Improving the Performance of Seabirds Detection Combining Multiple Semantic Segmentation Models

The loss of seabird habitats makes many precious seabirds on the brink of extinction. Monitoring seabirds by unmanned aerial vehicle (UAV) and high-resolution aerial imagery is of vital importance in the long term. The development of convolutional neural network brings great vitality to the detection of small object just like seabird. Due to the complex background and tiny size of seabird in UAV imagery, it is difficult for single semantic segmentation model to detect seabirds accurately. Ensemble learning combines multiple individual learners with a certain strategy to form a committee, which can greatly improve the performance. To realize highly precise and automatic seabird detection from UAV imagery, this paper proposes a deep learning method which combines two semantic segmentation models, as distinct from those generic object detection methods using bounding boxes. The experimental results show that compared to the models such as U-Net and UNet++, the ensemble model performed better overall.