You are on page 1of 9

Original research paper

Journal of Near Infrared Spectroscopy


2018, Vol. 26(1) 44–52

Use of LOCAL algorithm with near infrared ! The Author(s) 2017


Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
spectroscopy in forage resources for grazing DOI: 10.1177/0967033517746900
journals.sagepub.com/home/jns
systems in Colombia

C Ariza-Nieto1, OL Mayorga1, B Mojica2, D Parra1 and G Afanador-Tellez2

Abstract
This study used a total of 2020 Colombian forage resources of three families (Grass forages, legume forages, and other
forage plants) to develop near infrared spectroscopy calibrations for predicting the nutritional value. Spectra were collected
at 2 nm increments using a scanning visible/near infrared spectrometer. The reference data used for each forage were crude
protein, crude ash, neutral detergent fiber, acid detergent fiber, acid detergent lignin, measured according to the Association
of Official Analytical Chemists. Two chemometric tools for developing near infrared spectroscopy prediction models were
compared: the GLOBAL modified partial least squares, and the calibration strategy known as LOCAL. The LOCAL procedure is
designed to select, from a large database, samples with spectra resembling the sample being analyzed. Selected samples
were used as calibration sets after one-tenth of the samples were selected for validation from each database. Predictions of
nutrition indicators in validation samples using generic and specific calibrations were compared with both GLOBAL and
LOCAL procedures. For all predicted forages, LOCAL resulted in a significant improvement in both standard error of pre-
diction and bias values compared with GLOBAL. Determination coefficient values (r2) also improved using the LOCAL
algorithm, exceeding 0.9 for most forage sets. LOCAL calibration was then used with only one database (n 2020) comprising
all the forage samples and SEP and r2 were similar to those obtained in the three databases using LOCAL algorithm.
Therefore, LOCAL can accurately predict the composition of different forages using only one database, and could offer a
practical way to develop robust equations taking into account the biodiversity of Colombian forages.

Keywords
Colombian forages, chemometric, GLOBAL, LOCAL, calibration, near infrared
Received 28 September 2017; accepted 16 November 2017

topography; the characterization of tropical environ-


Introduction ments presents general patterns of climate and soils.
In Colombia, Forage-based livestock production plays This environmental condition affects the nutritional
a key role in national and regional economies, for food composition of the forages; these zones included: (1)
security and poverty alleviation. Additionally, forages The Eastern Savannahs have a low carrying capacity
are the cheapest nutrient resources to feed ruminants, and the natural herbage has a low nutritional value
and the amount of forage ingested by an animal is the associated with low digestibility energy; (2) The
most determinant on its nutritional value, which Andean Valleys hare more variable forages than in
depends on its availability, palatability, and digestibil- the eastern reflecting a different endowment of
ity.1–3 The production and conservation of high-quality resources and a different socio-economic context,
forages contribute to reduce the costs of production of although pasture-based, is more intensive that in
meat and milk and restrict under grazing conditions,
the use of concentrates and supplements for animals. 1
Colombian Corporation for Agricultural Research CORPOICA, Department
Forage resources of grasses and legumes play an of Animal Science, Bogota, Colombia
2
important role as a substitute for cereal grains in feed- Colombian National University, Department of Animal Science, Bogota
ing systems for meat and milk cattle, especially in Colombia
developing countries, such as Colombia.
Corresponding author:
Ruminant production systems are closely associated C Ariza-Nieto, Corporacion Colombiana de Investigacion Agropecuaria, Km
with the respective agro-ecological zones associated 14 Via Mosquera Bogota, Cundinamarca 7984, Colombia.
with its complex pattern of climate, soil, and Email: cariza@corpoica.org.co
Ariza-Nieto et al. 45

