5. Discussion and Conclusions
Understanding heterogeneity in human settlements is important for linking development patterns to ecological, economic, and social health (Stokes and Seto, 2019). While daytime optical remote sensing has tracked urban land cover change for decades (Zhu et al., 2019), there is a need for new data and analytical tools to be able to monitor whether changes in urban infrastructure are keeping pace with key factors tied to global sustainability trends and opportunities. NTL data derived from the Suomi-NPP and NOAA-20 VIIRS DNB, have the potential to add our understanding of urban infrastructural transitions. But to date, the high temporal variation of the VIIRS DNB observations (Elvidge et al., 2022; Li et al., 2019; Wang et al., 2021) has made it challenging to monitor and detect these factors, at the characteristic temporal scales necessary to capture urban development processes. With the major environmental effects of atmosphere and moonlight removed, the atmospheric- and Lunar-BRDF-corrected NASA’s Black Marble products provide new opportunities to use extremely dense (e.g., daily) NTL time series. In this study, we developed a VZA stratified algorithm for continuous monitoring of NTL changes based on the daily Black Marble products. With the observations stratified into four VZA intervals, this approach estimated an individual time series model for each VZA interval. Based on the predicted and the observed values, consistent breakpoints were identified and applied to all VZA stratified models as potential NTL changes when any of the VZA interval models detected consecutive anomaly observations (can tolerate one exception). Quantitative evaluation of the NTL change detection result showed that the algorithm can detect NTL changes with omission and commission error rates < 30%. In addition to the detected breakpoints of the NTL patterns, slopes of the estimated models can also provide extra information on the trend of gradual NTL radiances changes.
We developed new methods to filter, fit, and mitigate the high temporal uncertainties of the daily DNB time series observations. The applied 5x5 pixel buffer filtered potential outliers surrounding cloud and snow pixels to mitigate the cloud and snow contamination. The single term harmonic model can estimate the intra-annual patterns of the DNB time series as a result of joint factors including vegetation phenology, winter snow coverage, and cyclical human activity changes. We mitigated and simplified the complex and combined effects of the wide ranges of sensor viewing angle and different surface geometry conditions by stratifying the DNB time series based on four sets of VZA intervals. Compared with correcting the viewing angle effect (Tan et al., 2022), the stratification strategy avoided the need for collecting local 3-dimensional structure data and would not be influenced by the potential bias introduced in the correction approaches. The combined knowledge derived from VZA interval subsets, and all data models reduced the variance across VZAs without decreasing the temporal frequency of the data. This new VZA stratification approach would be particularly useful for pixels with strong angular effects and changes that can only be observed under specific viewing angles. For pixels with homogeneous emission among different viewing angles and short-term NTL changes, all data models with 0-60 VZAs could provide accurate and timely results.
The VZA-COLD algorithm can accurately detect different types of human-related NTL changes, either with or without land cover conversion (e.g., urban expansion). For NTL changes with land cover conversion, the concurrent changes of urbanization processes, construction actions, and land cultivation can be well detected with the potential in providing additional information about the land cover conversion stages. For the NTL changes without land cover conversion, the algorithm can identify the long-term and short-term land cover condition change, such as de-electrification of the electric infrastructure, electric grid variations, and human behavior related disturbances. Compared with the daytime optical remote sensing imagery, the NTL change metrics can provide complementary information of the human activity patterns that do not necessarily cause changes in land cover and provide a better understanding of coupling human-environment systems with a unique perspective.
Although VZA-COLD can model and detect the abrupt, cyclical, and gradual NTL changes for both the view angle affected and unaffected regions, it also has some limitations. For example, high latitude areas with large coverage of winter snow/ice could still lead to many commission errors if the cloud/snow QA and their buffers did not exclude them well. This error could be reduced in the upcoming Collection 2 Black Marble products after reprocessing all the data with an improved snow flag derived from the VIIRS snow product with higher spatial resolution. Moreover, the change probability threshold for each time series model is currently calculated based on the same RMSE for all periods, and the temporally varying fluctuations caused by the winter weather are more likely to be identified as NTL changes. To overcome this issue, further adjustments could be performed to consider the changing temporal variation of NTL data within a year, such as applying the temporally adjusted RMSE values for change detection (Zhu et al., 2015). Moreover, further post-processing steps such as filtering or including NTL changes based on their spatial patterns could also be explored.
In conclusion, we developed a VZA stratified COLD algorithm to continuously monitor NTL changes based on the daily atmospheric- and Lunar-BRDF-corrected Black Marble product. This method enabled the dense daily time series analysis of the DNB data by reducing the viewing angle introduced variations without decreasing the temporal frequency of the data. The results indicate that this method could be applied for operational mapping of global NTL changes at a spatial resolution of 15-arc-second with a daily updating frequency.