = 0.30, RPIQ = 2.13 However, spectrometers in the comparison studies had distinct classifications, be- longing to near infrared and mid infrared instruments (Table 1) [14,17,19]. As shown in Table 1, Agilent 4200, Agilent 4300 (Agilent Technologies, Santa Clara, CA, USA) and Bruker are all mid infrared spectrometers (mid-IR); the others are near infrared spectrome- ters (near-IR). The same samples produced different reflectance spectra from the mid-IR spectrometers compared with near-IR. Therefore, because the spectrometers were different in type, the comparison could not really be shown as to which spectrometer was better in their performance on spectral analysis. Sensors 2022, 22, 5638 3 of 12 Studies also showed that ASD spectrometers are the most commonly used in the Vis- NIR spectral analysis due to their good performance and reputation [6–15,17–19]. However, ASD spectrometers are expensive, being a barrier to studies on Vis-NIR spectroscopy [13,18]. Fortunately, there are many other successful spectral analysis studies using other Vis-NIR spectrometers [14–19]. Crucil et al. (2019) analyzed SOC in the Rothamsted long-term field experiments by Ocean Optics STS, showing a good performance on spectral analysis [15]. Li et al. (2021) analyzed the total carbon of the intertidal sediments in the Laizhou Bay of China using QE65000 (Ocean Optics, USA) and achieved good prediction results, with r > 0.87 and RPD = 2.8 [4]. These studies suggested the importance of selecting a suitable spectrometer for a cer- tain specific research [10]. However, how to evaluate the performance of a spectrometer on spectral analysis is not clear. Evaluating a spectral model usually utilizes three parameters, including the determination coefficient (R ), the mean root square error (RMSE) and the residual predictive deviation (RPD) in the test set [20,21]. A good spectral model usually has high R and RPD, and low RMSE in the validation set [20,21]. If the RPD is higher than 2.0, the spectral model is considered to be excellent and can be used for accurate quantifi- cation [20]. This evaluation can only resolve whether the spectral model is appropriate for spectral analysis, but can not resolve for what samples the spectral analysis is suitable, such as the basic technical parameters, including the detection range, error, and resolution. Eskildsen et al. (2016) proposed a method to evaluate spectral analysis across spec- trometers [22]. Different from the old method comparing the difference between predicted and measured values, this method compared the median of the predicted values to evaluate spectral analysis, showing that this method was more useful than the old method [22]. These results suggest the importance of exploring the predicted values when evaluating spectral analysis [22]. Inspired by Eskildsen et al. (2016) [22], we make full use of the predicted SOC in spectral models to evaluate the spectral analysis of SOC in the Yellow Sea and Bohai Sea, China. We also compare the performance of three common near-IR spectrometers on the spectral analysis of SOC, including Agilent Cary 5000 (bench-top), ASD FieldSpec 4 and Ocean Optics QEPro. Suggested by the past studies (Table 1) [14,17,19], we hypothesized that the performance on spectral analysis would be ordered as Cary 5000 > FieldSpec 4 > QEPro. 2. Materials and Methods 2.1. Sampling and Preparation Ninety-six sediments were sampled in the stations of the Yellow Sea and Bohai Sea, China. For detailed sampling sites see Qiu et al. 2022 [23]. After freeze drying, the sediments were ground and sieved to pass a 0.5 mm mesh. Each sample was divided into two parts, one for chemical determination and the other for collecting the Vis-NIR reflectance spectra. SOC concentrations were measured using the Vario EL III Elemental Analyzer after the inorganic carbon was removed by hydrochloric acid. Specifically, 1 g of dried sample mixed with 1 mL hydrochloric acid (1 mol/L) was ultrasonic for 3 h to remove inorganic carbon, which was repeated 2 times. Organic carbon concentrations in these sediments ranged from 0.25% to 1.73%, with an average of 0.75% and a median of 0.67%. 2.2. Reflectance Spectra Collection The reflectance spectra were collected using Cary 5000 (Agilent Technology, Santa Clara, CA, USA), FieldSpec 4 (Analytical Spectral Devices, Boulder, CO, USA) and QEPro (Ocean Optics, Dunedin, FL, USA). The specifications of these spectrometers are shown in Table 2. Cary 5000 is a bench-top lab spectrometer, thus had the best specification. Both FieldSpec 4 and QEPro are portable and can be used both in the lab and field. The three spectrometers had different specifications with the price ordered as QEPro < Cary 5000 < FieldSpec 4. The wavelength range of both Cary 5000 and FieldSpec4 is 350–2500 nm, while that of QEPro is Sensors 2022, 22, 5638 4 of 12 200–1000 nm. QEPro had the largest stray light, without the detail of wavelength repeatability and accuracy. Table 2. The specifications of spectrometers used in this study. Features Cary 5000 FieldSpec4 QEPro Hamamatsu Photodiode and TE CCD (<1000 nm), Sensor back–thinned cooled PbS InGaAs (>1000 nm) FFT–CCD Detector Quartz window Probe, fiber optic Probe, fiber optic Wavelength range 350–2500 nm 350–2500 nm 200–1100 nm 3 nm (700 nm) Optical resolution 1 nm 0.3 nm 10 nm (1400 nm–2100 nm) Signal-to-noise >30,000 >10,000 1000 Integration time 100 ms 8 ms–15 min 100 ms 0.02% (<1000 nm) <0.08% (600 nm) Stray light <0.0002% (1420 nm) 0.01% (>1000 nm) 0.4% (435 nm) Wavelength <0.02 nm (>750 nm) 0.1 nm —- repeatability <0.005 nm (<750 nm) Wavelength <0.4 nm (>750 nm) 0.5 nm —- accuracy <0.08 nm (<750 nm) Spectra collection was controlled by the software of each spectrometer. Cary 5000 has a special diffuse reflectance module for powder samples (DRA-2500, Agilent), which can be used to collect the reflectance spectra of soils or sediments. Different from Cary 5000, both FieldSpec 4 and QEPro are fiber spectrometers, which collect reflectance spectra by optical fibers. In QEPro, the optical fiber was inserted into the 45 hole of the probe bracket (RPH-1, Ocean Optics). The sample box was under the probe bracket and formed a 45 angle between samples and the probe. The vertical distance between the probe and the sample was 0.5 mm. Some parameters were set up by software. In Cary 5000, the wavelength interval was 1 nm, the scanning speed was 600 nm/min, and 5 replicates were set for each sample. In FieldSpec 4, 3 replicates were set each time, with 100 times of dark current, 10 times of reference spectrum and 5 times per sample to collect the reflectance spectra, and the white board was used for calibration every 15 min while collecting the spectra. In QEPro with the slit of 10 m, the integration time was 600 ms, the sampling interval was 1 nm and each sample had 5 replicates of reflectance spectra. 2.3. Spectral Analysis Each sample had a spectral curve, whose x-axis was wavelength (nm) and y-axis was reflectance (%), and each sample had its SOC concentrations. Therefore, 96 samples had 96 curves and corresponding SOC concentrations. Using these data, the relationship (spectral model) between reflectance spectra (many x) and SOC concentrations (y) could be established. Specifically, 96 samples were divided into a calibration set and a validation set by Kennard–Stone algorithm (K-S) at the ratio of 2:1 [24]. Each sample had many x (reflectance), i.e., x1, x2, x3, . . . xn were the reflectances at each wavelength, and one y (SOC concentration). Qiu et al. (2022) reported that the spectral model established by the whole spectra had much better prediction ability of total carbon and nitrogen concen- trations, maybe because SOC concentrations were reflected by all the wavelengths [25]. Therefore, the spectral model was also established by the whole spectra without charac- teristic spectra extraction in this study [25]. Confalonieri et al. (2001) reported that partial least squares regression (PLSR) was successful in modelling soil nutrients by near infrared spectroscopy [25], as many other studies found [19,25]. PLSR was also found to be perfect for total carbon and nitrogen modelling in the same sediments, especially using the whole Sensors 2022, 22, x FOR PEER REVIEW 5 of 12 least squares regression (PLSR) was successful in modelling soil nutrients by near infrared spectroscopy [25], as many other studies found [19,25]. PLSR was also found to be perfect for total carbon and nitrogen modelling in the same sediments, especially using the whole Sensors 2022, 22, 5638 5 of 12 spectra [25]. Therefore, in this study PLSR was used to establish the spectral model of SOC. Lastly, the spectral model was evaluated by the validation set using the determina- tion coefficient (R ), the mean root square error (RMSE) and the residual predictive devi- spectra [25]. Therefore, in this study PLSR was used to establish the spectral model of SOC. ation (RPD) [21,22]. Lastly, the spectral model was evaluated by the validation set using the determination RPD = SD⁄RMSE = ∑ (𝑦 −𝑦 ) ∑ (𝑦 −𝑦 ) , in which 𝑦 is the coefficient (R ), the mean root square error (RMSE) and the residual predictive deviation (RPD) [21,22]. measured value of the samq ple, 𝑦 is the predicted value by the model, 𝑦 is the average 1 n 1 n RPD = SD/RMSE = (y y ) / (y y ) , in which y is the measured value of the sample and å n is the number of sa å mples. Rossel et al. (2006) pro- i i i i i n1 i=1 n1 i1 measur posed a qua ed value ntitati of ve rela the sample, tionship y ˆ between is the pr R edicted PD andvalue model q by u the alitmodel, y: RPD > 2. y is 0,the the average models i i measur or predict ed ivalue ons are per of thefsample ect and can and be use n is thed number for rapiof d an samples. alysis [21]. Rossel et al. (2006) proposed a quantitative All analysrielationship s was perform between ed in MAT RPD L and AB 2 model 017b. quality: Specific progr RPD >ams had been 2.0, the models puor b- predictions are perfect and can be used for rapid analysis [21]. lished in the Chinese software copyright (2018SR325853). A new program in R is being All analysis was performed in MATLAB 2017b. Specific programs had been published prepared. in the Chinese software copyright (2018SR325853). A new program in R is being prepared. 