Identifying landscape features within upland thermo-erosion gullies on the Tibetan Plateau via UAV-based machine learning

Thermo-erosion gullies (TGs) are the most typical thermokarst feature in the upland permafrost of the Tibetan Plateau. Accurate identification of all major landscapes within TGs over space are important for permafrost stability, and the stage divisions for TGs are still a research gap due to the complexity of landscape distributions and their difficulty in being detected by satellites. To fill this gap, we conducted this study.

The landscape patterns of TGs reflect their development potential, with each stage displaying unique characteristics. As TGs develop, the proportions of margins and cracks decrease, while the proportions of water and exposed increase.

Firstly, the 6 common landscapes from 8 different TGs on the northeast Tibetan Plateau were identified through visual interpretation and expert experience. Secondly, a well performed landscapes identification model, based on machine learning and a 10-fold cross-validation method repeated 3 times, was established using vegetation indices and microtopographic parameters calculated from unmanned aerial vehicles (UAV) images with a resolution of 0.1 m.

The prediction accuracy of the final landscapes identification model reached 0.82 in the training dataset and 0.85 in the independent test set (with 4 sub-TGs). The relative importance of the Green Normalized Difference Vegetation (GNDVI) was the highest, reaching 37.27%, and the total relative importance of the three selected microtopographic parameters reached 14.51%. The different proportion of the specific landscape represented the different stage of TGs and was controlled by erosion type. With the weakening of abrupt permafrost thawing and the enhancement of water erosion along the gully, the areas of landscape named margin decreased, and the areas of exposed and water increased until the vulnerable permafrost was exhausted.

Our results enhance the understanding of how thermo-erosion gully development affects regional landscape patterns and offer a new insight into the impact of abrupt permafrost thawing on soil stability in upland thermo-erosion gullies on the Tibetan Plateau under climate change.