Analyzing Nighttime Lights Using Multi-Temporal Imagery from Luojia-1 and the International Space Station with In Situ and Land Use Data

Highlights

What are the main findings?

• Multitemporal imagery within a single night shows dynamic urban lighting, with commercial areas brighter early and ports brighter later.
• Nighttime remote sensing correlates strongly (R = 0.8–1.0) with in situ night sky brightness measurements, with variations in images taken seconds apart.

What is the implication of the main finding?

• Imaging timing is crucial for evaluating urban nighttime activity.
• Viewing angles, color sensitivity, and atmospheric conditions significantly affect observed brightness in nighttime remote sensing.

Abstract

Remotely sensed nighttime lights (NTLs) have become essential in urban and environmental research but are typically captured at fixed local times by sun-synchronous satellites, limiting their ability to capture changes throughout the night. In contrast, in situ measurements of night sky brightness (NSB) can provide continuous records over time, but direct comparisons with NTLs have remained rare. This study first examines the relationship between in situ NSB and remotely sensed NTLs using multi-temporal imagery from Luojia-1 and the International Space Station (ISS), focusing on 10 sites in Hong Kong and Macau. We find moderate to strong correlations between NSB and Luojia-1 (R = 0.73) and between NSB and ISS imagery (R = 0.8–1.0), though notable spatial and temporal variations persist. Even images captured within seconds differ in brightness across locations (R = 0.88–0.96), driven by factors such as changing viewing angles in dense urban areas, variations in light transmission paths, and atmospheric conditions, all influenced by satellite position. Our further analysis reveals distinct temporal patterns across land use categories: port facilities and airports are brightest late at night, whereas commercial districts peak earlier and gradually dim throughout the night. Within individual ISS images, transportation-related lighting tends to be red, and commercial areas appear blue compared to other urban areas, which may be due to lamp type differences (high pressure sodium, LED). This study highlights the need to cross-examine in situ and remotely sensed data in NTL research, emphasizing that factors such as local pass time, viewing geometry, color sensitivity, and atmospheric conditions can influence observations and ultimately affect the conclusions.

1. Introduction

As a result of rapid urbanization in recent decades, urban areas now occupy approximately 3% of the Earth’s land surface [1], while this number may appear modest, it represents a substantial transformation of our planet, with its environmental consequences extending far beyond this spatial footprint. One of the most pervasive effects is the rise of artificial light at night (ALAN), resulting in a 10% yearly brightness increase in night sky background between 2011 and 2022 [2]. Consequently, over 99% of the global population now lives under a light-polluted night sky [3]. Among the world’s urbanized areas, Hong Kong—along with the Greater Bay Area, home to a population of 86 million in South China—is one of the most light-polluted metropolitan areas in the world [4,5,6].

Satellite imagery from outer space can monitor the Earth’s changing nighttime profile. One of the earliest satellite projects, the Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS), enabled the first-ever large-scale monitoring of the nighttime Earth dating back to 1992 [7,8,9]. Using these data, researchers have identified strong correlations between nighttime lights (NTLs) and various socioeconomic indicators, including population density, economic development, and energy consumption [10,11,12]. However, as one of the earliest satellite missions, DMSP-OLS data has two major limitations: (a) data storage is constrained to 6-bit binary, with values ranging from 0 to 63, and (b) the lack of on-board calibration means that all values are relative [13]. Its successor, the Visible Infrared Imaging Radiometer Suite Day/Night Band (VIIRS/DNB), addresses these limitations by increasing data storage to 16-bit and incorporating on-board calibration using a solar diffuser system [14]. However, its spatial resolution—approximately 742m at the Equator—still poses challenges for detailed analysis within urban areas [5].

With enhanced sensor technology, the past decade has seen the development of numerous nanosatellites specifically designed for NTL monitoring [15,16]. In 2018, Wuhan University launched a small satellite, Luojia-1, capable of capturing NTL data at a ground sampling distance (GSD) of 130 m [17]. The Jilin-1 satellite achieves even finer resolution, with a GSD of 0.92 m [18]. Beyond conventional satellite systems, the International Space Station (ISS)—the often-overlooked and largest human-made satellite—occasionally provides high-resolution RGB imagery captured by astronauts using digital cameras [19,20]. Leveraging this new generation of satellite imagery, a variety of novel applications have emerged to analyze the nighttime Earth, including studies on housing prices [21], urban expansion [22], ghost cities [23], housing vacancy [24], urban structures [25], and freight traffic [26].

A basic assumption of existing studies is that brighter regions in satellite imagery indicate more human activities at night, while in general, this assumption should hold, there are other factors, often playing a dominant role, influencing the NTL data and their observations. To begin with, some economic activities create disproportionate NTL signals, such as gas flaring in Siberia [27]. Although NTL data have demonstrated good correlations with economic developments, evidence is often based on past experiences over the last three decades, primarily with the rise of developing countries in Asia [9]. Other studies have shown a break in such a relationship in East European and post-Soviet countries, where a stable or declining population corresponds to stable and increasing NTLs [28]. Scale is another problem. As most conclusions are drawn based on the coarse-resolution DMSP/OLS or VIIRS/DNB data, they may be only representative at the country level and may not apply to the city scale [29]. Apart from the socioeconomic factor and the scale problem, other physical factors such as angular effects, aerosols, and relative humidity all have an influence on the NTL data observed from outer space [5,30,31,32].

