Automated spectral feature extraction from hyperspectral images to differentiate weedy rice and barnyard grass from a rice crop.

Highlights • Line-scanning hyperspectral imaging for weedy rice and rice recognition. • Spectral features were automated extracted. • The performances of random forests and support vector machines were compared. • Important spectral features were selected for weed and rice recognition. • Recognition rates ranged from 0.92 to 1.0. Abstract Barnyard grass (Echinochloa crusgalli) and weedy rice (Oryza sativa f. spontanea) are two common and troublesome weed species in rice (Oryza sativa L.) crop. They cause significant yield loss in rice production while it is difficult to differentiate them for site-specific weed management. In this paper, we aimed to develop a classification model with important spectral features to recognize these two weeds and rice based on hyperspectral imaging techniques. There were 287 plant leaf samples in total which were scanned by the hyperspectral imaging systems within the spectral range from 380 nm to 1080 nm. After obtaining hyperspectral images, we first developed an algorithmic pipeline to automatically extract spectral features from line scan hyperspectral images. Then the raw spectral features were subjected to wavelet transformation for noise reduction. Random forests and support vector machine models were developed with the optimal hyperparameters to compare their performances in the test set. Moreover, feature selection was explored through successive projection algorithm (SPA). It is shown that the weighted support vector machine with 6 spectral features selected by SPA can achieve 100%, 100%, and 92% recognition rates for barnyard grass, weedy rice and rice, respectively. Furthermore, the selected 6 wavelengths (415 nm, 561 nm, 687 nm, 705 nm, 735 nm, 1007 nm) have the potential to design a customized optical sensor for these two weeds and rice discrimination in practice. [ABSTRACT FROM AUTHOR]