Snow Cover Detection {Methodology Using Landsat}
NASA MODIS Land February 2004
Methodology Using Landsat
Snow cover mapping by way of satellite imagery is a complicated process that must take into account the reflectance values of some of the bands of the electromagnetic spectrum captured by the imager. For this study, images are taken from the Landsat Thematic Mapper and Enhanced Thematic Mapper which takes images in seven different bandwidths.
To see the particular GIS analysis steps taken and tools used in ArcMap, see the Tool Map.
1) Data Collection and Import
The first step was to gather data from archived Landsat imagery using the Global Visualization Viewer (Glovis) on the USGS Earth Resources Observation and Science Centre (EROS) website. As cloud cover blocks the satellite's view of the Earth's surface, cloud free days were found and downloaded for a tile in the Coast Mountain region. This tile was downloaded for several different years in the month of October by requesting the images from the USGS.
Scenes used by Landsat Scene ID:
The seven bands making up each image for each timestamp were then imported into the GIS program ArcMap in order to conduct further analysis. A separate geodatabase was constructed for each timestamp to work on each set of images separately. The initial snow cover detection methodology was created using an image from October 25, 2011. A true color image was made using bands 1, 2 and 3 in order to visually compare the detected snow extent with the actual snow extent.
Hillshade mask with non-snow pixels in dark green.
There is no perfectly accurate way to filter out water bodies using band reflectances without also filtering out some snow cover. Therefore a 'Lakes and Rivers' shape file for British Columbia was downloaded from the University of British Columbia Geography files, imported into ArcMap and made into a Lakes and Rivers mask. Small rivers are more difficult to identify as they are very narrow and any spatial error in the feature would translate into mask inaccuracy. Small rivers were therefore absorbed into the total error of this method.
Border mask - A final mask was created which defined the exact study area for snow cover detection within the image. The edges of the each band image do not align perfectly, causing some pixels to be inaccurately represented in the NDSI mask. Using a defined border mask allows comparisons to be made between images taken on different days or years, as well as allowing comparison of an exact location between Landsat and MODIS images.
Masks applied:
Brightness Threshold - In TM Band 1 snow is reflects more light than almost all other surfaces. Using 0.16 as a minimum threshold reflectance ratio value of 0.16, which correlates to a digital number value of 41, low reflectance pixels are masked out of the original NDSI image.
Hillslopes in shade - NDSI has trouble distinguishing between shaded hillslopes with snow and without snow. Band 4 captures light at near infrared wavelengths which indicate temperature. As pixels in non-snow shaded hillslopes have a value less than 11, they can be masked out of the original NDSI image. The Hillshade mask is shown in this image where dark green represents the pixels that are subtracted from the possible snow covered area.
Water bodies - Snow and water have very similar reflectance values and both vary based on their characteristics. Snow varies with grain size, water content, depth and the amount of radiation absorbing impurities such as silt and sand it contains. Water surfaces vary with depth and turbidity.
2) Initial NDSI Calculation - Error Detection
An initial potential snow covered area is detected using the NDSI calculation and a threshold of 0.4. This area is larger than the actual snow cover as identified by visual comparison, and the misassigned pixels are located. Four Masks were developed to eliminate these errors using research into Landsat bandwidth details and methods trial and error.
3) Masking Application and Reclassification
Tools in ArcMap such as 'Raster Calculator' and 'Reclassify' are used in order to apply the masks to the original NDSI image. Pixels in the final image are classified into non-snow and snow pixels, and the area covered by these pixels is calculated based on the 30 meter resolution of the image.
NDSI histogram,
0.4 threshold.