Evaluation of cut alfalfa moisture content and operative conditions by hyperspectral imaging combined with chemometric tools: In-field application

Abstract

High quality of hay or silage can be obtained by monitoring the moisture level at mowing or harvest time to reduce the losses of yield and nutrients and preserve feeds in the long term. Mechanical conditioning is one of the most common ways to increase the rate of water loss from forage during drying. Near Infrared (NIR) spectroscopy is a good alternative to the classical gravimetric method to timely provide information on forage moisture content. The aim of this work was to evaluate the potential of in-field Vis/NIR hyperspectral imaging combined with chemometric tools to monitor alfalfa parameters after conditioning. Partial LeastSquares Discriminant Analysis (PLSDA) models were developed to discriminate samples according to several operative conditions (level of conditioning, field type, time after conditioning, and time of day) and yielded good discrimination power (mean sensitivity = 87%, mean probability = 89%). Moisture content was estimated by Partial LeastSquares (PLS) models obtaining determination coefficient (R 2 ) = 0.86 and Root Mean Square Error (RMSE) = 2.00% (external validation). The number of spectral bands was reduced by using the Variable Importance in Projection (VIP) method, passing from 272 to 74 bands. Further PLS models were built considering the reduced variable numbers and achieved comparable results to those obtained with the full spectra, demonstrating that reduction of the number of variables did not affect the goodness of the models. Finally, the best PLS model was applied to each pixel of the hyperspectral images to obtain false colour images. Pixels having similar colours were characterized by comparable moisture content. The models developed may be useful to determine moisture content of alfalfa remotely and in real-time during mowing or harvest procedures. • Quality of alfalfa after cutting was evaluated. • In-field Vis/NIR hyperspectral imaging was used. • Regression models to predict moisture content were developed. • The samples were discriminated according to several operative parameters. • The models could be used remotely or in real-time during mowing or harvest procedures.