It is estimated that the quantity of plant species of our planet is about 400,000. In Costa Rica, there have been identified about 11,000 and it is expected to have a total number of 12,000. Plant identification is fundamental for biological richness studies of a region, inventories, plant population monitoring and endangered animals, climate change impact in the forests, modeling of invading species, among others.
Costa Rica not only has an extraordinary biodiversity (about 4% of the world biodiversity) but has also been distinguished for the innovative use of information technology dedicated to biodiversity conservation, and also for its national conservation parks, which is a role model worldwide.
Nevertheless, plant species identification is a process that normally requires expect knowledge and use of dichotomous keys, interactive keys, or just the experience of an expert. This makes the process tedious, inefficient and error prone.
We aim then to support the efficient, semi-automatic identification of plant species of Costa Rica, based on its plant images. We are working on the creation of new loss functions that take into account the taxonomy of plants as a hierarchy in multi-label deep learning architectures. This because the taxonomy has several levels such as species, genera, families, in such a way that we can classify not only at one level but several. We are working in new loss layers that make use of this prior knowledge to calculate the loss.