Revealing the Unknown: Real-Time Recognition of Galápagos Snake Species Using Deep Learning

SIMPLE SUMMARY: The snakes in Galápagos are the least studied group of vertebrates in the archipelago. The conservation status of only four out of nine recognized species has been formally evaluated, and preliminary evidence suggests that some of the species may be entirely extinct on some islands. Moreover, nearly all park ranger reports and citizen/science photographic identifications of Galápagos snakes are spurious, given that the systematics of the snakes in the archipelago have just recently been clarified. Our solution is to provide park rangers and tourists with easily accessible applications for species identification in real time through automatic object recognition. We used deep learning algorithms on collected images of the snake species to develop the artificial intelligence platform, an application software, that is able to recognize a species of a snake using a user’s uploaded image. The application software works in the following way: once a user uploads an image of a snake into the application, the algorithm processes it, classifies it into one of the nine snake species, gives the class of the predicted species, as well as educates users by providing them with information about the distribution, natural history, conservation, and etymology of the snake. ABSTRACT: Real-time identification of wildlife is an upcoming and promising tool for the preservation of wildlife. In this research project, we aimed to use object detection and image classification for the racer snakes of the Galápagos Islands, Ecuador. The final target of this project was to build an artificial intelligence (AI) platform, in terms of a web or mobile application, which would serve as a real-time decision making and supporting mechanism for the visitors and park rangers of the Galápagos Islands, to correctly identify a snake species from the user’s uploaded image. Using the deep learning and machine learning algorithms and libraries, we modified and successfully implemented four region-based convolutional neural network (R-CNN) architectures (models for image classification): Inception V2, ResNet, MobileNet, and VGG16. Inception V2, ResNet and VGG16 reached an overall accuracy of 75%.