Developing a new cross-sensor calibration model for DMSP-OLS and Suomi-NPP VIIRS night-light imageries

Highlights

• Seasonal effect and noises in VIIRS data were removed by time series decomposition.

• A cross-sensor calibration model was proposed to generate NTL data from 1996 to 2017.

• The consistency of the calibrated NTL time series significantly improved.

• More coherent application results could be obtained by the calibrated data.

Abstract

Night-time light (NTL) data provides a great opportunity to monitor human activities and settlements. Currently, global-scale NTL data are acquired by two satellite sensors, i.e., DMSP-OLS and VIIRS, but the data collected by the satellites are not compatible. To address this issue, we proposed a method for generating long-term and consistent NTL data. First, a logistic model was employed to estimate and smooth the missing DMSP-OLS data. Second, the Lomb-Scargle Periodogram technique was used to statistically examine the presence of seasonality of monthly VIIRS time series. The seasonal effect, noisy and unstable observations in VIIRS were eliminated by the BFAST time-series decomposition algorithm. Then, we proposed a residuals corrected geographically weighted regression model (GWRc) to generate DMSP-like VIIRS data. A consistent NTL time series from 1996 to 2017 was formed by combining the DMSP-OLS and synthetic DMSP-like VIIRS data. Our assessment shows that the proposed GWRc model outperformed existing methods (e.g., power function model), yielding a lower regression RMSE (6.36), a significantly improved pixel-level NTL intensity consistency (SNDI = 82.73, R² = 0.986) and provided more coherent results when used for urban area extraction. The proposed method can be used to extend NTL time series, and in conjunction with the upcoming yearly VIIRS data and Black Marble daily VIIRS data, it is possible to support long-term NTL-based studies such as monitoring light pollution in ecosystems, and mapping human activities.

Introduction

Due to a unique capability in detecting lights at the Earth’s surface, nighttime light (NTL) imageries have been widely utilized as an indicator of human activities (Elvidge et al., 1997, Li et al., 2018, Zheng et al., 2018b). Since the beginning of data archiving in 1992, NTL images, collected from Operational Line-scan System of Defense Meteorology Satellite Program (DMSP-OLS) and the subsequent Visible Infrared Imaging Radiometer Suite (VIIRS) of Suomi NPP satellite, have triggered extensive researches in multiple disciplines (Cao et al., 2019, Elvidge et al., 2009, Zheng et al., 2018a). A great number of studies have utilized either DMSP-OLS or VIIRS data time series to monitor urban growth (Cinzano et al., 2001, Zhou et al., 2014), estimate and spatialize social-economic and environmental variables (N. Zhao et al., 2018), and assess the environmental consequences of urbanization (Gaston et al., 2012).

Despite a valuable long-term archive of DMSP-OLS and VIIRS, the potential of the entire historical archive has not been fully explored. The key reason is that severe inconsistency occurs between DMSP-OLS and VIIRS, making the NTL time series incompatible. The inconsistency is embodied in: (1) NTL intensity for the same year; (2) the trend and variability of NTL time series; and (3) application results, such as urban area extraction and Gross Domestic Product (GDP) estimation. Thus, NTL data cannot be used directly for multi-temporal studies. Fig. 1 demonstrates the NTL time series of different datasets at the pixel level and the sum of total light (SOTL) level. A more noticeable inconsistency can be seen at the pixel level than at the SOTL level. Specifically, the inconsistency of NTL data exists in two aspects – among different DMSP-OLS sensors and between DMSP-OLS and VIIRS.

• Inconsistency among different DMSP-OLS satellite sensors.

