Mapping hyperspectral NIR images using supervised self-organizing maps: Discrimination of weedy rice seeds

Abstract

Weedy rice is one of the most problematic weeds in rice-growing regions, particularly in Southeast Asia. Unlike other types of weeds, it is extremely difficult to distinguish weedy rice from cultivated rice directly from paddy seeds as they exhibit common morphological features. As such weed management can be a difficult, time-consuming, and inaccurate process that is often carried out manually. This study offers a novel classification approach based on an artificial neural network (ANN), utilizing self-organizing maps (SOMs), to directly discriminate weedy rice using the near-infrared (NIR) hyperspectral imaging (HSI) technique. The physical attributes, thermal behavior, and chemical profiles of the weedy and cultivated rice were thoroughly investigated by a range of analytical techniques, including optical microscopy, scanning electron microscopy, thermogravimetric analysis, and direct analysis in real-time mass spectrometry (DART-MS). For direct sample analysis by HSI, a global self-organizing map was generated with optimized parameters (scaling value and map size). The color indices (Red, Green, Blue values) of the sample image were defined to obtain a color ratio that can be used to classify unknown samples. The optimal threshold for classification was carefully determined using a receiver operating characteristic (ROC) curve. Performance metrics (sensitivity, specificity, precision, accuracy, and misclassification error) were used to evaluate the performance of the model. Classification accuracies of 98% (Weedy vs PL2) and 88% (Weedy vs RD49) were obtained with balanced sensitivity and specificity. The classification was assessed from the whole sample image, which was completely independent of the selected region of interest. As far as we know, this is the first instance where SOMs have been utilized to appraise seed quality by means of authentic HSI images.