Optical Features for Automated Determination of Agricultural Product Varieties

This paper studies methods to determine varieties of agricultural specimens using features extracted from optical images generated by low-cost commodity hardware and simple, efficient algorithms. It presents a framework for this and some related tasks of agricultural informatics, with a focus on data-intensive aspects. It describes a system implementation that permits such data to be iteratively and interactively explored and studied while also permitting efficient programmatic access. The core classification problem of determining a raisin variety is studied experimentally and the quantitative results are competitive with prior work. Some of the methods generate simple, human-understandable classifiers, of which a few examples are presented. Data exploration and visualization is implemented using self-organizing maps (SOMs) and several examples of useful visualizations are described.