Ensemble Deep Learning Approach for Object Recognition in Large-scale Remote Sensing Imagery

remote sensing data processing based on deep learning methods has been widely used in several cross-domain tasks, such as logistics chains and resource management. However, it faces several challenges and threats. The state-of-theart object recognition approaches used in urban environments are less sensitive to images with sparsely distributed objects. A representative example is maritime vessel detection on the water surface. The models become insensitive to images with single maritime vessel or none at all, as the satellite shot from a top view captures poor feature diversity of objects and the heterogeneous noise background. To address these challenges, this paper proposes an ensemble deep learning-based approach for maritime vessel recognition in large-scale remote sensing imagery. The proposed approach includes classifiers based on DenseNet121 and ResNet152, coupled with a YOLOv8-based detector and a semi-supervised learning strategy. The model was developed, trained and validated using the Airbus SPOT Satellite Imagery dataset. The results demonstrate the effectiveness of large-scale scanning remote sensing and the recognition of sparsely distributed objects, achieving a score of 0.91. Finally, the promising integration of this approach with other computer vision tasks and its application range are discussed.