2.4. Evaluation of the Spectral Analysis 2.4. Evaluation of the Spectral Analysis Four parameters were proposed to evaluate the spectral analysis from different spec- Four parameters were proposed to evaluate the spectral analysis from different spec- trometers, including the detection resolution, range (minimum and maximum) and error. trometers, including the detection resolution, range (minimum and maximum) and error. These parameters were calculated according to the distribution of predicted SOC concen- These parameters were calculated according to the distribution of predicted SOC concen- trations (Figure 1). trations (Figure 1). Figure 1. The concept map of the parameters for evaluating spectral models. Figure 1. The concept map of the parameters for evaluating spectral models. These four parameters were calculated as shown in Figure 2. The distributions of pre- These four parameters were calculated as shown in Figure 2. The distributions of dicted SOC concentrations were fitted as a function for both calibration sets and validation predicted SOC concentrations were fitted as a function for both calibration sets and vali- sets, whose x-axis was SOC concentrations. Under the condition that the function is near dation sets, whose x-axis was SOC concentrations. Under the condition that the function to or equals a constant (e.g., 0), the x-values (SOC concentrations) are the minimum and is near to or equals a constant (e.g., 0), the x-values (SOC concentrations) are the minimum maximum (detection range). The detection resolution is the average difference between and maximum (detection range). The detection resolution is the average difference be- the two neighboring SOC concentrations. The detection errors are the difference between tween the two neighboring SOC concentrations. The detection errors are the difference the predicted SOC and the measured SOC, which can be plotted by the quartile map. The between the predicted SOC and the measured SOC, which can be plotted by the quartile quartile map can directly show the actual error including the largest error, the smallest map. The quartile map can directly show the actual error including the largest error, the error and the median error. smallest error and the median error. Sensors 2022, 22, x FOR PEER REVIEW 6 of 12 Sensors 2022, 22, 5638 6 of 12 Sensors 2022, 22, x FOR PEER REVIEW 6 of 12 Figure 2. The flow chart for calculating the proposed parameters. Figure 2. The flow chart for calculating the proposed parameters. Figure 2. The flow chart for calculating the proposed parameters. All analysis was performed in SPSS 13.0 (SPSS Inc., Cary, NC, USA); a new program All analysis was performed in SPSS 13.0 (SPSS Inc., Cary, NC, USA); a new program All analysis was performed in SPSS 13.0 (SPSS Inc., Cary, NC, USA); a new program in R is being prepared. Here, we stress the procedure of how to calculate detection range. in R is being prepared. Here, we stress the procedure of how to calculate detection range. in R is being prepared. Here, we stress the procedure of how to calculate detection range. In SPSS, the predicted SOC was first analyzed by “Graphs”-“Histogram”. Then, the dis- In SPSS, the predicted SOC was first analyzed by “Graphs”-“Histogram”. Then, the dis- In SPSS, the predicted SOC was first analyzed by “Graphs”-“Histogram”. Then, the tribution of predicted SOC was performed by “Analyze”-“Regression”-”Nonlinear”- tribution of predicted SOC was performed by “Analyze”-“Regression”-”Nonlinear”- distribution of predicted SOC was performed by “Analyze”-“Regression”-”Nonlinear”- “Peak-Gaussian” or “Analyze”-“Fit curve”-”Cubic”. Based on the function, the extreme “Peak-Gaussian” or “Analyze”-“Fit curve”-”Cubic”. Based on the function, the extreme “Peak-Gaussian” or “Analyze”-“Fit curve”-”Cubic”. Based on the function, the extreme could be calculated. could couldbe be c calculated. alculated. 3. Results 3. Results 3. Results 3.1. Reflectance Spectra of the Same Sediments in Different Spectrometers 3.1. Reflectance Spectra of the Same Sediments in Different Spectrometers 3.1. Reflectance Spectra of the Same Sediments in Different Spectrometers The reflectance spectra collected by the three spectrometers were different (Figures 3 The reflectance spectra collected by the three spectrometers were different (Figures 3 and 4). The reflectance spectra collected by the three spectrometers were different (Figures 3 and 4). As shown in Figure 3, QEPro only collected the reflectance spectra below 1000 nm, As shown in Figure 3, QEPro only collected the reflectance spectra below 1000 nm, while both and 4). As shown in Figure 3, QEPro only collected the reflectance spectra below 1000 nm, while both Cary 5000 and FieldSpec 4 collected the complete visible and near infrared Cary 5000 and FieldSpec 4 collected the complete visible and near infrared reflectance spectrum while both Cary 5000 and FieldSpec 4 collected the complete visible and near infrared reflectance spectrum (350–2500 nm). At 500–1000 nm, the reflectance was mainly ranged (350–2500 nm). At 500–1000 nm, the reflectance was mainly ranged at 20–50% in Cary 5000, reflectance spectrum (350–2500 nm). At 500–1000 nm, the reflectance was mainly ranged at 20–50% in Cary 5000, 15–45% in FieldSpec 4 and 20–60% in QEPro. All spectral curves 15–45% in FieldSpec 4 and 20–60% in QEPro. All spectral curves in Cary 5000 were more at 20–50% in Cary 5000, 15–45% in FieldSpec 4 and 20–60% in QEPro. All spectral curves in Cary 5000 were more compact than the portable spectrometers. compact than the portable spectrometers. in Cary 5000 were more compact than the portable spectrometers. Figure Figure 3. 3. TT hh ee rere flefclte ac ntcaenc spe s ectp raec otfrsae o dif m see ndi tsme in H nt usa i nn g H haiu aa nn dghai and Bohai Sea Bohai Sea of China by of China by different spec different trometers. Figure 3. The reflectance spectra of sediments in Huanghai and Bohai Sea of China by different spectrometers. As shown in Figure 4, the differences among these three spectrometers were direct spectrometers. and simple. The shapes of the spectral curves were the same, especially for Cary 5000 and As shown in Figure 4, the differences among these three spectrometers were direct FieldSpec 4. The reflectance was ordered as QEPro > Cary 5000 > FieldSpec 4. As shown in Figure 4, the differences among these three spectrometers were direct and simple. The shapes of the spectral curves were the same, especially for Cary 5000 and and simple. The shapes of the spectral curves were the same, especially for Cary 5000 and FieldSpec 4. The reflectance was ordered as QEPro > Cary 5000 > FieldSpec 4. FieldSpec 4. The reflectance was ordered as QEPro > Cary 5000 > FieldSpec 4. Sensors 2022, 22, x FOR PEER REVIEW 7 of 12 Sensors 2022, 22, x FOR PEER REVIEW 7 of 12 Sensors 2022, 22, 5638 7 of 12 QEPro QE Cary Pr5000 o Cary FieldS 5000 pec4 FieldSpec4 500 1000 1500 2000 2500 500 1000 1500 2000 2500 Wavelength (nm) Wavelength (nm) Figure 4. The average reflectance spectra of sediments in Huanghai and Bohai Sea of China by dif- Figure 4. The average reflectance spectra of sediments in Huanghai and Bohai Sea of China by dif- Figure 4. The average reflectance spectra of sediments in Huanghai and Bohai Sea of China by ferent spectrometers. ferent spectrometers. different spectrometers. 3.2. Spectral Analysis from Different Spectrometers 3.2. Spectral Analysis from Different Spectrometers 3.2. Spectral Analysis from Different Spectrometers The spectral analysis was highly different across spectrometers. In the calibration The spectral analysis was highly different across spectrometers. In the calibration sets, The spectral analysis was highly different across spectrometers. In the calibration sets, Cary 5000 had the best model with the largest R and the smallest RMSE, while QEPro Cary 5000 had the best model with the largest R and2the smallest RMSE, while QEPro had sets, Cary 5000 had the best model with the largest R and the smallest RMSE, while QEPro had the worst model with the smallest R and the largest RMSE (Figure 4). In the valida- the worst model with the smallest R and the 2 largest RMSE (Figure 4). In the validation sets, had the worst model with the smallest R and the largest RMSE (Figure 4). In the valida- tion sets, Cary 5000 also had the best model with the largest R , the smallest RMSE and Cary 5000 also had the best model with the largest R , the smallest RMSE and the highest tion sets, Cary 5000 also had the best model with the largest R , the smallest RMSE and the highest RPD, while FieldSpec 4 had the worst model with the smallest R , the largest RPD, while FieldSpec 4 had the worst model with the smallest R , the largest RMSE and the highest RPD, while FieldSpec 4 had the worst model with the smallest R , the largest RMSE and the lowest RPD (Figure 5). These results showed that Cary 5000 had the best the lowest RPD (Figure 5). These results showed that Cary 5000 had the best performance RMSE and the lowest RPD (Figure 5). These results showed that Cary 5000 had the best performance on the spectral analysis of SOC. on the spectral analysis of SOC. performance on the spectral analysis of SOC. a Calibration set in QEPro Validation set in QEPro 1.5 Calibration set in QEPro Validation set in QEPro 1.5 1.0 1.0 0.5 0.5 0.0 r =0.59, RMSE=0.18 r =0.75, RMSE=0.18 2 2 0.0 r =0.59, RMSE=0.18 r =0.75, RMSE=0.18 RPD=1.6 RPD=1.6 Calibration set in FieldSpec 4 Validation set in FieldSpec 4 1.5 Calibration set in FieldSpec 4 Validation set in FieldSpec 4 1.5 1.0 1.0 0.5 0.5 0.0 r =0.52, RMSE=0.19 r =0.98, RMSE=0.06 0.0 r =0.52, RMSE=0.19 RPD=1.4 r =0.98, RMSE=0.06 RPD=1.4 Calibration set in Cary 5000 Validation set in Cary 5000 1.5 Calibration set in Cary 5000 Validation set in Cary 5000 1.5 1.0 1.0 0.5 0.5 r =0.86, RMSE=0.11 0.0 r =0.99, RMSE=0.04 r RPD =0.86, RM =2.6 SE=0.11 r =0.99, RMSE=0.04 0.0 RPD=2.6 0.00.5 1.01.5 0.00.5 1.01.5 0.00.5 1.01.5 0.00.5 1.01.5 Measured concentrations (%) Measured concentrations (%) Measured concentrations (%) Measured concentrations (%) Figure 5. Results of spectral analysis on organic carbon concentrations in sediments of the Yellow Figure 5. Results of spectral analysis on organic carbon concentrations in sediments of the Yellow Sea Figure 5. Results of spectral analysis on organic carbon concentrations in sediments of the Yellow Sea and Bohai Sea, China (spectral models for both the calibration set and validation set in QEPro and Bohai Sea, China (spectral models for both the calibration set and validation set in QEPro (a,b), Sea and Bohai Sea, China (spectral models for both the calibration set and validation set in QEPro (a,b), FieldSpec 4 (c,d), and Cary 5000 (e,f), respectively). FieldSpec 4 (c,d), and Cary 5000 (e,f), respectively). (a,b), FieldSpec 4 (c,d), and Cary 5000 (e,f), respectively). Reflectance rates (%) Mode Mode lled co lled co ncen ncen trat tr io at ns ( ions ( %)%)Mo M de old led elled conc conc entent rati rons ations (%) (%)Mod Mod elled co elled co ncen ncen tratio tratio ns ( ns ( %)%) Reflectance rates (%) Sensors 2022, 22, x FOR PEER REVIEW 8 of 12 Sensors 2022, 22, x FOR PEER REVIEW 8 of 12 Sensors 2022, 22, 5638 8 of 12 3.3. Evaluation of the Spectral Analysis 3.3. Evaluation of the Spectral Analysis 3.3. Evaluation of the Spectral Analysis Firstly, the errors of spectral models both in the calibration set and the validation set Firstly, the errors of spectral models both in the calibration set and the validation set Firstly, the errors of spectral models both in the calibration set and the validation set were the smallest from Cary 5000 but the largest from QEPro (Figure 6). These results were the smallest from Cary 5000 but the largest from QEPro (Figure 6). These results were the smallest from Cary 5000 but the largest from QEPro (Figure 6). These results were were consistent with results of the spectral analysis (Figure 5). For the portable spectrom- were consistent with results of the spectral analysis (Figure 5). For the portable spectrom- consistent with results of the spectral analysis (Figure 5). For the portable spectrometers, eters, the errors in FieldSpec 4 increased from the calibration set to the validation set, but eters, the errors in FieldSpec 4 increased from the calibration set to the validation set, but the errors in FieldSpec 4 increased from the calibration set to the validation set, but errors errors decreased in QEPro (Figure 6), showing that QEPro had a consistent performance errors decreased in QEPro (Figure 6), showing that QEPro had a consistent performance decreased in QEPro (Figure 6), showing that QEPro had a consistent performance on on spectral analysis. on spectral analysis. spectral analysis. a b Calibration set Validation set Calibration set Validation set 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 −0.2 −0.2 −0.4 −0.4 Cary5000 QEPro Cary5000 QEPro Cary5000 FieldSpec4 QEPro Cary5000 FieldSpec4 QEPro FieldSpec4 FieldSpec4 Figure 6. The quartile map of the absolute error for predicted organic carbon concentrations in the Figure Figure 6. 