Data-driven automatic labelling of land cover classes from remotely sensed images

With the developing technology and automation, automatic labelling of images is of great importance for automatic mapping. However, the most significant disadvantage of this method is that the classes' labels cannot be generated automatically. In the current remote sensing literature, understanding and automatically labelling clusters obtained from the clustering process without a training phase is a problem that requires effective solutions. In this study, in order to solve this problem, we present a methodology that creates labels without any training phase. We use the bands in the image and Corine data in this process. The methodology uses a database created by examining the spectral characteristics of land classes from sample images collected from various geographies and time periods. The spectral index values of the unlabelled classes obtained are evaluated using this database, and the relevant label is assigned to each class. This database was created by analyzing Sentinel-2 Level-1 images of the Mediterranean and the Black Sea regions in Turkey. Then, these labels compare with the Corine classes corresponds to each pixel according to the ruleset. This developed approach aims to automatically label land, a green agricultural area, forest, urban area, and uncultivated agricultural area. The reason for choosing these areas is that they are the areas that generally make up the environment and a large part of the ecosystem, which are important areas that many researchers frequently use in their studies. The methodology developed was tested with Sentinel 2 images of Gemlik, Hatay regions from Turkey, and Agioi Apostoli region from Greece. The results of the accuracy analysis are 80%, 83%, and %82 for Gemlik, Hatay, and Agioi Apostoli areas.

Automated summary

This study proposes a method for automatically labeling images without a training phase. By using bands in the image and Corine data, a database is created by examining spectral characteristics of land classes from sample images. The unlabelled classes are evaluated using this database, and the relevant label is assigned. The developed approach is tested in several regions in Turkey and Greece and achieves high accuracy.