YOLO-HARVEST: A hybrid ViT architecture with locality-enhanced attention for automated wildlife species classification

Wildlife conservation efforts increasingly depend on automated species classification for processing large-scale camera trap data, yet existing approaches struggle with accuracy and computational efficiency in resource-constrained environments. This paper introduces HARVEST (Hierarchical Attention for Robust Vision Enhancement with Shifted Tokenization), a novel hybrid architecture integrating YOLOv8 object detection with transformer-based classification. The architecture incorporates three key innovations: Shifted Patch Tokenization (SPT) for boundary information preservation, Local Information Enhancer (LIFE) for spatial feature extraction, and Locality-Enhanced Attention (LEA) for adaptive feature integration. The model is evaluated on two comprehensive datasets: a challenging 45-species Ohio State University (OSU) Small Animals dataset exhibiting an extreme class imbalance (6320:1 ratio) and a balanced 6-species African wildlife dataset. The HARVEST demonstrates excellent performance and achieves 85.27% accuracy on the OSU dataset and 94.74% accuracy on the Wildlife dataset with only 13.0M parameters, representing an 85% reduction compared to standard Vision Transformers while maintaining superior performance. The OSU evaluation demonstrates robust performance across highly imbalanced real-world conditions with species sample sizes ranging from 1 to 6320 images, validating practical applicability for conservation scenarios. Qualitative analysis reveals biologically meaningful attention patterns focusing on taxonomically relevant features. The efficient architecture enables real-world deployment in conservation applications, providing a practical solution for automated wildlife monitoring and biodiversity surveillance.