6. The The q quartile uartile map of the absolute er map of the absolute err ror for pr or for predicted edicted org organic anic carb carbon on concentrati concentrations ons in the in the calibration set and validation set (absolute errors in the calibration set (a) and validation set (b)). calibration set and validation set (absolute errors in the calibration set (a) and validation set (b)). calibration set and validation set (absolute errors in the calibration set (a) and validation set (b)). Secondly, the spectral curves were close to a normal distribution, except for the cali- Secondly, the spectral curves were close to a normal distribution, except for the Secondly, the spectral curves were close to a normal distribution, except for the cali- calibration bration set of set FieldSpec 4 (Figure of FieldSpec 4 (Figur7 e). C 7). a Cary ry 5000 had 5000 had a a wider wider detection ran detection ran gge e oof f SOC than SOC than bration set of FieldSpec 4 (Figure 7). Cary 5000 had a wider detection range of SOC than the the portable portable spectr spectrometers ometers (Figu (Figure re 7). Th 7). The e calibr calibration ation set set in QEP in QEPr ro o h had ad a a w wider ider de detection tection the portable spectrometers (Figure 7). The calibration set in QEPro had a wider detection range than the validation set, showing a consistent detection range of SOC (Figure 7). These range than the validation set, showing a consistent detection range of SOC (Figure 7). range than the validation set, showing a consistent detection range of SOC (Figure 7). rThese re esults arsults e consistent are cons with istent with those those in Table in 3 Table . 3. These results are consistent with those in Table 3. QEPro Cary5000 FieldSpec4 Cary5000 FieldSpec4 QEPro a c a c d f d f 8 e TOC concentrations (%) TOC concentrations (%) Figure 7. The distribution curves of predicted organic carbon concentrations by different spectral Figure 7. The distribution curves of predicted organic carbon concentrations by different spectral Figure 7. The distribution curves of predicted organic carbon concentrations by different spectral models models established by both calibration set and validation set using different spectrometers (distri- models established by both calibration set and validation set using different spectrometers (distri- established by both calibration set and validation set using different spectrometers (distribution curves bution curves from the calibration set by Cary 5000 (a), FieldSpec 4 (b), and QEPro (c), respectively; bution curves from the calibration set by Cary 5000 (a), FieldSpec 4 (b), and QEPro (c), respectively; from the calibration set by Cary 5000 (a), FieldSpec 4 (b), and QEPro (c), respectively; distribution curves distribution curves from the validation set by Cary 5000 (d), FieldSpec 4 (e), and QEPro (f), respec- distribution curves from the validation set by Cary 5000 (d), FieldSpec 4 (e), and QEPro (f), respec- from the validation set by Cary 5000 (d), FieldSpec 4 (e), and QEPro (f), respectively). tively). tively). 0 0.2 .2 0 0.4 .4 0.6 0.6 0 0.8 .8 1 1..0 0 1.2 1.2 1 1..4 4 1 1.6 .6 1.8 1.8 0.2 0.2 0 0.4 .4 0 0.6 .6 0.8 0.8 1 1.0 .0 1 1..2 2 1.4 1.4 1 1..6 6 1 1.8 .8 0 0.2 .2 0.4 0.4 0 0.6 .6 0 0.8 .8 1.0 1.0 1 1.2 .2 1 1..4 4 1.6 1.6 1 1..8 8 Calibration set Absolute error (%) Absolute error (%) Validation set Calibration set Validation set Sensors 2022, 22, 5638 9 of 12 Table 3. The parameters to evaluate spectral models in both calibration set (C) and validation set (V), respectively. Cary 5000 FieldSpec 4 QEPro Parameters C V C V C V Max 2.41 2.38 2.90 1.69 2.14 2.03 Min 0.01 0.00 0.00 0.00 0.01 0.00 Error 0.03 0.09 0.05 0.14 0.14 0.14 Resolution 0.006 0.005 0.005 0.004 0.006 0.004 Thirdly, the detection resolutions were all less than 0.01% for the three spectrometers, and were not significantly different across spectrometers (Table 3). All spectrometers increased their resolution from the calibration set to the validation set, especially QEPro (Table 3). Clearly, Cary 5000 could predict a larger range of SOC concentrations with the smallest error and highest resolution (Table 3). FieldSpec 4 had better performance in the calibration set but lost the advantage in the validation set (Table 3). Although the technical parameters of spectral analysis in QEPro had no advantages, they were consistent in both the calibration set and validation set (Figure 7, Table 2). 4. Discussion We studied the spectral analysis of SOC in the Yellow Sea and Bohai Sea of China using Agilent Cary 5000, ASD FieldSpec 4 and Ocean Optics QEPro. We proposed four parameters to evaluate the spectral analysis of SOC, including the detection range (max and min), resolution and error. These parameters were used to evaluate the performance of the spectrometers on spectral analysis. 4.1. Advantages of the Proposed Method on Evaluation of Spectral Analysis by Spectrometers Results showed a large difference in the spectral analysis among the three spectrome- ters. Consistent with our hypothesis, the portable spectrometers had the worse spectral analysis than the bench-top spectrometer (Figures 6 and 7, Table 3). This was mainly due to the large difference in specification, including spectral resolution, system signal- to-noise ratio, stray light, detector, wavelength repeatability, wavelength accuracy and so on (Table 2) [15,16,19]. In this study, Cary 5000 had better specifications than the other two portable spectrometers, especially in terms of the signal-to-noise ratio, stray light, wavelength repeatability and wavelength accuracy (Table 2). Although it was easily hy- pothesized, we showed the quantitative differences between the spectrometers, making the evaluation quantitatively. These results are consistent with the past studies (Table 1) [15,16,19]. Sharififar et al. (2019) analyzed SOC on an Australian farm using Brucker Optik Vertex 70 (bench-top) and Agilent 4200 (portable) and found that the spectral model derived from Vertex 70 was better than that from Agilent 4200 (Table 1) [19]. Linderhom et al. (2019) used ASD LabSpec and two other portable spectrometers to classify different soils and found that the lab spectrometer had a better classification ability than the portable spectrometers, mainly because the lab spectrometer had a wider spectral range and a better detector [26]. Unfortunately, these were qualitative studied, so we could not compare their performance on the spectral analysis by the technical parameters shown in Table 3. This study also provided a scientific basis for evaluating spectral analysis. Although the combination of R , RMSE and RPD can evaluate whether a spectral analysis is better, it is difficult to quantify how much better it is. Using the proposed parameters, we could clearly understand which samples would be suitable for the spectral technique we established, such as its detection range, resolution and error. That is, we could accurately define samples by the concentrations that could be determined (detection range) and differentiated (detection resolution) at what errors. In this study, if the SOC concentration of a sediment sampled from the Yellow Sea and Bohai Sea of China is about 2.15%, it would be more accurately Sensors 2022, 22, 5638 10 of 12 analyzed by Cary 5000. If the SOC concentration is about 2.00% and the detection error is not concerning, QEPro would be more suitable. Furthermore, if we develop a technique or an equipment that can accurately and rapidly analyze SOC based on vis-NIR spectroscopy, we should present its technical parameters, in which the proposed four parameters are the most basic. Therefore, the proposed four parameters are valuable for the equipment or a technique based on Vis-NIR spectroscopy. Spectral models are highly dependent on modelling samples; therefore, further study should pay attention to the accurate and rapid simulation of the distribution curve of the predicted SOC concentrations. 4.2. Performance of the Portable Spectrometers on Spectral Analysis Inconsistent with our hypothesis, the portable spectrometers had a similar perfor- mance on spectral analysis. The detection range, error and resolution in FieldSpec 4 were not significantly different from those in QEPro (Figures 6 and 7, Table 3), but QEPro had a more consistent detection range (Figure 7, Table 3) and errors (Figure 6, Table 3) in both the calibration set and the validation set than FieldSpec 4. This suggests that QEPro has a more stable performance than FieldSpec 4, at least in this study. Yahaya et al. (2015) found that the spectral model established by the QE65000 spec- trometer could be transferred to the FieldSpec 3 spectrometer more easily and in a better way than another portable spectrometer, suggesting that the two spectrometers could be replaced by each other [27]. In this study, QEPro and FieldSpec 4 were the upgraded spec- trometers of QE65000 and FieldSpec 3, respectively, suggesting that QEPro and FieldSpec 4 also could replace each other. These results indicate that if the error is not the main consideration, QEPro is a good choice due to its high stability and low price, suitable for long-term and in situ (online) use. However, if we want to establish a rapid spectral measurement technology to replace the traditional lab chemical analysis, Cary 5000 is undoubtedly a better choice. This study provided a direct reference for selecting different spectrometers to conduct effective scientific research. 5. Conclusions We studied the spectral analysis of organic carbon in sediments by three spectrometers and proposed four parameters to evaluate the performance of these spectrometers. The results showed that the bench-top spectrometer had better performance on spectral analysis than the portable spectrometers, mainly due to their differences in specification, such as signal-to-noise ratio, stray light, detector, wavelength repeatability, wavelength accuracy and so on. For the two portable spectrometers, FieldSpec 4 had the best performance in the calibration set, but lost its advantage in the validation set. QEPro had ordinary technical parameters on spectral analysis, but was consistent in both the calibration set and validation set. These results suggest that if the stability is the main consideration, QEPro is a good choice for long-term and in situ use due to its low price. This study provides a direct reference for selecting different spectrometers to conduct effective scientific research, and also provides a scientific basis for evaluating the spectral analysis. 