As research on NTL continues to advance, a critical question remains, as in other areas of remote sensing: how can we effectively ground truth observed NTL data, a challenge that remains unresolved [9,15,29,33]. On the ground, sky quality meters (SQMs, Unihedron, Grimsby, ON, Canada) have been widely deployed to measure the brightness of the night sky [34,35,36,37,38]. Several studies have examined the correlation between SQM measurements and satellite-observed NTL data, reporting a wide range of correlation coefficients, from 0.34 to 0.77 [29,39,40,41]. Notably, the strength of the correlation appears to vary considerably depending on the ambient brightness of the area. For instance, Ref. [41] found that, when analyzing coarse-resolution DMSP data, only SQM readings darker than 19 ${\mathrm{mag}/\mathrm{arcsec}}^2$ showed strong correlations with diffuse light emissions. This finding is consistent with other studies [42,43,44], which also reported stronger relationships between satellite observations and ground measurements primarily in relatively dark environments. However, these studies have been predominantly based on coarse-resolution satellite data. Research on medium-resolution observations, such as those from Luojia-1 and the International Space Station (ISS), remains limited. Given that conclusions often differ depending on spatial resolution, addressing this gap is increasingly important. Furthermore, because these studies rely on data from sun-synchronous satellites, their analyses are constrained to observations captured at approximately the same local time (around 2 a.m.). As a result, temporal variations in NTL within a single night have largely been overlooked [33].

Multitemporal information within a single night is critical for NTL studies, as nighttime activities are known to vary substantially over time. In Hong Kong and other major metropolitan areas across East Asia, decorative lighting on buildings—a primary source of NTL—is often turned off after 10 or 11 p.m. [6,16,34]. However, these early-night lighting activities are not captured by data from either DMSP or VIIRS satellites, due to their fixed local overpass times. Understanding the contributions of different land use types to NTL emissions is a central question in the field. Several studies have attempted to quantify these contributions using land use data [45,46,47]. For instance, using a panoramic image from EROS-B with a ground sampling distance finer than 1 m, Ref. [45] found that land use types explained 89% of the variability in NTL, with arterial roads, commercial areas, and service areas identified as the brightest sources. In Berlin and Massachusetts, multi-family residential and commercial areas were found to contribute the most to NTL emissions, while in Shanghai, commercial areas dominated, and in Quzhou, a smaller city in China, transportation facilities were the primary contributors [47]. Global studies consistently identify commercial and transportation-related land uses as major sources of NTL. However, because these conclusions are drawn primarily from data collected after midnight—when many commercial activities have ceased—analyzing NTL data from earlier in the night and across multiple time points is essential for developing a more comprehensive understanding of nighttime urban dynamics.

In this study, we conduct a comprehensive analysis using multi-temporal imagery from Luojia-1 and ISS to analyze NTL, primarily from two aspects.

• First, we analyze the relationship between NTL data and ground-based NSB data across two multitemporal datasets. We find that even uncalibrated ISS images can achieve good agreements with NSB, and we find that the consistency across images taken within seconds is not as perfect as we thought.
• Second, we analyze the changing NTLs overnight across land use categories in Hong Kong, identifying categories that contribute the most NTLs. We highlight the necessity of analyzing NTLs at different local times that most sun-synchronous data cannot fulfill.

The remainder of this paper is organized as follows. In the next section, we introduce the study area, the satellite data, and the NSB data. We then present methods used in this study in Section 3. In Section 4.1, we show the correlation between field and remote sensing data and analyze the changing NTL data across different hours from the two satellite sources in Section 4.2. Finally, we draw conclusions and provide discussion in Section 5.

2. Study Area and Data

2.1. Study Area

Hong Kong, with a population of over seven million in 2016 [48], is one of the most light-polluted cities in the world [34]. It comprises three major regions: Hong Kong Island, the Kowloon Peninsula, and the New Territories, as shown in Figure 1. Downtown areas occupy only a small portion of the total landmass, concentrated primarily in northern Hong Kong Island and the Kowloon Peninsula. Located in the southwest is the Hong Kong International Airport, an area characterized by relatively stable lighting conditions, which is included in this study. To the west of Hong Kong lies Macau, where three additional in situ stations (introduced later) provide data for correlation analysis. However, while Macau data are used to supplement the correlation analysis, detailed analyses of nighttime light sources are conducted only for Hong Kong, where comprehensive land use data are available.

2.2. Multi-Temporal Nighttime Satellite Imagery

Two sets of multi-temporal imagery from the Luojia-1 satellite and the ISS are used in this study, as shown in Figure 2. These two datasets were selected for two main reasons: (a) their spatial resolutions are comparable—approximately 130 m for Luojia-1 and between 5 and 200 m for the ISS, depending on the specific image [15]—and both are substantially finer than that of VIIRS; and (b) the ISS images include color information, which is particularly valuable for light pollution studies. Urban lighting has increasingly transitioned from high-pressure sodium (HPS) lamps to light-emitting diodes (LEDs), with LEDs typically emitting a bluer spectrum compared to HPS [49]. However, current single-band nighttime satellite sensors, such as VIIRS-DNB, have limited sensitivity to wavelengths shorter than 500 nm [50], thereby underrepresenting blue light emissions and potentially underestimating NTLs [33]. As a result, multispectral imagery, such as that from the ISS, provides important advantages for investigating urban nighttime lighting conditions.

2.2.1. Luojia-1 Data

The Luojia-1 satellite, launched by Wuhan University in June 2018, provides calibrated nighttime imagery with a medium resolution of 130 m [51], while higher-resolution nighttime satellites, such as the commercial Jilin-1 with a 0.92-m ground sampling distance (GSD), are available, Luojia-1 stands out as the first satellite offering free-of-charge calibrated nighttime imagery, making it a valuable resource for night light analysis. The Luojia-1 satellite captures multiple snapshots within seconds, and we obtained all available cloudless image snapshots for Hong Kong (from http://59.175.109.173:8888/app/login_en.html, accessed on 15 June 2020). As shown in

Table 1. Available cloudless Luojia-1 images used in this study, their captured time, and the operating GaN-MN stations in the scene.

