Vectorial dataset for thermokarst detection in the European Arctic region by Earth Observation and AI

The dataset comprises the vector database developed to train an image segmentation model aimed at detecting and mapping thermokarst-affected areas across the European Arctic. Training polygons were randomly selected from several representative sites undergoing permafrost thaw and were delineated through detailed geomorphological mapping. High-resolution Maxar satellite imagery (0.6 m) served as the primary source for identifying thermokarst-induced features within Arctic peatlands.
To construct the training dataset, regions exhibiting both thermokarst activity and stable terrain were manually labelled in ArcGIS Pro using its image classification tools. This process involved digitizing polygons over distinct thermokarst landforms, taking advantage of the high spatial resolution provided by the World Imagery service to ensure accurate feature delineation.
In total, 133 polygon features representing thermokarst landsystems were extracted from northern Finland, 205 from the Abisko region and northeastern Sweden, and 59 from northern Norway. No fixed sample size was predetermined. The dataset aimed to include as many representative features as possible.
The final dataset contains vector data for both thermokarst landsystems and independently mapped thermokarst ponds, along with full spatial references and metadata. Additionally, shapefiles delineating the areas where thermokarst and non-thermokarst features were detected are provided.