Mapping retrogressive thaw slumps using deep neural networks

•Trained high-accuracy deep neural networks to map retrogressive thaw slumps (RTS).•Tested the impact of negative data in training RTS segmentation models.•Evaluated the impact of ‘within-class’ and ‘between-class’ variances on RTS models.•Developed a Lightweight workflow for training deep learning RTS segmentation models.•Developed an effective RTS data fusion method for multi-source satellite imageries. Retrogressive thaw slumps (RTS) are thermokarst features in ice-rich hillslope permafrost terrain, and their occurrence in the warming Arctic is increasingly frequent and has caused dynamic changes to the landscape. RTS can significantly impact permafrost stability and generate substantial carbon emissions. Understanding the spatial and temporal distribution of RTS is a critical step to understanding and modelling greenhouse gas emissions from permafrost thaw. Mapping RTS using conventional Earth observation approaches is challenging due to the highly dynamic nature and often small scale of RTS in the Arctic. In this study, we trained deep neural network models to map RTS across several landscapes in Siberia and Canada. Convolutional neural networks were trained with 965 RTS features, where 509 were from the Yamal and Gydan peninsulas in Siberia, and 456 from six other pan-Arctic regions including Canada and Northeastern Siberia. We further tested the impact of negative data on the model performance. We used 4-m Maxar commercial imagery as the base map, 10-m NDVI derived from Sentinel-2 and 2-m elevation data from the ArcticDEM as model inputs and applied image augmentation techniques to enhance training. The best-performing model reached a validation Intersection over Union (IoU) score of 0.74 and a test IoU score of 0.71. Compared to past efforts to map RTS features, this represents one of the best-performing models and generalises well for mapping RTS in different permafrost regions, representing a critical step towards pan-Arctic deployment. The predicted RTS matched very well with the ground truth labels visually. We also tested how model performance varied across different regional contexts. The result shows an overall positive impact on the model performance when data from different regions were incorporated into the training. We propose this method as an effective, accurate and computationally undemanding approach for RTS mapping.