Rapeseed seeds quality classification with usage of VIS-NIR fiber optic probe and artificial neural networks

• Published in: 2016 International Conference on Optoelectronics and Image Processing (ICOIP) pp. 44 - 48
• Main Authors: Wojciechowski, Tomasz, Niedbala, Gniewko, Czechlowski, Miroslaw, Nawrocka, Janina Rudowicz, Piechnik, Leszek, Niemann, Janetta
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2016
• Subjects: artificial neural networks; classification; Fats; Neural networks; Optical fiber networks; Optical fiber sensors; Proteins; rapeseed; Testing; VIS-NIR spectroscopy
• DOI: 10.1109/OPTIP.2016.7528517

The development of non-destructive methods like VIS-NIR reflection spectroscopy and artificial neural networks to analyse the rape seeds content of fat and protein was the subject of this work. The research material contained the seeds of 46 winter rapeseed lines obtained from interspecies crossing male sterile lines of MS-8 and 6 control forms. The seeds were pre-cleaned and crude fat and crude protein content were determined using a laboratory spectrometer NIRS 6500 (FOSS) with factory calibration. Acquisition of VIS-NIR spectra were carried out on five gram seeds samples. Spectra in the range of 400-2200 nm with 10 nm and 16 nm resolution respectively, and interpolated to 2 nm resolution were obtained using Agrospec spectrometer (Tec5, Germany) equipped with the RP-7 fiber optic measuring head. Obtained spectra, after pretreatment, were analysed by artificial neural networks in the Statistica software. 901 independent features network input were presented in total. At the output of the network one independent features were presented indicating the level of seeds crude fat and crude protein as LOW (up to 42% of crude fat and 19% crude protein level), and HIGH (above 42% of crude fat and above 19% crude protein level). Two neural networks were prepared with RBF topology and structure of 901: 901-4-1: 1 (for crude fat) and 901: 901-22-1: 1 (for crude protein). Data set separation was done based on the structure of 139 training set cases and 60 cases in the validation and testing set. In total 259 cases were analyzed. Analysis of classification accuracy for fat reached 76.19% level of correct classification for HIGH and 68.90% for LOW. Analysis of classification accuracy for crude protein was at 74.72% of correct classification level for HIGH and 69.04% correct for LOW. The proposed method of rapeseed seeds classification is appropriate for this type of issues. The further development of this method must consider the automatic analysis during samples scanning process.