Weed Detection and Control Using Mask-R-CNN

Recent transdisciplinary deep learning research has increased interest in using deep learning in the agricultural industry. Weed management and control are some of the most crucial tasks in agriculture to maintain optimum crop productivity. Correctly identifying the undesirable plants must come before presenting a practical management plan in order to manage and control weeds. Agriculture images are complex and contain elements like similar color and texture, therefore we need to apply a deep neural network that makes use of pixel-wise grouping to identify the plant type. In this thesis, using field pictures and aerial images, we assessed the performance of Mask R-CNN, one of the most used deep neural networks, for weed plant recognition (detection and classification). Mask R-CNN was created to address issues with instance segmentation (pixel-wise analysis). We employed the Crop/Weed Field Image Dataset (CWFID) and Mask R-CNN to distinguish between crop plants and weed plants during the field image analysis. The CWFID’s limitations, however, are that it groups all weed plants into a single class and requires that all crop plants come from a single organic carrot field. To solve this problem, we created a synthetic dataset of 80 different species of weed plants and assessed it using Mask R-CNN.