"We applied Simple Non-Iterative Clustering (SNIC) (Achanta and Susstrunk 2017) to develop image segments or super-pixels from the WorldView mosaic (Figure 4). Input bands for the segmentation were Blue, Green, Red, and NIR reflectance (range ~0–1.0), and NDVI (range -1–1.0). Parameters for SNIC were compactness=0.0001, connectivity=4, neighborhoodSize=128, and size=6. While the original input imagery was 2 m resolution, we performed the segmentation on a 6 m resolution version of the input imagery generated by averaging. Thus, the initial seeds were spaced about 36 m apart (size times scale). We used 6 m resolution for the segmentation both to reduce computational demands and because it was a good resolution to capture meaningful tundra vegetation patches while minimizing the delineation of small features such as troughs in patterned ground that are better represented in a vegetation mosaic.

We performed 2 rounds of merging adjacent segments together, if they met a tight threshold of spectral similarity. If the mean value of reflectance and NDVI was within 0.005 in all five bands (50 for scaled values), adjacent segments were merged. The second round allowed for merged segments created in the first round to be merged with neighbors, if the similarity criteria were met.

Finally, we renumbered the segments so that each contiguous patch was assigned a unique code. The segmentation algorithm relies on randomly generated 64bit integers to initially define unique values. The additional renumbering step ensured that we did not end up with spatially separated segments with the same code."