Plant Weed Detection Using Deep Learning

Abstract

Weeds are the plants that grow along with the primary plant in agricultural crops. These undesirable plants compete with the main crop for core elements like water, sunlight, and sometimes also for fertilizers. This causes losses to the crop quality as well as to the crop yield. The conventional solution to this weed menace is hand weeding but this process being labor-intensive, costly time-consuming, farmers have moved towards the use of herbicides. The latter method is effective but causes environmental as well as health concerns for humans who consume these vegetable crops. Hence, Precision Agriculture suggests the variable spraying of herbicides so that the primary plants are not affected by herbicide chemicals. So, site-specific weed management has been introduced for weed control by using Artificial Intelligence. In this project, the Eggplant (Brinjal) vegetable crop has been taken into consideration for weed detection through semantic segmentation of the plant and non-plant (weed) parts from images. The dataset collection for the project was done manually by taking images from a private farm in Gandhinagar, Gujarat. The images also required ground truth for the learning purposes which were generated using external software tools. Deep learning models such as UNet & LinkNet with different backbone models were utilized for the segmentation purpose.LinkNet with backbone Mo bilenetv2 and Resnet34 were used and UNet with backbone Inceptionv3 and Resnet18 were used. The best results are achieved using UNet with backbone Resnet18 for a mean IoU score of 0.89. With the help of this segmentation, the precise location of weeds from images can be hence achieved.