A neural network approach to predicting viability of native seeds from their optic RGB images

In this study, an application of neural networks with RGB optic images is explored for predicting the viability of individual seeds. The initial experiments use a small set of seeds, and implement different neural net architectures, from a simple feed-forward neural network to multi-layer and convolutional networks, aided by image augmentation techniques. Those techniques are necessary for implementing deep learning algorithms with small data. Here we try simple rotation of images with different zooming ranges, and the resulting models for the neural nets are tested for controlling that the complex models are not just over-fitting the data, but generalizing their properties well. As a result, a model for a deep convolutional net achieves the best results, with an accuracy of 90% in validation and 100% in an out-of-sample (small) control-test set. These results present a promising line for research on predicting seed viability based only on RGB information (visible attributes such as shape, color and size), being a relevant application for establishing precision crops with native seeds, or for monitoring and reforesting the paramo ecosystem at The Andes mountains in South America.