6. Patents Chinese invention: A method for obtaining the performance of a spectrometer (CN202011288620.1). Author Contributions: Y.L. conducted the experiment and draft preparation, X.L. taught the data analysis, H.Q. and G.H. prepared the samples, P.F. designed the experiment and improved the manuscript. All authors have read and agreed to the published version of the manuscript. Funding: This research was supported by the National Natural Science Foundation of China (32171578, U2006209), and Shandong Natural Science Foundation, China (ZR 2021MD093, ZR2021MD103, ZR2021FQ028). Sensors 2022, 22, 5638 11 of 12 Data Availability Statement: https://www.researchgate.net/publication/362290727_Reflectance_rates_ and_orgainc_carbon_concentrations_of_the_sediments_in_the_Yellow_Sea_and_Bohai_Sea_China (accessed on 18 April 2022). Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. References 1. Rossel, R.A.V.; Webster, R. Predicting soil properties from the Australian soil visible-near infrared spectroscopic database. Eur. J. Soil Sci. 2012, 63, 848–860. [CrossRef] 2. Rial, M.; Cortizas, M.; Rodríguez-Lado, L. Understanding the spatial distribution of factors controlling topsoil organic carbon content in European soils. Sci. Total Environ. 2017, 609, 1411–1422. [CrossRef] [PubMed] 3. Liu, S.; Shen, H.; Chen, S.; Zhao, X.; Biswas, A.; Jia, X.; Shi, Z.; Fang, J. Estimating forest soil organic carbon content using vis-NIR spectroscopy: Implications for large-scale soil carbon spectroscopic assessment. Geoderma 2019, 348, 37–44. [CrossRef] 4. Li, X.; Li, Z.; Hou, G.; Qiu, H.; Lv, H.; Chen, G.; Fan, P. Carbon content detection of marine sediments based on multispectral fusion. Spectrosc. Spect. Anal. 2021, 41, 2898–2903. 5. Gholizadeh, A.; Rossel, R.A.V.; Saberioon, M.; Boruvka, L.; Kratina, J.; Pavlu, ˚ L. National-scale spectroscopic assessment of soil organic carbon in forests of the Czech Republic. Geoderma 2021, 385, 114832. [CrossRef] 6. Knox, N.M.; Grunwald, S.; McDowell, M.L.; Bruland, G.L.; Myers, D.B.; Harris, W.G. Modelling soil carbon fractions with visible near-infrared (VNIR) and mid-infrared (MIR) spectroscopy. Geoderma 2015, 239–240, 229–239. [CrossRef] 7. Vestergaard, R.-J.; Vasava, H.B.; Aspinall, D.; Chen, S.; Gillespie, A.; Adamchuk, V.; Biswas, A. Evaluation of optimized preprocessing and modelling algorithms for prediction of soil properties using vis-NIR spectroscopy. Sensors 2021, 20, 6745. [CrossRef] 8. Sarathjith, M.C.; Das, B.S.; Wani, S.P.; Sahrawat, K.L. Variable indicators for optimum wavelength selection in diffuse reflectance spectroscopy of soils. Geoderma 2016, 267, 1–9. [CrossRef] 9. Dotto, A.C.; Dalolin, R.S.D.; Grunwald, S.; Caten, A.; Filho, W.P. Two preprocessing techniques to reduce model covariables in soil property predictions by Vis-NIR spectroscopy. Soil Till. Res. 2017, 172, 59–68. [CrossRef] 10. Dhawale, N.M.; Adamchuk, V.I.; Prasher, S.O.; Rossel, R.A.V.; Ismail, A.A. Evaluation of two portable hyperspectral-sensor-based instruments to predict key soil properties in Canadian soils. Sensors 2022, 22, 2556. [CrossRef] 11. Ng, W.; Minasny, B.; Montazerolghaem, M.; Padarian, J.; Ferguson, R.; Bailey, S.; McBratney, A.B. Convolutional neural network for simultaneous prediction of several soil properties using visible/near-infrared, mid-infrared, and their combined spectra. Geoderma 2019, 352, 251–267. [CrossRef] 12. Zheng, G.; Ryu, D.; Jiao, C.; Xie, X.; Shang, G. Visible and near-infrared reflectance spectroscopy analysis of a coastal soil chronosequence. Remote Sens. 2019, 11, 2336. [CrossRef] 13. Xu, S.; Zhao, Y.; Wang, M.; Shi, X. Comparison of multivariate methods for estimating selected soil properties from intact soil cores of paddy fields by Vis-NIR spectroscopy. Geoderma 2018, 310, 29–43. [CrossRef] 14. Semella, S.; Hutengs, C.; Seidel, M.; Ulrich, M.; Schneider, B.; Ortner, M.; Thiele-Bruhn, S.; Ludwig, B.; Vohland, M. Accuracy and reproducibility of laboratory diffuse reflectance measurements with portable VNIR and MIR spectrometers for predictive soil organic carbon modelling. Sensors 2022, 22, 2749. [CrossRef] 15. Crucil, G.; Castaldi, F.; Aldana-Jague, E.; van Wesemael, B.; Macdonald, A.; Oost, K.V. Assessing the performance of UAS- Compatible multispectral and hyperspectral sensors for soil organic carbon prediction. Sustainability 2019, 11, 1889. [CrossRef] 16. Karyotis, K.; Angelopoulou, T.; Tziolas, N.; Palaiologou, E.; Samarinas, N.; Zalidis, G. Evaluation of a micro-electro mechanical systems spectral sensor for soil properties estimation. Land 2021, 10, 63. [CrossRef] 17. Greenberg, I.; Seidel, M.; Vohland, M.; Ludwig, B. Performance of field-scale lab vs in situ visible/near- and mid-infrared spectroscopy for estimation of soil properties. Eur. J. Soil Sci. 2022, 73, e13180. [CrossRef] 18. Sharififar, A.; Singh, K.; Jones, E.; Ginting, F.I.; Minasny, B. Evaluating a low-cost portable NIR spectrometer for the prediction of soil organic and total carbon using different calibration models. Soil Use Manag. 2019, 35, 607–616. [CrossRef] 19. Li, S.; Rossel, R.A.V.; Webster, R. The cost-effectiveness of reflectance spectroscopy for estimating soil organic carbon. Eur. J. Soil Sci. 2022, 73, e13202. [CrossRef] 20. Rossel, R.A.V.; Mcglynn, R.N.; Mcbratney, A.B. Determining the composition of mineral-organic mixes using UV-vis-NIR diffuse reflectance spectroscopy. Geoderma 2006, 137, 70–82. [CrossRef] 21. Summers, D.; Lewis, M.; Ostendorf, B.; Lewis, M. Visible near-infrared reflectance spectroscopy as a predictive indicator of soil properties. Ecol. Indic. 2011, 11, 123–131. [CrossRef] 22. Eskildsen, C.E.; Hansen, P.W.; Skov, T.; Marini, F.; Nøgaarda, L. Evaluation of multivariate calibration models transferred between spectroscopic instruments: Applied to near infrared measurements of flour samples. J. Near Infrared Spec. 2016, 24, 151–156. [CrossRef] Sensors 2022, 22, 5638 12 of 12 23. Qiu, H.; Fan, P.; Hou, G.; Li, X.; Wang, Y. Analysis and model comparison of carbon and nitrogen concentrations in sediments of the Yellow Sea and Bohai Sea by visible-near infrared spectroscopy. Bull. Environ. Contam. Toxicol. 2022, 108, 1124–1131. [CrossRef] [PubMed] 24. Galvão, R.K.; Araujo, M.C.; José, G.E.; Pontes, M.J.C.; Silva, E.C.; Saldanha, T.C.B. A method for calibration and validation subset partitioning. Talanta 2005, 67, 736–740. [CrossRef] 25. Confalonieri, M.; Fornasier, F.; Ursino, A.; Boccardi, F.; Pintus, B.; Odoardi, M. The potential of near infrared reflectance spectroscopy as a tool for the chemical characterisation of agricultural soils. J. Near Infrared Spec. 2001, 9, 123–131. [CrossRef] 26. Linderholm, J.; Geladi, P.; Gorretta, N.; Bendoula, R.; Gobrecht, A. Near infrared and hyperspectral studies of archaeological stratigraphy and statistical considerations. Geoarchaeology 2019, 34, 311–321. [CrossRef] 27. Yahaya, O.K.M.; MatJafri, M.Z.; Aziz, A.A.; Omar, A.F. Visible spectroscopy calibration transfer model in determining pH of Sala mangoes. J. Instrum. 2015, 10, T05002. [CrossRef] http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Sensors Multidisciplinary Digital Publishing Institute

A Method to Evaluate Spectral Analysis by Spectroscopy

Liu, Yan; Fan, Pingping; Qiu, Huimin; Li, Xueying; Hou, Guangli
Sensors , Volume 22 (15) – Jul 28, 2022
Download PDF
12 pages

Loading next page...
 
/lp/multidisciplinary-digital-publishing-institute/a-method-to-evaluate-spectral-analysis-by-spectroscopy-w03JhYn03W

References (27)

Publisher
Multidisciplinary Digital Publishing Institute
Copyright
© 1996-2022 MDPI (Basel, Switzerland) unless otherwise stated Disclaimer The statements, opinions and data contained in the journals are solely those of the individual authors and contributors and not of the publisher and the editor(s). MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Terms and Conditions Privacy Policy
ISSN
1424-8220
DOI
10.3390/s22155638
Publisher site
See Article on Publisher Site

Abstract

sensors Article 1 , 2 1 , 2 , 1 , 2 1 , 2 1 , 2 Yan Liu , Pingping Fan * , Huimin Qiu , Xueying Li and Guangli Hou Institute of Oceanographic Instrumentation, Qilu University of Technology, Qingdao 266061, China; [email protected] (Y.L.); [email protected] (H.Q.); [email protected] (X.L.); [email protected] (G.H.) Shandong Key Lab of Marine Environment Monitoring Technology, Qingdao 266061, China * Correspondence: [email protected]; Tel.: +86-186-6982-3823 Abstract: Visible and near infrared spectroscopy has been widely used to develop a method for rapidly determining organic carbon in soils or sediments (SOC). Most of these studies concentrated on how to establish a good spectral model but ignored how to evaluate the method, such as the use of detection range (max and min), resolution and error for SOC spectral analysis. Here, we proposed a method to evaluate the spectral analysis of SOC. Using 96 sediments sampled in the Yellow Sea and Bohai Sea, China, we established three spectral models of SOC after collecting their spectral reflectance by Agilent Cary 5000, ASD FieldSpec 4 and Ocean Optics QEPro, respectively. For both the calibration set and validation set in each spectrometer, the predicted SOC concentrations followed a distribution curve (function), in which the x-axis was the SOC concentrations. Using these curves, we developed these four technical parameters. The detection ranges were the SOC concentrations where the curve was near to or crossing with the lateral axis, while the detection resolution was the average difference between the two neighboring SOC concentrations. The detection errors were the differences between the predicted SOC and the measured SOC. The results show that these technical parameters were better in the bench-top spectrometer (Cary 5000) than those in the portable spectrometers when analyzing the same samples. For the portable spectrometers, QEPro had a broader detection range and more consistent detection error than FieldSpec 4, suggesting that the low-cost QEPro performed as well as the high-cost FieldSpec 4. This study provides a good example for evaluating spectral analysis by spectroscopy, which can support the development of the spectral method. Citation: Liu, Y.; Fan, P.; Qiu, H.; Li, X.; Hou, G. A Method to Evaluate Keywords: detection resolution; organic carbon; sediment; spectrometer; spectral analysis Spectral Analysis by Spectroscopy. Sensors 2022, 22, 5638. https:// doi.org/10.3390/s22155638 1. Introduction Academic Editor: Marc Brecht Visible and near infrared reflectance (Vis-NIR) spectroscopy has been widely used Received: 19 April 2022 to rapidly analyze many properties in various types of soils or sediments [1–12]. As the Accepted: 25 July 2022 primary chemical attribute, organic carbon in soils or sediments (SOC) is directly related Published: 28 July 2022 with Vis-NIR reflectance and has been mainly investigated for 20 years [13]. Past