SEGMENTATION OF SENTINEL AND LANDSAT SATELLITE IMAGES: CURRENT APPROACHES AND PROSPECTS

Polyakov Artem Nikolaevich, Polyakova Kristina Eduardovna

Segmentation of Sentinel and Landsat satellite images represents a relevant task in the field of remote sensing of the Earth. Modern machine learning methods open new possibilities for automated processing of large datasets and extracting valuable information. This article provides a systematic analysis of existing segmentation approaches, including the use of convolutional neural networks, clustering algorithms, and ensemble models. Based on a sample of 2500 high-resolution images obtained from Sentinel-2 and Landsat-8 satellites during the period from 2018 to 2022, a comparative testing of 5 different segmentation models was conducted. The best results were demonstrated by an ensemble of the U-Net convolutional neural network and the K-means clustering algorithm, achieving a segmentation accuracy of 94.7% according to the IoU metric. A new approach was also proposed, based on the pre-trained EfficientNet-B4 model and an original loss function, Focal Tversky Loss, which allows improving accuracy to 96.2% while reducing computational complexity. The obtained results have high theoretical and practical significance, opening prospects for the development of intelligent next-generation geoinformation systems. Further research should be aimed at developing transfer learning methods and adapting models to images from different sources.

Keywords:image segmentation, satellite images, Sentinel, Landsat, machine learning, neural networks, clustering