The economic importance of the wine industry in Australia and worldwide is driving the development and application of innovative methods and technologies for monitoring vineyards. The harvest can significantly vary from year to year and also within the vineyard due to soil conditions, diseases, pests, climate and variation in vineyard management practices. Current practices for yield and quality estimation are destructive, expensive, inaccurate, and largely subjective. These factors make the production of high quality grapes for wine making challenging. Image processing and analysis together with machine learning have the potential to provide an inexpensive, non-destructive way of capturing precise information about the entire vineyard.
The use of image processing to enhance vineyard practices and forecast the success and/ or quality of wine grapes is in its infancy. While much of the work to date is promising researchers and practitioners have not yet achieved the "vineyard of the future", where image analysis (image and video) is a powerful tool that is adopted by viticulturists to inform the management of their vineyards. This research project will develop new research methodologies using image processing and machine learning to detect and recognize grape quality, pests and diseases for improved precision viticulture. New pre-processing and feature extraction techniques will be investigated together with supervised and unsupervised neural networks for machine learning.
This is a collaborative research project with researchers from CM3 and NWGIC. The project fits within the CM3 research direction of machine learning & machine vision under FoR code 080106 "Image Processing & Machine Learning" and will form one of the project initiatives within the new research group being formed within CM3 of "Multimedia Analytics & Sensing for Agriculture & Environment" proposed by the sub-dean of research, Prof. Mark Morrison. The focus of the project would be to generate significant journal/conference publications, strengthen the collaborative linkages between CM3 and NWGIC, serve as pilot project(s) to establish new linkages with new researchers from NWGIC and Faculty of Science, increase research collaborative activities amongst staff on different CSU campuses, and target future research grant from wine industries or Wine Australia (WA). The project is divided into three sub-projects with a time frame of six months:
The main expenditures for this project is to employ a programmer/research associate (RA) for six months to assist the researchers and to accelerate the research work. This project is strongly supported by NWGIC which has agreed to contribute a $10,000 grant to purchase equipment. This proposal is to request CM3 to fund the personnel costs to employ the programmer/RA for six months. Dr. Lu has the necessary background in machine learning/vision having completed his PhD in this area. To accelerate the research work and meet the schedule, the CM3 researchers will focus on performing the algorithms and techniques designs, leaving the programmer/RA to focus on the development and programming aspects. It is expected that all developmental work will use the Matlab platform which is familiar to all researchers and to Dr. Lu.
This sub-project builds on the current project between CM3 and NWGIC on "Developing a phone-based imaging tool to inform on fruit volume and potential optimal harvest time" (funded by Wine Australia). The current project focuses on reducing memory requirements and computational/processing power to be feasible for real time implementation which is limited by the processing and sensor capabilities of the mobile phone. This new sub-project will focus on developing new techniques which give good performance and is not constrained or limited by processing power or hardware constraints. The sub-project will serve two objectives: (1) it would serve as the new versions and future-proofing for the mobile imaging app with Wine Australia as more powerful mobile phones get on the market; and (2) it gives a novel application for developing new machine learning/vision for spherical detection and colour segmentation. For (2), the sub-project will develop new imaging pipelines using Bayesian, neural networks and intelligent nature-inspired techniques (e.g. swarm intelligence techniques).
Rapid, accurate identification of diseases in the vineyard is key to preventing serious outbreaks and losses in yield and quality. This sub-project will develop novel methods that allow for early detection of grapevine downy/powdery mildew and other diseases by the use of non-destructive techniques based on imaging and machine learning/vision. The later part of this work will investigate the use of new hyperspectral techniques. This technology would be suitable to be deployed on field as a portable system allowing the detection of the disease even when not detectable by human eye and, in the future, as part of the farming machinery for automated detection and spraying.
While some work has been undertaken towards the detection of powdery and downy mildew and botrytis in grapes there is still significant scope for improvement. Detection in the field is difficult as the grapes may be covered by bloom and disease can exhibit different signs and symptoms depending on the grape variety and the stage of development of the disease. Moreover, more than one disease might be present. Current work does not address the issues of more than one disease or identification of the disease at various stages in its development. Furthermore, the most successful work to date has not been on in-field images. This work will focus the research on early stage detection and analysis (before Inflorescences and the berry stage). New pre-processing and feature extraction techniques will be investigated together with supervised and unsupervised neural networks for machine learning.
Downy Mildew Powdery mildew
Accurate identification of pests is critical to an effective management program that provides optimal control while minimizing pesticide use. Timely inspection and good cultural practices will aid in reducing the need for insecticide applications. This sub-project will develop novel image processing/machine learning methods that allow for detection of grape flower and fruits pests such as grape flea beetle, grape berry moth, multicolored Asian lady beetle, yellowjackets. There has been little research work which has been undertaken towards the detection and recognition for pests for grapes in viticulture. Some examples of pests are shown below. New techniques will be investigated together with supervised and unsupervised neural networks for machine learning.