Automated identification of thermokarst lakes using machine learning in the permafrost landscape of Central Yakutia (Eastern Siberia)

Séminaire de Lara Hughes-Allen, postdoc au GEOPS (Géosciences Paris Saclay)

Permafrost landscapes cover 20 million km² of the northern hemisphere and are particularly abundant in Siberia, Alaska, and northern Canada. An important feature of permafrost is its storage of enough organic carbon (OC) to significantly impact global climate if released into the atmosphere as greenhouse gas (GHG). In the ice-rich permafrost area of Central Yakutia (Eastern Siberia, Russia), climate warming and other natural and anthropogenic disturbances have caused permafrost degradation and soil subsidence, resulting in the formation of numerous thermokarst (thaw) lakes. These lakes are hotspots of greenhouse gas emissions, but with substantial spatial and temporal heterogeneity across the Arctic. This talk presents work done to combine in-situ greenhouse gas measurements from lakes near Yakutia, Russia, with remote sensing analysis using machine learning. Mask Region-Based Convolutional Neural Networks (R-CNN) instance segmentation was used to automate lake detection in Satellite pour l’Observation de la Terre (SPOT) and declassified US military (CORONA) images. Using these techniques, we quantify changes in lake surface area in the Yakutsk region since the 1960s and upscale in-situ greenhouse gas measurements to a larger spatial area.