Hyperspectral proximal sensing of the botanical composition and nutrient content of New Zealand pastures : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Earth Science

Abstract

The potential of hyperspectral proximal sensing to quantify sward characteristics important in making critical decisions on the management of sheep and dairy pastures in New Zealand has been investigated. Hyperspectral data were acquired using an ASD FieldSpec® Pro FR spectroradiometer attached to the Canopy Pasture Probe (CAPP). The CAPP was developed to enable the collection of in situ reflectance data from New Zealand pasture canopies independent of ambient light conditions. A matt white ceramic tile was selected as a reflectance standard to be used with the CAPP, after testing a variety of materials. Pasture reflectance factor spectra between 350-2500 nm (with spectral resolutions of 3 nm between 350-1000 nm and 10 nm between 1000-2500 nm) and pasture samples were collected from six hill country and lowland areas, across all seasons (August 2006 to September 2007) in a number of regions in the North Island of New Zealand. After pre-processing (e.g. spectral averaging, de-stepping, elimination of noisy wavelengths, smoothing) the spectral data collected from sites were correlated against pasture botanical composition (expressed as proportions of grass, legume and weed) and pasture nutrients (nitrogen, phosphorus, potassium, calcium, magnesium, sodium and sulphur) expressed in percentage of dry matter (%) and amount (kg ha-1) using partial least squares regressions (PLSR). The accuracy and precision of the calibrations were tested using either the full cross-validation leave-one-out method or testing datasets. Regressions were carried out using the reflectance factor data per se and after mathematical transformation, including first derivative, absorbance and continuum-removed spectra. Overall best results were obtained using the first derivative data. The quality of predictions varied greatly with the pasture attribute, site and season. Some reasonable results were achieved for the prediction of pasture grass and legume proportions when analysing samples collected during autumn (grass: R2 > 0.81 and SD/RMSEP 2.3 and legume: R2 > 0.80 and SD/RMSEP 2.2), but predicting pasture weed content was poor for all sites and seasons (R2 = 0.44 and SD/RMSEP = 1.2). The inaccurate predictions might be explained by the fact that the diversity found in the field and observed in the pasture spectral data was not taken into account in the pasture botanical separation. The potential for using proximal sensing techniques to predict pasture nutrients in situ was confirmed, with the sensing of pasture N, P and K increased by the procedure of separating the data according to the season of the year. The full potential of the technology will only be realised if a substantial dataset representing all the variability found in the field is gathered. The importance of obtaining representative datasets that embrace all the biophysical factors (e.g. pasture type, canopy structure) likely to affect the relat ionship, when building prediction calibrations, was highlighted in this research by the variance in the predictions for the same nutrient using different datasets, and by the inconsistency in the number of common wavelengths when examining the wavelengths contributing to the relationship. The ability to use a single model to predict multiple nutrients, or indeed individual nutrients, will only come through a good understanding of the factors likely to influence any calibration function. It has been demonstrated in this research that reasonably accurate and precise pasture nutrient predictions (R2 > 0.74 and SD/RMSEP 2.0) can be made from fresh in situ canopy measurements. This still falls short of the quality of the predictions reported for near infrared reflectance spectroscopy (NIRS) for dried, ground samples analysed under controlled laboratory conditions