Animal Species Recognition with Deep Convolutional Neural Networks from Ecological Camera Trap Images.

Area of Science:

• Ecology
• Computer Vision
• Machine Learning

Background:

• Accurate species identification is crucial for biodiversity assessment, conservation efforts, and understanding climate change impacts.
• Camera traps generate vast ecological image datasets, necessitating automated analysis for efficiency.
• Deep learning offers advanced solutions for image-based species identification.

Purpose of the Study:

• To train and evaluate machine learning models for classifying snakes, lizards, and toads from camera trap images.
• To compare the performance of pretrained models (VGG16, ResNet50) against a self-trained convolutional neural network (CNN-1).

Main Methods:

• Experimentation with two pretrained deep learning models: VGG16 and ResNet50.
• Development and testing of a self-trained convolutional neural network (CNN-1) with adjustable layers and augmentation.
• Multiclassification task to differentiate between snakes, lizards, and toads.

Main Results:

• ResNet50 achieved 86% accuracy, and VGG16 reached 87% accuracy in species classification.
• The self-trained CNN-1 model attained 72% accuracy.
• Transfer learning models significantly outperformed the self-trained model.

Conclusions:

• Transfer learning approaches demonstrate superior performance for automated species identification in camera trap data.
• Deep learning models show promise for identifying species, even those with challenging visual characteristics like small body size or camouflage.
• Automated analysis using deep learning is essential for managing and interpreting large-scale ecological image datasets.