Machine learning algorithms require large amounts of data to perform tasks with the same performance as humans. The typical size of training data for image recognition applications lies in the order of hundreds of thousands of images. While pure image data is relatively easy to acquire, training data must be labeled, meaning that each image must have an associated label. Those labels classify the objects in the image. For plant applications generating such data is difficult, since labeling an image correctly requires expert knowledge.
This presentation discusses our progress toward generating large numbers of labeled image with a robotic system. Over the last months we increased the system's throughput and are now able to generate tens of thousands of images per day. Since we have full control over which objects to present to the system, we can generate exactly the kind of dataset needed for an application. For example, we can create datasets including crops and their associated weeds, as to distinguish them from each other. We are already training Convolutional Neural Networks (CNN) with our own data. Those CNNs are going to be deployed on autonomous platforms to identify weeds and evaluate crops' well-being.
The project's goal is to facilitate the application of machine learning in agriculture and plant sciences. Thus, our datasets and algorithms will be open for use by academics and industry. Furthermore, we aim to develop a plug-and-play system, that can be acquired and used by researchers in their own labs and greenhouses, kickstarting an even faster process of generating labeled data.
The University of Winnipeg's digital agriculture project is strongly connected to local industrial partners and Manitoba's growers associations. Dozens of researchers are involved from the UofW and partner institutions. In 2019 the project received additional funding of $2.4 million from the government of Canada.