Soil organic carbon predictions in Subarctic Greenland by visible–near infrared spectroscopy

Figures & data

Figure 1. Study area and sampling sites (bold) in South Greenland.

Figure 2. Data subdivision and modeling scheme. UPA = Upernaviarsuk data set; SIA = Søndre Igaliku data set; 2^nd SG = second Savitzky-Golay derivative; s.p. = smoothing points; CV = cross-validation.

Table 1. General statistics on SOC (%) reference data for the two complete datasets, data used for calibration model development, and validation datasets used in this study.

Data set/SOC (%)	N	Min	Max	Mean	Median	SE	SD	CV	Q1	Q3
UPA-all	173	0.6	40.9	8.7	4.7	0.7	8.9	102	2.4	12.2
SIA-all	157	0.3	9.8	2.0	1.6	0.1	1.5	74	1.1	2.4
Model 1^a	130	0.6	40.9	8.8	4.7	0.8	9.1	103	2.4	12.2
Model 1.1^b	160	0.3	40.9	7.5	3.9	0.7	8.6	115	1.7	10.7
Model 1.2^c	160	0.3	40.9	7.5	3.9	0.7	8.6	114	1.8	10.7
Model 2^d	118	0.3	9.8	2.0	1.6	0.1	1.6	78	1.1	2.4
Model 2.1^e	151	0.3	40.9	4.0	1.8	0.6	7.0	175	1.2	3.1
Model 2.2^f	151	0.3	40.9	3.5	1.8	0.4	5.4	155	1.2	3.1
UPA-validation ^g	43	0.8	37.2	8.4	4.7	1.3	8.5	101	2.4	11.7
SIA-validation ^h	39	0.4	5.4	1.9	1.6	0.2	1.2	63	1.1	2.3

SOC = soil organic carbon; CV = coefficient of variation; UPA = Upernaviarsuk; SIA = Søndre Igaliku.

^aSeventy-five percent of UPA data used for calibration Model 1.

^bSeventy-five percent of UPA data + 25 percent of SIA data, where 25 percent of SIA data was selected with Kennard-Stone selection.

^cSeventy-five percent of UPA data + 25 percent of SIA, where 25 percent SIA was selected with SOC selection.

^dSeventy-five percent of SIA data used for calibration Model 2.

^eSeventy-five percent of SIA + 25 percent of UPA, where 25 percent UPA was selected with Kennard-Stone selection.

^fSeventy-five percent of SIA + 25 percent of UPA, where 25 percent SIA was selected with SOC selection.

^gTwenty-five percent of UPA used for validation of all models.

^hTwenty-five percent of SIA used for validation of all models.

Figure 3. Mean spectra of the complete vis-NIRS data set from Upernaviarsuk (UPA) and Søndre Igaliku (SIA).

Figure 4. Principal component analysis of the investigated vis-NIRS data sets reported in scores (a) and (c) and plot of PC1 (b) and PC3 (d) ladings.

Table 2. Calibration statistics for the optimal calibration models used in this study.

Model	Pretreatment	Number of factors	R^2	RMSECV
Model 1	2nd SG, 20 s.p.	4	0.94	2.24
Model 2	2nd SG, 20 s.p.	4	0.89	0.51
Model 1.1	2nd SG, 20 s.p.	4	0.92	2.45
Model 1.2	2nd SG, 20 s.p.	4	0.92	2.46
Model 2.1	2nd SG, 20 s.p.	5	0.90	2.18
Model 2.2	2nd SG, 20 s.p.	6	0.83	2.26

RMSECV = root mean squared error of cross-validation; 2^nd SG = second Savitzky-Golay derivative; s.p. = smoothing points.

Figure 5. Regression coefficients on the Upernaviarsuk and Søndre Igaliku calibration model vis-NIRS data set.

Figure 6. Independent validation of calibration Model 1 (Upernaviarsuk) and Model 2 (Søndre Igaliku).

Figure 7. Independent validation of the enlarged calibration models when using selection according to soil organic carbon values for enlargement.

Figure 8. Principal component analysis results to support model performance.

Table A1. Independent validation model statistics.

Model	R^2	RMSEP	Bias	SD	IQ	RPD	RPIQ
Model 1 site-specific	0.95	1.8	−0.12	8.53	9.55	4.74	5.31
Model 1 site-wise	0.78	2.82	−0.83	1.17	1.33	2.00	0.47
Model 2 site-specific	0.82	0.64	0.24	1.17	1.33	1.83	2.08
Model 2 site-wise	0.9	7.13	−4.13	8.53	9.55	1.20	1.34
Model 1.1 site-specific	0.95	1.81	−0.25	8.53	9.55	4.71	5.28
Model 1.1 site-wise	0.78	2.83	−0.34	1.17	1.33	0.41	0.47
Model 1.2 site-specific	0.94	2.05	−0.14	8.53	9.55	4.16	4.66
Model 1.2 site-wise	0.77	2.14	0.05	1.17	1.33	0.55	0.62
Model 2.1 site-specific	0.67	1.56	0.35	1.17	1.33	0.75	0.85
Model 2.1 site-wise	0.94	2.01	−0.64	8.53	9.55	4.24	4.75
Model 2.2 site-specific	0.7	1.25	0.03	1.17	1.33	0.94	1.06
Model 2.2 site-wise	0.95	2.03	0.06	8.53	9.55	4.20	4.70

Note that Models 1.1 and 2.1 were enlarged with a subset selected by the Kennard-Stone algorithm and Models 1.2 and 2.2 with a subset selected according to reference SOC values. RMSEP = root mean square error of prediction; IQ = interquartile distance; RPD = ratio of standard error of prediction to standard deviation; RPIQ = ratio of performance to interquartile distance.