Year: 2026
Type: article
Abstract: Objective. The purpose of this research is to search for and analyze solutions in the field of semantic segmentation of land cover on high-resolution satellite images, which provide high quality and reduce dependence on the expensive process of pixel-by-pixel data annotation. Method. The study uses three methods: classical supervised learning based on the U-Net architecture with a ResNet101 encoder, the unsupervised STEGO method that uses features extracted by a self-learning transformer, and the state-of-the-art third-generation DINOv3 model, which is pre-trained on the large-scale satellite dataset SAT-493M and adapted to the segmentation task by adding a segmentation head and a knowledge distillation procedure to the convolutional network. The experimental evaluation is performed on the open DeepGlobe dataset using the IoU metric for land cover classes. Result. The results demonstrate that a baseline supervised approach yields strong performance when a sufficient amount of labeled data is available, while STEGO confirms the feasibility of generating semantically consistent regions without the direct use of annotations. High metric scores were achieved by transferring DINOv3 features into a convolutional architecture via distillation, which improved segmentation quality when training on a limited subset of the data. Conclusion. Self-supervised methods achieve high segmentation accuracy even when labeled data is scarce, which justifies their use and further development within the field of computer vision as applied to Earth remote sensing data.
Authors: S. D. Shibaykin, P. A. Mishina, P. A. Mishin
Institution: National Research Mordovia State University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225916208
