Use of near-infrared spectroscopy and multivariate approach for estimating silage fermentation quality from freshly harvested maize

Figures & data

Table 1. Specifications of the calibration curve used for fresh harvested pre-ensiling maize (n = 205), by the use of a poliSPEC^NIR near infrared system.

Constituents (g/kg DM)	Mean ($\overline{X}$)	Standard deviation (SD)	Standard error of calibration (SEC)	R^2^a
DM (g/kg)	357	54.6	19.8	0.93
Ash	39.6	5.88	3.47	0.82
Crude protein	72.4	13.0	7.67	0.82
Ether extract	24.6	8.99	5.91	0.77
aNDF	424	75.6	28.1	0.93
ADF	225	40.5	16.5	0.92
ADL	23.4	5.46	4.89	0.48
Water-soluble carbohydrates	65.1	26.3	17.4	0.77
Starch	267	93.3	43.1	0.89

^aCoefficient of determination.

DM: Dry matter; aNDF: Neutral detergent fibre; ADF: Acid detergent fibre; ADL: Acid detergent lignin.

Figure 1. PC1 and PC2 biplot from PCA algorithm.

Table 2. Specifications of the fresh whole corn plant, pre-ensiling dataset.

Variables (g/kg DM)	Shapiro test (w)	Minimum	Maximum	Mean ($\overline{X}$)	Standard deviation (SD)	Coefficient of variation (CV)
DM (g/kg)	0.97	238	552	352	53.8	15.3%
Ash	0.98	30.8	59.3	38.9	3.70	9.51%
Crude protein	0.98	59.4	81.8	67.1	2.86	4.26%
Ether extract	0.98	13.6	32.9	23.7	3.09	13.0%
aNDF	0.99	376	515	438	22.8	5.20%
ADF	0.99	181	281	228	18.9	8.27%
ADL	0.98	13.0	34.0	22.3	3.84	17.2%
Starch	0.99	173	375	281	37.4	13.3%
Water-soluble Carbohydrates	1.00	6.90	92.7	51.9	12.9	24.8%

DM: Dry matter; aNDF: Neutral detergent fibre; ADF: Acid detergent fibre; ADL: Acid detergent lignin.

Table 3. Specifications of the post-ensiled dataset.

Variables (g/kg DM)	Shapiro test (w)	Minimum	Maximum	Mean ($\overline{X}$)	Standard deviation (SD)	Coefficient of variation (CV) (%)
DM (g/kg)	0.97	230	544	342	53.0	15.5
Ash	0.98	26.0	46.6	33.9	3.44	10.2
Crude protein	1.00	55.4	92.3	72.3	5.91	8.17
Ether extract	0.98	8.40	31.6	21.9	3.69	16.9
aNDF	0.98	327	496	399	31.3	7.85
ADF	0.98	160	285	211	21.9	10.3
Starch	0.99	186	444	322	43.6	13.6
Water-soluble carbohydrates	0.97	0.20	13.8	5.24	2.23	42.5
pH^a	0.99	3.56	4.18	3.83	0.09	2.25
Ammonia^b	0.99	3.93	7.85	5.68	0.67	11.7
Glucose	0.83	0.00	11.6	3.00	1.93	64.5
Fructose	0.98	0.00	5.40	2.24	0.93	41.3
Mannitol	0.92	1.30	29.1	11.1	4.48	40.2
Ethanol	0.97	0.40	32.0	10.3	5.04	49.0
Lactic acid	0.99	24.1	99.2	49.4	10.4	21.0
Acetic acid	0.96	2.40	28.4	11.1	4.34	39.3
Propionic acid	0.77	0.00	4.70	0.55	0.68	124
Butyric acid	0.62	0.10	2.50	0.72	0.41	56.8

^aExpressed as a pure number.

^b% of the total nitrogen.

DM: Dry matter; aNDF: Neutral detergent fibre; ADF: Acid detergent fibre.

Table 4. Fermentation quality index (FQI) of maize silage belonging to different FAO classes (class 200 vs. class 600–700), grown in different yield potential areas (low vs. medium vs. high) and harvested at different Maturity (early, EH vs. medium, MH vs. late, LH).

	Year					Level input field
	2016	2017	2018	p		Low	Medium	High	p
FQI value
Average	54.4b	57.1a	43.4c	<.0001		50.5b	55.4a	48.9b	<.0001
Minimum	32.2	28.7	29.7			28.7	31.6	37.4
Maximum	84.6	81.5	63.0			77.8	84.6	67.0
SEM^a	0.33	0.59	0.83			0.41	0.40	0.59	0.41
Interaction years ×	p	p	p		Interaction level input field ×	p	p	p
Year	—	—	—			<.0001	<.0001	only 2016
Level input field	<0.0001	<0.0001	0.0005			—	—	—
Class cultivar	0.1039	<.0001	0.0102			<0.0001	0.0501	<0.0001
Maturity	<.0001	<.0001	<.0001			<0.0001	<0.0001	<0.0001
	Class cultivar				Maturity
	Early (200)	Late (600–700)	P		1/3 milk line	2/3 milk line	+ 5d 2/3 milk line	P
FQI value
Average	51.7	51.5	0.71		56.3a	52.0b	46.6c	<.0001
Minimum	31.1	28.7			39.6	37.1	28.7
Maximum	84.6	77.3			84.6	77.3	74.5
SEM^a	0.41	0.42			0.46	0.46	0.46
Interaction class cultivar ×	p	p		Interaction maturity ×	p	p	p
Year	<0.0001	<0.0001			<0.0001	<0.0001	<0.0001
Level Input Field	<0.0001	<0.0001			<0.0001	<0.0001	<0.0001
Class Cultivar	—	—			0.0469	0.0140	0.2607
Maturity	<0.0001	<0.0001			—	—	—

^aStandard error of means.

