American Geophysical Union
During 2016 I was lucky to work with Twila Moon from the University of Bristol to create a program which identifies clouds in satellite imagery. In December I presented this research at a poster session at the American Geophysical Union (AGU) in San Francisco with other students from across the country.
The Github repository for this project can be found here.
Below you will find a more detailed description of the research as well as the event.
Documentation of the large changes taking place in Greenland outlet glaciers has been greatly enhanced by using large satellite imagery collections to map changing glacier termini and glacier speeds. Cloud cover in the polar regions, however, presents a significant obstacle to the use of visible-band satellite imagery for glacier change measurements. As the number of images in archives has increased significantly, the time required to visually inspect individual images for mapping suitability has also increased. To reduce the time spent combing through the archive for Landsat 8 imagery, we are building a set of tools that combines metadata records with image-to-image spatial correlation to rapidly identify images that capture glacier termini, even if the images are partially cloud covered. This system uses previously mapped terminus positions to identify the areas of images that must be cloud free, and verifies with image to image correlation that these areas are not cloud covered, all without significant operator interaction. The result, for any glacier terminus considered for mapping, is a record of useful satellite images from the archive. The automatically identified scenes can then be used for operator-involved terminus mapping. This frees the operator from the task of scene identification, increasing the level of efficiency of the process significantly. The tools will be integrated into efforts to automate scene selection for seasonal coastal mapping in Greenland and Antarctica.
One of the figures the program generated was a side-by-side comparison between the L8 image and the cloud mask. This made it easy to see how accurate the current algorithm was working. An example of this is below.
As mentioned on the poster, one of the utilities of these cloud masks is the generation of cloud free timelapses. The one below does have a few bad frames but overall worked well.
This December I had the opportunity to present a poster at the American Geophysical Union (AGU) in San Francisco. With over 24,000 people attending, the meeting was spread across four enormous buildings called the Moscone Center.
Presenting my poster.
On the first day, the Bright Stars Program - an initiative supporting high school research - hosted a field trip to various scenes around San Francisco. We started by driving down to a national park, where we learned about whaling and hiked a trail or two before heading back towards the city. Next, we toured a facility doing ocean research with car-sized ROVs. The tour included a visit to their manufacturing space, where they were in the process of building autonomous torpedos to map the seafloor. After the tour, our guide talked about some of the discoveries their ROVs had made. He explained that there was a backlog of species waiting to be officially “discovered” because they didn’t have the time to do all of the paperwork.
I also stopped by the Tesla dealership in my free time.
That night, I presented my poster at a “practice session” of sorts. I talked to around 40 people over the next few hours; many talked about the interesting research they were involved in. One woman had spent the last three years working on a very similar problem as the one my research addressed, with a very different approach.
I suppose I should explain my project. Over the last few months, I have been attempting to create cloud masks for satellite imagery. NASA has launched several satellites under the LandSat program to image the earth and researchers use these images to watch the earth change. A problem with spaced-based earth imagery is the presence of clouds rendering an image useless. Spending time sorting through these images take a lot of time and makes their use difficult for researchers.
My goal was to ease this process by writing software that can identify clouds in images taken specifically over Greenland for the purpose of glacial termini tracing. I worked with a glaciologist based at the University of Bristol and data from project GO-LIVE to accomplish this. In summary, I used GO-LIVE’s data on how well images correlated with each other to build cloud masks for images. These masks can then be averaged over a specific area to determine if clouds are covering the landform of interest in the image. This is described in more detail in the poster above.
A NASA researcher presenting on CO2 emissions.
The following day I manned my poster at the main session (first picture in this post). After presenting in the morning, I was free to look at other’s research and explore the exhibit hall. I met other high school students doing everything from analyzing soil’s effect on plant growth to predicting solar flares using tensor flow and machine learning. The exhibition hall was full of presentations from big companies like NASA and private companies selling research equipment.
After a busy couple of days, it was time to say goodbye to San Francisco.