Detection of thermokarst landforms in the European Arctic region using Earth observation and Artificial Intelligence

This is the authors' abstract. We don't add key points for this paper.

Permafrost thawing is a striking environmental issue of global concern that affects arctic environments and their alpine counterparts due to accelerated climate change. Thermokarst landforms and ecosystems are progressively displacing permafrost, creating a variety of mixed permafrost-thermokarst landscapes. In Northern Europe, all permafrost classes (continuous, discontinuous, sporadic, and isolated) can be observed and monitored. We here propose a model for thermokarst extension assessment in Northern Fennoscandia (Finland, Sweden and Norway) by training an image segmentation model to recognize and map thermokarst-affected areas. Training polygons were randomly selected from several key sites experiencing thawing permafrost and were sampled by geomorphological mapping. High-resolution Maxar imagery were employed for training a deep-learning model to detect thermokarst-induced features in the Arctic peatlands, by applying combined capabilities of ResNet and U-Net. The achieved model enabled the correct detection of thermokarst ponds in more than 80 % cases. Similarly, the model correctly detected thermokarst in 77 % cases based on textural differentiation of permafrost-hosting and thermokarst features within thawing landsystems. The results of this study confirm that deep learning models generated by Artificial Intelligence, supported by geomorphological mapping, can reliably detect and extract thermokarst features from available satellite imagery. Future performance may be attained by i) increasing the number and quality of diagnostic features employed in training, and ii) using high-resolution satellite or aerial imagery of wider coverage for sampling.