Time Series based Gastropod Classification

This paper presents Series-K, an automatic Gastropod classification system based on Time Series and k-Nearest Neighbor. Species evolve over time. Biologists have discovered, attempted to describe and put them into categories. Our proposed method automatically processes gastropods' images to help the malacologists accurately identify gastropods into their natural families. The method would further be applied to identify whether certain species have never been discovered before. We first transform gastropod images into time series. Specifically, we detect the edge of each gastropod shape, and measure the distances either from 1) the shape's centroid or 2) its center point to the edge. The distances are in the form of time series data, which serve as the image signatures. We then leverage the dynamic time warping (DTW, [1]) method to determine the similarities/distances between pairs of the time series data. The k-nearest neighbor (k-NN) method is used to classify each gastropod into its family. We conduct experiments to determine which k value yields better performance, and to compare the accuracies of two scenarios: when centroid-based distances and when center point-based distances are used. The results from the experiments show that k-NN with k = 3 performs best (with the error of less than 8%) when time series data consist of the distances from the centroid.