SummaryTaxonomic experts classify millions of specimens, but this is very time-consuming and therefore expensive. Image analysis is a way to automate identification and was previously done at Naturalis Biodiversity Center for slipper orchids (Cypripedioideae) by the program ‘OrchID’. This program operated by extracting a pre-defined number of features from images, and these features were used to train artificial neural networks (ANN) to classify out-of-sample images. This program was extended to work for a collection of Javanese butterflies, donated to Naturalis by the Van Groenendael-Krijger Foundation. Originally, for the orchids, an image was divided into a pre-defined number of horizontal and vertical bins and the mean blue-green-red values of each bin were calculated (BGR method) to obtain image features. In the extended implementation, characteristic image features were extracted using the SURF algorithm implemented in OpenCV and clustered with the BagOfWords method (SURF-BOW method). In addition, a combination of BGR- and SURF-BOW was implemented to extract both types of features in a single dataset (BGR-SURF method). A selection of the butterfly and orchid images was made to create datasets with at least 5 and at most 50 specimens per species. The SURF-BOW and BGR-SURF methods were applied to both selected datasets, and the original BGR method was applied to the selected butterfly dataset. PCA plots were made to inspect visually how well the applied methods discriminated among the species. For the butterflies, both genus and species appeared to cluster together in the plots of the SURF-BOW method. However, no obvious clustering was noticeable for the orchid plots. The performance of the ANNs was validated by a stratified k-fold cross validation. For the butterflies, the BGR-SURF method scored best with an accuracy of 77%, against 71% for the SURF-BOW method and 66% for the BGR method, all for chained genus and species prediction with k = 10. The new methods could not improve the accuracy of the orchid classification with k = 10, which was 75% on genus, 52% on genus and section and 48% on genus, section and species in the original framework and now less than 25% for all. The validation results also showed that at least about 15 specimens per species were necessary for a good prediction with the SURF-BOW method. The BGR-SURF method was found to be the best of these methods for butterflies, but the original BGR method was best for the slipper orchids. In the future these methods may be tested with other datasets, for example with mosquitoes. In addition, other classifiers may be tested for better performance, like support vector machines.
A selection of AKARI FIS BSC extragalactic objects
