Identification of Snake Species in Sri Lanka Using Convolutional Neural Networks

Abstract

Snake bites in Sri Lanka cause death to nearly 100 people annually. Administering the appropriate anti-venom treatment for snake bite victims gets delayed causing complications as a result of the inability of people to identify the snake. Incorrect identification of snakes also causes threats to the existence of harmless snakes threatening the biodiversity of Sri Lanka. As a remedial measure to these problems, the first automatic snake identification from a given image using convolutional neural networks (CNN) is described in this study using 2000 images from each of six snake species found in Sri Lanka to train five CNN models. Four of the models were developed using the pre-trained architectures InceptionV3, VGG16, ResNet50 and MobileNet using transfer learning while the fifth model was developed from scratch. This study revealed that MobileNet with transfer learning yielding an accuracy of 90.5% is the most suitable model for automatic snake identification.