studies Publisher’s Note: MDPI stays neutral show that more effort was concentrated on how to establish an accurate spectral model of with regard to jurisdictional claims in SOC [1,3,13]. Therefore, many studies paid attention to the use of varied algorithms in the published maps and institutional affil- whole process of spectral analysis to accurately predict SOC using Vis-NIR spectroscopy. iations. For example, Liu et al. (2019) studied the spectral analysis of SOC in Chinese forests at the large (country) scale and found that the use of cubist after clustering could greatly improve the accuracy of SOC prediction compared with partial least squares regression (PLSR) [3]. At the local scale, PLSR was usually used to establish spectral models but Copyright: © 2022 by the authors. showed distinctive results in different studies [13]. Licensee MDPI, Basel, Switzerland. Obviously, spectrometers play an important role in spectral analysis [10,14–19]. During This article is an open access article the past three years, many studies have investigated the influences of spectrometers on distributed under the terms and the spectral analysis of SOC (Table 1). Results showed that different spectrometers could conditions of the Creative Commons produce varying spectral models, with distinctive R , RMSE and RPD or RPIQ, even for the Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ same samples (Table 1). For example, SOC in the Rothamsted long-term field experiments 4.0/). was analyzed by ASD FieldSpec 3 and Ocean Optics STS, respectively [15]. The spectral Sensors 2022, 22, 5638. https://doi.org/10.3390/s22155638 https://www.mdpi.com/journal/sensors Sensors 2022, 22, 5638 2 of 12 model derived from FieldSpec 3 had higher R and RPD, and lower RMSE than that from STS, and the difference became smaller outdoors compared with the laboratory (Table 1) [15]. The results also showed that laboratory spectrometers had a better performance than the field ones due to their technical advantages in terms of signal-to-noise ratio, stability of the measurement conditions, spectral resolution and wide spectral range [15,17,19]. For example, as mid-IR spectrometers, Vertex 70 (Bruker Optics, Ettlingen, Germany) and Agilent 4200 (Agilent Technologies, Santa Clara, CA, USA) are bench-top and portable, respectively [20]. SOC in an Australian farm was analyzed by Vertex 70 and Agilent 4200, respectively. The results showed that the spectral model derived from Vertex 70 was better than that from Agilent 4200 (Table 1) [20]. Table 1. Recent studies on spectral analysis of SOC by different spectrometers. Unit Spectrometers Spectral Models in Validation Set References ASD FieldSpec 4 R = 0.89, RMSE = 2.57, RPD = 3.04 gkg [14] Agilent 4300 R = 0.98, RMSE = 1.12, RPD = 7.01 ASD FieldSpec 3 R = 0.96, RMSE = 2.1, RPD = 5.4 gkg [15] Ocean Optics STS R = 0.94, RMSE = 2.4, RPD = 3.9 ASD FieldSpec 3 R = 0.89, RMSE = 3.9, RPD = 2.9 gkg [15] Ocean Optics STS R = 0.85, RMSE = 4.2, RPD = 2.6 Spectral Evolution R = 0.91, RMSE = 0.32, RPIQ = 1.31 FSR + 3500 % [16] Spectral Engines OY R = 0.80, RMSE = 0.46, RPIQ = 1.47 MEMS S2.2 Silver Spring RMSE = 0.23, RPIQ = 9.94 Foss XDS Ettlingen RMSE = 0.29, RPIQ = 8.01 gkg Bruker-TENSOR [17] ASD FieldSpec 3 RMSE = 0.83, RPIQ = 2.87 Agilent 4300 RMSE = 1.02, RPIQ = 2.28 East Norwalk R = 0.54, RMSE = 4.1 Ocean Optics USB2000 + gkg [10] Hamamatsu Photonics R = 0.49, RMSE = 4.5 C9914GB ASD AgriSpec R = 0.89, RMSE = 0.12 [18] NeoSpectra R = 0.78, RMSE = 0.16 Bruker Optics R = 0.96, RMSE = 0.17, RPIQ = 3.70 Vertex 70 % [19] Agilent 4200 R = 0.91, RMSE = 0.26, RPIQ = 2.46 ASD Labspec R = 0.88, RMSE = 0.30, RPIQ = 2.13 However, spectrometers in the comparison studies had distinct classifications, be- longing to near infrared and mid infrared instruments (Table 1) [14,17,19]. As shown in Table 1, Agilent 4200, Agilent 4300 (Agilent Technologies, Santa Clara, CA, USA) and Bruker are all mid infrared spectrometers (mid-IR); the others are near infrared spectrome- ters (near-IR). The same samples produced different reflectance spectra from the mid-IR spectrometers compared with near-IR. Therefore, because the spectrometers were different in type, the comparison could not really be shown as to which spectrometer was better in their performance on spectral analysis. Sensors 2022, 22, 5638 3 of 12 Studies also showed that ASD spectrometers are the most commonly used in the Vis- NIR spectral analysis due to their good performance and reputation [6–15,17–19]. However, ASD spectrometers are expensive, being a barrier to studies on Vis-NIR spectroscopy [13,18]. Fortunately, there are many other successful spectral analysis studies using other Vis-NIR spectrometers [14–19]. Crucil et al. (2019) analyzed SOC in the Rothamsted long-term field experiments by Ocean Optics STS, showing a good performance on spectral analysis [15]. Li et al. (2021) analyzed the total carbon of the intertidal sediments in the Laizhou Bay of China using QE65000 (Ocean Optics, USA) and achieved good prediction results, with r > 0.87 and RPD = 2.8 [4]. These studies suggested the importance of selecting a suitable spectrometer for a cer- tain specific research [10]. However, how to evaluate the performance of a spectrometer on spectral analysis is not clear. Evaluating a spectral model usually utilizes three parameters, including the determination coefficient (R ), the mean root square error (RMSE) and the residual predictive deviation (RPD) in the test set [20,21]. A good spectral model usually has high R and RPD, and low RMSE in the validation set [20,21]. If the RPD is higher than 2.0, the spectral model is considered to be excellent and can be used for accurate quantifi- cation [20]. This evaluation can only resolve whether the spectral model is appropriate for spectral analysis, but can not resolve for what samples the spectral analysis is suitable, such as the basic technical parameters, including the detection range, error, and resolution. Eskildsen et al. (2016) proposed a method to evaluate spectral analysis across spec- trometers [22]. Different from the old method comparing the difference between predicted and measured values, this method compared the median of the predicted values to evaluate spectral analysis, showing that this method was more useful than the old method [22]. These results suggest the importance of exploring the predicted values when evaluating spectral analysis [22]. Inspired by Eskildsen et al. (2016) [22], we make full use of the predicted SOC in spectral models to evaluate the spectral analysis of SOC in the Yellow Sea and Bohai Sea, China. We also compare the performance of three common near-IR spectrometers on the spectral analysis of SOC, including Agilent Cary 5000 (bench-top), ASD FieldSpec 4 and Ocean Optics QEPro. Suggested by the past studies (Table 1) [14,17,19], we hypothesized that the performance on spectral analysis would be ordered as Cary 5000 > FieldSpec 4 > QEPro. 2. Materials and Methods 2.1. Sampling and Preparation Ninety-six sediments were sampled in the stations of the Yellow Sea and Bohai Sea, China. For detailed sampling sites see Qiu et al. 2022 [23]. After freeze drying, the sediments were ground and sieved to pass a 0.5 mm mesh. Each sample was divided into two parts, one for chemical determination and the other for collecting the Vis-NIR reflectance spectra. SOC concentrations were measured using the Vario EL III Elemental Analyzer after the inorganic carbon was removed by hydrochloric acid. Specifically, 1 g of dried sample mixed with 1 mL hydrochloric acid (1 mol/L) was ultrasonic for 3 h to remove inorganic carbon, which was repeated 2 times. Organic carbon concentrations in these sediments ranged from 0.25% to 1.73%, with an average of 0.75% and a median of 0.67%. 2.2. Reflectance Spectra Collection The reflectance spectra were collected using Cary 5000 (Agilent Technology, Santa Clara, CA, USA), FieldSpec 4 (Analytical Spectral Devices, Boulder, CO, USA) and QEPro (Ocean Optics, Dunedin, FL, USA). The specifications of these spectrometers are shown in Table 2. Cary 5000 is a bench-top lab spectrometer, thus had the best specification. Both FieldSpec 4 and QEPro are portable and can be used both in the lab and field. The three spectrometers had different specifications with the price ordered as QEPro < Cary 5000 < FieldSpec 4. The wavelength range of both Cary 5000 and FieldSpec4 is 350–2500 nm, while that of QEPro is Sensors 2022, 22, 5638 4 of 12 200–1000 nm. QEPro had the largest stray light, without the detail of wavelength repeatability and accuracy. Table 2. The specifications of spectrometers used in this study. Features Cary 5000 FieldSpec4 QEPro Hamamatsu Photodiode and TE CCD (<1000 nm), Sensor back–thinned cooled PbS InGaAs (>1000 nm) FFT–CCD Detector Quartz window Probe, fiber optic Probe, fiber optic Wavelength range 350–2500 nm 350–2500 nm 200–1100 nm 3 nm (700 nm) Optical resolution 1 nm 0.3 nm 10 nm (1400 nm–2100 nm) Signal-to-noise >30,000 >10,000 1000 Integration time 100 ms 8 ms–15 min 100 ms 0.02% (<1000 nm) <0.08% (600 nm) Stray light <0.0002% (1420 nm) 0.01% (>1000 nm) 0.4% (435 nm) Wavelength <0.02 nm (>750 nm) 0.1 nm —- repeatability <0.005 nm (<750 nm) Wavelength <0.4 nm (>750 nm) 0.5 nm —- accuracy <0.08 nm (<750 nm) Spectra collection was controlled by the software of each spectrometer. Cary 5000 has a special diffuse reflectance module for powder samples (DRA-2500, Agilent), which can be used to collect the reflectance spectra of soils or sediments. Different from Cary 5000, both FieldSpec 4 and QEPro are fiber spectrometers, which collect reflectance spectra by optical fibers. In QEPro, the optical fiber was inserted into the 45 hole of the probe bracket (RPH-1, Ocean Optics). The sample box was under the probe bracket and formed a 45 angle between samples and the probe. The vertical distance between the probe and the sample was 0.5 mm. Some parameters were set up by software. In Cary 5000, the wavelength interval was 1 nm, the scanning speed was 600 nm/min, and 5 replicates were set for each sample. In FieldSpec 4, 3 replicates were set each time, with 100 times of dark current, 10 times of reference spectrum and 5 times per sample to collect the reflectance spectra, and the white board was used for calibration every 15 min while collecting the spectra. In QEPro with the slit of 10 m, the integration time was 600 ms, the sampling interval was 1 nm and each sample had 5 replicates of reflectance spectra. 