Thesis - Open Access
Master of Science (MS)
Electrical Engineering and Computer Science
Sung Y. Shin
Automatic Greenness Identification, Histogram of Oriented Gradiant, K-Means Clustering, Precision Agriculture, Support Vector Machine, Weed Identification
This paper proposes a sophisticated classification process to segment the leaves of carrots from weeds (mostly Chamomile). In the early stages, of the plants’ development, both weeds and carrot leaves are intermixed with each other and have similar color texture. This makes it difficult to identify without the help of the domain experts. Therefore, it is essential to remove the weed regions so that the carrot plants can grow without any interruptions. The process of identifying the weeds become more challenging when both plant and weed regions overlap (inter-leaves). The proposed system addresses this problem by creating a sophisticated means for weed identification. The major components of this system are composed of three processes: Image Segmentation, Feature Extraction, and Decision-Making. In the Image Segmentation process, the input images are processed into lower units where the relevant features are extracted. In the second proposed method, K-Means clustering is applied to extract the images that will be used for the identification process. The images are then normalized into a binary image using Otsu’s Thresholding. Next, in the Feature Extraction stage, relevant information of the weed and leaves are extracted from the lower unit images. Furthermore, to extract the information from the Region of Interest (ROI), Histogram of Oriented Gradient (HoG) is used to locate and label all the weed and carrot leaves regions. In the Decision-Making process, the system makes use of Support Vector Machine (SVM), which is a supervised learning algorithm, is used to analyze and segregate the weeds from the plants. Afterward, the findings are used to dictate which plants receive herbicides and which do not. The main priority for the Image Segmentation process is on overlapping images where weeds need to be isolated from plants; otherwise, in the later stages, those plants cannot be used for cultivation purposes. These methods of weed detection are effective as it automates the identification process and fewer herbicides will be used, which in turn is beneficial to the environment. The evaluation of the approach was done using an open dataset of images consisting of carrot plants. The system was able to achieve 88.99% accuracy for weed classification using this dataset. Further improvement of the proposed method successfully classifies the plant regions at a success rate of 92%. These methodologies will help reduce the use of herbicides while improving the performance and costs of Precision Agriculture.
Number of Pages
South Dakota State University
In Copyright - Educational Use Permitted
Saha, Dheeman, "Development of Enhanced Weed Detection System with Adaptive Thresholding, K-Means and Support Vector Machine" (2019). Electronic Theses and Dissertations. 3374.