Transfer learning in Plant Phenotyping

08 March 2017

Plant Phenotyping, Innovation

Image analysis and classification techniques are advancing rapidly with developments such as deep learning, neural networks and artificial intelligence. These developments are of upmost importance for plant phenotyping, as they will play an increasingly prominent role in the future.

Several studies have been concluded where researchers apply deep learning and transfer learning to a dataset of plant images in order to identify species, as well as diseases. These approaches operate on the concept of feeding data to a neural network, which is turn learns from the data and will be able to draw conclusions based on its ‘acquired knowledge’.

Transfer learning is particularly interesting, as an already trained model is applied to the new dataset, and can evaluate the dataset based on its ‘previous experience’. Think of it like this: if we can teach a robot to catch a baseball, it can teach itself to how catch a basketball. In case of plant phenotyping: using an existing Convolutional Neural Network through transfer learning and feeding it a large amount of images from a public dataset, a study concluded it can identify 14 different crops and 26 diseases at 99.35% accuracy. In another study, the Inception-v3 neural network developed by Google was 98% accurate in identifying Arabidopsis from tobacco plants after only fine-tuning the model by remodeling the final layers.

Transfer learning is a development that has the potential of being widely used in plant phenotyping. Applications could be anywhere from plant development analysis, evaluating quality/ yield of newly developed strings to creating a smartphone app that is able to diagnose plant disease anywhere in the world.

Mohanty, S., Hughes, D., & Salathé, a. M. (2016). UsingDeepLearningforImage-BasedPlantDiseaseDetection. Frontiers in Plant Science, 2-10.
Tapas, A. (2016). Transfer Learning for Image Classification and Plant Phenotyping. International Journal of Advanced Research in Computer Engineering & Technology, 2665-2669.