SEM: Standard Error of Means.

^a,b,cMean values within a row bearing different superscripts differ significantly (p < .05).

Table 5. Composition of the fresh whole plant maize belonging to different FQI classes (NEQ, EQ).

	FQI^a classification
Variables (g/kg DM)	NEQ, FQI < 57.4	EQ, FQI ≥ 57.4	P
DM (g/kg)	36.5a	32.3b	<0.0001
Ash	3.79b	4.05a	<0.0001
Crude protein	6.67b	6.77a	<0.0001
Ether extract	2.42a	2.18b	<0.0001
aNDF	43.1b	44.4a	<0.0001
ADF	22.4b	23.3a	<0.0001
ADL	2.20a	2.09b	0.0002
Water-soluble carbohydrates	5.01b	5.48a	<0.0001
Starch	29.2a	25.8b	<0.0001

^aFermentative qualitative index (FQI).

NEQ: Not excellent quality; EQ: Excellent quality; DM: Dry matter; aNDF: Neutral detergent fibre; ADF: Acid detergent fibre; ADL: Acid detergent lignin.

a,b, : Mean values within a row bearing different superscripts differ significantly (p < .05).

Table 6. Results of stepwise regression of constituents for FQI: estimated regression coefficients, intercept of the model and P (R² = 0.48 for the training dataset and 0.55 for reduced dataset).

	FQI^a training dataset coefficients	p	FQI^a reduced dataset coefficients	p
Constituents (g/kg DM)
DM (g/kg)	−1.10	<0.0001
Ash	6.60	0.001
Ether extract	19.2	0.0002	20.5	0.0181
aNDF	0.72	0.0301
ADL	−15.9	<0.0001
Starch	−0.76	0.0611	−1.89	0.0043
WSC			1.44	0.0473
The year = 2018			v10.6	<0.0001
Maturity phases = late (EH)			−3.56	0.0364
Level input field = ‘high’			3.75	0.0058
Intercept	45.1	0.0004	47.5	<0.0001

^aFermentative qualitative index (FQI).

R²: Coeffficient of determination; FQI: Fermentative qualitative index; DM: Dry matter; aNDF: Neutral detergent fibre; ADL: Acid detergent lignin; WSC: Water-soluble carbohydrates.

Figure 2. scatter plot for the outcomes of the CDA (Can1 and Can2) for the samples of the original dataset, considering the three classes for FQIc. (A) Green filled area = ‘lower’; red filled area = ‘good’; blue filled area = ‘excellent’. (B) Green filled area = ‘lower’; red filled area = ‘good or excellent’. Ellipses represent a 95% confidence interval.

Table 7. Confusion matrix for CDA in the original dataset (n = 797, 25 incomplete records removed) and in the reduced dataset (n = 101, 7 incomplete records removed) for FQIc, classified in three levels.

		Original dataset			Reduced dataset
	Actual Class	FQI^a < 47.7	47.7 ≤ FQI^a < 57.4	FQI^a ≥ 57.4	FQI^a < 47.7	47.7 ≤ FQI^a < 57.4	FQI^a ≥ 57.4
Predicted as	FQI < 47.7	153	84	14	36	21	6
Predicted as	47.7 ≤ FQI < 57.4	59	148	64	2	18	14
Predicted as	FQI ≥ 57.4	9	93	173	0	1	3
Sensitivity		0.69	0.46	0.69	0.95	0.45	0.13
Specificity		0.83	0.74	0.81	0.57	0.74	0.99
Accuracy		0.79	0.62	0.77	0.71	0.62	0.79
Precision		0.61	0.55	0.63	0.57	0.53	0.75
MCC^b		0.55	0.35	0.55	0.54	0.33	0.25

^aFermentative qualitative index (FQI); ^bMatthew Correlation Coefficient (MCC).

Figure 3. scatter plot for the outcomes of the CDA (Can1 and Can2) for the samples of the reduced dataset, in the three classes for FQIc. (A) green filled area = “lower”; red filled area = “good”; blue filled area = “excellent”. (B) Green filled area = “lower”; red filled area = “good or excellent”. Ellipses represent a 95% confidence interval.

Table 8. confusion matrix for PLS-DA, SVMs and KNN algorithms, validated on the testing set of FQIc.

		PLS-DA^e		SVMs^f		KNN^g
	Actual Class	NEQ^c, FQI^a < 57.4	EQ^d, FQI^a ≥ 57.4	NEQ^c, FQI^a < 57.4	EQ^d, FQI^a ≥ 57.4	NEQ^c, FQI^a < 57.4	EQ^d, FQI^a ≥ 57.4
Predicted as	NEQ^c, FQI^a < 57.4	119	15	137	37	140	36
Predicted as	EQ^d, FQI^a ≥ 57.4	28	54	10	32	7	33
Sensitivity		0.81	0.78	0.93	0.46	0.95	0.48
Specificity		0.78	0.76	0.46	0.93	0.48	0.95
Accuracy		0.75	0.81	0.78	0.78	0.8	0.8
Precision		0.89	0.66	0.79	0.76	0.8	0.83
MCC^b		0.57	0.57	0.51	0.50	0.55	0.54

^aFermentative Qualitative Index (FQI); ^bMatthew Correlation Coefficient (MCC); ^cnot excellent quality (NEQ); ^dexcellent quality (EQ); Partial Least Square Discriminant Analysis (PLS-DA); Support Vector Machine (SVMs); K-Nearest Neighbours (KNN).