Prediction of Forage Quality from Remotely Sensed Data: Comparison of Cultivar‐Specific and Cultivar‐Independent Equations Using Three Methods of Calibration

Abstract

RESEARCHT he study findings suggested that cultivar-specifi c (CS) equations were not necessary for prediction of N concentration. Both cultivar-independent (CI) and CS neutral detergent fi ber (NDF) and acid detergent fi ber (ADF) equations revealed poor to moderate performance. Comparison of partial least square (PLS), multiple linear regression with maximum r 2 improvement (MAXR), and artifi cial neural network (ANN) calibration and validation results revealed that ANN rarely outperformed PLS and MAXR. Autocorrelation was apparent in the MAXR CS equations for NDF and ADF. From this study it appears that when using hyperspectral data, PLS is the preferred approach since it minimizes the eff ect of autocorrelation among wavelengths in close proximity to each other. Although ANN rarely outperformed PLS or MAXR it is a useful tool in identifying or isolating those wavelengths that may be useful in predicting forage quality.Laboratory chemical analysis methods have long been used for quantifying forage nutritive quality (Kellems and Church, ABSTRACT Remote sensing of pasture canopies could reduce laborious fi eld sampling procedures and provide timely information to ranch and livestock managers. This study evaluated partial least squares (PLS) regression, multiple linear regression with maximum r 2 improvement (MAXR), and artifi cial neural networks (ANN) for development of robust equations (i.e., cultivar-independent [CI]) to predict concentrations of nitrogen (N), neutral detergent fi ber (NDF), and acid detergent fi ber (ADF) from hyperspectral remotely sensed refl ectance data collected from six cultivars of live bermudagrass canopies. Specifi c study objectives were to: i) develop both cultivar-independent and cultivar-specifi c (CS) equations to predict N, NDF, and ADF from remotely sensed canopy refl ectance, ii) validate the accuracy of these equations on a independent data set, iii) compare the results from the CS and CI equations to determine how well equations developed across cultivar perform relative to equations developed for a specifi c cultivar of bermudagrass, and iv) compare contributing wavelengths across calibration method.

Development of Leaf Spectral Models for Evaluating Large Numbers of Sugarcane Genotypes

“…Our results confirmed that sugarcane leaf relative chlorophyll and N contents could be well estimated across a wide range of genotypes and throughout the growing season (Table 4). This is consistent with early reports in other crops (Carter and Estep, 2002; Zhao et al, 2005a, 2005b; Starks and Brown, 2010).…”

“…Leaf chemical composition (SPAD, N and C contents, and C to N ratio), yield components ( juice sucrose, TCH, CRS, and TSH), and leaf refl ectance data were randomly assigned into two groups (i.e., two independent datasets with equal numbers of genotypes). The trailing dataset was used to develop leaf refl ectance models for prediction of leaf chemical and yield traits using the methods of maximum r 2 improvement (MAXR) regression (Zhao et al, 2007) and partial least-square (PLS) regression (Starks and Brown, 2010) under PROC REG and PROC PLS, respectively, in SAS statistical software (SAS Institute, 2007). The second dataset was used for model validations.…”

“…The r 2 values continued increasing considerably with increases up to 10 wavebands for most of the leaf chemical and yield traits except for leaf relative chlorophyll level. Earlier studies suggested that leaf chlorophyll and/or leaf N contents could be well predicted using leaf refl ectance values in a few wavebands or refl ectance ratios in two wavebands (Chappelle et al, 1992;Daughtry et al, 2000;Carter and Spiering, 2002;Zhao et al, 2005aZhao et al, , 2005b; however, PLS regression has been reported as the preferred approach for analysis of hyperspectral data because it minimizes the eff ect of autocorrelation among wavelengths in close proximity to each other (Lindberg et al, 1983;Starks and Brown, 2010). Based on these previous reports, we agreed that PLS regression was probably the best approach for developing refl ectance models to assess these leaf-and yield-traits for sugarcane genotypic evaluation.…”

Bermudagrass, Wheat, and Tall Fescue Crude Protein Forage Estimation using Mobile‐Platform, Active‐Spectral and Canopy‐Height Data

“…An R 2 = 0.82 was reported by Starks et al (2008) for N concentration to spectral reflectance in the 705‐ to 1685‐nm range in warm‐season grass pastures. Starks and Brown (2010) noted no cultivar‐specific model offered an advantage in improving estimation accuracy for N concentrations of three bermudagrass cultivars using hyperspectral reflectance measurements from 350 to 1125 nm. Guo et al (2010) observed R 2 values of 0.63 for forage CP estimation when hyperspectral data in the 350‐ to 2500‐nm range were collected in a semiarid, mixed‐prairie ecosystem.…”

Rapid Determination of Carbon, Nitrogen, and Phosphorus Contents of Field Crops in China Using Visible and Near‐Infrared Reflectance Spectroscopy

“…The variation inherent in the sample population studied could be increased when the sample set consists of a number of heterogeneous samples. However, the development of separate vis-NIR calibrations for different crop species requires more efforts and resources for calibrations and periodic validations than the use of a global calibration model (Shenk et al, 1979;Starks and Brown, 2010). Use of the local (e.g., species-specific) calibrations may therefore be preferable in applications requiring maximum precision.…”