No abstract

It is widely known that the visible and near infrared (VIS-NIR) spectroscopy has the potential of estimating soil total nitrogen (TN), organic carbon (OC) and moisture content (MC) due to the direct spectral responses these properties have in the NIR region. However, improving the predication accuracy requires advanced modelling techniques, particularly when measurement is planned for fresh (wet and un-processed) soil samples. The aim of this work is to compare the predictive performance of two linear multivariate and two machine learning methods for TN, OC and MC. The two multivariate methods investigated included Principal Component Regression (PCR) and Partial Least Squares Regression (PLSR), whereas the machine learning methods included Least-Squares Support Vector Machines (LS-SVM), and Cubist. A mobile, fibre type, VIS-NIR spectrophotometer was utilised to collect soil spectra (305-2200 nm) in diffuse reflectance mode from 140 wet soil samples collected from one field in Germany. The results indicate that machine learning methods are capable of tackling non-linear problems in the dataset. LS-SVMs and the Cubist method over-performed the linear multivariate methods for the prediction of all three soil properties studied. LS-SVM provided the best prediction for MC (root mean square error of prediction (RMSEP) = 0.457 % and residual prediction deviation (RPD) = 2.24) and OC (RMSEP = 0.062 % and RPD = 2.20), whereas the Cubist method provided the best prediction for TN (RMSEP = 0.071 and RPD=1.96).

No abstract

This paper assesses the potential use of a hyperspectral camera for measurement of yellow rust and fusarium head blight in wheat and barley canopy under laboratory conditions. Scanning of crop canopy in trays occurred between anthesis growth stage 60, and hard dough growth stage 87. Visual assessment was made at four levels, namely, at the head, at the flag leaves, at 2 nd and 3 rd leaves, and at the lower canopy. Partial least squares regression (PLSR) analyses were implemented separately on data captured at four growing stages to establish separate calibration models to predict the percentage coverage of yellow rust and fusarium head blight infection. Results showed that the standard deviation between 500 and 650 nm and the squared difference between 650 and 700 nm wavelengths were found to be significantly different between healthy and infected canopy particularly for yellow rust in both crops, whereas the effect of water-stress was generally found to be unimportant. The PLSR yellow rust models were of good prediction capability for 6 out of 8 growing stages, a very good prediction at early milk stage in wheat and a moderate prediction at the late milk development stage in barley. For fusarium, predictions were very good for seven growing stages and of good performance for anthesis growing stage in wheat, with best performing for the milk development stages. However, the root mean square error of predictions for yellow rust were almost half of those for fusarium, suggesting higher prediction accuracies for yellow rust measurement under laboratory conditions.

There is a lack of information on optimal measurement configuration of hyperspectral imagers for on-line measurement of a wheat canopy, this paper aims at identifying this configuration using a passive sensor (400-750 nm). The individual and interaction effects of camera height and angle, sensor integration time and light source distance and height on the spectra's signal-to-noise ratio (SNR) were evaluated under laboratory scanning conditions, from which an optimal configuration was defined and tested under on-line field measurement conditions. The influences of soil total nitrogen (TN) and moisture content (MC) measured with an on-line visible and near infrared (vis-NIR) spectroscopy sensor on SNR were also studied. Analysis of variance and principal component analysis (PCA) were applied to understand the effects of the laboratory considered factors and to identify the most influencing components on SNR. Results showed that integration time and camera height and angle are highly influential factors affecting SNR. Among integration times of 10, 20 and 50 ms, the highest SNR was obtained with 1.2 m, 1.2 m and 10° values of light height, light distance and camera angle, respectively. The optimum integration time for on-line field measurement was 50 ms,