YOLOv11-Lite architecture for wildlife detection from drone images

IntroductionDrones equipped with cameras are helpful in wildlife tracking. Deep learning has great potential for detecting wildlife, but is constrained by the challenge of detecting tiny objects, especially from higher altitudes.MethodsThese limitations are addressed by an enhanced You Only Look Once 11 (YOLOv11-Lite) model. YOLOv11-Lite is a lightweight, edge-friendly variant of YOLOv11 that reduces computational complexity while maintaining high detection accuracy. Standard Convolution + Batch Normalization + SiLU (CBS) blocks are replaced with Depthwise-CBS units, which reduce the number of parameters and FLOPs. The enhanced version employs a Spatial Reasoning-Enhanced Coordinate Attention-based Simple Attention Module (CA-SimAM) for improved feature representation, Dynamic Sampling (DySample) for adaptive sampling, and a bounding-box IoU for accurate localization. The C2 block with the Parallel Split Attention (C2PSA) module is also replaced with a Ghost-ELAN block, as it enables ghost feature generation and multi-branch ELAN aggregation, achieving good performance with fewer computations.ResultsThe multiscale detection head aids in detecting smaller animals. The enhanced model achieves an mAP@0.5 of 98.5% and an mAP@0.5:0.95 of 94.7% on the WAID dataset.DiscussionThe performance of the model is assessed through comparative tests, which demonstrate the superiority of the enhanced YOLOv11-Lite model over existing algorithms. The proposed approach supports UAV-based wildlife monitoring and improves detection performance and generalization under real-world conditions.