An Assessment of Cloud Detection Methods Concerning High Altitude Snow and Glacial Environments With Sentinel-2

Abstract

Glacier fluctuations are regarded as one of the most significant indicators of climate change. The expansion and contraction of glaciers can be observed by outlining glacier boundaries or measuring snow lines with optical Earth observation satellites. New satellites, such as Sentinel-2A/B, provide high spatial resolution images and short revisit times that can be used to make ample measurements to accurately determine glacial variability. Likewise, ever increasing volumes of satellite data make automated boundary and snow line detection a desirable solution for researchers.

Clouds, however, present a challenge to obtaining useful optical image data. Mountainous regions are often surrounded or covered by clouds. And clouds can be a menacing phenomenon in remote sensing because they greatly attenuate and reflect the short wavelengths used by optical Earth observation satellites. Thus, prior to snow cover or glacier analysis, clouds must be separated from cloud-free pixels. Currently, many techniques exist to automatically detect clouds and classify them. However, they are not perfect. Many techniques have encountered difficulties in cases where snowy and icy landscapes share similar properties with clouds. The mountainous topography also adds to the challenge. Steep slopes and topographic shadows have a profound effect on surface reflectances and can lead to misclassifications.

To address these issues, this study presents an assessment of multiple cloud detection techniques. Several sentinel-2A scenes were acquired at various times between 12/07/15 and 31/12/16. The scenes are centered on the Bara Shigri (glacier), a large 11 km long glacier in Himachal Pradesh, India. Eventually, 6 scenes were processed using 6 different cloud detection techniques: Default L1C, Sen2Cor, Fmask, Temporal Averaging, Maximum Likelihood Classification, and a human-based manual mask which was used as a reference mask.

This study found that automatic cloud detection methods over mountainous terrain performed poorly. Results showed that Maximum Likelihood Classification was the most accurate automatic technique, followed by Sen2Cor. Cloud detection methods often misclassified mountain peaks and glaciers as clouds. At lower altitudes, cloud detection often missed marking known clouds by classifying them as cloud-free instead. Furthermore, it was found that a majority of snow and ice pixels are relatively easy to distinguish from clouds using Sentinel-2's band B3 and B11, or NDSI feature. However, a minority of snow and ice pixels, particularly near cloud edges, are almost indistinguishable from clouds. Results also strongly suggest that mountainous topography negatively affects cloud detection performance. Finally, this study came to the conclusion that cloud detection methods have the potential to improve, but currently are not yet beneficial to glacier research.