Harnessing Computer Vision to Identify Pests and Predators: A Comparative Model Analysis

Abstract

This study focuses on developing an intelligent system using machine learning algorithms and motion-sensor camera traps to achieve real-time classification of pest predators in the setting of the bushlands of New Zealand. The primary goal is to provide an optimized tool for ecological monitoring, aiding in the preservation of native Kiwi bird habitats and aligning with the eco-city initiatives. By training and comparing various deep learning models, including Convolutional Neural Networks (CNNs), Multi-Layer Perceptron (MLP), and Vision Transformers (ViT), the system aims to assist in accurately identifying and managing pest populations. We found that our best-performing model on our data was ResNet-50 with an overall accuracy of 98.15% and an average f1-score of 0.982 across the five classes. This was closely followed by DenseNet- 121 with an overall accuracy of 97.95% and an average F1 score of 0.978 and our CNN with a Vision Transformer model with an overall accuracy of 97.58% and an average F1 score of 0.976.