The Defense Meteorological Satellite Program (DMSP)/Operational Linescan System (OLS) stable nighttime light (NTL) data have been extensively used in a range of study fields, including socioeconomic research, electricity generation activity, population distribution, urban extent mapping, and environmental assessment (e.g., light pollution) [1
]. The global coverage and long-temporal record (i.e., 1992–2013) advances the application of NTL datasets in related dynamic studies, e.g., the global dynamics of urban expansion, electricity consumption, and economy growth [10
]. Acquiring the pathway (or trend) of historical socioeconomic dynamics is of great importance to future projections for sustainable development as well as deepening our understanding of drivers (or reasons) behind these changes [4
]. However, due to the lack of on-board calibration, varied atmospheric conditions, satellite shift or sensor degradation, the obtained digital number (DN) of NTL time series cannot be directly used across years to detect these dynamics [15
]. Therefore, a calibration procedure should first be implemented for the NTL datasets, before applying these to relevant dynamic studies [11
A variety of studies have been carried out to calibrate the NTL dataset for achieving a temporally consistent and comparable time series. Elvidge et al. [16
] proposed a general framework for calibrating NTL series using the second order regression model. In particular, a reference region that is relatively stable over years and a reference image derived in a particular year and satellite, are defined when implementing the calibration procedure. More specifically, the widely known Sicily, Italy was regarded as an invariant area in terms of socioeconomic activities over recent decades, and the image (featured as year plus satellite) with the maximum luminance (i.e., sum of lit DN values) over this island was selected as a reference for calibration [17
]. This framework has been extensively adopted over the world due to its easy implementation and robust performance [21
]. Moreover, it is flexible and can therefore be adjusted according to different study areas for particular applications, with reference regions such as Jixi county in China [22
], Swain county in the United States [23
], Lucknow and Nawabganj in India [24
], and Okinawa in Japan [25
]. The selection of reference region should meet the following requirements: (1) relative stability over time without dramatic changes; and (2) with a wide range of DN values to build the model [25
]. In addition, several studies were performed based on the calibration framework of Elvidge et al. [16
] with further modification. For instance, Zhang et al. [15
] used the image that lies in the middle of the NTL temporal interval as the reference to minimize the bias caused by images that are temporally far away. Moreover, instead of a reference region, a ridgeline sampling approach was adopted to manually select representative samples for model calibration in their study. Wu et al. [25
] employed a power function to calibrate the NTL time series using a radiance-calibrated image as a reference. The power function used is different with the conventional second order regression model, and the reference region was manually selected based on the luminance change of NTL time series in 1990, 2000, and 2010.
Until now, many calibration studies focused on regional or national scales (e.g., China and USA), and there have only been limited attempts at a global coverage and the whole temporal range (1992–2013) [15
]. Therefore, there are several issues that remain to be addressed for NTL calibration at the global scale. First is the sensitivity of the selected reference image, including the selection of a stable region and the appropriate year (and satellite). There are two assumptions in the widely used framework of Elvidge et al. [16
]: (1) the selected reference regions are invariant over time in terms of the nighttime luminosity; and (2) the selected reference image is unbiased and serves as a benchmark for calibration. However, these are always subjectively determined based on local knowledge, with diverse spatial and temporal references [15
], which may result in systematic biases, and may therefore be incomparable among results derived from different reference datasets. Second, compared to the raw NTL time series, there is likely a notable modification in the calibrated NTL time series, because only one reference image is being used to calibrate the whole time series. For example, after the calibration using the approaches proposed by Elvidge et al. [16
] and Zhang et al. [15
], the range of total lights (DN values) in the world over the last two decades was narrowed down from an initial 242~470 (million) to around 280~380 (million). The range of total NTL DN values was considerably reduced with a flattened change of NTL time series for each satellite compared to the initial pattern [25
]. Due to the unknown actual trend of NTL, the calibration with large modifications may lose the considerable amount information we can gain from the raw NTL series. In fact, these issues are crucial for achieving a reliable and consistent DMSP/OLS NTL time series to support subsequent dynamic studies.
To address the challenging issues (e.g., difficulties in selecting reference regions) in previous studies, we developed a stepwise calibration scheme for NTL data derived from different satellites and years. This study aims to provide a calibrated DMSP/OLS NTL time series (1992–2013) with the following criteria: (1) to improve the systematic consistency of NTL time series derived from different satellites (or periods) instead of the whole NTL dataset; (2) to make full use of the temporally neighbored image as a reference for calibration at the global level; and (3) to generate a temporally consistent long-term NTL time series with less modifications of the original DN values. The remaining parts of this paper are organized as: Section 2
briefly introduces the DMSP/OLS NTL dataset; Section 3
provides a detailed introduction of the stepwise calibration procedure; Section 4
presents calibrated results for comparison and validation over global and selected regions; the concluding remarks are given in Section 5
In this study, we proposed a stepwise approach to calibrate the NTL time series for achieving a temporally more consistent NTL observation. Each step was designed to calibrate particular satellites (or years) with systematic over- or under-estimation. First, the overall underestimation of F14 was adjusted using referred NTL images derived from F12. Then, we modified the subset (2003–2007) of NTL time series of F15 to calibrate its inconsistent trend. Thereafter, a hybrid calibration scheme was implemented on F16, which used the calibration framework proposed by Elvidge et al. [16
]. Furthermore, we systematically adjusted them to make the trend as consistent as the historical trajectory. Finally, we specifically improved the abnormal NTL image of F182010 to mitigate its overly-high DN values. Our calibrated results show a temporally more consistent pattern of NTL time series, as well as a good agreement with socioeconomic data, such as GDP and EC.
Different from the general calibration approaches using a reference image and same scheme for the entire NTL time series [15
], the proposed stepwise calibration approach is aimed at calibrating systematic biases in each satellite (or period), with less modification of raw DN values. The calibrated results are improved significantly at both global and country scales, when compared to the raw NTL time series. Moreover, it also maintains a similar range of SNTL as the raw NTL data, which is different from other widely used approaches with narrowed ranges of SNTL. In addition, the stepwise calibration approach is more successfully aligned with socioeconomic indicators (e.g., GDP and EC), as compared to other approaches. Equally important, the stepwise calibration approach can be repeated without rigorous requirements on reference regions, which were always subjectively selected in different studies [15
]. For instance, we used global references derived from different satellites at the same year mark for calibration, and only incorporated the commonly used example of Sicily in the final step of calibration for F182010.
The calibrated NTL time series using the stepwise calibration approach provides a useful input for future dynamic studies such as urbanization and electricity consumption, which highly rely on the calibrated NTL results to reflect the actual change [10
]. It should be noted that our stepwise calibration scheme follows the order from the 1990s to 2010s. That means we assumed that the night light brightness increased from the 1990s to 2010s, and NTL series of F10 and F12 serve as a reference in the sequential stepwise calibration. Although we found that our calibrated results are reliable with regard to the change of SNTL, reduced NDI, and correlation with GDP and EC at a global scale, the trends of calibrated NTL series (e.g., F10 and F12) may not show a consistent increase in some countries. In addition, our results using the stepwise calibration still keep the dimming trend of nighttime light brightness in some countries (e.g., some of the former Soviet Union countries and Syria). This is because our calibration is a systematic adjustment of NTL data at the global level, i.e., the calibration is consistent for countries with increasing or decreasing NTL. With the newly launched Visible/Infrared Imager/Radiometer Suite (VIIRS) [33
], the improved DMSP/OLS NTL time series can be further extended, with a more consistent trend and a longer temporal coverage.