Hybrid YOLOv11–EfficientNetB4 Framework for Wildlife Detection and Species Classification in Himalayan Ecosystems

Abstract

With the increasing rates of habitat fragmentation, climate change, and human encroachment in Himalayan habitats, the demand for an intelligent wildlife monitoring system to aid in conserving biodiversity and mitigating human–wildlife conflict has become even greater. Traditional monitoring methods, such as manual observation and camera-trap analysis, are time-consuming, labour-intensive, and less effective in complex environments. Existing automated wildlife monitoring systems are further hampered by challenges, including occlusion, low illumination, background clutter, and different animal poses. To overcome these drawbacks, this study introduces a hybrid deep learning system using the YOLOv11 and EfficientNet-B4 models to detect and classify wildlife in the Himalayas and its challenging climate. The architecture merges the real-time object localization capabilities of YOLOv11 with the fine-grained feature extraction and classification power of EfficientNet-B4, enhancing the recognition accuracy for endangered species and very similar ones. To create a unified wildlife dataset, we used publicly available Kaggle data for seven Himalayan wildlife species and primary camera-trap images from the region of Ladakh and Trans-Himalayas, which amounted to a total of 35,163 images. To overcome the problem of class imbalance, data augmentation was performed using mosaic and cutmix and class-weighted training. Experimental results show that the proposed hybrid approach achieves an overall classification accuracy of 91.15% and macro F1-score of 0.9109, which is higher than that of the two sole baselines, YOLOv11 baseline (84.7%) and EfficientNet-B4-only baseline (87.9%). The suggested system offers a scalable, AI-driven solution for monitoring and conserving wildlife in ecologically vital areas of the Himalayas, including endangered species, and supporting decision-making processes. The present invention falls under the category of wildlife detection, particularly for detection in the context of the Himalayan ecosystem, as well as camera-trap imagery, species classification, and deep learning.