Weed Recognition in Agriculture: A Mask R-CNN Approach

Abstract

Recent interdisciplinary collaboration on deep learning has led to a growing interest in its application in the agriculture domain. Weed control and management are some of the crucial tasks in agriculture to maintain high crop productivity. The inception phase of weed control and management is to successfully recognize the weed plants, followed by providing a suitable management plan. Due to the complexities in agriculture images, such as similar colour and texture, we need to incorporate a deep neural network that uses pixel-wise grouping for identifying the plant species. In this thesis, we analysed the performance of one of the most popular deep neural networks aimed to solve the instance segmentation (pixel-wise analysis) problems: Mask R-CNN, for weed plant recognition (detection and classification) using field images and aerial images. We have used Mask R-CNN to recognize the crop plants and weed plants using the Crop/Weed Field Image Dataset (CWFID) for the field image study. However, the CWFID's limitations are that it identifies all weed plants as a single class and all of the crop plants are from a single organic carrot field. We have created a synthetic dataset with 80 weed plant species to tackle this problem and tested it with Mask R-CNN to expand our study. Throughout this thesis, we predominantly focused on detecting one specific invasive weed type called Persicaria Perfoliata or Mile-A-Minute (MAM) for our aerial image study. In general, supervised model outcomes are slow to aerial images, primarily due to large image size and scarcity of well-annotated datasets, making it relatively harder to recognize the species from higher altitudes. We propose a three-level (leaves, trees, forest) hierarchy to recognize the species using Unmanned Aerial Vehicles(UAVs) to address this issue. To create a dataset that resembles weed clusters similar to MAM, we have used a localized style transfer technique to transfer the style from the available MAM images to a portion of the aerial images' content using VGG-19 architecture. We have also generated another dataset at a relatively low altitude and tested it with Mask R-CNN and reached ~92% AP50 using these low-altitude resized images.