Date	Local Time	Operating Station
3 September 2018	22:44:02	HKU, HKn, TST, Col, Tai
	22:44:07	HKU, HKn, TST, Col, Tai
	22:44:12	HKU, HKn, TST, Col, Tai
	22:44:17	HKU, HKn, TST, Col, Tai
	22:44:22	HKU, HKn, TST, Col, Tai
	22:44:27	HKU, HKn, TST, Col, Tai
24 November 2018	22:46:52	HKU, HKn, TST, AP, iObs, Mac, Col, Tai
	22:46:57	HKU, HKn, TST, AP, iObs, Mac, Col, Tai
	22:47:02	HKU, HKn, TST, AP, iObs, Mac, Col, Tai
	22:47:07	HKU, HKn, TST, AP, iObs, Mac, Col, Tai
	22:47:12	HKU, HKn, TST, AP, iObs, Mac, Col, Tai
29 January 2019	22:56:32	HKU, HKn, TST, iObs, Mac, Col, Tai
	22:56:37	HKU, HKn, TST, iObs, Mac, Col, Tai
	22:56:42	HKU, HKn, TST, iObs, Mac, Col, Tai
	22:56:47	HKU, HKn, TST, iObs, Mac, Col, Tai
	22:56:52	HKU, HKn, TST, iObs, Mac, Col, Tai
	22:56:57	HKn, iObs
11 March 2019	22:56:18	HKU, HKn, TST, iObs, Mac
	22:56:23	HKU, HKn, TST, iObs, Mac

, four sets of images were used in this study, captured on 3 September 2018, 24 November 2018, 29 January 2019, and 11 March 2019. We also list the operating GaN-MN stations (to be discussed in Section 2.3) in the table.

2.2.2. ISS Imagery

ISS images are captured by astronauts using commercial digital single-lens reflex (DSLR) cameras onboard the ISS and have been utilized in several studies to investigate NTLs [19,34,52,53]. These images consist of three RGB channels but are not calibrated. Unlike most satellite systems, the local pass time of the ISS is not fixed, as the ISS orbits the Earth approximately every 90 min. A total of four image sets were retrieved (two from 2015, one from 2018, and one from 2020), as shown in

Table 2. Available ISS images and the corresponding georeferencing error evaluated by the third-order polynomial interpolation in terms of RMSE. The last column lists the operating in situ stations in the scene.

Date	Local Time	RMSE (Degree)	Operating Station
19 January 2015	0:57:46+1	0.00108	HKU, TST, Cap
	0:57:47+1	0.00118	HKU, TST, Ap, Cap
	0:57:58+1	0.00159	HKU, TST, Cap
	0:58:01+1	0.00102	HKU, TST, Ap, Cap
	0:58:04+1	0.00123	HKU, TST, Ap, Cap
	0:58:08+1	0.00135	HKU, TST, Ap, Cap
	0:58:25+1	0.00141	HKU, TST, Ap, Cap
23 January 2015	23:07:22	0.00156	HKU, TST, Ap, Cap
	23:07:26	0.00132	HKU, TST, Ap, Cap
	23:07:29	0.00177	HKU, TST, Ap, Cap
	23:07:31	0.00127	HKU, TST, Ap, Cap
	23:07:34	0.00103	HKU, TST, Ap, Cap
	23:07:39	0.00119	HKU, TST, Ap, Cap
	23:07:43	0.00141	HKU, TST, Ap, Cap
	23:07:55	0.00077	HKU, TST, Ap, Cap
	23:08:02	0.00063	HKU, TST, Ap, Cap
	23:08:05	0.00077	HKU, TST, Ap, Cap
	23:08:11	0.00078	HKU, TST, Ap, Cap
	23:08:15	0.00060	HKU, TST, Ap, Cap
	23:08:19	0.00106	HKU, TST, Ap, Cap
28 February 2018	4:35:42+1	0.00150	HKU, HKn, TST, iObs, Mac, Col, Tai
	4:35:47+1	0.00146	HKU, HKn, TST, iObs, Mac, Col, Tai
	4:35:53+1	0.00100	HKU, HKn, TST, iObs, Mac, Col, Tai
	4:35:57+1	0.00119	HKU, HKn, TST, iObs, Mac, Col, Tai
	4:36:03+1	0.00121	HKU, HKn, TST, iObs, Mac, Col, Tai
	4:36:05+1	0.00140	HKU, HKn, TST, iObs, Mac, Col, Tai
26 February 2020	3:32:36+1	0.00128	HKU, HKn, TST, AP, iObs, KP, Mac, Col
	3:32:37+1	0.00149	HKU, HKn, TST, AP, iObs, KP, Mac, Col

. Similar to the Luojia-1 images, the ISS images consist of multiple snapshots, though these snapshots differ slightly as the ISS orbits while the images are captured. For the correlation analysis, we used all available image snapshots, and for the land use analysis, we selected the clearest image.

We had a list of available nighttime ISS images with thumbnail previews available. We first conducted a visual inspection and selected candidate images from a large pool of these ISS images. As Hong Kong has a unique nighttime footprint, this step helped filter out a large pool of candidates. We then obtained the raw ISS image files (in 16-bit) in NEF format and converted them to TIF format for visualization. Some candidate images were filtered out due to low image quality in terms of clouds or noise. Additionally, we cross-checked clouds by visualizing Himawari-8 geostationary satellite imagery within the hour of observation, which shows largely cloud-free conditions within the study area (see Supplementary Materials).

