Location: Crop Production Systems Research Unit
Title: Discrimination of crop and weeds on visible and visible/near-infrared spectrums using support vector machines, artificial neural network and decision tree Authors
|Deng, Wei -|
|Zhao, Chunjiang -|
|Wang, Xiu -|
Submitted to: Sensors and Transducers
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: March 18, 2014
Publication Date: April 1, 2014
Citation: Deng, W., Huang, Y., Zhao, C., Wang, X. 2014. Discrimination of crop and weeds on visible and visible/near-infrared spectrums using support vector machines, artificial neural network and decision tree. Sensors and Transducers. 26:26-34. Interpretive Summary: To discriminate weeds from crop is important for farming practice to achieve effective weed management. Methods are available to characterize the weeds in crop field but a rapid, non-contact method is needed for high-efficient management. The scientists at National Engineering Research Center for Information Technology in Agriculture (NERCITA), China and USDA-ARS Crop production Systems Research Unit, Stoneville, MS, have collaboratively conducted a study to develop a spectral method to evaluate and compare the visible and visible/near-infrared spectrums in discrimination of weeds from crop. The results indicated that the visible spectrum can be used in the discrimination better than the visible/near-infrared spectrum. This suggests that the less expensive visible sensor could be used in practice rather than the more expensive visible/near-infrared sensor, which is useful for practical weed management.
Technical Abstract: Weeds are regarded as farmers' natural enemy. In order to avoid excessive pesticide residues, the destruction of ecological environment, and to guarantee the quality and safety of agricultural products, it is urgent to develop highly-efficient weed management methods. Amongst, weed discrimination is the key part. There have been a lot of researches on weed detection/discrimination using spectral measurement on plant leaf/canopy. However, as reported so far the spectral ranges from the researches were not consistent and no research was reported to determine more efficient wavelength range for weed classification. Some researchers adopted visible spectrum, some adopted near-infrared spectrum, the others adopted both visible and near-infrared spectrum. The purpose of this study was to compare the classifications of the spectral reflectance in range of 350 ~ 760 nm and in 350 ~ 2500 nm for crop/weed discrimination. Through spectral analysis of these data respectively using three kinds of modeling methods of Support Vector Machines (SVMs), Artificial Neural Network (ANN), and Decision Tree (DT), the results showed that the three classifiers could differentiate crop and weeds better in 350 ~ 760 nm wavelength range than in 350 ~ 2500 nm. Therefore, the visible wavelength range could be good enough to meet the requirement for crop/weed spectral discrimination, which might reduce the cost of weed detect sensors.