eastern Colombia; (3) Caribbean lowlands equals the additionally, these procedures use polluting chemicals.
inter-Andean valleys in importance in terms of its In this context, alternative methods should be gener-
stock of cattle, but production systems are more diver- ated that overcome the disadvantages of wet chemical
sified than elsewhere in the country, since extensive beef methods and allow the analysis of a large number of
breeding operations coexist with dual purpose systems. samples associated with the dynamics of the decision-
Pastures, both sown and natural, differ considerably making processes that producers take to optimize milk
from those in the other regions of the country and a and meat production based on strategic management of
great diversity of species can be found, depending on forages available on farms.
soil drainage and fertility; (4) The Andean plateau near The near infrared (NIR) spectroscopy is a physical
Bogotá (Sabana de Bogotá) is characterized by fertile method, which depends on the measurement of the
soils with abundant organic matter; the area is used for absorption by a sample of wavelengths in the NIR
high value crops, such as flowers and horticulture, but region (780–2500 nm) of the electromagnetic spectrum.
also for intensive, pasture-based dairying with high NIR technology has been utilized for the analysis of
grade Holstein cattle; and (5) Amazon basin, dual pur- animal feeds because of its potential benefits under rou-
pose production systems evolved based almost exclu- tine conditions of rapidity, comprehension and low
sively on pastures, currently dominated by Brachiaria cost.10 To obtain a satisfactory result of the nutrient
decumbens which alternates with degraded paddocks content of animal feed by NIR spectroscopy, it is neces-
with a variety of low quality naturalized grasses.4 sary to calibrate the apparatus against standard refer-
These classifications have been used effectively to ence samples, which have been analyzed by traditional
analyze the range of adaptation of different tropical methods of wet chemistry.
forage resources. Thus, the genus Panicum adapts to The first report on the use of the NIR spectroscopy
fertile soils with a wide range of precipitation, while to evaluate forage quality was made in 197611 with
the genus Pennisetum is associated with fertile soils subsequent studies conducted in Chile,12 France,13
and humid climates. The genus Cynodon has been Uruguay,14 and Belgium.1 The literature on NIR appli-
established particularly in fertile soils with humid envir- cations focused primarily on three types of local
onments, while the genus Brachiaria adapts to acid soils approaches: comparative analysis using structured
with a low fertility.5 This relationship of the tropical and constituted data (CARNAC), locally weighted
grasses with the environment has allowed its expansion regressions (LWR), and the LOCAL algorithm used
and even naturalization, when they have been intro- in this study.
duced to different microregions of Colombia. On the The objective of this study was to compare the reli-
other hand, tropical forage legumes have special adap- ability of the NIR spectroscopy approaches: LOCAL
tation requirements associated in some cases with the (linear PLS on selected samples) and GLOBAL MPLS
effective fixation of nitrogen and in others by competi- to predict the chemical characteristics of forage
tion with aggressive pastures and defoliation by grazing resources used in grazing systems in Colombia. The
of the animals. However, tropical legumes increase the study used a biodiverse database to predict chemical
protein supply of the diet, the rate of passage and sub- characteristics associated with the nutritional value of
sequently the animal performance when they are forage resources and also to test the influence of this
associated with forage grasses.6 type of validation on the parameters used to evaluate
Fertilization is required in tropical conditions to the prediction performance of the two approaches.
maintain high levels of animal production, since most
soils, especially those located in the humid tropics, have
low fertility.7,8 In these conditions a lot of nutrients are Material and methods
removed and must be continually replaced. On the
Samples and reference methods
other hand, proper management of grazing is essential
for an efficient use of forage produced and to ensure The study used a total of 2020 of forage resources avail-
long-term pasture productivity. Seasonal variations able in feeding systems from different types of cattle
(rainfall, drought) affect the growth of forages and a production in Colombia that were collected from
solution to the abundance of forage resources produced 2012 to 2016. Colombia is a country with a large
during the period of rapid growth is the conservation of plant biodiversity associated with its complex pattern
forages through the production of hay or silage. The of climate, soil, and topography Three geographic
composition of hay or silage is similar to that of green regions were characterized: the Andean Highlands,
forage, although 10%–20% of nutrients may be lost consisting of the three Andean ranges and intervening
during processing.9 valley lowlands; the Caribbean lowlands coastal region,
In order to monitor the great variability of the chem- and the Eastern region, the great plain that lies to the
ical characteristics of tropical forages in accordance east of the Andes and that includes the Amazon basin.
with the previously described scenarios, analytical In Colombia, climate is tropical and isothermal and its
methods of wet chemistry (proximal analysis and fiber differences are related to altitude. Tropical rainforest is
analysis) are used, which are characterized by a great in the Amazon basin and central Magdalena Valley
amount of time and resources from a laboratory; River, which is marked by an annual rainfall greater
46 Journal of Near Infrared Spectroscopy 26(1)

