Automatic classification of seed species by neural networks and optic RGB images

Anonymous

TL;DR: RGB image-based seed phenotyping (species recognition) with neural networks under a double-cross validation scheme.

Abstract: In this paper we propose a neural network approach aiming at automatically detecting native seeds species from the paramo ecosystem based on optic RGB images. The network architectures which are explored for this purpose consist in shallow feed-forward networks with one hidden layer holding up to 15 neurons, and deep convolutional neural networks (CNNs). Focus is placed on four species which are commonly found under natural conditions in the paramo ecosystem, namely Espeletia congestiflora, Bucquetia glutinosa, Calamagrostis effusa and Puya santosii. Images of the individual seeds were taken on both a white and a black-soil background, the latter simulating the natural conditions where the seeds can be found. We show that relevant knowledge for classification of the seeds' species can be extracted only from their optical information. Under a double cross-validation scheme, a 14-neuron shallow network achieves an 88% test accuracy, while a CNN achieves 94%. On the other hand, after augmenting the available image-data, a CNN is built with a 100% accuracy on validation and a small control-test set. Overall, this neural network approach suggests a promising methodology for seed species prediction based on optical RGB images.

Keywords: Neural networks, Robust learning, Image augmentation, Deep learning application, Seed phenotyping