We worked with a team of researchers from the Miller Lab in order to determine whether machine learning could be used to classify grape leaves as originating from one of four given rootstocks based on data composed of images of these leaves. Our goal is to create a program that is able to gather the data needed to make these predictions from images of leaves and predict which of the four rootstocks the leaves were grafted onto.

The first approach we used was the implementation of a deep convolutional neural network to train on cropped images of each leaf and predict which of the four rootstocks each leaf had been grafted onto. It was our hope that deep learning would be an immediate solution, but quickly our experimentation proved otherwise. We tried using our algorithm to compare all four rootstocks at once, then one rootstock versus the others, but neither of these approaches yielded a result higher than random change accuracy.

We then decided to try two other approaches that are not machine learning based: General Procrustes Analysis and Persistent Homology Analysis. Both of these methods used plotted points that we gathered by automatically extracting the outline of each of the leaves and attempted to find a signal from these points that would differentiate a leaf from one rootstock from a leaf of another. These results were inconclusive and did not yield the viable results we had hoped for.

Although this project yielded negative results, there is important information to be gained from the research and experimentation. The research shows that we could not find any visual signals that come from the leaf images provided. In future iterations of this project, it might be helpful to look at other physical attributes to make a meaningful prediction. This project provided valuable insight on the limitations of machine learning techniques and gave us the opportunity to acknowledge that not all questions can be answered with machine learning.

Examples of our data from each of the four rootstocks can be found at: https://cs.slu.edu/~astylianou/grapeleaves/GrapeLeavesDisplay.html

Dr. Stylianou has been an extremely valuable resource for this project. She has consistently given the group her time and guidance in order for us to achieve our goal. One of the main issues we ran into was that the machine learning approach did not yield positive results. At this point, she was able to bring us new ideas and put us in contact with other faculty members who were able to offer new insights and potential solutions. Dr. Stylianou went above and beyond to help with major and minor issues we ran into throughout the project. We are very grateful to have had such an encouraging and knowledgeable faculty sponsor.