than 2500 mm, with an average annual temperature studied parameters in the GLOBAL approach.18 After
above 23 C. A tropical monsoon climate, which is char- developing MPLS equations, cross-validation was per-
acterized by one or two periods of summer during the formed to select the optimal number of factors and to
year, is located on the coast of the Caribbean region avoid overfitting. Calibration performance for each
and in some areas of the Andean region. Tropical model was assessed by standard error of calibration
savannahs are located in the lowlands of the (SEC), coefficient of determination of calibration
Caribbean region where wet and dry seasons alternate (R2C ), standard error of cross-calibration (SECV), coef-
during the year. This climate is also found in the plains ficient of determination in cross-validation (R2CV ), resi-
of the East and the upper part of the Magdalena River dual predictive deviation (RPD).19
with rainfall of 1000–1800 mm. In general, forage sam- All the equations were obtained using the standard
ples were taken during dry and rainy seasons at differ- normal variate and detrend (SNVD) methods for
ent days of regrowth; previously, the samples scatter correction. Four mathematical treatments
were analyzed for wet chemistry. Forage samples were were tested (1,4,4,1; 1,8,8,1; 2,4,4,1; and 2,8,8,1), in
grouped by family in three databases: (1) grass forage which the first digit is the number of the derivative,
(n ¼ 1418); (2) legume forage (n ¼ 320); (3) other forage the second digit is the gap over which the derivative is
plants (n ¼ 282). calculated, the third digit is the number of data points
Chemical analysis: Samples were oven-dried at 65 C in the first smoothing and the fourth digit is the
for 48 h to a constant weight, and then ground through a number of data points in the second smoothing. No
1 mm screen. They were stored in ambient laboratory outlier elimination was performed during calibration,
conditions. The chemical composition was determined and the best predictive model obtained with each
by AOAC15 reference methods: Dry matter (DM) database was subsequently subjected to external
105 C by AOAC2001.12 method; ashes by 942.05 validation.
AOAC984.13 method; crude protein (CP) by In the LOCAL approach the specific factors for each
AOAC984.13 method; neutral-detergent fiber (NDF) model were optimized according to the software
by AOAC2002.04 method; acid-detergent fiber (ADF) WinISI 4.7.0. The aim of the optimization was to
by AOAC973.18, and acid-detergent lignin (ADL) con- build models with the best prediction performance in
centrations was determined from the acid detergent terms of standard error of prediction corrected for
insoluble residue treated with 12 M sulfuric acid.16 mean bias (SEPc) to minimize the under and over fit-
Samples were measured twice and the standard error ting of the model. These factors were the optimum
of the laboratory (SEL) was estimated according to the number of selected samples and the minimum and max-
equation SEL ¼ sqrt((sum(r1  r2)2)/2k) where r1 ¼ rep- imum PLS components for the LOCAL prediction
licate 1, r2 ¼ replicate 2, and k ¼ number of samples). models.20
NIR spectroscopy analysis: After sample homogen- In order to evaluate the predictive capacity of the
ization, dry and round forage samples were placed in a calibrations four strategies of validations were
50 mm diameter ring cup and were scanned from 400 to performed: (1) GLOBAL generic ¼ global equations
1098 and from 1100 to 2498 nm, in 2 nm increments, using the entire forage database, (2) GLOBAL
using a scanning VIS/NIR spectrometer (Foss specific ¼ global equations using each single product
NIRSystems model 6500; www.foss.com). Spectra database, (3) LOCAL generic ¼ local procedure using
were recorded with the WinISI 4.7.0 (www.foss.com). the entire forage database, (4) LOCAL specific ¼ local
Every sample was measured twice and the average of procedure using each single product database.
the replicated spectra, obtained as log (1/R), was used
in the calibration analysis.
Calibrations and validations: Calibrations were devel-
Results and discussion
oped using WinISI 4.7.017 software. The CENTER The number of samples, mean, standard deviation,
algorithm was applied before calibration development, standard error of laboratory (SEL), minimum, and
samples were ranked according to their Mahalanobis maximum values of the chemical composition variables
distance (H) to the average sample. Samples with H for both calibration and validation sets in the generic
values greater than 3 were considered outliers and dis- database are given in Table 1. The database with all
carded. The validation files were generated by selecting forages contained large range of values for each con-
every 10th sample of the center file, with the remaining stituent. For instance, CP ranged from 2.3% DM to
samples making up the calibration data set. Four data- 36.0% DM, CA ranged from 2.8% DM to 19.4%
base calibration and validation sets were generated: A DM, NDF ranged from 18.7% DM to 76.2% DM,
generic database (n ¼ 1817 and n ¼ 203, respectively) ADF ranged from 5.2% DM to 50.3% DM, ADL
that included all the forage samples, grass forage ranged from 0.8% DM to 12.3% DM. Differences
(n ¼ 1276 and n ¼ 142, respectively), legume for- between minimum and maximum values of all the ref-
age (n ¼ 288 and n ¼ 32, respectively) and, other erence methods confirmed that the calibration set was
forage plants (n ¼ 253 and n ¼ 29, respectively). constructed with a diverse set of Colombian forages
The modified partial least squares (MPLS) regres- that were collected in different geographic regions, alti-
sion method was used to obtain equations for all the tude, seasons (dry and wet), at different days of
Ariza-Nieto et al. 47

Table 1. Descriptive statistics of the chemical characteristics of samples used in the calibration and validation sets for the generic
database.

Calibration set (n ¼ 1817) Validation set (n ¼ 203)

N Minimum Maximum Mean s.d. SEL n Minimum Maximum Mean s.d.

DM % 1237 84.2 97.8 92.1 1.9 3.19 139 84.4 96.9 92.0 2.1
CP % DM 1550 2.3 36.0 14.4 6.4 0.50 180 2.4 33.4 15.3 7.0
CA % DM 1279 2.8 19.4 10.1 2.5 0.20 148 4.0 15.7 10.2 2.3
NDF % DM 1768 18.7 76.2 54.1 10.8 1.55 201 18.7 72.2 54.2 10.5
ADF % DM 1714 5.2 50.3 26.5 6.7 1.45 194 6.3 49.7 26.5 7.3
ADL % DM 606 0.8 12.3 4.8 2.2 1.23 78 0.9 11.5 4.6 2.0
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; n: number of samples;
s.d.: standard deviation; SEL: standard error of laboratory.

Table 2. Descriptive statistics of the chemical characteristics of samples used in the calibration and validation sets for the specific
databases: grass forage, legume forage and other forage plants.

Nutrient n Minimum Maximum Mean s.d. n Minimum Maximum Mean s.d.

