Deep learning-based detection for thermokarst topography using the chopped picture method

Abstract The polar region is sensitive to climate change, with concerns about the effects on ecosystems and human society. In particular, the thawing of permafrost associated with rising temperatures accelerates the microbial decomposition of organic carbon in the soil, leading to greenhouse gas emissions. Thermokarst is a landform process formed by thawing ice-rich permafrost and subsidence of the ground surface. This landform is an indicator of permafrost degradation; thus, evaluating the distribution of thermokarst is essential for understanding the impact on Arctic regions. Although assessing the thermokarst has been a labor-intensive task because of its widespread occurrence in the Arctic, automatic detection using deep learning and remote sensing techniques has been applied. However, the cost of creating training data for the specific area was challenging because thermokarst size and shape varied by region. Here, we classified thermokarst from satellite images using a recently developed method, the chopped picture method, which is suitable for identifying ambiguous and amorphous objects such as the thermokarst. This study uses high-resolution panchromatic and pan-sharpened images in eastern Siberia to evaluate the effects of differences in satellite images on classification accuracy. The training and test images were divided into 60 pixels in height and width, and each cell was classified into two categories: thermokarst topography or others. In addition, we used Global Map Data to calculate the percentage of thermokarst topography for each slope orientation (south or north) to identify the environmental conditions that facilitate the development of this topography. Results showed that our approach could clearly and automatically distinguish developed thermokarst from other landforms such as forests, lakes, and urban. Classification of thermokarst topography in panchromatic and pan-sharpened images indicated that automatic detection was possible in both images. Additionally, thermokarst topography was distributed on south-oriented slopes rather than north. This method will achieve low-cost automatic detection of thermokarst through the use of satellite data and AI. With the increase of small satellites, opportunities to utilize satellite images for observations in Arctic research will expand. Our approach will contribute to environmental monitoring in the Arctic by enabling the automatic mapping of thermokarst.