17BEC018
Weed detection is an important aspect in enhanced crop production. These systems are a crucial part for finding solutions to the present problem. In order to remove weed different pesticides are used. But they can also be sprayed accidentally on the crop. Hence, using weed detection lesser
harm to the crops are ensured. Research in the field of wheat crops, pulses, and potato and tomato plants have been done. But research for weed detection in Brinjal crops is not yet done as per the literature survey.
Accurate agriculture selective spraying of pesticides on weed densities in order to lessen the negative impact on the crops. In order to do so, in this thesis, Machine Learning based solution is proposed. Machine learning ensures precision in detecting weed and thus this has also been proved
with the efficiency parameters found in this thesis.
There are two approaches implemented in this thesis.
1. Classification Approach
2. Segmentation approach
Classification approach states whether the image has ‘weed’ in it whereas segmentation approach shows where the ‘weed’ is present in the image. Classification and segmentation approach states the position of the weed in the image in this thesis.
For this process, dataset has been procured manually in the fields of Gandhinagar, Gujarat, India. There are three types of Brinjal crops which are used in the entire system. For the pre-processing of the dataset, data augmentation is done. Augmentation of the data is the process for elevating the variety corresponding to the dataset. The images are blurred, rotated and horizontally flipped. Background from the image is segmented in order to obtain the features of weed and crop. This is done using edge detection technique.
After the pre-processing of the dataset, it is fed to two approaches. Feature Extraction Approach (FEA) and Feature Segmentation Approach (FSA). In FEA, K-Nearest Neighbor (KNN) + Gray Level Co-occurrence Matrix (GLCM) is implemented. And for FSA, Convolutional Neural Network (CNN) + SegNet is implemented. Their efficiency parameters are obtained and compared with the existing literature.
The high efficiency and statistical parameters are obtained in this thesis. KNN+GLCM is used as a classification approach and CNN+SegNet is used as a segmentation approach. The accuracy and IoU score obtained is 97.86 and 97.69 respectively. Thus, a new outlook for having elevated
production in agriculture sector has been proposed.