Calibration set (n ¼ 1276) Validation set (n ¼ 142)

Grass forage
DM % 804 84.2 97.8 92.1 2.0 90 84.4 96.9 91.9 2.3
CP % DM 1156 2.3 35.2 13.1 6.2 132 2.4 33.4 14.2 6.6
CA % DM 884 2.8 17.9 10.2 2.5 100 4.0 15.7 10.4 2.3
NDF % DM 1247 33.8 76.2 59.0 6.6 140 39.9 72.2 58.6 6.5
ADF % DM 1223 12.0 40.6 27.4 4.8 136 14.1 38.2 27.0 4.8
ADL % DM 337 1.7 12.3 5.3 2.5 46 2.0 11.5 4.9 2.2
Calibration set (n ¼ 288) Validation set (n ¼ 32)
Legume forage
DM % 230 86.3 97.0 92.3 1.7 26 88.5 96.5 92.6 1.9
CP % DM 211 6.9 31.3 19.2 5.4 26 8.3 31.8 18.7 6.5
CA % DM 201 4.4 15.0 9.4 2.2 25 5.8 15.4 9.0 2.3
NDF % DM 274 28.3 72.5 44.5 10.1 32 32.7 66.4 46.4 10.5
ADF % DM 262 9.3 50.3 27.2 9.2 31 14.0 49.7 27.4 11.0
ADL % DM 124 1.7 9.0 4.7 1.7 12 3.0 7.7 5.0 1.5
Calibration set (n ¼ 253) Validation set (n ¼ 29)
Other forage plants
DM % 203 87.7 94.8 91.9 1.3 23 89.0 94.2 91.9 1.2
CP % DM 183 6.2 36.0 17.0 5.7 22 6.8 33.1 18.1 7.9
CA % DM 194 5.6 19.4 10.8 2.8 23 7.2 14.1 10.2 2.0
NDF % DM 247 18.7 72.4 39.9 9.2 29 18.7 65.6 41.7 11.1
ADF % DM 229 5.2 48.3 21.0 9.1 27 6.3 45.8 23.1 10.8
ADL % DM 145 0.8 8.0 3.8 1.6 20 0.9 6.6 3.5 1.6
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; N: number of samples;
s.d.: standard deviation.

regrowth. The reduction of the database to smaller GLOBAL calibrations were developed to predict the
data sets (by family) generated more homogenous nutritional value of Colombian forages. Tables 3 to 6
matrices. The mean, minimum, maximum and stand- show the calibration statistics of the different calibra-
ard deviation values of the calibration set were also tion sets evaluated (grass forage, specific legume forage,
similar to those of the validation set for the family of specific other forage plants and generic forages data-
grass forages, legume forages and other forage plants base, respectively), together with the math treatments
(Table 2). and the number of PLS factors used. According to the
48 Journal of Near Infrared Spectroscopy 26(1)

Table 3. Equation development statistics of GLOBAL using a specific grass forage database with different math treatments.

Calibration Cross-validation

Nutrient Math PLS terms n s.d. SEC R2C SECV R2CV RPD

DM % 2441 9 804 2.05 1.12 0.70 1.24 0.63 1.65


CP % DM 2881 10 1156 6.16 0.84 0.98 0.86 0.98 7.12
CA % DM 2441 10 884 2.46 0.87 0.87 0.94 0.85 2.63
NDF % DM 2441 9 1247 6.56 2.20 0.89 2.32 0.88 2.83
ADF % DM 2881 9 1223 4.83 1.90 0.85 1.97 0.83 2.45
ADL % DM 2441 9 337 2.48 0.77 0.90 0.91 0.86 2.72
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; N: number of samples;
s.d.: standard deviation; SEC: standard error of calibration; R2C : coefficient of determination in calibration; SECV: standard error of cross-calibration; R2CV :
coefficient of determination in cross-validation; RPD: residual predicted deviation.

Table 4. Equation development statistics of GLOBAL using a specific legume forage database with different math treatments.

Calibration Cross-validation

Nutrient Math PLS terms n s.d. SEC R2C SECV R2CV RPD

DM % 2881 6 230 1.68 0.95 0.68 1.02 0.63 1.64


CP % DM 2881 10 211 5.38 0.77 0.98 0.87 0.97 6.17
CA % DM 2441 8 201 2.24 0.90 0.84 1.15 0.74 1.95
NDF % DM 1441 8 274 10.09 2.79 0.92 3.13 0.90 3.23
ADF % DM 2441 8 262 9.18 1.95 0.96 2.49 0.93 3.68
ADL % DM 2441 6 124 1.72 0.78 0.80 1.00 0.66 1.73
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; N: number of samples;
s.d.: standard deviation; SEC: standard error of calibration; R2C : coefficient of determination in calibration; SECV: standard error of cross-calibration; R2CV :
coefficient of determination in cross-validation; RPD: residual predicted deviation.

Table 5. Equation development statistics of GLOBAL using a specific other forage plants database with different math treatments.

Calibration Cross-validation

Nutrient Math PLS terms n s.d. SEC R2C SECV R2CV RPD

DM % 2881 7 203 1.23 0.68 0.70 0.77 0.60 1.58


CP % DM 2881 10 183 5.66 0.77 0.98 0.94 0.97 6.05
CA % DM 2441 10 194 2.77 0.79 0.92 1.06 0.85 2.62
NDF % DM 2441 7 247 9.19 2.76 0.91 3.23 0.88 2.85
ADF % DM 2441 5 229 9.09 2.05 0.95 2.32 0.93 3.93
ADL % DM 2881 6 145 1.60 0.61 0.85 0.74 0.80 2.17
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; N: number of samples;
s.d.: standard deviation; SEC: standard error of calibration; R2C : coefficient of determination in calibration; SECV: standard error of cross-calibration; R2CV :
coefficient of determination in cross-validation; RPD: residual predicted deviation.