After these steps, we conducted georeferencing using ArcGIS 10.4 on the final candidate images. By selecting twenty to forty control points depending on the image, with high-resolution optical images and the OpenStreetMap road network as references, we obtained the georeferenced images using third-order polynomial interpolation—an interpolation chosen purposely to allow for uncertainty over control points. The obtained root mean square errors (RMSEs) are shown in Table 2. They are roughly between 0.00060^∘ and 0.00177^∘ (about 66–196 m). To reduce error and standardize comparison, the ISS images were exported to a spatial resolution of 0.001^∘ (about 111 m) using the bilinear interpolation algorithm.

2.3. Field Measurements

The Globe at Night - Sky Brightness Monitoring Network (GaN-MN) was launched in 2014, aiming to record the skyglow brightness at an interval of 30 s to minutes. The network uses SQMs to measure the zenith night sky brightness (NSB). As of October 2025, the network has more than 90 stations around the world. Twelve stations are located in Hong Kong (seven were available for dates when the studied remote sensing images were captured), and an additional three are located in Macau, a city closely connected to Hong Kong (Figure 1) [4]. We use available sunlight- and moonlight-free data collected from these stations to compare with remote sensing data. Details and characteristics of the Hong Kong stations are listed in

Table 3. Station details and characteristics.

Station	Location	Lat	Lon	Ele	Characteristics
HKU	HKU Observatory Dome	22.28	114.14	100	Urban. Situated on an urban campus in a residential neighborhood. Within a 1.5 km circle, major land use classes: institutional open space, private residential, roads and transport facilities, and natural vegetation.
TST	Hong Kong Space Museum	22.29	114.17	10	Urban. Situated in a planetarium in commercial neighborhood. Within 1.5 km major land use classes: commercial, other urban (park), roads and transport facilities, and water bodies.
KP	King’s Park Meteorological Station, Hong Kong Observatory	22.31	114.17	65	Urban. Situated in urban park in mixed residential/commercial neighborhood. Within 1.5 km major land use classes: other urban (park), institutional open space, commercial, and roads and transport facilities.
HKn	Ho Koon Nature Education cum Astronomical Centre (Sponsored by Sik Sik Yuen)	22.38	114.11	140	Situated between urban areas and mountains. Major land use classes: natural vegetation, institutional open space, and closest station to port facilities.
Cap	The Swire Institute of Marine, HKU	22.21	114.26	5	Situated in a research facility far away from built environment. Major land use: natural vegetation.
iObs	iObservatory, Hong Kong Space Museum	22.41	114.32	80	Situated in an institutional facility far away from built environment. The 2nd farthest from urban areas. Major land use: natural vegetation.
AP	Astropark, Hong Kong Space Museum	22.38	114.34	5	Situated in an institutional facility far away from built environment. The farthest from urban areas. Major land use: natural vegetation.
Mac	Macau	22.19	113.54	30	Macau urban center. Major land use classes: Mixed Commercial (Casinos and hotels), Private Residential, Public Residential.
Tai	Taipa	22.15	113.57	40	Macau new town. Major land use class: commercial (casinos and hotels).
Col	Coloane	22.11	113.56	10	Major land use classes: natural vegetation, rural settlement.

Note: Lat/Lon in decimal degrees; Ele: Approximate elevation (units: meters); Lat/Lon in decimal degrees. Explore these stations at http://globeatnight-network.org/map.html, accessed on 15 June 2020.

. For more information, please refer to the GaN-MN website (http://globeatnight-network.org/global-at-night-monitoring-network.html, accessed on 15 June 2020).

The NSB value is in logarithmic scale, and a larger value indicates a darker sky, and vice versa. We show some sample data collected on the same dates as the images were recorded in Figure 3. Some data were discarded due to moonlight, sunlight, etc. When it is cloudy, the backscattered ALAN from the sky and clouds becomes strong, increasing the measured brightness (a smaller number) [54]. In a cloudless condition, the brightness is relatively stable over time, and its temporal changes mainly reflect the changes in the local lighting environment [34]. Among the eight dates, the 11 March 2019 NSB value was quite stable. At around 23:00, the NSB value increased (darkened) sharply, because most decorative lights in the city were turned off at 23:00, as recommended in the Charter on External Lighting (https://www.charteronexternallighting.gov.hk/en/index.html, accessed on 15 June 2020).

2.4. Land Use Data

To analyze night lights in different city zones, we obtained the Hong Kong land use map in 2018 from the Planning Department (https://www.pland.gov.hk/pland_en/info_serv/open_data/landu/index.html, accessed on 15 June 2020). The original land use data are organized as a two-level classification system. A total of 10 classes with 27 subclasses are available. We grouped them into 14 categories as shown in Figure 4 and

Table 4. Hong Kong land use data and their corresponding pixel numbers (#). The original classes are defined in the map of Land Utilization 2018. The first ten classes are categorized as urban classes, and the last four are natural classes.

Class	#Pixel	Original Class
Private Residential	2291	Private Residential
Public Residential	1478	Public Residential
Rural Settlement	2998	Rural Settlement
Commercial	389	Commercial
Industrial	2291	Industrial Land
		Industrial Estates/SciTech Parks
		Warehouse and Open Storage
Institutional Open Space	4742	Government, Institutional and Community Facilities
		Open Space and Recreation
Roads and Railways	4501	Roads and Transport Facilities
		Railways
Airport	1127	Airport
Port Facilities	404	Port Facilities
Other Urban	3900	Cemeteries/Funeral Facilities
		Utilities
		Vacant Land/Construction in Progress
		Others
Agriculture	5862	Agricultural Land
		Fish Ponds/Gel Wais
Natural Vegetation	64,174	Woodland
		Shrubland
		Grassland
		Mangrove/Swamp
Barren	533	Badland
		Rocky Shore
Water Bodies	2692	Reservoirs
		Streams and Nullahs

. As our study concerns human-perceived night lights in cities, we used subclasses in residential and commercial areas, and grouped the natural classes. We also grouped two subclasses—roads and railways—but kept the airport and port facilities in the transportation category separated. The land use pattern was stable in Hong Kong from 2015 to 2020, and land use changes were negligible.

