 |
Aerial view of the landscape typical of the
Hudson Bay Lowlands, Ontario, Canada.
Image taken by K. Rühland. |
"The term thermokarst describes the processes and landforms that involve collapse of the land surface as a result of the melting of ground ice." (Kokelj & Jorgenson, 2013). Thermokarst comes in a variety of shapes and forms: pingos, ice wedge polygons, thermokarst lakes and ponds, thaw slumps, thermal erosion gullies .... thermokarst features range in size from one meter squared to thousands of hectares. But my interest is in thermokarst lakes. Thermokarst lakes are very dynamic in nature. They form as the underlying permafrost (perennially frozen ground) thaws. The lake formation can be caused by climatic factors, such as rising temperatures, or other disturbances, for instance, forest fires or human activity. As the thermokarst lakes progress through their life-cycle they develop laterally by means of thermal and mechanical erosion. In their last stage, thermokarst lakes normally drain either through catastrophic drainage events caused by shoreline breaching or through subsurface drainage (infiltration through the ground). Tracing the thermokarst lake dynamics through time allows us to get a glimpse of the changes occurring in the underlying permafrost. For the purposes of this project, I selected Hudson Bay Lowlands as my study area. This region is rich in thermokarst lakes, and therefore perfect for analyzing their dynamics over time. (Bouchard et al., 2016)
 |
| Sentinel-1 |
Synthetic Aperture Radar (SAR) is a perfect tool for earth observation in the Canadian subarctic. Compared to optical sensors that rely on natural sunlight, SAR is an active sensor that transmits microwaves and listens for the response. SAR can penetrate clouds an operate in all weather conditions (Brisco, 2015). The data that I am using in this project is provided by Sentinel-1, a constellation of two satellites operated by European Space Agency. The Sentinel-1 product that I obtained includes two polarizations: HH (Horizontal Transmit and Horizontal Receive) and HV (Horizontal Transmit and Vertical Receive). The images that you see below are false color composites on two different summer days, where red and blue channels were given HH back-scatter images, while the green channel was given an HV image. As you can see in the image on the left the water is very dark. Mirror-like surface of the water reflects most of the electromagnetic energy away from the sensor creating very low back-scatter values in the HH and HV channels which explains its dark appearance. However, in the image on the right you can see that water surfaces appear as bright magenta. Magenta is a combination of high red and blue values. What is causing such a difference? Wind. Wind creates waves on the surface of the water resulting in a very high back-scatter in HH polarization.
 |
| R:HH G:HV B:HH (left:low wind; right: high wind) |
My plan is to create water masks for a time-series of images and compare them to identify changes. Taking into account the effect of the wind, I am going to use only the HV polarization for extracting accurate water masks. (The example that I present below uses HH polarization, which is suitable only for days with low wind speeds). After pre-processing the SAR images I can analyze the distribution of the back-scatter values in the form of a histogram.
Looking at the histogram, we can see two peaks - land, the smaller peak on the left, and water on the right. To create a water mask we just need to find where to draw the boundary between the two classes. Easy task? I attempted guessing the threshold value as you can see in the images below.
In the image on the left I used a threshold of -22 dB and we can see that there are a lot of water pixels present in the lakes, which is obviously incorrect. In my second guess I used -17.5 as a threshold value between land and water. This time I seem to be doing better.
 |
| Left: Threshold at -22 dB, (HH Polarization); Right: Threshold at -17.5 dB, (HH Polarization) |
However, how would I tell if I have captured all the water pixels? The most common algorithm used for finding the right threshold value is called Otsu Algorithm. It was proposed by a Japaneses scientist Nobuyuki Otsu in 1979. His method allows to subdivide a black and white image into foreground and background automatically, with no a priori knowledge. The algorithm iterates through all possible threshold values and selects the value that maximizes the between-class variance. (Otsu, 1079). Implementing Otsu algorithm on my data is my next step...
Resources used:
Bouchard, F., MacDonald, L. A., Turner, K. W., Thienpont, J. R., Medeiros, A. S., Biskaborn, B. K., ... & Wolfe, B. B. (2016). Paleolimnology of thermokarst lakes: a window into permafrost landscape evolution. Arctic Science, 3(2), 91-117.
Brisco, B. (2015). Mapping and monitoring surface water and wetlands with synthetic aperture radar. Remote Sensing of Wetlands: Applications and Advances, 119-136.
Kokelj, S. V., & Jorgenson, M. T. (2013). Advances in thermokarst research. Permafrost and Periglacial Processes, 24(2), 108-119.
Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics, 9(1), 62-66.
Looking at the histogram, we can see two peaks - land, the smaller peak on the left, and water on the right. To create a water mask we just need to find where to draw the boundary between the two classes. Easy task? I attempted guessing the threshold value as you can see in the images below. 