DMSP-OLS consists of different satellite sensors. Because of the absence of an onboard calibration system and sensor degradation, the NTL intensity is not comparable within satellite sensors across different years and among different satellite sensors in the same year. Substantial efforts have been made to develop inter-annual calibration methods to address inconsistency among DMSP-OLS sensors (Li et al., 2013, Pandey et al., 2017). Most of the methods adopted the pseudo-invariant feature (PIF) paradigm. After selecting one/several calibration sites as PIFs and a reference image, the PIF paradigm assumed that the NTL intensity was stable for the calibration sites and NTL changes were only induced by system biases. Then, all other images were calibrated to the reference image with the empirical model estimated in the calibration sites. The differences between these PIF-based methods lied in minor details, such as calibration sites, reference images, empirical models and parameter estimation approaches. Elvidge et al. (2009) selected Sicily, Italy, as the calibration site, and used a 2nd order polynomial regression model to calibrate other images into DMSP-OLS F12 in 1999 (as a reference image). Wu et al. (2013) picked Mauritius, Puerto Rico, and Okinawa as calibration sites, and a power function model was employed. Other studies used a globally consistent bias (i.e., the statistical characteristic of NTL images). For example, Zhang et al. (2016) proposed a ridgeline sampling—selecting the densest part of the density plot—and a regression method to generate consistent DMSP-OLS time series globally.

• Inconsistency between DMSP-OLS series and VIIRS.

Compared to the inconsistency among DMSP-OLS sensors, the inconsistency problem between DMSP-OLS and VIIRS is far more severe (Li et al., 2017). This issue can be ascribed to two main reasons: (1) The sensor performance differences between DMSP-OLS and VIIRS. VIIRS is superior to DMSP-OLS in terms of a better spatial, temporal, and radiometric resolution, as well as low light detection limit (Elvidge et al., 2013). DMSP-OLS is unit-less, while VIIRS is presented with radiance. In addition, the overpass time of DMSP-OLS (∼19:30) and VIIRS (∼01:30) are also different (Elvidge et al., 2013). (2) The inherent drawbacks of both satellite sensors. First, subjected to system biases, the inconsistency of DMSP-OLS series cannot be fully solved even after inter-annual calibration. Second, the dynamic range of VIIRS images (i.e., 14-bit) is large, but the DMSP-OLS imagery is quantized with 6-bit DN values. Thus, DMSP-OLS images are often saturated in highly urbanized regions, while the NTL intensity variation is clearly presented by VIIRS. Version 1 VIIRS, the most commonly-used VIIRS data, is a monthly average composite product and is affected by insufficiently suppressed noise signals (e.g., high energy particles) and strong seasonal effect induced by vegetation and snow coverage (Levin, 2017). Only a few studies have addressed the seasonal effect before using VIIRS data for applications (Xie et al., 2019, Zhao et al., 2018b). Many VIIRS-based studies used the average of monthly VIIRS data in a specific year or even arbitrarily selected a single month (Levin and Zhang, 2017, Ma et al., 2018). As shown in Fig. 1, if the seasonal effect and noise signals were not removed, it would lead to strong data fluctuations. Thus far, only a few cross-sensor calibration models have been proposed. Shao et al. (2014) used the raw daily VIIRS and DMSP-OLS data over Dome C in the Antarctic and employed a linear regression model to calibrate DMSP-OLS into VIIRS-like data. Similarly, Li et al. (2017) used a monthly averaged VIIRS and specially ordered monthly DMSP-OLS archive, and employed a power function model to generate VIIRS-like DMSP-OLS data. Neither methods were satisfactory because the datasets used were inaccessible to the general public, and neither method dealt with the noise in VIIRS or the saturation problem in DMSP-OLS. Because of the lack of feasible cross-sensor calibration methods, almost all the existing studies have been confined to using either DMSP-OLS data (1992–2013) or VIIRS data (2012 to present). In other words, the long-term dynamic pattern of NTL images and the associated proxies (e.g., urban extent, GDP) becomes intractable.

In this study, we intended to develop a new method for generating consistent NTL time series by combining DMSP-OLS and VIIRS. To achieve this goal, the following two research problems must be addressed:

• How to remove the seasonal effect and abnormal noises in monthly VIIRS data; and

• How to establish an effective cross-sensor calibration model and to generate pixel-level consistent NTL images from 1996 to 2017;

The specific objectives of this study are: (1) to estimate missing data of global radiance calibrated DMSP-OLS data; (2) to eliminate the seasonal effect and noise signals of VIIRS by time series decomposition algorithm; and (3) to establish a residuals-corrected geographically weighted regression (GWRc) cross-sensor calibration model to produce consistent NTL time series.