Remote Sensing

Ivan Barton, School of Environmental and Forest Sciences

The first Sentinel-2 satellite was sent to orbit in 2015 as part of the European Copernicus program. It provides free, high resolution Earth observation data on global scale, collecting an incredible amount of data (1.6 TB/orbit). In the past, image interpretation was done by human operators, but in the last few decades computer vision has played a significant role in mapping due to the increasing data volume. With object based image analysis (OBIA) the mapping quality could approach or even overtake the man-made maps with high resolution images. The first and most important step in OBIA is the image segmentation, wherein the image is taken apart into homogenous regions. The state of art image segmentation methods are not frequently used in satellite remote sensing. Most modern methods are data driven, which perform well in artificial environment, but fail in natural environment due to the lack of training data. Therefore the older, unsupervised image segmentation algorithms are preferred by users. There are several commercial and open source OBIA solutions for processing long time series in order to extract thematic information. However, all of them have strengths and weaknesses which limits their research and operational applications. To overcome this problem, we implemented a modified multiresolution image segmentation algorithm which could be utilized for Sentinel-2 data processed in High Performance Computing (HPC) environments.