University of Twente Student Theses
Mapping of glacier extend using deep learning method
Prashant, P. (2022) Mapping of glacier extend using deep learning method.
|
PDF
6MB |
| Abstract: | Glaciers are extremely sensitive to changing climatic conditions. Thus, they are among the 54 Essential Climate Variables (ECV) identified by the Global Climate Observation System (GCOS). The catastrophic consequences of the continuous rise in temperature are glacial retreat and loss of glacial mass, leading to sea-level rise, fresh water loss, hydrological shift, habitat loss, etc. Glaciers from the Arctic, Antarctic, Himalayas, or Alps are continuously receding. It is essential to monitor glacier parameters such as length, mass, area, and extent. This research aimed to map the glacier extents of Jostedalsbreen and Svalbard, Norway using open-source Sentinel-1 and Sentinel-2 satellite data using a novel deep learning-based data fusion method. We investigated the sequence of band combinations of SAR and optical data in pre-designed fully convolutional networks, FCNDK-6, SegNet, UNet, and ResUnet, to delineate accurate glacier extents. In Sentinel-1 experiments, our 3-band experiments provided a comparatively higher F1 score, but at the same time, SegNet has almost 20% less accurate prediction value than the UNet model. After multiple trials of the model, we observed that, in the case of Sentinel-1 VV and VH polarization data, FCNDK and the UNet have stable output. However, the accuracy from the UNet was more consistent, delivering a higher F1 score. Sentinel-2-based model performance was higher than the Sentinel-1-based model. We ran all networks in two different input setups (spectral band combinations); first, we used Blue, Green, Red, and Near Infrared (NIR) bands and achieved an F1 score of 0.88. Later we included shortwave infrared (SWIR) and other NIR bands and used twelve bands as input. In the 12-band experiment, we observed that the same UNet model (used for the 4-band experiment) increased the accuracy by 2% and managed to have an F1 value of 0.90. In our novel experiment, we fused the Sentinel-1 and Sentinel-2 datasets and checked their influence on the resultant mapping of glacier extents. We realized that fusing the data from these two optical and SAR satellites improved the F1 value. We also observed that the same model with a standalone method gave better results with fusion data. For example, FCNDK with Sentinel-1 3-band experiment has an F1 score of 0.73, and Sentinel-2 4-band experiment has 0.83, and when we fused the same three-channel derived from VV and VH polarization of S1 and four bands from S2 and performed a 7-band fusion experiment, we achieved 0.88 of F1 score. Our final attempt experiment concluded that the 18-band fusion experiments provided the best result. This research concluded that UNet is a robust model for accurate glacier extent mapping and can contribute to building and updating glacier databases. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Programme: | Spatial Engineering MSc (60962) |
| Link to this item: | https://purl.utwente.nl/essays/90934 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page