Fig. 2. Flowchart of the VZA-COLD algorithm. NTL: nighttime
light; VZA: View Zenith Angle; COLD: COntinuous monitoring of Land
Disturbance; DNB: Day/Night Band; BRDF: Bidirectional Reflectance
Distribution Function.
3.1. Remove remaining cloud and snow impacted observations
Considering that cloud and snow edge pixels are very likely to be
influenced by thin clouds and snow, a spatial buffer was applied to
remove these edge pixels. Confident/probable cloud, cirrus cloud,
snow/ice, and observations were firstly removed according to the QA
flags of the standard NASA’s Black Marble product. The cloud/snow edge
pixels removal was tested by dilating cloud/snow pixels (at 8-connected
directions) from 0 to 11 pixels to find the optimal moving window size
based on our calibration samples. For the moving window size equal to or
less than five pixels, both omission and commission errors dropped
gradually along with the increase of the window sizes (Fig. 3a). A
decrease in F1 scores and an increase in omission/commission errors were
observed when the moving window size was larger than five pixels (Fig.
3a), which is mostly due to the removal of too many clear observations
for change detection. Thus, the 5x5 pixel moving window was selected as
the optimal buffer size for masking potential cloud/snow-influenced
pixels. Fig. 3b shows the cloud/snow masks and their 5x5 pixel buffer
for an example image collected at tile h10v04 on Day-of-Year (DOY) 45 in
2015, in which the red pixels are the ones that were captured by the
cloud/snow buffer.