Fig. 6. Change detection results with different numbers of consecutive anomaly observations, and different change probabilities. Where the shapes of the marker represent the selected number of consecutive anomaly observations (from 12 to 16), the marker sizes indicate the applied change probability (70%, 75%, 80%, 85%, and 90%), and the marker face colors show the F1 scores.
3.4. Consistent dark pixel removal
Change probability in VZA-COLD was normalized by the RMSE calculated from robust regression, which could be still sensitive to pixels that are consistently dark due to their relatively low temporal uncertainties and resulted very small RMSE values in model estimation. Therefore, a slight change in the DNB values caused by outliers could lead to a substantial increase in the final normalized change probability, and result in commission errors. To mitigate this issue, the detected breakpoints with low overall values of DNB and small change magnitudes are considered as low confidence changes that are more likely to be commission errors caused by the outliers. Changes detected over the consistent dark area were identified based on the model predicted overall DNB values before and after the change, and the corresponding change magnitude. A threshold of 1.0 \(nW*m^{-2}*\text{sr}^{-1}\), which is two times the breakthrough value of the NTL detection limit (\(L_{\min}=\ 0.5\ nW*m^{-2}*\text{sr}^{-1}\)) defined in the daily Black Marble product (Román et al., 2018), was applied to exclude the low confidence changes in consistent dark areas. The detected breakpoints, with before-break overall value, after-break overall value, and change magnitude value all less than 1.0\(nW*m^{-2}*\text{sr}^{-1},\ \)would be identified as consistent dark pixels and would be removed from the final change detection results. Commissions caused by the scattering light (Fig. 7) and salt-and-pepper noise (Fig. 8) were substantially removed by this approach.