results, when using the generic calibration with the entire lower (R2CV 0.66 and RPD 1.73), whereas CP, NDF,
forage database, spectral variations had a low correl- ADF, and CA calibration equations explained more
ation with DM (R2CV 0.58 and RPD 1.55) values, while than 80% the variance of the nutritional value of
in the specific calibration of each database, DM correl- Colombian forages, except for CA in the legume
ations improved in the grass forage (R2CV 0.63 and RPD forage calibration equation (R2CV 0.74).
1.65), legume forage (R2CV 0.63 and RPD 1.64), and The math treatment of the best predictive models
other forage plants (R2CV 0.60 and RPD 1.58). The obtained with each database was subsequently sub-
ADL in the generic forage calibration also showed a jected to external validation. For the specific grass
low correlation (R2CV 0.78 and RPD 2.14); however, in forage database math treatment selected and PLS fac-
the specific calibration of grass forage database and in tors used for each nutrient were: DM (2441 and 9),
the specific other forage plants database, the correlation CP (2881 and 10), CA (2441 and 10), NDF (2441
improved (R2CV 0.86 and RPD 2.72; R2CV 0.80 and RPD and 9), ADF (2881 and 9), ADL (2441 and 9)
2.17), but in the legume forage database, this value was (Table 3). For the specific legume forage database
Ariza-Nieto et al. 49

Table 6. Equation development statistics of GLOBAL using a generic forages database with different math treatments.

Calibration Cross-validation

Nutrient Math PLS terms n s.d. SEC R2C SECV R2CV RPD

DM % 2441 8 1237 1.88 1.17 0.61 1.21 0.58 1.55


CP % DM 2441 10 1550 6.42 0.89 0.98 0.93 0.98 6.91
CA % DM 2441 9 1279 2.51 1.02 0.83 1.07 0.82 2.34
NDF % DM 2441 9 1768 10.83 3.01 0.92 3.13 0.92 3.46
ADF % DM 2441 9 1714 6.71 2.55 0.86 2.68 0.84 2.50
ADL % DM 2441 9 606 2.24 0.93 0.83 1.05 0.78 2.14
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; N: number of samples;
s.d.: standard deviation; SEC: standard error of calibration; R2C : coefficient of determination in calibration; SECV: standard error of cross-calibration; R2CV :
coefficient of determination in cross-validation; RPD: residual predicted deviation.

Table 7. Monitoring statistics of the validation set of the grass forage database using GLOBAL and LOCAL.

Nutrient Math n Bias SEPc r2 Best locala Bias SEPc r2

Global generic Local generic

DM % 2441 90 0.16 1.20 0.73 1,501,010 0.20 0.77 0.89


CP % DM 2441 132 0.10 0.89 0.98 250,810 0.09 0.82 0.99
CA % DM 2441 100 0.06 1.08 0.78 2001,010 0.02 0.95 0.84
NDF % DM 2441 140 0.06 2.63 0.84 200,810 0.04 1.72 0.93
ADF % DM 2441 136 0.17 2.18 0.80 150,110 0.09 1.52 0.90
ADL % DM 2441 46 0.11 0.98 0.80 100,510 0.09 0.69 0.90
Global specific Local specific

DM % 2441 90 0.15 1.04 0.79 150,910 0.18 0.81 0.88


CP % DM 2881 132 0.11 0.91 0.98 250,910 0.08 0.75 0.99
CA % DM 2441 100 0.19 0.98 0.82 20,039 0.02 0.89 0.85
NDF % DM 2441 140 0.11 2.15 0.89 10,069 0.08 1.75 0.93
ADF % DM 2881 136 0.08 1.57 0.90 150,110 0.04 1.51 0.90
ADL % DM 2441 46 0.14 0.90 0.83 10,017 0.18 0.71 0.89
a
Local settings: number of selected samples, number of PLS factors (minimum number, maximum number).
Generic: calibration using the entire forage database. Specific: calibration using the single grass forage database. DM: dry matter; CP: Crude protein; CA:
crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; n: number of samples; SEPc: standard error of prediction
corrected for bias; r2: coefficient of determination.

math treatments and PLS factors used were: DM prediction in order to increase the robustness of the
(2881 and 6), CP (2881 and 10), CA (2441 and 8), calibration, although this usually reduces the predic-
NDF (1441 and 8), ADF (2441 and 8), ADL (2441 tion.20 As could be expected, the GLOBAL generic cali-
and 6) (Table 4). For the specific other forage plants brations of each database had the largest SEPc values
database math treatments selected and PLS factors in the majority of nutrients evaluated, when compared
used in each parameter were: DM (2881 and 7), CP to specific GLOBAL calibrations, but in the case of the
(2881 and 10), CA (2441 and 10), NDF (2441 and 7), grass forage database, the GLOBAL generic for the CP
ADF (2441 and 5), ADL (2881 and 6) (Table 5). In nutrient was slightly lower compared to the GLOBAL
the generic forage database, the math treatment specific (0.89% DM vs. 0.91% DM) (Table 7). For the
selected and PLS factors used for each nutrient legume forage database, a similar behavior was
were: DM (2441 and 8), CP (2441 and 10), CA observed for the CA nutrient (1.12% DM vs. 1.50%
(2441 and 9), NDF (2441 and 9), ADF (2441 and DM) (Table 8).
9), ADL (2441 and 9) (Table 6). Within the LOCAL algorithm, in the generic and
In the GLOBAL calibration, the selection of the specific calibrations, SEPc values varied in each nutri-
samples is a critical step with greater effects on the pre- ent depending on the database used. In the grass forage
cision and accuracy of the process. In this sense, the database DM, NDF, and ADL presented lower SEPc
selection of samples for calibration should include all in the generic LOCAL calibration compared to the spe-
possible sources of variation encountered during the cific LOCAL calibration, while, in the other nutrients
50 Journal of Near Infrared Spectroscopy 26(1)

Table 8. Monitoring statistics of the validation set of the legume forage database using GLOBAL and LOCAL.

