Application of Discriminant Analysis to Classify Pangola Hay with Digital Images and NIR Spectra

Abstract. Pangola grass is a good source of forage and can be fed by fresh, hay or silage format. In proper environments, pangola grass can support better ruminant performance in terms of milk yield and composition, body weight gain, feed conversion ratio and meat quality. To ensure its nutrients, quick and quantization methods used in field are strongly demanded from the forage industry. Currently three main approaches have been conducted for determination on forage quality and they are organoleptic observations, chemical composition analysis, and feed trial evaluations. Visual evaluation determines the quality on physical parameters such as maturity, colour, leaf to stem ratio, and foreign objects. Chemical analysis mainly reveals the NDF (Neutral detergent fiber), ADF (acid detergent fiber), and CP (crude protein). In this study, we applied computer vision and NIR spectroscopy on pangola hay so as to classify them based on their quality. When 20 color and texture features were used in the LDA (Linear discriminant analysis) model, about 97% classification accuracy both on training and testing process was achieved. The 20 features were suggested by SFS (Sequential forward selection) algorithm among the 79 feature candidates. When 10 scores of spectral features were used in LDA model, the classification accuracy was about 75%. When these two sensors information were joined by Dempster-Shafer theory, the classification accuracy could be improved.