3. Methods
We proposed an algorithm called “View Zenith Angle (VZA) Stratified
COLD” (hereafter referred to as VZA-COLD), which is built on the
Landsat-based COntinuous monitoring of Land Disturbance (COLD) algorithm
(Zhu et al., 2020). It includes three major innovations that are: (i)
cloud/snow buffer; (ii) change detection based on observations from four
stratifications of VZA; and (iii) consistent dark pixel (DNB radiance
< 1.0 \(nW*m^{-2}*\text{sr}^{-1}\)) removal. The workflow of
the VZA-COLD algorithm is illustrated in Fig. 2. The first component
involved removing the remaining cloud and snow observations based on the
exclusion of cloud and snow edge pixels that are partially influenced by
clouds or snow. In the second component, DNB observations were
stratified into four groups based on the VZA intervals to mitigate the
variance caused by the compounded impacts from the different viewing
geometry and surface condition, and thus, reduce the overall time-series
variation. For each VZA interval, NTL changes were monitored based on
the change detection framework similar to the COLD algorithm to obtain
the individual sets of estimated time series models and detect a unified
set of breakpoints among all the VZA intervals continuously. The third
component involved filtering the consistent dark pixels with a minimum
detectable NTL radiance threshold. The VZA-COLD algorithm was calibrated
based on the calibration samples, in which the changes detected within
the period of ± six months of the calibration sample change intervals
were determined as the right call. Metrics of omission rates, commission
rates, and F1 score of NTL change were used to evaluate algorithm
performance. The final map accuracy was evaluated based on independent
validation samples following the “good practice” protocols (Olofsson
et al., 2014).