3. Methodology

3.1. Luojia-1 Data Calibration

To calibrate the Luojia-1 data, we covert the raw digital number (DN) values to at-sensor radiance using the following equation [51]:

$$L={10}^{-10}{d}^{3/2}\mathrm{W}/({\mathrm{m}}^2 · \mathrm{sr}·\mathsf{\mu }\mathrm{m}),$$

(1)

where L and d are at-sensor radiance and raw DN value, respectively. The radiance data are then multiplied by ${10}^5$ in our study so that the data range is roughly 0–10,000.

3.2. Luojia-1 Satellite’s Position

The received radiance is influenced by the elevation angle $\theta$, whereas the satellite view is affected by the azimuth angle $\alpha$ (Figure 5). As multiple snapshots exist, these angles play an unneglectable role in recording the night light data [55]. We here denote the satellite height from the horizontal plane (sea level) as H and the projected point of the satellite on the horizontal plane as $(x,y)$. Then, assuming a flat-land case, for a given point (ground target) on the plane $(x_i,y_i)$, its elevation angle ${\theta }_i$ to the satellite is formulated as follows:

$${\theta }_i=arctan{\displaystyle \frac{H}{d_i}},$$

(2)

where $d_i$ is the distance between $(x,y)$ and $(x_i,y_i)$ is as follows:

$$d_i=\sqrt{{(x-x_i)}^2+{(y-y_i)}^2}.$$

(3)

In the header, the satellite azimuth $\alpha$ and elevation $\theta$ to the center point of the image are given. We here denote the center point as $(x_c,y_c)$. Then, the projected point $(x,y)$ of satellite on the horizontal plane is calculated by solving the following equations:

$${(x-x_c)}^2+{(y-y_c)}^2={\left({\displaystyle \frac{H}{tan\theta }}\right)}^2,$$

(4)

$$tan\alpha ={\displaystyle \frac{y-y_c}{x-x_c}}.$$

(5)

For convenience, we denote $x-x_c$ as $\delta$ and $tan\alpha$ as a; the formula becomes the following:

$${\delta }^2+{\left(a\delta \right)}^2={\left({\displaystyle \frac{H}{tan\theta }}\right)}^2,$$

(6)

$$\delta =\sqrt{{\displaystyle \frac{H^2}{(1+a^2){tan}^2\theta }}}={\displaystyle \frac{H}{tan\theta \sqrt{1+a^2}}}.$$

(7)

We can now obtain the projected point of the satellite $(x,y)$ as follows:

$$x=\delta +x_c,$$

(8)

$$y=a\delta +y_c.$$

(9)

After obtaining the projected satellite location on the horizontal plane, we can calculate individual elevation ${\theta }_i$, azimuth ${\alpha }_i$, and the distance ${\tau }_i$ from a location i to the satellite as follows:

$${\tau }_i=\sqrt{d_i^2+H^2},$$

(10)

$${\theta }_i=arctan{\displaystyle \frac{H}{d_i}},$$

(11)

$${\alpha }_i=arctan{\displaystyle \frac{y-y_c}{x-x_c}}.$$

(12)

Calculating the satellite’s position is beneficial to understanding the anisotropic characteristic of night lights [55]. This becomes important as we have multiple snapshots within seconds on a single date.

3.3. Pearson R Correlation Coefficient

We use the Pearson correlation coefficient to analyze the relationship between satellite observed night lights and ground NSB values. The typical NSB value observed in Hong Kong is between 14 and 20 mag/arcsec² and is in logarithmic scale. Therefore, we converted the satellite’s readings in the linear unit (radiance in remote sensing) to the logarithmic unit (magnitude) before analyzing their correlation.

In the analysis, both the calibrated Luojia-1 and raw ISS images (16-bit) are in linear scale. Denote their values as x, and their correlation coefficient with the NSB values is calculated as follows:

$$R={\displaystyle \frac{1}{N-1}}\sum _{i=1}^N\left({\displaystyle \frac{\overline{{\widehat{x}}_i-{\mu }_{\widehat{x}}}}{{\sigma }_{\widehat{x}}}}\right)\left({\displaystyle \frac{y_i-{\mu }_y}{{\sigma }_y}}\right),$$

(13)

where,

$$\widehat{x}=lo{g}_{10}\left(x\right),$$

(14)

where ${\mu }_{\widehat{x}}$ and ${\sigma }_{\widehat{x}}$ are the mean and standard deviation of the logarithmic scale satellite observations, respectively, and ${\mu }_y$ and ${\sigma }_y$ are the mean and standard deviation of NSB values, respectively.

4. Results and Analysis

4.1. Correlation Analysis

4.1.1. Multiple Snapshots and Satellite Locations

Figure 6 shows the Luojia-1 satellite’s positions when it was capturing the image snapshots. Take the data taken in September as an example. Although the images mainly cover Hong Kong and the Pearl River Delta, the satellite was to the southwest of the captured area near the Hainan Island (>700 km), going from south to north. The satellite was about 647 km high, and the path from the University of Hong Kong (HKU station) to the satellite is about 774 km. The azimuth changed from southwest 36.28^∘ to 61.41^∘ (from the position ’0’ to the position ’5’ in Figure 6). The difference in azimuth angle will lead to different light transmission paths. As Hong Kong is a highly populated city with many high-rise buildings, the angle effect is unneglectable [30].

