A support vector machine embedded weed identification system

machine vision;Pixelwise method;SVM;Weed identification

Lin, Chufan ; Grift, Tony E. ; Grift ; Tony E.

关键词: machine vision;    Pixelwise method;    SVM;    Weed identification;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/14615/Lin_Chufan.pdf?sequence=2&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Over the past decades, the over-reliance on herbicides during corn production has causedsevere environmental and biological problems such as pollution in the soil and undergroundwater, and the emergence of the herbicide-resistant weed species. A potential solution to reducethe use of herbicides while maintaining adequate weed control lies in the combined use ofchemical and mechanical weeding, in which weeds are controlled adaptively according to theirreaction to herbicides. Accurate weed identification is a prerequisite for accomplishing such acontrol strategy.A machine vision system for weed identification, which utilized the morphologicalproperties of weed leaves, was developed in this research. The system incorporated a newimage segmentation algorithm, termed the ‘Pixelwise method’ to binarize the color weedimages for subsequent image processing and feature extraction procedures. Subsequently, aSupport Vector Machine (SVM) based classifier was constructed to distinguish various weedspecies using seven morphological features.2,325 indoor images consisting of six weed species were acquired during the first fiveweeks after emergence of the plants. Among 1,006 test images, the SVM system achieved over94% accuracy in crop (corn) versus weed discrimination and 95% in grass versus broadleafweed discrimination. The average classification accuracy for individual weed species wasapproximately 86%. In addition, the system obtained the best classification result after thesecond week after plant emergence. In field tests, the SVM classifier based on the indoor imagelibrary was able to identify 71.1% of 270 weed plants in the field. With an adaptive medianfilter to enhance the image quality, the accuracy was raised to 75.9% at the expense of extraimage processing time.Both of the laboratory and field tests showed that the SVM method with reasonableaccuracy is feasible for weed identification during their early growth season.