Nutrient Math n Bias SEPc r2 Best locala Bias SEPc r2

Global generic Local generic

DM % 2441 26 0.04 1.39 0.45 10,099 0.02 1.25 0.56


CP % DM 2441 27 0.10 1.32 0.96 15,088 0.03 0.96 0.98
CA % DM 2441 25 0.06 1.12 0.77 10,057 0.07 0.89 0.86
NDF % DM 2441 32 0.94 4.41 0.83 10,099 0.75 3.08 0.92
ADF % DM 2441 31 0.90 2.63 0.95 10,099 0.81 1.53 0.98
ADL % DM 2441 12 0.56 0.73 0.80 5012 1.04 0.80 0.73
Global specific Local specific

DM % 2881 26 0.07 1.32 0.52 15,066 0.01 1.22 0.59


CP % DM 2881 27 0.03 0.83 0.98 10,099 0.09 0.83 0.98
CA % DM 2441 25 0.01 1.50 0.59 10,044 0.07 1.25 0.71
NDF % DM 1441 32 0.21 3.36 0.90 10,088 0.05 3.38 0.90
ADF % DM 2441 31 0.20 1.79 0.98 10,099 0.55 1.47 0.98
ADL % DM 2441 12 0.18 0.64 0.84 5058 0.07 0.67 0.83
a
Local settings: number of selected samples, number of PLS factors (minimum number, maximum number).
Generic: Calibration using the entire forage database. Specific: Calibration using the single legume forage database.
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; n: number of samples;
SEPc: standard error of prediction corrected for bias; r2: coefficient of determination.

higher SEPc values were obtained when using the spe- r2 0.88) and the lowest increase was for the CP variable
cific LOCAL calibration (Table 7). In the legume (r2 0.98 vs. r2 0.99).
forage database all the nutrients presented lower The specific LOCAL calibration using samples in the
SEPc values in the specific LOCAL calibration, grass forage database showed lower SEPc values com-
except for CA and NDF that showed a lower SEPc pared to using samples in the only one database; how-
value when using the generic LOCAL calibration ever, SEPc values with the only one database were
(0.89% DM vs. 1.25% DM and 3.08% DM vs. lower than those obtained using the specific LOCAL
3.38% DM, respectively) (Table 8). In the other calibration of the legume forage, except for CP
forage plants database CA (0.66% DM vs. 0.74% (0.87% DM vs 0.83% DM). Using samples in the
DM), NDF (3.16% DM vs. 3.35% DM), and ADF other forage plants database, SEPc values were lower
(2.31% DM vs. 2.49% DM) had lower SEPc values using the only one database compared to the specific
in the generic LOCAL calibration, while DM (0.90% LOCAL approach for CP (0.87% DM vs. 0.91% DM),
DM vs. 0.79% DM) and ADL (1.09% DM vs. 0.57% NDF (2.40% DM vs. 3.35% DM), and ADF (1.71%
DM) had lower SEPc values when using the generic DM vs. 2.49% DM).
LOCAL calibration, CP presented similar SEPc The comparison of the bias values between the
values in both calibrations (Table 9). GLOBAL and LOCAL algorithm showed lower
The use of the LOCAL algorithm obviates the need values for the nutrients evaluated in the whole forage
to select between precision and robustness of a calibra- database (Tables 10). Lower values were more consist-
tion; Consequently, when the algorithm is used for an ent for nutrients related to fiber fractions of forage
unknown sample, it must be predicted that a specific resources: NDF (0.26% DM vs. 0.13% DM), ADF
equation is obtained from a group of samples selected (0.10% DM vs. 0.07% DM), and ADL (0.14%
from a database based on its spectral similarity to the DM vs. 0.07% DM). On the other hand, considering
unknown sample, which allows obtaining a high degree that the total errors of a prediction by NIR spectros-
of accuracy in prediction; On the other hand, robust- copy can be divided in the sum of the errors,22 it can be
ness is achieved by building a comprehensive database inferred that the greater precision of the LOCAL algo-
that must have thousands of samples.21 In this study, rithm is attributable to a considerable reduction in the
the accuracy of LOCAL algorithm, was tested on the error due to the chemometric model used.
validation files using only the entire forage database NIR spectroscopy calibration studies have tradition-
(Table 10). Again, LOCAL SEPc values were lower ally used what are known as the GLOBAL equations
than those reported by the GLOBAL calibration and by which a GLOBAL calibration group is used to
the r2 value also was improved with the LOCAL obtain predictive models; however, the potential of
algorithm. The greatest increases were observed in the using a local algorithm was evident in the literature,
estimates of DM (r2 0.65 vs. r2 0.81), CA (r2 0.78 vs. r2 but the major limitation was initially the large
0.86), ADF (r2 0.87 vs. r2 0.95) and ADL (r2 0.78 vs. number of computations required for the prediction
Ariza-Nieto et al. 51

Table 9. Monitoring statistics of the validation set of the other forage plants database using GLOBAL and LOCAL.