2.3. Spectral Analysis Each sample had a spectral curve, whose x-axis was wavelength (nm) and y-axis was reflectance (%), and each sample had its SOC concentrations. Therefore, 96 samples had 96 curves and corresponding SOC concentrations. Using these data, the relationship (spectral model) between reflectance spectra (many x) and SOC concentrations (y) could be established. Specifically, 96 samples were divided into a calibration set and a validation set by Kennard–Stone algorithm (K-S) at the ratio of 2:1 [24]. Each sample had many x (reflectance), i.e., x1, x2, x3, . . . xn were the reflectances at each wavelength, and one y (SOC concentration). Qiu et al. (2022) reported that the spectral model established by the whole spectra had much better prediction ability of total carbon and nitrogen concen- trations, maybe because SOC concentrations were reflected by all the wavelengths [25]. Therefore, the spectral model was also established by the whole spectra without charac- teristic spectra extraction in this study [25]. Confalonieri et al. (2001) reported that partial least squares regression (PLSR) was successful in modelling soil nutrients by near infrared spectroscopy [25], as many other studies found [19,25]. PLSR was also found to be perfect for total carbon and nitrogen modelling in the same sediments, especially using the whole Sensors 2022, 22, x FOR PEER REVIEW 5 of 12 least squares regression (PLSR) was successful in modelling soil nutrients by near infrared spectroscopy [25], as many other studies found [19,25]. PLSR was also found to be perfect for total carbon and nitrogen modelling in the same sediments, especially using the whole Sensors 2022, 22, 5638 5 of 12 spectra [25]. Therefore, in this study PLSR was used to establish the spectral model of SOC. Lastly, the spectral model was evaluated by the validation set using the determina- tion coefficient (R ), the mean root square error (RMSE) and the residual predictive devi- spectra [25]. Therefore, in this study PLSR was used to establish the spectral model of SOC. ation (RPD) [21,22]. Lastly, the spectral model was evaluated by the validation set using the determination RPD = SD⁄RMSE = ∑ (𝑦 −𝑦 ) ∑ (𝑦 −𝑦 ) , in which 𝑦 is the coefficient (R ), the mean root square error (RMSE) and the residual predictive deviation (RPD) [21,22]. measured value of the samq ple, 𝑦 is the predicted value by the model, 𝑦 is the average 1 n 1 n RPD = SD/RMSE = (y y ) / (y y ) , in which y is the measured value of the sample and å n is the number of sa å mples. Rossel et al. (2006) pro- i i i i i n1 i=1 n1 i1 measur posed a qua ed value ntitati of ve rela the sample, tionship y ˆ between is the pr R edicted PD andvalue model q by u the alitmodel, y: RPD > 2. y is 0,the the average models i i measur or predict ed ivalue ons are per of thefsample ect and can and be use n is thed number for rapiof d an samples. alysis [21]. Rossel et al. (2006) proposed a quantitative All analysrielationship s was perform between ed in MAT RPD L and AB 2 model 017b. quality: Specific progr RPD >ams had been 2.0, the models puor b- predictions are perfect and can be used for rapid analysis [21]. lished in the Chinese software copyright (2018SR325853). A new program in R is being All analysis was performed in MATLAB 2017b. Specific programs had been published prepared. in the Chinese software copyright (2018SR325853). A new program in R is being prepared. 2.4. Evaluation of the Spectral Analysis 2.4. Evaluation of the Spectral Analysis Four parameters were proposed to evaluate the spectral analysis from different spec- Four parameters were proposed to evaluate the spectral analysis from different spec- trometers, including the detection resolution, range (minimum and maximum) and error. trometers, including the detection resolution, range (minimum and maximum) and error. These parameters were calculated according to the distribution of predicted SOC concen- These parameters were calculated according to the distribution of predicted SOC concen- trations (Figure 1). trations (Figure 1). Figure 1. The concept map of the parameters for evaluating spectral models. Figure 1. The concept map of the parameters for evaluating spectral models. These four parameters were calculated as shown in Figure 2. The distributions of pre- These four parameters were calculated as shown in Figure 2. The distributions of dicted SOC concentrations were fitted as a function for both calibration sets and validation predicted SOC concentrations were fitted as a function for both calibration sets and vali- sets, whose x-axis was SOC concentrations. Under the condition that the function is near dation sets, whose x-axis was SOC concentrations. Under the condition that the function to or equals a constant (e.g., 0), the x-values (SOC concentrations) are the minimum and is near to or equals a constant (e.g., 0), the x-values (SOC concentrations) are the minimum maximum (detection range). The detection resolution is the average difference between and maximum (detection range). The detection resolution is the average difference be- the two neighboring SOC concentrations. The detection errors are the difference between tween the two neighboring SOC concentrations. The detection errors are the difference the predicted SOC and the measured SOC, which can be plotted by the quartile map. The between the predicted SOC and the measured SOC, which can be plotted by the quartile quartile map can directly show the actual error including the largest error, the smallest map. The quartile map can directly show the actual error including the largest error, the error and the median error. smallest error and the median error. Sensors 2022, 22, x FOR PEER REVIEW 6 of 12 Sensors 2022, 22, 5638 6 of 12 Sensors 2022, 22, x FOR PEER REVIEW 6 of 12 Figure 2. The flow chart for calculating the proposed parameters. Figure 2. The flow chart for calculating the proposed parameters. Figure 2. The flow chart for calculating the proposed parameters. All analysis was performed in SPSS 13.0 (SPSS Inc., Cary, NC, USA); a new program All analysis was performed in SPSS 13.0 (SPSS Inc., Cary, NC, USA); a new program All analysis was performed in SPSS 13.0 (SPSS Inc., Cary, NC, USA); a new program in R is being prepared. Here, we stress the procedure of how to calculate detection range. in R is being prepared. Here, we stress the procedure of how to calculate detection range. in R is being prepared. Here, we stress the procedure of how to calculate detection range. In SPSS, the predicted SOC was first analyzed by “Graphs”-“Histogram”. Then, the dis- In SPSS, the predicted SOC was first analyzed by “Graphs”-“Histogram”. Then, the dis- In SPSS, the predicted SOC was first analyzed by “Graphs”-“Histogram”. Then, the tribution of predicted SOC was performed by “Analyze”-“Regression”-”Nonlinear”- tribution of predicted SOC was performed by “Analyze”-“Regression”-”Nonlinear”- distribution of predicted SOC was performed by “Analyze”-“Regression”-”Nonlinear”- “Peak-Gaussian” or “Analyze”-“Fit curve”-”Cubic”. Based on the function, the extreme “Peak-Gaussian” or “Analyze”-“Fit curve”-”Cubic”. Based on the function, the extreme “Peak-Gaussian” or “Analyze”-“Fit curve”-”Cubic”. Based on the function, the extreme could be calculated. could couldbe be c calculated. alculated. 3. Results 3. Results 3. Results 3.1. Reflectance Spectra of the Same Sediments in Different Spectrometers 3.1. Reflectance Spectra of the Same Sediments in Different Spectrometers 3.1. Reflectance Spectra of the Same Sediments in Different Spectrometers The reflectance spectra collected by the three spectrometers were different (Figures 3 The reflectance spectra collected by the three spectrometers were different (Figures 3 and 4). The reflectance spectra collected by the three spectrometers were different (Figures 3 and 4). As shown in Figure 3, QEPro only collected the reflectance spectra below 1000 nm, As shown in Figure 3, QEPro only collected the reflectance spectra below 1000 nm, while both and 4). As shown in Figure 3, QEPro only collected the reflectance spectra below 1000 nm, while both Cary 5000 and FieldSpec 4 collected the complete visible and near infrared Cary 5000 and FieldSpec 4 collected the complete visible and near infrared reflectance spectrum while both Cary 5000 and FieldSpec 4 collected the complete visible and near infrared reflectance spectrum (350–2500 nm). At 500–1000 nm, the reflectance was mainly ranged (350–2500 nm). At 500–1000 nm, the reflectance was mainly ranged at 20–50% in Cary 5000, reflectance spectrum (350–2500 nm). At 500–1000 nm, the reflectance was mainly ranged at 20–50% in Cary 5000, 15–45% in FieldSpec 4 and 20–60% in QEPro. All spectral curves 15–45% in FieldSpec 4 and 20–60% in QEPro. All spectral curves in Cary 5000 were more at 20–50% in Cary 5000, 15–45% in FieldSpec 4 and 20–60% in QEPro. All spectral curves in Cary 5000 were more compact than the portable spectrometers. compact than the portable spectrometers. in Cary 5000 were more compact than the portable spectrometers. Figure Figure 3. 3. TT hh ee rere flefclte ac ntcaenc spe s ectp raec otfrsae o dif m see ndi tsme in H nt usa i nn g H haiu aa nn dghai and Bohai Sea Bohai Sea of China by of China by different spec different trometers. Figure 3. The reflectance spectra of sediments in Huanghai and Bohai Sea of China by different spectrometers. As shown in Figure 4, the differences among these three spectrometers were direct spectrometers. and simple. The shapes of the spectral curves were the same, especially for Cary 5000 and As shown in Figure 4, the differences among these three spectrometers were direct FieldSpec 4. The reflectance was ordered as QEPro > Cary 5000 > FieldSpec 4. As shown in Figure 4, the differences among these three spectrometers were direct and simple. The shapes of the spectral curves were the same, especially for Cary 5000 and and simple. The shapes of the spectral curves were the same, especially for Cary 5000 and FieldSpec 4. The reflectance was ordered as QEPro > Cary 5000 > FieldSpec 4. FieldSpec 4. The reflectance was ordered as QEPro > Cary 5000 > FieldSpec 4. Sensors 2022, 22, x FOR PEER REVIEW 7 of 12 Sensors 2022, 22, x FOR PEER REVIEW 7 of 12 Sensors 2022, 22, 5638 7 of 12 QEPro QE Cary Pr5000 o Cary FieldS 5000 pec4 FieldSpec4 500 1000 1500 2000 2500 500 1000 1500 2000 2500 Wavelength (nm) Wavelength (nm) Figure 4. The average reflectance spectra of sediments in Huanghai and Bohai Sea of China by dif- Figure 4. The average reflectance spectra of sediments in Huanghai and Bohai Sea of China by dif- Figure 4. The average reflectance spectra of sediments in Huanghai and Bohai Sea of China by ferent spectrometers. ferent spectrometers. different spectrometers. 