Using Texture‐Based Image Segmentation and Machine Learning With High‐Resolution Satellite Imagery to Assess Permafrost Degradation Landforms in the Russian High Arctic

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Inauen, Cornelia M., Nitze, Ingmar, Langer, Moritz, Morgenstern, Anne, Hajnsek, Irena, Grosse, Guido
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000550

Amplified climate change across the Arctic causes significant permafrost thaw and an increase of permafrost degradation landforms. These landforms range from fine‐scale degrading ice wedge‐polygon‐networks to large‐scale features such as thermo‐erosional gullies and reshape entire landscapes. In particular the expansion of thermo‐erosional gullies could have far‐reaching consequences by restructuring drainage pathways. Our study aims at finding a suitable remote sensing‐based approach for quantifying landscape‐scale permafrost degradation in gully‐dominated Arctic landscapes. We use historical and recent high‐resolution panchromatic satellite imagery allowing multi‐decadal analysis of degradation trajectories. Given that degradation stages are characterized by distinct but subtle textural characteristics in satellite imagery, we tested texture‐based machine learning segmentation methods including Random Forest (RF) using gray level co‐occurrence matrix (GLCM) texture features and deep learning Convolutional Neural Networks (CNNs) using a UNet architecture. For CNN, we tested various framework adjustments. Our results showed that CNN outperforms RF particularly for complex texture‐defined classes. CNN reached a micro mIoU of 0.71 (accuracy 83.2%) compared to 0.61 (accuracy 75.9%) for RF. Well‐developed baydzherakhs, an advanced stage of ice‐wedge‐polygon degradation, were detected with high confidence (recall of 0.78–0.96 for CNN). Data augmentation and the use of GLCM features within CNN enhanced robustness against domain shifts. However, the most efficient way to adapt the trained model for additional sites was achieved through targeted fine‐tuning. In conclusion, CNN segmentation demonstrated satisfying performance in quantifying fuzzy permafrost degradation stages. It can be expanded in space and time and therefore enables studying long‐term permafrost degradation dynamics. Climate change is particularly strong in the Arctic, causing permafrost (permanently frozen ground) to thaw. Permafrost can have high ice contents, whose melting results in localized surface subsidence as the soil collapses into the space previously occupied by the ice. This forms permafrost thaw landforms ranging from meter‐scale melting ice wedge‐polygon networks to features such as erosional gullies extending over hundreds of meters in length. The development of such landforms can reshape landscapes, impact ecosystems, and alter drainage pathways. On high‐resolution satellite imagery, degradation structures can be identified according to their distinct patterns. In our study, we tested machine learning methods to map these structures. To enable long‐term analysis of permafrost thaw, we tested these methods on historical and recent greyscale satellite imagery. The tested methods included pixel‐based classical segmentation (Random Forest) using texture metrics as inputs and deep learning Convolutional Neural Networks (CNNs). Our results showed that CNNs performed best, providing good results in delineating permafrost degradation across large areas. The model can be adapted and improved for other sites by retraining it with a small amount of site‐specific training data. This research is important because it enables understanding how permafrost is changing across the Arctic. Image segmentation enables landscape‐scale mapping of permafrost degradation stages based on their texture in panchromatic imagery Convolutional Neural Networks outperform feature‐based Random Forests in identifying subtle target classes with high intra‐class variability Site specific fine‐tuning is an effective way to allow transferring the model to other study sites