University of Twente Student Theses
Cloud masking over glacial snow cover using Sentinel-2 cloud products and semi-supervised image classification in the Indian Western Himalayas
Mishra, Somdutta (2022) Cloud masking over glacial snow cover using Sentinel-2 cloud products and semi-supervised image classification in the Indian Western Himalayas.
|
PDF
9MB |
| Abstract: | Cloud contamination in satellite images are a major obstacle for using these for glacial studies. A number of cloud masking algorithms using rule based or machine learning techniques are developed and still an active field of research. However, one common challenge in cloud making algorithms is the detection of clouds over bright surfaces like snow. The recent study on intercomparison exercise (CMIX) on cloud masking algorithms found that the popular cloud products of Sentinel-2, scene classification layer (SCL) and S2cloudless suffers from accurate detection of clouds over snow cover. These cloud masks are frequently used by the remote sensing community for cloud screening before image analysis. This study evaluates the three cloud masking methods available for Sentinel-2 images: level-1C cloud masks, SCL and S2cloudless for their cloud masking capabilities over the snow cover in high mountain glaciers of the Indian Western Himalayas. The results show that these cloud masking methods fail to distinguish snow from clouds in winter images. Level-1C performed the worst compared to SCL and S2cloudless. These cloud products also show wide overlap in their spectral signals in the Green and SWIR wavelengths which might explain the poor cloud masking by these products. The study also attempted to develop a cloud mask using the spectral properties of landcover features in the study area. The convex and non-convex clustering methods were used to find spectral classes belonging to cloud and landcover features in the Green and SWIR wavelength feature space. These cluster labels were then used to classify the images using nearest-neighbouring classifier. The resulting classified images were assessed for their accuracy over manually labelled cloud pixels. The non-convex cluster labels of Spectral clustering showed >85% cloud detection in both a summer and winter image. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/93112 |
| 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