Imaging spectroscopy for characterisation of grass swards

Keywords: Imaging spectroscopy, imaging spectrometry, remote sensing, reflection, reflectance, grass sward, white clover, recognition, characterisation, ground cover, growth monitoring, stress detection, heterogeneity quantification

The potential of imaging spectroscopy as a tool for characterisation of grass swards was explored with respect to growth monitoring, detection of nitrogen and drought stress, and assessment of dry matter yield, clover content, nutrient content, feeding value, sward heterogeneity and production capacity. To this end, an experimental imaging spectroscopy system was developed. The system detects reflection in image lines in the wavelength range from 405-1659 nm with three different sensors at 1.3 m above the soil surface. Spectral resolution varies between 5-13 nm, and spatial resolution between 0.28-1.45 mm ²per pixel at the soil. As a result of system design, reflection intensity is a function of leaf height and leaf angle. The system was tested on mini swards grown in containers. For each mini sward, 42 image lines were recorded in a regular sampling pattern per recording event.

Five experiments were conducted with Lolium perenne L. and/or Trifolium repens L. mini swards. In these experiments degree of sward damage, level of nitrogen (N) application (two experiments), water supply and white clover content were varied. In the sward damage experiment and in one N experiment light interception was recorded regularly; at harvest, also crop height and canopy reflectance (with a Cropscan) was measured. Mini swards were harvested at a fixed level and in one of the N experiments in three strata. During the experiments, hyperspectral reflectance was recorded 2-4 times per week.

Image lines were classified to separate pixels containing soil, dead material and green leaves. These classes were subdivided into reflection intensity classes. Ground cover (GC), reflection intensity, image line texture, spatial heterogeneity and patterns, and spectral characteristics of green leaves were quantified. An index of reflection intensity (IRI) measured the distribution of green pixels over intensity classes and quantified vertical canopy geometry. Horizontal sward heterogeneity was quantified with the spatial standard deviation of GC (GC-SSD) and logarithmically transformed GC (TGC-SSD), and image line texture and spatial patterns with wavelet entropy (WE). Spectral characteristics were quantified with shifts of various spectral edges. Partial least squares (PLS) models combining spectral and spatial information were calibrated and validated on two separate data-sets from the sward damage and one N experiment, in order to predict dry matter (DM) yield, feeding quality and nutrient content. Effects of replicate observations on reduction of prediction error were studied for different fractions of model bias.

GC was differently related to light interception under a cloudy sky and under a clear sky (R ²_adj = 0.87-0.94) and also for dense and open swards. Growth was accurately monitored with evolution of GC and IRI, and GC and IRI at harvest were strongly related to DM yield (R ²_adj = 0.75-0.82). Seasonal means of GC and IRI were strongly (R ²_adj = 0.77-0.93) related to annual DM matter yield and light interception capacity. There was a clear (R ²_adj = 0.69) relation between seasonal mean GC-SSD and tiller density. Seasonal means of GC-SSD differentiated dense from damaged swards. The WE of image line texture robustly differentiated clover from grass swards, while mixtures had intermediate values. Position of spectral edges was strongly related to reflection intensity. This relation differed for grass and clover swards, varied with N supply level and changed after harvesting canopy strata. Leaf angle was identified as the most important factor affecting this relationship. Drought stress was detected in an early stage, when DM content of leaves was still below 20%, from shifts of edges near water absorption features. A combination of shifts of the green and red edge was strongly related (R ²=0.95) to DM yield reduction due to N shortage. The prediction errors relative to the mean (of validation sets) of the PLS models were 6.2-11.7 % for N content, 5.5-9.1 % for DM content, 13.6-18.7 % for sugar content, 6.0-7.5 % for ash content, and 3.5-4.8 % for crude fibre content. Predictions of P, K, S, Mg, Na and Fe were robust in both experiments. Combining GC and IRI with mean sward spectra resulted in a prediction error of 235-268 kg DM ha ^-1for yields of less than 1000 up to 4000 kg DM ha ^-1. Multiple observations may reduce the mean prediction error for DM yield with 27 to 54%, depending on model bias and number of observations. The accuracy of DM yield assessment with imaging spectroscopy was better than with the disk plate meter or Cropscan. It is concluded that imaging spectroscopy is a powerful tool in grassland research and may provide valuable information for fine-tuning of grassland management. In this study it provided fast, automatic and non-destructive means for monitoring and quantification of growth, and estimating dry matter yield, spatial heterogeneity and sward damage, nitrogen and water deficiency, clover content, feeding quality and nutrient content of swards. Finally, system requirements for application of imaging spectroscopy in the field are discussed.