Unsupervised Ensemble Semantic Segmentation for Foreground-Background Separation on Satellite Image

Image segmentation is of great importance in computer vision and deep learning (DL) techniques especially supervised DL outperforms other methods for image segmentation. However, a large amount of annotated/labeled data is needed for training supervised DL models, while such big annotated data is typically unavailable in practice such as in satellite imagery analytics. In order to address this challenge, a novel ensemble unsupervised semantic segmentation method is proposed for image segmentation on satellite images. Specifically, an unsupervised semantic segmentation model is employed to implement foreground-background separation and then be placed within an ensemble model to increase the prediction accuracy further. Experimental results demonstrated that the proposed method outperforms baseline models such as k-means on a satellite image benchmark, the XView2 dataset. The proposed approach provides a promising solution to semantic segmentation in images that will benefit many mission critical applications such as disaster relief using satellite imagery analytics.