Nutrient Math n Bias SEPc r2 Best locala Bias SEPc r2

Global generic Local generic

DM % 2441 23 0.16 1.14 0.21 1,501,010 0.09 0.90 0.43


CP % DM 2441 23 0.14 1.15 0.98 150,410 0.32 0.91 0.99
CA % DM 2441 23 0.15 1.10 0.71 50,410 0.01 0.66 0.89
NDF % DM 2441 29 0.47 4.11 0.87 15,022 0.52 3.16 0.92
ADF % DM 2441 27 0.04 4.26 0.85 150,510 0.23 2.31 0.96
ADL % DM 2441 20 0.03 1.02 0.65 10,058 0.02 1.09 0.84
Global specific Local specific

DM % 2881 23 0.13 0.94 0.39 10,055 0.07 0.79 0.57


CP % DM 2881 23 0.04 0.90 0.99 10,068 0.19 0.91 0.99
CA % DM 2441 23 0.25 0.84 0.82 5077 0.01 0.74 0.87
NDF % DM 2441 29 0.49 3.35 0.91 10,015 0.82 3.35 0.92
ADF % DM 2441 27 0.26 2.95 0.93 10,066 0.17 2.49 0.95
ADL % DM 2881 20 0.12 0.68 0.84 10,034 0.15 0.57 0.87
a
Local settings: number of selected samples, number of PLS factors (minimum number, maximum number).
Generic: Calibration using the entire forage database. Specific: Calibration using the single other forage plants database.
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; n: number of samples;
SEPc: standard error of prediction corrected for bias; r2: coefficient of determination.

Table 10. Monitoring statistic of the validation set using GLOBAL and LOCAL calibrations with the entire forage database.

Global Local

Nutrient Math n Bias SEPc r2 Best locala Bias SEPc r2

DM % 2441 139 0.07 1.23 0.65 1,501,010 0.10 0.92 0.81


CP % DM 2441 182 0.08 0.99 0.98 150,810 0.04 0.87 0.99
CA % DM 2441 148 0.05 1.08 0.78 1,001,010 0.04 0.89 0.86
NDF % DM 2441 201 0.26 3.21 0.91 150,110 0.13 2.40 0.95
ADF % DM 2441 194 0.10 2.62 0.87 100,110 0.07 1.71 0.95
ADL % DM 2441 78 0.14 0.96 0.78 100,910 0.07 0.71 0.88
a
Local settings: number of selected samples, number of PLS factors (minimum number, maximum number).
DM: dry matter; CP: crude protein; CA: crude ash; NDF: neutral detergent fiber, ADF: acid detergent fiber; ADL: acid detergent lignin; n: number of samples;
SEPc: standard error of prediction corrected for bias; r2: coefficient of determination.

of routine analyses.23–25 This limitation was overcome forages better than using the GLOBAL approach. In
when the LOCAL algorithm was presented, which was this sense, LOCAL can accurately predict the compos-
able to carry out fast and accurate predictions.13,20 In ition of different forages using only one database, and
this study, the ability of LOCAL to select forage sam- could offer a practical way to develop a robust equation
ples was not affected by the number and variety of for- taking into account the biodiversity of Colombian
ages include in the database because diverse forage forages.
families were included in the only one database. This The LOCAL algorithm combines the advantages of
is because the LOCAL algorithm selects forage samples the GLOBAL strategy as it uses a simple database that
and creates a unique calibration equation for a single covers all the expected variability of the population
sample, in addition is not necessary to identify the type sample under study with an accuracy that is obtained
of sample that is going to be analyzed, however accur- through specific calibrations. Therefore, a greater pre-
acy of the prediction is guaranteed because if the GH cision and accuracy of the predictions of the quality of
and NH of the unknown forage sample are within spe- forages using the LOCAL algorithm is obtained by
cifications the analytical result can be reliable. evaluating the averages of the statistical data for SEP,
bias, and r2 compared to the predictions obtained using
GLOBAL calibrations.
Conclusions The LOCAL algorithm has proven to be an invalu-
The accuracy of prediction of LOCAL algorithm has able strategy to minimize errors in calibration equa-
proved to predict the nutritional value of Colombian tions to predict the nutritional value of forages given
52 Journal of Near Infrared Spectroscopy 26(1)

