Premium
Imaging Spectroscopy for On‐Farm Measurement of Grassland Yield and Quality
Author(s) -
Schut A. G. T.,
Heijden G. W. A. M.,
Hoving I.,
Stienezen M. W. J.,
Evert F. K.,
Meuleman J.
Publication year - 2006
Publication title -
agronomy journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.752
H-Index - 131
eISSN - 1435-0645
pISSN - 0002-1962
DOI - 10.2134/agronj2005.0225
Subject(s) - yield (engineering) , calibration , partial least squares regression , hyperspectral imaging , environmental science , sampling (signal processing) , approximation error , grassland , linear regression , fraction (chemistry) , mathematics , statistics , soil science , remote sensing , agronomy , chemistry , computer science , materials science , geography , biology , organic chemistry , filter (signal processing) , metallurgy , computer vision
Grassland management has a large influence on the operating cost and environmental impact of dairy farms and requires accurate, detailed, and timely information about the yield and quality of grass. Our objective was to evaluate imaging spectroscopy as a means to obtain accurate, detailed, and rapid measurements of grass yield and quality. The work consisted of three steps. In the first step, a new mobile measurement system comprising several hyperspectral sensors was constructed and calibrated on measurements collected in six field experiments in the Netherlands in 2 yr. A partial least squares regression model was used to fit parameters derived from hyperspectral images to values of DM (dry matter) yield and quality obtained through destructive sampling. Leave‐k‐out cross validation showed relative errors of prediction of 8 to 22% (167–477 kg DM ha −1 absolute) for DM yield, 21% (0.07 absolute) for the fraction of clover in DM, 6 to 12% for nutrient concentration, 15 to 16% for sugar concentration, and 3 to 5% for feeding values. In the second step, the measurement system was used to predict grassland yield and quality on fields from two farms. In the third step, the absence of calibration data for a specific date was simulated with a leave‐harvest‐out procedure. Predictions of absolute values were strongly biased due to system instability. Prediction of relative values was good, with a mean absolute error of 183 kg ha −1 for DM yield. The instability of the measurement system requires that duosampling must be used for prediction of absolute values.