3.2. Spectral Analysis from Different Spectrometers 3.2. Spectral Analysis from Different Spectrometers 3.2. Spectral Analysis from Different Spectrometers The spectral analysis was highly different across spectrometers. In the calibration The spectral analysis was highly different across spectrometers. In the calibration sets, The spectral analysis was highly different across spectrometers. In the calibration sets, Cary 5000 had the best model with the largest R and the smallest RMSE, while QEPro Cary 5000 had the best model with the largest R and2the smallest RMSE, while QEPro had sets, Cary 5000 had the best model with the largest R and the smallest RMSE, while QEPro had the worst model with the smallest R and the largest RMSE (Figure 4). In the valida- the worst model with the smallest R and the 2 largest RMSE (Figure 4). In the validation sets, had the worst model with the smallest R and the largest RMSE (Figure 4). In the valida- tion sets, Cary 5000 also had the best model with the largest R , the smallest RMSE and Cary 5000 also had the best model with the largest R , the smallest RMSE and the highest tion sets, Cary 5000 also had the best model with the largest R , the smallest RMSE and the highest RPD, while FieldSpec 4 had the worst model with the smallest R , the largest RPD, while FieldSpec 4 had the worst model with the smallest R , the largest RMSE and the highest RPD, while FieldSpec 4 had the worst model with the smallest R , the largest RMSE and the lowest RPD (Figure 5). These results showed that Cary 5000 had the best the lowest RPD (Figure 5). These results showed that Cary 5000 had the best performance RMSE and the lowest RPD (Figure 5). These results showed that Cary 5000 had the best performance on the spectral analysis of SOC. on the spectral analysis of SOC. performance on the spectral analysis of SOC. a Calibration set in QEPro Validation set in QEPro 1.5 Calibration set in QEPro Validation set in QEPro 1.5 1.0 1.0 0.5 0.5 0.0 r =0.59, RMSE=0.18 r =0.75, RMSE=0.18 2 2 0.0 r =0.59, RMSE=0.18 r =0.75, RMSE=0.18 RPD=1.6 RPD=1.6 Calibration set in FieldSpec 4 Validation set in FieldSpec 4 1.5 Calibration set in FieldSpec 4 Validation set in FieldSpec 4 1.5 1.0 1.0 0.5 0.5 0.0 r =0.52, RMSE=0.19 r =0.98, RMSE=0.06 0.0 r =0.52, RMSE=0.19 RPD=1.4 r =0.98, RMSE=0.06 RPD=1.4 Calibration set in Cary 5000 Validation set in Cary 5000 1.5 Calibration set in Cary 5000 Validation set in Cary 5000 1.5 1.0 1.0 0.5 0.5 r =0.86, RMSE=0.11 0.0 r =0.99, RMSE=0.04 r RPD =0.86, RM =2.6 SE=0.11 r =0.99, RMSE=0.04 0.0 RPD=2.6 0.00.5 1.01.5 0.00.5 1.01.5 0.00.5 1.01.5 0.00.5 1.01.5 Measured concentrations (%) Measured concentrations (%) Measured concentrations (%) Measured concentrations (%) Figure 5. Results of spectral analysis on organic carbon concentrations in sediments of the Yellow Figure 5. Results of spectral analysis on organic carbon concentrations in sediments of the Yellow Sea Figure 5. Results of spectral analysis on organic carbon concentrations in sediments of the Yellow Sea and Bohai Sea, China (spectral models for both the calibration set and validation set in QEPro and Bohai Sea, China (spectral models for both the calibration set and validation set in QEPro (a,b), Sea and Bohai Sea, China (spectral models for both the calibration set and validation set in QEPro (a,b), FieldSpec 4 (c,d), and Cary 5000 (e,f), respectively). FieldSpec 4 (c,d), and Cary 5000 (e,f), respectively). (a,b), FieldSpec 4 (c,d), and Cary 5000 (e,f), respectively). Reflectance rates (%) Mode Mode lled co lled co ncen ncen trat tr io at ns ( ions ( %)%)Mo M de old led elled conc conc entent rati rons ations (%) (%)Mod Mod elled co elled co ncen ncen tratio tratio ns ( ns ( %)%) Reflectance rates (%) Sensors 2022, 22, x FOR PEER REVIEW 8 of 12 Sensors 2022, 22, x FOR PEER REVIEW 8 of 12 Sensors 2022, 22, 5638 8 of 12 3.3. Evaluation of the Spectral Analysis 3.3. Evaluation of the Spectral Analysis 3.3. Evaluation of the Spectral Analysis Firstly, the errors of spectral models both in the calibration set and the validation set Firstly, the errors of spectral models both in the calibration set and the validation set Firstly, the errors of spectral models both in the calibration set and the validation set were the smallest from Cary 5000 but the largest from QEPro (Figure 6). These results were the smallest from Cary 5000 but the largest from QEPro (Figure 6). These results were the smallest from Cary 5000 but the largest from QEPro (Figure 6). These results were were consistent with results of the spectral analysis (Figure 5). For the portable spectrom- were consistent with results of the spectral analysis (Figure 5). For the portable spectrom- consistent with results of the spectral analysis (Figure 5). For the portable spectrometers, eters, the errors in FieldSpec 4 increased from the calibration set to the validation set, but eters, the errors in FieldSpec 4 increased from the calibration set to the validation set, but the errors in FieldSpec 4 increased from the calibration set to the validation set, but errors errors decreased in QEPro (Figure 6), showing that QEPro had a consistent performance errors decreased in QEPro (Figure 6), showing that QEPro had a consistent performance decreased in QEPro (Figure 6), showing that QEPro had a consistent performance on on spectral analysis. on spectral analysis. spectral analysis. a b Calibration set Validation set Calibration set Validation set 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 −0.2 −0.2 −0.4 −0.4 Cary5000 QEPro Cary5000 QEPro Cary5000 FieldSpec4 QEPro Cary5000 FieldSpec4 QEPro FieldSpec4 FieldSpec4 Figure 6. The quartile map of the absolute error for predicted organic carbon concentrations in the Figure Figure 6. 6. The The q quartile uartile map of the absolute er map of the absolute err ror for pr or for predicted edicted org organic anic carb carbon on concentrati concentrations ons in the in the calibration set and validation set (absolute errors in the calibration set (a) and validation set (b)). calibration set and validation set (absolute errors in the calibration set (a) and validation set (b)). calibration set and validation set (absolute errors in the calibration set (a) and validation set (b)). Secondly, the spectral curves were close to a normal distribution, except for the cali- Secondly, the spectral curves were close to a normal distribution, except for the Secondly, the spectral curves were close to a normal distribution, except for the cali- calibration bration set of set FieldSpec 4 (Figure of FieldSpec 4 (Figur7 e). C 7). a Cary ry 5000 had 5000 had a a wider wider detection ran detection ran gge e oof f SOC than SOC than bration set of FieldSpec 4 (Figure 7). Cary 5000 had a wider detection range of SOC than the the portable portable spectr spectrometers ometers (Figu (Figure re 7). Th 7). The e calibr calibration ation set set in QEP in QEPr ro o h had ad a a w wider ider de detection tection the portable spectrometers (Figure 7). The calibration set in QEPro had a wider detection range than the validation set, showing a consistent detection range of SOC (Figure 7). These range than the validation set, showing a consistent detection range of SOC (Figure 7). range than the validation set, showing a consistent detection range of SOC (Figure 7). rThese re esults arsults e consistent are cons with istent with those those in Table in 3 Table . 3. These results are consistent with those in Table 3. QEPro Cary5000 FieldSpec4 Cary5000 FieldSpec4 QEPro a c a c d f d f 8 e TOC concentrations (%) TOC concentrations (%) Figure 7. The distribution curves of predicted organic carbon concentrations by different spectral Figure 7. The distribution curves of predicted organic carbon concentrations by different spectral Figure 7. The distribution curves of predicted organic carbon concentrations by different spectral models models established by both calibration set and validation set using different spectrometers (distri- models established by both calibration set and validation set using different spectrometers (distri- established by both calibration set and validation set using different spectrometers (distribution curves bution curves from the calibration set by Cary 5000 (a), FieldSpec 4 (b), and QEPro (c), respectively; bution curves from the calibration set by Cary 5000 (a), FieldSpec 4 (b), and QEPro (c), respectively; from the calibration set by Cary 5000 (a), FieldSpec 4 (b), and QEPro (c), respectively; distribution curves distribution curves from the validation set by Cary 5000 (d), FieldSpec 4 (e), and QEPro (f), respec- distribution curves from the validation set by Cary 5000 (d), FieldSpec 4 (e), and QEPro (f), respec- from the validation set by Cary 5000 (d), FieldSpec 4 (e), and QEPro (f), respectively). tively). tively). 0 0.2 .2 0 0.4 .4 0.6 0.6 0 0.8 .8 1 1..0 0 1.2 1.2 1 1..4 4 1 1.6 .6 1.8 1.8 0.2 0.2 0 0.4 .4 0 0.6 .6 0.8 0.8 1 1.0 .0 1 1..2 2 1.4 1.4 1 1..6 6 1 1.8 .8 0 0.2 .2 0.4 0.4 0 0.6 .6 0 0.8 .8 1.0 1.0 1 1.2 .2 1 1..4 4 1.6 1.6 1 1..8 8 Calibration set Absolute error (%) Absolute error (%) Validation set Calibration set Validation set Sensors 2022, 22, 5638 9 of 12 Table 3. The parameters to evaluate spectral models in both calibration set (C) and validation set (V), respectively. Cary 5000 FieldSpec 4 QEPro Parameters C V C V C V Max 2.41 2.38 2.90 1.69 2.14 2.03 Min 0.01 0.00 0.00 0.00 0.01 0.00 Error 0.03 0.09 0.05 0.14 0.14 0.14 Resolution 0.006 0.005 0.005 0.004 0.006 0.004 Thirdly, the detection resolutions were all less than 0.01% for the three spectrometers, and were not significantly different across spectrometers (Table 3). All spectrometers increased their resolution from the calibration set to the validation set, especially QEPro (Table 3). Clearly, Cary 5000 could predict a larger range of SOC concentrations with the smallest error and highest resolution (Table 3). FieldSpec 4 had better performance in the calibration set but lost the advantage in the validation set (Table 3). Although the technical parameters of spectral analysis in QEPro had no advantages, they were consistent in both the calibration set and validation set (Figure 7, Table 2). 4. Discussion We studied the spectral analysis of SOC in the Yellow Sea and Bohai Sea of China using Agilent Cary 5000, ASD FieldSpec 4 and Ocean Optics QEPro. We proposed four parameters to evaluate the spectral analysis of SOC, including the detection range (max and min), resolution and error. These parameters were used to evaluate the performance of the spectrometers on spectral analysis. 4.1. Advantages of the Proposed Method on Evaluation of Spectral Analysis by Spectrometers Results showed a large difference in the spectral analysis among the three spectrome- ters. Consistent with our hypothesis, the portable spectrometers had the worse spectral analysis than the bench-top spectrometer (Figures 6 and 7, Table 3). This was mainly due to the large difference in specification, including spectral resolution, system signal- to-noise ratio, stray light, detector, wavelength repeatability, wavelength accuracy and so on (Table 2) [15,16,19]. In this study, Cary 5000 had better specifications than the other two portable spectrometers, especially in terms of the signal-to-noise ratio, stray light, wavelength repeatability and wavelength accuracy (Table 2). Although it was easily hy- pothesized, we showed the quantitative differences between the spectrometers, making the evaluation quantitatively. These results are consistent with the past studies (Table 1) [15,16,19]. Sharififar et al. (2019) analyzed SOC on an Australian farm using Brucker Optik Vertex 70 (bench-top) and Agilent 4200 (portable) and found that the spectral model derived from Vertex 70 was better than that from Agilent 4200 (Table 1) [19]. Linderhom et al. (2019) used ASD LabSpec and two other portable spectrometers to classify different soils and found that the lab spectrometer had a better classification ability than the portable spectrometers, mainly because the lab spectrometer had a wider spectral range and a better detector [26]. Unfortunately, these were qualitative studied, so we could not compare their performance on the spectral analysis by the technical parameters shown in Table 3. This study also provided a scientific basis for evaluating spectral analysis. Although the combination of R , RMSE and RPD can evaluate whether a spectral analysis is better, it is difficult to quantify how much better it is. Using the proposed parameters, we could clearly understand which samples would be suitable for the spectral technique we established, such as its detection range, resolution and error. That is, we could accurately define samples by the concentrations that could be determined (detection range) and differentiated (detection resolution) at what errors. In this study, if the SOC concentration of a sediment sampled from the Yellow Sea and Bohai Sea of China is about 2.15%, it would be more accurately Sensors 2022, 22, 5638 10 of 12 analyzed by Cary 5000. If the SOC concentration is about 2.00% and the detection error is not concerning, QEPro would be more suitable. Furthermore, if we develop a technique or an equipment that can accurately and rapidly analyze SOC based on vis-NIR spectroscopy, we should present its technical parameters, in which the proposed four parameters are the most basic. Therefore, the proposed four parameters are valuable for the equipment or a technique based on Vis-NIR spectroscopy. Spectral models are highly dependent on modelling samples; therefore, further study should pay attention to the accurate and rapid simulation of the distribution curve of the predicted SOC concentrations. 