To analyze the consistency of Luojia-1 multiple snapshots in a short period, we resampled these Luojia-1 images to 0.001^∘ (approximately 111 m) and conducted a scatter correlation analysis. The scatter plots are shown in Figure 7. We used two image snapshots taken within 10 s from November as an example. It should be safe to assume that ground lighting was unchanged within 10 s. We found that the correlation in urban areas (direct lights) is better than that in the natural environment (diffuse lights). However, even in the bright urban regions, the consistency is not exactly perfect, but R = 0.96 within 10 s.

4.1.2. Correlation Between the Luojia-1 Images and Field Measurements

After analyzing multiple snapshots, we show the scatter plot between the calibrated Luojia-1 images and the field measurements from four different dates in Figure 8. As multiple snapshots exist, they are all used in the analysis. For the Luojia-1 data, we extracted the mean value from an 11 × 11 window (i.e., 1100 × 1100 m) of the centered pixel of the station. The mean standard deviation of the satellite data of these stations is 15.47%, or 10.14% if we exclude the three suburban stations (AP, iObs, and Col). In our analyses, we also tested for 1 × 1 window, 5 × 5 windows and with Gaussian kernels, their results were worse than 11 × 11.

For the NSB data, we used the mean of 5 min records from the network. The scatter plot is shown in Figure 8. The R coefficient of this plot is 0.7282 ($R^2=$ 0.5302, p < 0.0001; we also conducted an analysis based on a 1 × 1 window and a 5 × 5 mean window, and the achieved R correlation coefficients are 0.6781 and 0.7176, respectively). Note that as the NSB values are in magnitude and reverse order, we use the negative values to calculate the coefficient. The obtained result is acceptable, considering in some images some clouds appeared in the central areas near the harbor (Figure 3g,h) and the images are slightly blurred. The result indicates that there is a positive linear correlation between the NSB values and remote sensing night lights: brighter in ground measurements leads to brighter remote sensing images. However, some deviations exist. For example, the brightest in satellite images is the Tai station, but it is not the brightest in our network. The ground data indicate that the TST station is the brightest among Hong Kong and Macau.

4.1.3. Correlation Between ISS Images and Field Measurements

In this section, we analyze the relationship between NSB values and ISS imagery. The correlation plots of the four dates’ data in a channel-wise fashion are shown in Figure 9. The plots are obtained using the mean value of an 11 × 11 (i.e., 1100 × 1100 m) window from the best image snapshot. The first to the fourth row shows the results from the ISS-20150119, ISS-20150123, ISS-20180228, and ISS-20200226 data, respectively. A more detailed table including all obtained correlation coefficients using a single pixel and a 5 × 5 window (i.e., 110 × 110 m and 550 × 550 m, respectively) from other image snapshots of the same date is shown in

Table 5. Summary of R correlation coefficients with different channels and integral windows across multiple snapshots (# denotes the number of snapshots).

		11 × 11				5 × 5	1 × 1
Date	#	Red	Green	Blue	Average	Average	Average
19 January 2015	7
23 January 2015	13
28 February 2018	6
26 February 2020	2

.

From Figure 9, the obtained coefficients are between 0.85 and 0.99, indicating a high linear correlation between the NSB values and remote sensing night lights. For all data from four individual dates, the highest correlation is achieved at the blue channel and the lowest at the red channel.

To test the robustness of the correlation, we also analyze image snapshots captured on the same day, as shown in Table 5. The obtained correlation coefficients are all higher than 0.81 for the 28 image snapshots. For the blue channel, the coefficients are between 0.91 and 0.9984, the highest among the three channels. The coefficients for the red and green channels are between 0.81 and 0.96 and between 0.85 and 0.97. The obtained coefficients via a 5 × 5 mean window and 1 × 1 mean window are similar to (but slightly lower than) the ones via an 11 × 11 mean window.

Finally, we use all available ISS images from the four dates to conduct a correlation analysis. Note we should treat such comparisons carefully, as these ISS images are not calibrated. The correlation plots in a channel-wise fashion are shown in Figure 10. Even though the ISS images are not calibrated, an excellent linear correlation can still be achieved with the multitemporal data (R > 0.81).

4.2. Analysis with City Zones

4.2.1. Land Use Brightness Ranking Based on Land Use Data

The brightness rankings of land use classes based on the ISS images and Luojia-1 images are shown in Figure 11. Both datasets have four available dates for analysis. For the ISS images, two were captured in 2015 at 0:58⁺¹ and 23:08 local time, one was captured in 2018 at 4:36⁺¹ local time, and another one was captured in 2020 at 3:32⁺¹ local time. For the Luojia-1 images, they were captured between 22:44 and 22:56 local time and between September 2018 and March 2019. For reference purposes, we show the land use map with only residential and commercial areas in Figure 12 with sample nighttime photos in Figure 12. The complete land use map can be found in Figure 4.

We first analyze the ranking based on ISS images. It is consistent across the four-date ISS images: the class of port facilities ranks first, and barren ranks last. Within urban classes, public residential areas, and roads and railways are brighter than private residential areas throughout the night. Land use types that do not have nighttime activities, such as the institutional or open spaces, industrial areas, and other built-up areas, are darker. Rural settlements are the darkest of the urban classes. For the natural classes, agricultural areas are the brightest, followed by natural vegetation, water bodies, and barren. Commercial areas are brighter in early night (23:08, brighter than the airport) than late night after midnight (darker than the airport). This indicates that the brightness of ALAN changes with time in different city zones. In downtown early night, the city is lit up by human activities. With most people falling asleep and the decorative lighting near both sides of the Victoria Harbor turned off, commercial areas have almost the same brightness as residential areas.

