Cloud Masking for Landsat 8 and MODIS Terra Over Snow‐Covered Terrain: Error Analysis and Spectral Similarity Between Snow and Cloud. Issue 7 (29th July 2019)

• Record Type: Journal Article
• Authors: Stillinger, Timbo; Roberts, Dar A.; Collar, Natalie M.; Dozier, Jeff
• Is Part Of: Water resources research. Volume 55:Issue 7(2019)
• Page Start: 6169
• Page End: 6184
• Publication Date: 2019-07-29
• Subjects: snow -- remote sensing -- Landsat -- MODIS; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2019WR024932
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Automated, reliable cloud masks over snow‐covered terrain would improve the retrieval of snow properties from multispectral satellite sensors. The U.S. Geological Survey and NASA chose the currently operational cloud masks based on global performance across diverse land cover types. This study assesses errors in these cloud masks over snow‐covered, midlatitude mountains. We use 26 Landsat 8 scenes with manually delineated cloud, snow, and land cover to assess the performance of two cloud masks: CFMask for the Landsat 8 OLI and the cloud mask that ships with Moderate‐Resolution Imaging Spectroradiometer (MODIS) surface reflectance products MOD09GA and MYD09GA. The overall precision and recall of CFMask over snow‐covered terrain are 0.70 and 0.86; the MOD09GA cloud mask precision and recall are 0.17 and 0.72. A plausible reason for poorer performance of cloud masks over snow lies in the potential similarity between multispectral signatures of snow and cloud pixels in three situations: (1) Snow at high elevation is bright enough in the "cirrus" bands (Landsat band 9 or MODIS band 26) to be classified as cirrus. (2) Reflectances of "dark" clouds in shortwave infrared (SWIR) bands are bracketed by snow spectra in those wavelengths. (3) Snow as part of a fractional mixture in a pixel with soils sometimes produces "bright SWIR" pixels that look like clouds. Improvement in snow‐cloud discrimination in these cases will require more information than just the spectrum of the sensor's bands or will require deployment of a spaceborne imaging spectrometer, which could discriminate between snow and cloud under conditions where a multispectral sensor might not. Plain Language Summary: Both snow and clouds are brighter in the visible spectrum and colder in the thermal infrared than other parts of the landscape, so discriminating between them in passive optical satellite images poses a long‐standing problem in remote sensing of Earth. The operational cloud masks now provided for satellite imagery were chosen because they identify clouds correctly over most land covers. Over snow‐covered terrain in the mountains, however, errors in identification persist. This study evaluates those errors and identifies reasons why some snow and clouds appear very similar in our workhorse satellite sensors, specifically Landsat and Moderate Resolution Imaging Spectroradiometer: (1) At high elevations, snow can look like cirrus clouds in the specific spectral regions employed to detect cirrus. (2) Snow is usually dark in spectral bands where most clouds are bright, but thin clouds can appear similarly dark in those same bands. (3) Image pixels containing both snow and soil can appear bright in the same spectral regions where clouds are bright. Fixing these cases of misidentification will require analyses that use more information than just the spectrum of the sensor's bands or will require deployment of an imaging spectrometer in space. Key Points: Some snow and cloud pixels have similar multispectral signatures, so discrimination can be difficult The best available cloud masks for Landsat and MODIS consistently identify snowpack boundaries as clouds Better snow‐cloud discrimination with multispectral sensors requires information beyond just the pixel's spectral signature