Evaluating a deep-learning approach for mapping retrogressive thaw slumps across the Arctic
Retrogressive thaw slumps (RTS) are typical landforms indicating processes of rapid thawing and degrading permafrost. Their abundance is increasing in many regions and quantifying their dynamics is of high importance for assessing geomorphic, hydrologic, and biogeochemical impacts of climate change...
Main Authors: | , , , , |
---|---|
Format: | Conference Object |
Language: | unknown |
Published: |
American Geophysical Union
2021
|
Subjects: | |
Online Access: | https://epic.awi.de/id/eprint/55337/ https://hdl.handle.net/10013/epic.bf42dc2a-edd8-480e-829c-ffbfb003f1ce |
Summary: | Retrogressive thaw slumps (RTS) are typical landforms indicating processes of rapid thawing and degrading permafrost. Their abundance is increasing in many regions and quantifying their dynamics is of high importance for assessing geomorphic, hydrologic, and biogeochemical impacts of climate change in the Arctic. Here we present a deep-learning (DL) based semantic segmentation framework to detect RTS, using high-resolution multi-spectral PlanetScope, topographic (ArcticDEM elevation and slope), and medium-resolution multi-temporal Landsat Trend data. We created a highly automated processing pipeline, which is designed to allow reproducible results and to be flexible for multiple input data types. The processing workflow is based on the pytorch deep-learning framework and includes a variety of different segmentation architectures (UNet, UNet++, DeepLabV3), backbones and includes common data transformation techniques such as augmentation or normalization. We tested (training, validation) our DL based model in six different regions of 100 to 300 kmĀ² size across Canada, and Siberia. We performed a regional cross-validation (5 regions training, 1 region validation) to test the spatial robustness and transferability of the algorithm. Furthermore, we tested different architectures, backbones and loss-functions to identify the best performing and most robust parameter sets. For training the models we created a database of manually digitized and validated RTS polygons. The resulting model performance varied strongly between different regions with maximum Intersection over Union (IoU) scores between 0.15 and 0.58. The strong regional variation emphasizes the need for sufficiently large training data, which is representative of the diversity of RTS types. However, the creation of good training data proved to be challenging due to the fuzzy definition and delineation of RTS. We are further continuing to improve the usability and the functionality to add further datasets and classes. We will show first results from the upscaling beyond small test areas towards large spatial clusters of extensive RTS presence e.g. Peel Plateau in NW Canada. |
---|