The rankings based on the Luojia-1 images are slightly different from the ISS images. First, of all, as the Luojia-1 images were captured before 23:00, it is reasonable that commercial areas are the brightest in all classes, followed by port facilities. It is interesting that the airport is less bright in the Luojia-1 images and is even darker than roads and railways, which is inconsistent with the ISS images. As for residential areas, the brightness of public and private residential areas is similar. The rankings of the remaining land use classes are consistent with the ISS images.

Based on both remote sensing data sources, we found that commercial areas are brighter in early night, and port facilities are brighter in late night. Industrial areas, open space, and other build-up areas are the darkest within city zones. The natural classes are the darkest as expected. However, some inconsistency between the ISS images and Luojia-1 images should be emphasized. (1) The airport is brighter in ISS images and darker in Luojia-1 images. (2) Public residential areas are brighter than private residential areas in the ISS images, but they have similar brightness on the Luojia-1 images. The difference is likely due to their respective spectrum response functions as follows: Luojia-1 has a spectral response from about 500 to 900 nm [51], and ISS images were captured from Nikon D4 and D5 that have a spectral response that is closer to human eyes [19,56].

4.2.2. Private and Public Residential Areas

Private residential areas defined in the land use data include two types of buildings in Hong Kong: (1) typical residential areas such as houses in the west, villas, and condos; and (2) mixed-use residential buildings with ground-floor shopping like tong lau in Hong Kong and East Asia. From the nighttime photo of tong lau in Figure 12, it is apparent that the mixed-use private residential areas exhibit different external lighting features compared to typical residential areas (though mixed-use buildings are typical and common in Hong Kong). To further distinguish their difference, we selected some private residential areas we know for sure are not mixed-use (as shown in Figure 12), and then show the night light distributions of public, private, and the selected private residential areas in Figure 13. For simplicity, we show two results with ISS images and two with Luojia-1 images. Although the average brightness of private residential areas in the Luojia-1 images is brighter than public residential areas, the deviation within the former is larger as well, indicating that private residential areas are more heterogeneous. However, the selected private residential areas (without mixed-use buildings) are significantly darker with a lower deviation. Therefore, the brightness of private residential areas should be interpreted carefully in two parts. The selected private residential areas (pure residential) are darker than the mixed-use private residential areas and the public ones. On the other hand, the mixed-use private residential areas, e.g., tong lau (relatively poor living environment in Hong Kong [57]), receive more artificial lights.

Another interesting phenomenon is that the Luojia-1 data successfully captured the lighting changes in commercial and private residential areas (with mixed-use buildings) before and after external lighting was turned off. The Hong Kong Government has a task force on external lighting and recommends switching off external lighting (typically in commercial areas) after 23:00, following the switching-off recommendation outlined in the Charter on External Lighting. The turn-off time of individual buildings is not precisely 23:00 as measured from Figure 3, because some stakeholders enforce such action slightly before 23:00 (time alignment issues and other reasons). From Figure 13c,d observed by the Luojia-1 data at two different local times (22:44 and 22:56), it is apparent that commercial and private residential areas are significantly darker at 22:56 than at 22:44, e.g., the mean value of commercial areas dropped from 1505 to 813. It shows the effectiveness of the action, and also challenges remote sensing on how to monitor multitemporal night lights within one night.

4.2.3. The Airport

We show the normalized images of the airport in Figure 14, and of the downtown around the Victoria Harbor in Figure 15. Their locations are shown as blue rectangles in Figure 1. The image center of the latter is the major commercial areas in Hong Kong (Central and Tsim Sha Tsui), and the upper left is the Kwai Tsing Container Terminals.

Luojia-1 images and ISS images exhibit different features in the two areas. In the airport (Figure 14), runways are significantly darker in Luojia-1 images than in ISS images. In the Luojia-1 images, runways emit less than 10% of the light than the brightest spot of the airport; in the ISS images, runways show 40% to 60% brightness compared to the brightest spot. Based on the histograms (Figure 14), the ISS-20150123 image is saturated in the red and green channels, which is one of the reasons why the airport is brighter in the ISS images. The blue channel of the ISS-20150123 image is not saturated, and the locations of bright spots are similar to the Luojia-1 images at the lower right part of the airport (Figure 14).

As of the ISS-20200226 image captured at 3:32⁺¹ local time, all three channels are not saturated, but still, the airport is much brighter than Luojia-1 images. One possible reason may be that the ISS images are not with such high resolution (0.001^∘). The actual GSD may be three or five times larger, as the hot spot in ISS images is larger than that in the Luojia-1 images, and some details are missing. Therefore, an area of 5 × 5 pixels (i.e., 550 × 550 m) may be the same light source, making the ISS images much more averaged and blurred. If we calculate the average brightness in the airport, a lot of dark regions are included, and therefore the mean value will be small. This can be seen as a result of the modifiable areal unit problem [58,59]. We get different statistics from the two sets of images with different true spatial units.

An interesting observation is that light centers of the airport shift with time. In early night, the brightest regions are on the southeast of the airport. They are the passenger terminals and the roads to terminals. Some spotlights and LED signboards are located near the road. Late at night, after midnight, the brightest region shifts to the southwest of the airport, where the airfreight terminals are located. The Hong Kong International Airport is a busy transportation hub with more than 70 million passenger traffic and 5 million metric tons of cargo traffic in 2018 [60]. The observed light center shifting agrees with the fact that most cargo flights operate in late night [61].

