The DEM was corrected, however, in a different manner than the USGS imagery. Although the USGS imagery contains pixels of no data ("NA" values that appear as transparent holes on the map) and cannot be corrected, the DEM NA values can. The "holes" in the map were isolated and compared to neighboring pixels. Then the NA pixels were given the mean pixel values of these surrounding pixels. Next, the isolated clumps are then remapped with the rest of the DEM. This was then remapped to create the full DEM which could then be used to make a mask for the USGS imagery.  Lastly, a mask for each image was created and applied to minimize noise throughout the images.   

On certain days, pixels may reflect differently and thus have different color values despite representing the same land cover. To combat this, both files were obtained in the last week of October to reduce seasonal differences, however, the images are seven years apart.  Each image was also normalized based on the 2023 image's color scale and corrected so that the colors of the pixels reflect the same land cover when appropriate. After correcting the color of the differing images' pixels, the images were corrected topographically. Since Greenland is a higher latitude country, the Sun's shadow over high altitude areas, such as mountains or tall glaciers, is much larger and will darken the pixel's color which can confuse the classification. A topographic correction was assessed using the Sun's azimuth angle and elevation angle on each day the image was captured respectively.  Lastly, a mask for each corrected image was created using the corrected DEM and applied to minimize noise throughout the images.   

The next step taken was to classify the images. Both images were imported to  ESRI's ArcGIS Pro software. The images were then classified using training samples. For each image, samples were hand-drawn to give the algorithm test pixels. These test pixels are then used to classify pixels of similar values to create the entire classification. Three classes were assigned: Water, Snow, and Ice Sheet. 

Once the training samples were saved, the Random Trees classification tool was run in ArcGIS Pro. Random Trees classification takes each pixel, on a case-by-case basis, and compares it to the pixels in the training samples. The classification then used 50 decision trees with a maximum depth of 30 to vote on which class the pixel belongs in.  The majority vote of all 50 decision trees decides the classification. This iterated through every pixel of both images until a final classification was reached and mapped. 

Both images were then taken from ArcGIS Pro and used in R Studio for further analysis. These images were analyzed using a Sankey Network Diagram to show the flow of land cover classes. This diagram shows the change in land cover classes from 2015 and which land cover classes they are currently in 2023. A supplemental change plot was created to determine the extent of the land cover change in the form of a graph.