Some Theory and Observations Concerning the Combined Use of Computational Shape Analysis and Identification Keys to Identify Plants From Digital Images.

Computers provide the means to automate plant specimen characterization from digital images. However, multiple issues require attention before full achievement of this goal: parsing occluded and overlapping features in images, handling missing or inaccurate information from damaged or partial specimens, improving accuracy of identification, and fusing traditional and computational methods of identification. Here, I examined species identification accuracy with varying sizes of the leaf shape database. I focused on combining traditional key-based approaches to species identification with computational pattern recognition approaches. The leaf database consisted of 2,420 leaves from 151 species. Most of these species were from lowland Neotropical rainforests (notorious for similar leaf shapes), and the remaining 51 species were horticultural plants for xeric landscapes. Using shape analysis based on Fourier and Procrustes distance metrics, species identification accuracy was 72% when all 151 species were considered. I then applied the asymptotic equipartition property from information theory to predict identification accuracy when continuous shape metrics and discrete characters were used together. Empirical accuracies match theory precisely. An advantage to this approach is that biological information concerning shape and size characteristics is maintained and stored in a database. This information can be used by other investigators with interests in development, ecology, and evolution of plants. My long-term goal is to automate the population of databases with continuous and discrete plant characteristics (leaf sizes, shapes, margins, apices, bases, arrangement, lobation) based on the analysis of herbarium specimen images.