Mapping non-lake thermokarst landforms on the Tibetan Plateau using remote sensing and deep learning
Ph.D. Thawing of ice-rich permafrost can result in distinct landforms on the surface, known as thermokarst landforms. The number and extent of thermokarst landforms in permafrost areas have increased in the recent decades. However, their distribution and temporal changes, especially the non-lake one...
| Other Authors: | , , |
|---|---|
| Format: | Text |
| Language: | English Chinese |
| Published: |
2019
|
| Subjects: | |
| Online Access: | https://julac.hosted.exlibrisgroup.com/primo-explore/search?query=addsrcrid,exact,991039874559003407,AND&tab=default_tab&search_scope=All&vid=CUHK&mode=advanced&lang=en_US https://repository.lib.cuhk.edu.hk/en/item/cuhk-2398852 |
| Summary: | Ph.D. Thawing of ice-rich permafrost can result in distinct landforms on the surface, known as thermokarst landforms. The number and extent of thermokarst landforms in permafrost areas have increased in the recent decades. However, their distribution and temporal changes, especially the non-lake ones on the Tibetan Plateau are poorly understood or quantified. This knowledge gap is because (1) most of them are in the remote and inaccessible areas and (2) their diverse characteristics and similarity to the surroundings on satellite images make the automatic mapping extremely challenging. To map non-lake thermokarst landforms and obtain their distribution on the Tibetan Plateau, I apply geographic object-based image analysis (GEOBIA) and state-of- the-art deep learning algorithms to high-resolution remote sensing images. I innovatively develop a strategy (including delineating training polygons, preparing training images, merging inference results, and polygonising) to utilize deep learning in the processing of remote sensing images. Then I apply the methods to two local studies on the Tibetan Plateau: thermo-erosion gullies on the Eboling Mountain and retrogressive thaw slumps in the Beiluhe region. I utilized images from Google Earth and unmanned aerial vehicle (UAV) images for the Eboling case study, and Planet CubeSat images for the Beiluhe region. In the Eboling area, the GEOBIA results on UAV and Google images show that this method can detect the locations of these landforms but fails to delineate their boundaries and extents. Moreover, the GEOBIA results contain many false positives, which leads to a low accuracy. In contrast, the method based on a deep learning algorithm (i.e., DeepLab) has much better performance and the mapped boundaries are comparable to manual delineation. It allows us to delineate all the 16 thermo-erosion gullies with an F1 score of 0.74. Further analysis shows that these gullies are narrow and co-located with stream vectors, which implies a strong influence of surface streams ... |
|---|