Classification of plant somatic embryos by using neural network classifiers

The development of somatic embryos is characterized by a series of morphological changes. Quantitative kinetic studies have been hampered by the difficulties in enumerating and characterizing embryo populations. By employing neural networks, we have developed a pattern-recognition system for characterizing the morphological features of carrot somatic embryos. This pattern-recognition system employs a hierarchical decision tree to achieve optimal classification. It successfully classified carrot somatic embryos into normal and abnormal embryo classes. For normal embryo classes (globular, oblong, heart, and torpedo embryos), an accuracy of 90% or higher was achieved. The features identified by the neural network classifiers as most important for embryo classification are almost identical to those obtained by the branch-and-bound searching algorithm used previously. However, employing the neural networks shortens the system developing time greatly. Coupled with an image analysis system, this neural-network-based pattern-recognition system shows great potential in embryo sorting and automation of synthetic seed production.