Self-Supervised Learning Based Land Cover Semantic Segmentation of Satellite Imagery

Land Cover Semantic Segmentation plays a pivotal role in the field of remote sensing and geographical monitoring. By automated partitioning of image into different land cover classes, it can provide valuable insights into the diverse features on the Earth's surface which can be useful for a multitude of tasks including urban planning and infrastructure manage-ment, agricultural planning, crop monitoring, water resource utilization, forest, and woodland management, etc. Supervised deep learning techniques have seen enormous success in this field. However, their success can be attributed mostly to the availability of large-scale labeled datasets belonging to specific regions. However, this poses difficulties in the application of these techniques in the areas where labeled datasets are not readily available. This study uses self-supervised techniques to apply available supervised architectures on areas with unavailable data. Self-supervised training techniques - SimCLR and SimSiam have been the focus of this study. For fine-tuning and segmentation benchmark analysis, the popular architectures UNet, UNet++ and DeepLabV3have been studied. Ablation studies have been performend by applying self-supervision techniques on different quantities of labeled and unlabeled data. This study also explores the feasibility of the relatively newer LandCoverNet dataset released for the Asia region. This paper reveals how comparable results can be achieved with only a fraction of the labeled dataset using self-supervision techniques.