SerpentSleuth: A Deep Ensemble Learning Approach for Venomous and Non-Venomous Snake Identification

Abstract

"Snakes are over 3900 species distributed worldwide except in Antarctica, Iceland, Ireland, New Zealand, and Greenland. According to the World Health Organization, about 5.4 million people in the world are bitten by snakes per year, with half of them having venomous bites and 100,000 cases resulting in death. Snakes are essential as predators, ecosystem engineers, and of economic and medicinal importance. Snakebite is a pertinent medical emergency, that calls for urgent management, and identifying snakes is a top priority. However, the rate was lower because human populations have limited knowledge in identifying the snake based on the visual characteristics of body shape, eye shape, and color patterns. Snakes can be classified as venomous and non-venomous based on the images, which is done with the help of image classification techniques and Deep Learning (DL). An ensemble methodology with one novel Convolutional Neural Network (CNN) model and two transfer learning models, MobileNetV2 and ResNet50, is used to design the mobile application for public approach and camping enthusiasts. The author used three convolutional layers in the CNN model. The first, second, and third convolutional layer has 16, 32, and 64 filters respectively, and each of size 3x3, with 'same' padding and ReLU activation. These convolutional layers are followed by max pooling layers and other layers such as dropout, flatten, and dense layers to build the complete model. The tested model is on individual models and ensemble combinations, being robust against different parameter settings and hyperparameter settings.

The proposed ensemble approach achieved a test accuracy of 87% and 0.94 Receiver Operating Characteristic (ROC) during the testing phase. Novel CNN, Modified MobileNetV2, and ResNet50 achieved 64%, 82%, and 82% respectively. It takes only 5 seconds to detect snakes as venomous or non-venomous. The learning curves and confusion matrix are displayed to evaluate the model's performance and identify potential errors in the testing chapter."