Identifying thermokarst lakes using deep learning and high-resolution satellite images

Abstract

Thermokarst lakes play a critical role in hydrologic connectivity, permafrost stability, and carbon exchange from local to regional scales. Due to the typically small sizes and highly dynamic nature of thermokarst lakes, their identification in large regions remains challenging. This study presented a deep-learning model and applied it to high-resolution (1.2 m) satellite imagery to automatically delineate and inventory thermokarst lakes. The method was applied in the Yellow River source region in eastern Tibetan Plateau and identified 52,486 thermokarst lakes, with the majority (90.9%) smaller than 0.01 km2. It's the most comprehensive survey of thermokarst lakes within the region and more than 45% of these lakes were not covered by any existing lake datasets, thereby leading to a possible underestimation of the amount and effects of thermokarst lakes. Validation with visually interpreted data reported MIoU of 0.97, F1 score of 0.96, and PA of 0.97, confirming that thermokarst lakes we detected were matched very well with the reference. The experiment demonstrated great potential for investigating the distribution and impacts of thermokarst lakes in borad regions, such as the entire Tibetan Plateau or even the globe, to provide critical knowledge for their response to climate change and effects from their dynamics.

Keywords

• Thermokarst lake
• Deep learning
• High-resolution satellite imagery
• Yellow river source region