The deep learning revolution has brought substantial advances in a great number of fields such as robotics, real-time language translation and image recognition.
Deep learning is part of a broader family of artificial intelligence and machine learning methods based on learning data representations. Deep learning applications can learn from data such as images, video or text and can increase their predictive accuracy when provided with more data.
Scientist Ignacio Heredia and colleagues from the Spanish National Research Council tested the potential of deep learning for plant classification. Since most of this work is done manually by the biodiversity research community, they argued that an accurate deep learning algorithm could be of tremendous help for researchers.
To put the idea in practice, they focused their research on finding ways to relieve human experts of the burden of identifying thousands of images by hand, and so they tested deep learning techniques to automise species identification of plant images. The team built a state-of-the-art neural network to classify images of plants. They compiled a wide-ranging dataset of 240,000 images of European plants from the PlantNet initiative and used the ResNet model neural network architecture to help them perform classification tasks.
All this work ran on Cloud Compute resources at the IFCA CSIC site (member of the EGI Federated Cloud), under the LifeWatch Virtual Organisation. The training of the network involved one week of computations using accelerated computing. The image datasets amounted to up to 140 GB of storage.
The results show that the model can indeed classify plants accurately. Heredia and his team concluded that while technology is not yet able to completely replace human experts, it does represent a good start in providing reliable labels for species identification.
Their plant classifier is available online for free.
I Heredia et al., 2017. Proceedings of ACM CF’17. doi.org/10.1145/3075564.3075590. (full text)