Discrimination of weedy rice by using near-infrared spectroscopy combined with chemometrics

Abstract

Weedy rice is one of the most notorious weeds occurring in rice-growing areas, especially in South-East Asia. Weedy rice especially in form of paddy seed is difficult to manage and separate as they provide common features (morphological resemblance) to cultivated rice. This work presents a modification of self-organizing map (SOMs) for the classification of weedy rice from cultivated rice via in situ direct sample analysis from paddy seed using near-infrared (NIR) spectroscopy and hyperspectral NIR camera. The sample pretreatment was carried out by a cyclone vacuum machine to remove the contaminated particles and other impurities. The physical characteristics and the thermal behavior of rice samples were investigated by optical microscope and thermogravimetric analysis (TGA), respectively, and the volatile chemical profiles were monitored by using DART-MS. They provide the distinctive patterns between cultivated rice and weed rice. A near-infrared with reflectance accessory was used for direct sample analysis. The acquired NIR spectra were smoothed using Savitzky-Golay polynomial, baseline-aligned using standard normal variate (SNV), mean-centered and the second derivative was calculated to reveal the significant NIR regions. Self-organizing maps was�well-optimized and was�applied for the classification of weedy samples from four cultivated rice. The results were validated and were achieved very high predictive value in the range of 91% to 99% and 88% to 99% for precision and accuracy, respectively. Furthermore, the developed supervised SOMs was applied on the pair-wise hyperspectral image to generate the supervised global SOM map with different color scales as the representative of each sample class. Each hyperspectral pixel from the sample image was validated with the global map, then, the color of best map unit (BMU) was re-projected on the image pixel. The process was undergone until all image pixels was projected with the color of BMU. The classification was achieved by the ratio of the projected color on the sample image. The classification accuracy for weedy seeds was 90%, demonstrating the potential of a global model for seed quality assessment.