4.2. Performance of the Portable Spectrometers on Spectral Analysis Inconsistent with our hypothesis, the portable spectrometers had a similar perfor- mance on spectral analysis. The detection range, error and resolution in FieldSpec 4 were not significantly different from those in QEPro (Figures 6 and 7, Table 3), but QEPro had a more consistent detection range (Figure 7, Table 3) and errors (Figure 6, Table 3) in both the calibration set and the validation set than FieldSpec 4. This suggests that QEPro has a more stable performance than FieldSpec 4, at least in this study. Yahaya et al. (2015) found that the spectral model established by the QE65000 spec- trometer could be transferred to the FieldSpec 3 spectrometer more easily and in a better way than another portable spectrometer, suggesting that the two spectrometers could be replaced by each other [27]. In this study, QEPro and FieldSpec 4 were the upgraded spec- trometers of QE65000 and FieldSpec 3, respectively, suggesting that QEPro and FieldSpec 4 also could replace each other. These results indicate that if the error is not the main consideration, QEPro is a good choice due to its high stability and low price, suitable for long-term and in situ (online) use. However, if we want to establish a rapid spectral measurement technology to replace the traditional lab chemical analysis, Cary 5000 is undoubtedly a better choice. This study provided a direct reference for selecting different spectrometers to conduct effective scientific research. 5. Conclusions We studied the spectral analysis of organic carbon in sediments by three spectrometers and proposed four parameters to evaluate the performance of these spectrometers. The results showed that the bench-top spectrometer had better performance on spectral analysis than the portable spectrometers, mainly due to their differences in specification, such as signal-to-noise ratio, stray light, detector, wavelength repeatability, wavelength accuracy and so on. For the two portable spectrometers, FieldSpec 4 had the best performance in the calibration set, but lost its advantage in the validation set. QEPro had ordinary technical parameters on spectral analysis, but was consistent in both the calibration set and validation set. These results suggest that if the stability is the main consideration, QEPro is a good choice for long-term and in situ use due to its low price. This study provides a direct reference for selecting different spectrometers to conduct effective scientific research, and also provides a scientific basis for evaluating the spectral analysis. 6. Patents Chinese invention: A method for obtaining the performance of a spectrometer (CN202011288620.1). Author Contributions: Y.L. conducted the experiment and draft preparation, X.L. taught the data analysis, H.Q. and G.H. prepared the samples, P.F. designed the experiment and improved the manuscript. All authors have read and agreed to the published version of the manuscript. Funding: This research was supported by the National Natural Science Foundation of China (32171578, U2006209), and Shandong Natural Science Foundation, China (ZR 2021MD093, ZR2021MD103, ZR2021FQ028). Sensors 2022, 22, 5638 11 of 12 Data Availability Statement: https://www.researchgate.net/publication/362290727_Reflectance_rates_ and_orgainc_carbon_concentrations_of_the_sediments_in_the_Yellow_Sea_and_Bohai_Sea_China (accessed on 18 April 2022). Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. References 1. Rossel, R.A.V.; Webster, R. Predicting soil properties from the Australian soil visible-near infrared spectroscopic database. Eur. J. Soil Sci. 2012, 63, 848–860. [CrossRef] 2. Rial, M.; Cortizas, M.; Rodríguez-Lado, L. Understanding the spatial distribution of factors controlling topsoil organic carbon content in European soils. Sci. Total Environ. 2017, 609, 1411–1422. [CrossRef] [PubMed] 3. Liu, S.; Shen, H.; Chen, S.; Zhao, X.; Biswas, A.; Jia, X.; Shi, Z.; Fang, J. Estimating forest soil organic carbon content using vis-NIR spectroscopy: Implications for large-scale soil carbon spectroscopic assessment. Geoderma 2019, 348, 37–44. [CrossRef] 4. Li, X.; Li, Z.; Hou, G.; Qiu, H.; Lv, H.; Chen, G.; Fan, P. Carbon content detection of marine sediments based on multispectral fusion. Spectrosc. Spect. Anal. 2021, 41, 2898–2903. 5. Gholizadeh, A.; Rossel, R.A.V.; Saberioon, M.; Boruvka, L.; Kratina, J.; Pavlu, ˚ L. National-scale spectroscopic assessment of soil organic carbon in forests of the Czech Republic. Geoderma 2021, 385, 114832. [CrossRef] 6. Knox, N.M.; Grunwald, S.; McDowell, M.L.; Bruland, G.L.; Myers, D.B.; Harris, W.G. Modelling soil carbon fractions with visible near-infrared (VNIR) and mid-infrared (MIR) spectroscopy. Geoderma 2015, 239–240, 229–239. [CrossRef] 7. Vestergaard, R.-J.; Vasava, H.B.; Aspinall, D.; Chen, S.; Gillespie, A.; Adamchuk, V.; Biswas, A. Evaluation of optimized preprocessing and modelling algorithms for prediction of soil properties using vis-NIR spectroscopy. Sensors 2021, 20, 6745. [CrossRef] 8. Sarathjith, M.C.; Das, B.S.; Wani, S.P.; Sahrawat, K.L. Variable indicators for optimum wavelength selection in diffuse reflectance spectroscopy of soils. Geoderma 2016, 267, 1–9. [CrossRef] 9. Dotto, A.C.; Dalolin, R.S.D.; Grunwald, S.; Caten, A.; Filho, W.P. Two preprocessing techniques to reduce model covariables in soil property predictions by Vis-NIR spectroscopy. Soil Till. Res. 2017, 172, 59–68. [CrossRef] 10. Dhawale, N.M.; Adamchuk, V.I.; Prasher, S.O.; Rossel, R.A.V.; Ismail, A.A. Evaluation of two portable hyperspectral-sensor-based instruments to predict key soil properties in Canadian soils. Sensors 2022, 22, 2556. [CrossRef] 11. Ng, W.; Minasny, B.; Montazerolghaem, M.; Padarian, J.; Ferguson, R.; Bailey, S.; McBratney, A.B. Convolutional neural network for simultaneous prediction of several soil properties using visible/near-infrared, mid-infrared, and their combined spectra. Geoderma 2019, 352, 251–267. [CrossRef] 12. Zheng, G.; Ryu, D.; Jiao, C.; Xie, X.; Shang, G. Visible and near-infrared reflectance spectroscopy analysis of a coastal soil chronosequence. Remote Sens. 2019, 11, 2336. [CrossRef] 13. Xu, S.; Zhao, Y.; Wang, M.; Shi, X. Comparison of multivariate methods for estimating selected soil properties from intact soil cores of paddy fields by Vis-NIR spectroscopy. Geoderma 2018, 310, 29–43. [CrossRef] 14. Semella, S.; Hutengs, C.; Seidel, M.; Ulrich, M.; Schneider, B.; Ortner, M.; Thiele-Bruhn, S.; Ludwig, B.; Vohland, M. Accuracy and reproducibility of laboratory diffuse reflectance measurements with portable VNIR and MIR spectrometers for predictive soil organic carbon modelling. Sensors 2022, 22, 2749. [CrossRef] 15. Crucil, G.; Castaldi, F.; Aldana-Jague, E.; van Wesemael, B.; Macdonald, A.; Oost, K.V. Assessing the performance of UAS- Compatible multispectral and hyperspectral sensors for soil organic carbon prediction. Sustainability 2019, 11, 1889. [CrossRef] 16. Karyotis, K.; Angelopoulou, T.; Tziolas, N.; Palaiologou, E.; Samarinas, N.; Zalidis, G. Evaluation of a micro-electro mechanical systems spectral sensor for soil properties estimation. Land 2021, 10, 63. [CrossRef] 17. Greenberg, I.; Seidel, M.; Vohland, M.; Ludwig, B. Performance of field-scale lab vs in situ visible/near- and mid-infrared spectroscopy for estimation of soil properties. Eur. J. Soil Sci. 2022, 73, e13180. [CrossRef] 18. Sharififar, A.; Singh, K.; Jones, E.; Ginting, F.I.; Minasny, B. Evaluating a low-cost portable NIR spectrometer for the prediction of soil organic and total carbon using different calibration models. Soil Use Manag. 2019, 35, 607–616. [CrossRef] 19. Li, S.; Rossel, R.A.V.; Webster, R. The cost-effectiveness of reflectance spectroscopy for estimating soil organic carbon. Eur. J. Soil Sci. 2022, 73, e13202. [CrossRef] 20. Rossel, R.A.V.; Mcglynn, R.N.; Mcbratney, A.B. Determining the composition of mineral-organic mixes using UV-vis-NIR diffuse reflectance spectroscopy. Geoderma 2006, 137, 70–82. [CrossRef] 21. Summers, D.; Lewis, M.; Ostendorf, B.; Lewis, M. Visible near-infrared reflectance spectroscopy as a predictive indicator of soil properties. Ecol. Indic. 2011, 11, 123–131. [CrossRef] 22. Eskildsen, C.E.; Hansen, P.W.; Skov, T.; Marini, F.; Nøgaarda, L. Evaluation of multivariate calibration models transferred between spectroscopic instruments: Applied to near infrared measurements of flour samples. J. Near Infrared Spec. 2016, 24, 151–156. [CrossRef] Sensors 2022, 22, 5638 12 of 12 23. Qiu, H.; Fan, P.; Hou, G.; Li, X.; Wang, Y. Analysis and model comparison of carbon and nitrogen concentrations in sediments of the Yellow Sea and Bohai Sea by visible-near infrared spectroscopy. Bull. Environ. Contam. Toxicol. 2022, 108, 1124–1131. [CrossRef] [PubMed] 24. Galvão, R.K.; Araujo, M.C.; José, G.E.; Pontes, M.J.C.; Silva, E.C.; Saldanha, T.C.B. A method for calibration and validation subset partitioning. Talanta 2005, 67, 736–740. [CrossRef] 25. Confalonieri, M.; Fornasier, F.; Ursino, A.; Boccardi, F.; Pintus, B.; Odoardi, M. The potential of near infrared reflectance spectroscopy as a tool for the chemical characterisation of agricultural soils. J. Near Infrared Spec. 2001, 9, 123–131. [CrossRef] 26. Linderholm, J.; Geladi, P.; Gorretta, N.; Bendoula, R.; Gobrecht, A. Near infrared and hyperspectral studies of archaeological stratigraphy and statistical considerations. Geoarchaeology 2019, 34, 311–321. [CrossRef] 27. Yahaya, O.K.M.; MatJafri, M.Z.; Aziz, A.A.; Omar, A.F. Visible spectroscopy calibration transfer model in determining pH of Sala mangoes. J. Instrum. 2015, 10, T05002. [CrossRef]

Journal

Sensors – Multidisciplinary Digital Publishing Institute

Published: Jul 28, 2022

Keywords: detection resolution; organic carbon; sediment; spectrometer; spectral analysis

There are no references for this article.