Mapping Retrogressive Thaw Slumps Using Satellite Data With Deep Learning

Abstract

<p>Retrogressive thaw slumps (RTS) are thermokarst features in ice-rich hillslope permafrost terrain and can cause dynamic changes to the landscape. Their occurrence in the Arctic has become increasingly frequent. RTS can significantly impact permafrost stability and generate substantial carbon emissions. Understanding the spatial distribution of RTS is critical to understanding and modelling global warming factors 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 used 4-m Maxar commercial imagery as the base map, 10-m NDVI derived from Sentinel-2 as the vegetation feature and 2-m ArcticDEM as the elevation feature. 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. Our experiments shed light on the impact of within-class and between-class variances of RTS in different regions on the model performance and provided critical implications for our follow-up study. We propose this method as an effective, accurate and computationally undemanding approach for RTS mapping.</p>

A labeling intercomparison of retrogressive thaw slumps by a diverse group of domain experts

“…Previous studies reported maximum IoU values of 0.58 (Nitze et al, 2021), and 0.74 (Yang et al, 2023) or F1 values of 0.73 (Nitze et al, 2021), 0.75-0.85 (Witharana et al, 2022) and 0.85 (Huang et al, 2020). These previous studies typically used calibration and validation data from one or a few participants, thus the maximum IoUs reported reflect the degree to which DL frameworks can mimic the participants' ontological understandings and delineation tendencies.…”

“…(Nitze et al, 2021). However, they all achieved accuracy metrics using IoU (intersection over Union), which ranged from low to very good agreement: 0.15-0.58 (Nitze et al, 2021), 0.71-0.74 (Yang et al, 2023) or F1 of 0.25-0.73 (Nitze et al, 2021), 0.85 (Huang et al, 2020) and 0.75-0.85 (Witharana et al, 2022). As these values are relative to validation data based on self-created hand-drawn labels, and not independent benchmark datasets, accuracy metrics are difficult to compare across methods and geographical regions.…”

“…Recent DL initiatives for mapping RTS across different locations within the permafrost extent, such as Siberia (Yang et al, 2023), Tibetan Plateau (Huang et al, 2020(Huang et al, , 2021Xia et al, 2022), Northern Canada (Witharana et al, 2022), various localities across the Arctic (Nitze et al, 2021), or the entire Arctic (Huang et al, 2023), used different methods and imagery sources, to train and validate their models. For example, (Huang et al, 2020(Huang et al, , 2021Nitze et al, 2021;Witharana et al, 2022;Xia et al, 2022) created their own hand-drawn labels, while (Yang et al, 2023) used a combination of newly hand-drawn labels and already existing external sources e.g. (Nitze et al, 2021).…”