the complexity of the process and the wide range of feed CSSA, SSSA, Madison, WI. p. 277–300. 1974.
resources used in ruminant production processes. DOI:10.2134/1974.foragefertilization.c13.
NIR spectroscopy technology provides a rapid 9. Gallaher RN and Pitman WD. Conservation of forages
sensor for the routine analysis of the nutritional quality in the tropics and subtropics. In: Sotomayor-Rios A and
dynamics of grasses, legumes and other forage Pitman WD (eds) Tropical forage plants: development and
use, 2001.
resources used in grazing feeding systems across a
10. Workman J and Weyer L (eds.) In: Practical guide to
broad spectrum of the tropical landscape.
interpretive near infrared spectroscopy. Boca Raton:
CRC Press, Taylor and Francis Group, 2008.
Acknowledgements 11. Norris KH, Barnes RF, Moore JE, et al. Predicting
We acknowledge the support from livestock Network of forage quality by infrared reflectance spectroscopy.
Corpoica for sending the forage samples. We are also grateful J Anim Sci 1976; 43: 889–897.
to Alipio Loaiza, Doris Montañez and Enrique Castro for 12. Alomar D, Fuchslocher R, Cuevas J, et al. Prediction of
their technical support in the chemical and spectral analysis. the composition of fresh pastures by near infrared reflect-
ance or interactance-reflectance spectroscopy. J. Chilean
Declaration of conflicting interests J Agric Res 2009; 69: 198–206.
The author(s) declared no potential conflicts of interest with 13. Andueza D, Picard F, Jestin M, et al. NIRS prediction of
respect to the research, authorship, and/or publication of this the feed value of temperate forages: efficacy of four cali-
article. bration strategies. J Anim 2011; 5: 1002–1013.
14. Cozzolino D, Acosta YG and Garcı́a J. Application of
near infrared reflectance spectroscopy (NIRS) to forage
Funding
evaluation in Uruguay. In: International grasslands con-
The author(s) disclosed receipt of the following financial sup- gress, 19, San Pedro, 2001, Proceedings. Piracicaba:
port for the research, authorship, and/or publication of this FEALQ, p. 370.
article: We acknowledge funding from Agriculture Minister 15. Official methods of analysis of AOAC International. 19th
of Colombia. ed. 2012, Official Journal of the European Communities,
Gaithersburg, Maryland, USA.
References 16. Van Soest PJ and Robertson JB. Analysis of forages and
1. Decruyenaere V, Planchon V, Dardenne P, et al. fibrous foods a laboratory manual for animal science.
Prediction error and repeatability of near infrared reflect- Ithaca, NY: Cornell University, 1985.
ance spectroscopy applied to faeces samples in order to 17. Shenk JS and Westerhaus MO. The application of near
predict voluntary intake and digestibility of forages by Infrared Reflectance Spectroscopy (NIRS) to forage ana-
ruminants. Anim Feed Sci Technol 2005; 205: 49–59. lysis. In: Fahey GC Jr. (ed.) Forage quality, evaluation and
2. Hanna WW, Dujardin M and Monson WG. Using diverse utilization. Madison: Soil Science Society of America/
species to improve quality and yield in the Pennisetum American Society of Agronomy/Crop Science Society of
genus. In: Proc. XVI Int. Grassl. Congr. Nice, France, America, 1994, pp.406–449.
1989. 18. Shenk JS and Westerhaus MO. Calibration system for
3. Tassone S, Masoero G and Peiretti PG. Vibrational spec- spectrographic analyzing instruments. U.S. Patent
troscopy to predict in vitro digestibility and the maturity 5798526, 1998.
index of different forage crops during the growing cycle 19. Workman J and Mark H. Choosing the best regression
and after freeze- or oven-drying treatment. J. Anim Feed model. J. Spectrosc 2015; 30.
Sci Technol 2014; 194: 12–25. 20. Shenk JS, Westerhaus MO and Berzaghi P. Investigation
4. Vera RR. Country pasture/forage resource profiles, of a LOCAL calibration procedure for NIR instruments.
Colombia. Food and Agriculture Organization of the J. Near Infrared Spectrosc 1997; 5: 223–232.
United Nations (FAO), www.fao.org/ag/agp/agpc/doc/ 21. Berzaghi P, Shenk JS and Westerhaus MO. LOCAL pre-
counprof/PDF%20files/Colombia-English.pdf, 2006 diction with near infrared multi-product databases. J.
(accessed 1 November 2017). Near Infrared Spectrosc 2000; 8: 1–9.
5. Pitman WD Environmental constraints to tropical forage 22. Mark H and Workman J. Statistics in spectroscopy. New
plant adaptation and productivity. In: Sotomayor-Rios A; York: Academic Press, 1991.
Pitman WD, (eds.) Tropical forage plants: Development 23. Davies AM, Britcher HV, Franklin JG, et al. The appli-
and use. CRC Press, Boca Raton, FL, USA. 2000, p. 17– cation of Fourier-transformed near-infrared spectra to
23. DOI: 10.1201/9781420038781.ch2. quantitative analysis by comparison of similarity indices
6. Aiken GE, Pitman WD, Chambliss CG, et al. Responses (CARNAC). J Microchimica Acta 1988; 94: 61–64.
of yearling steers to different stocking rates on a subtrop- 24. Naes T, Isaksson T and Kowalski B. Locally weighted
ical grass-legume pasture. J. Anim Sci 1991; 69: 3348. regression and scatter correction for near-infrared reflect-
7. de Sant-Anna AC, Jantalia CP, Sá JM, et al. Changes in ance data. J Anal Chem 1990; 62: 664–673.
soil organic carbon during 22 years of pastures, cropping 25. Aastveit AH and Marum P. Near-infrared reflectance
or integrated crop/livestock systems in the Brazilian spectroscopy: different strategies for local calibrations
Cerrado. Nutr Cycl Agroecosyst 2017; 108: 101–120. in analyses of forage quality. J Appl Spectrosc 1993; 47:
8. Vicente-Chandler J. Fertilization of humid tropical grass- 463–469.
lands. In: Mays DA (ed) Forage fertilization., ASA,

You might also like