4.2.4. Downtown: Commercial Areas and Port Facilities

We show the normalized nighttime images in the downtown around the Victoria Harbor and the corresponding daytime image in Figure 15, where the upper left is the Kwai Tsing Container Terminals. In the Luojia-1 images, the central areas are so bright that other regions are relatively dark. On the other hand, the ISS-20150123 image is bright in all regions and is saturated in all channels (Figure 15). In the second row of Figure 15, only the ISS image before midnight is very bright in the central commercial areas. For images after midnight, the central commercial areas have a similar brightness to residential areas. The night lights in central areas are bluer than the surrounding environment, as shown in the third row of Figure 15. The fourth row of Figure 15 indicates that in late night, central commercial areas and other regions have a similar brightness. The light center shifted from the central areas to the container terminals.

Based on the Luojia-1 and ISS images before and after midnight, we can see that the lighting situation changes with time. Central commercial areas are brighter in early night, and the container terminals are brighter in late night. The airport is brighter at the passenger terminals and the roads with spotlights and LED signboards in early night, and is brighter at the airfreight terminals in late night. The lighting difference among times reminds us to use the nighttime satellite data wisely, as most of them only reflect a snapshot of cities’ nighttime profile at a particular local time.

4.2.5. Color Difference Within City Zones

The ISS imagery is captured by commercial DSLR cameras, and therefore provides us a valuable source to analyze color effects. We further divide the nightlights based on land use types and generate average nightlights in red and blue channels, as shown in Figure 16. The ratio between red and blue values is also plotted and overlapped on the color bar. We show four images as follows: two from relatively early night (near midnight at 23:08 and 00:58⁺¹ local time) and two from late night (at 3:32⁺¹ and 4:35⁺¹ local time). Of the four images, port facilities are the brightest in the red channel. But in the blue channel, commercial areas, for example, are the brightest before midnight in 2015. In Hong Kong, external lighting facilities are still dominated by the HPS type, as they were installed before the popularity of LEDs [49]. The public and private residential areas show similar lighting patterns and are bluer than commercial areas. Another interesting phenomenon is that the ratio between red and blue lights for the airport becomes less red and bluer in 2020, implying that a certain transformation of lighting technology might undergo.

5. Discussion and Conclusions

5.1. Findings and Strengths

In this paper, we presented a comprehensive analysis of remotely sensed nighttime lights (NTLs) using multi-temporal imagery from the Luojia-1 and International Space Station (ISS), two medium-resolution satellite data sources. We first tested the correlations between remotely sensed NTL data and night sky brightness (NSB) measured on the ground. Our results show moderate to strong correlations, with the Luojia-1 data over Hong Kong and Macau from four different dates in 2018–2019 yielding a correlation coefficient of $R=0.73$. For individual ISS imagery in 2015–2020, correlation coefficients ranged from 0.8 to 1.0. We observed notable differences in correlation by color channel, with the best correlations found in the blue channel and the weakest in the red channel.

Our further analysis highlights that NTL data changes throughout the night. In the early hours (before midnight), commercial areas are the brightest land-use category. These areas also appear bluer compared to other land-use types. After midnight, port facilities and airports become the brightest as commercial lighting diminishes. These areas also appear redder than other zones in the city. We also observe considerable heterogeneity in NTL across different residential areas, with mixed-use buildings (tong lau) being the brightest, followed by public housing. Private residential areas, associated with higher incomes, exhibited the least exposure to excessive lighting.

The usage of multi-temporal imagery from various local times makes this study possible, as existing sun-synchronous satellites can only obtain data at similar local times (2 a.m., for example). The comparison of two data sources, from their consistencies and differences, expands our understanding of NTL variations. The differences that occur among images obtained even seconds apart may expand considerable discussion and further analysis arising from this study.

5.2. Further Considerations and Limitations

There are a lot of factors that can affect the NTL values (

Table 6. Major factors affecting observed NTL values.

Factor	Effects
Sensor
Spectral response	Spectral response is a sensor’s specific wavelength detection; it impacts NTLs by filtering bands, causing brightness overestimation, or underestimation.
Signal-to-noise (SNR) ratio	A higher SNR reduces random noise interference, leading to more stable and accurate NTL brightness measurements.
Low light sensitivity	Sensors with higher low light sensitivity can capture dim NTL signals (e.g., rural areas) more effectively.
Saturation	When NTL brightness exceeds the sensor’s saturation threshold, the signal is clipped, resulting in underestimated brightness for high-light areas (e.g., city centers).
Aging	Mainly reduces sensitivity, raises noise, shifts spectrum, and biases long-term NTL data; not considered in most analyses.
False Changes (artifacts)
Clouds	Clouds block NTL signals, leading to artificially reduced brightness in covered areas.
Lunar interference	Moonlight can increase background brightness.
Real Changes (condition)
Local time	NTL brightness varies with local time (e.g., higher at night when lights are on, lower in early morning).
Atmospheric conditions	Even without clouds, atmospheric conditions affect scattering, leading to NTL changes (e.g., humidity is known to be associated with NTLs).
Observational angle	Mountainous and tall buildings can block NTL, leading to observed target change; also change the path length of NTL signals through the atmosphere, causing brightness underestimation.
Real Changes (source)
Addition	New light sources (e.g., new buildings, streetlights) increase local NTL brightness, reflecting regional development or expansion.
Replacement	Replacing old light sources (e.g., incandescent with LED) changes brightness intensity or spectral characteristics, altering measured NTL values.
Removal	Removing light sources (e.g., demolished buildings, turned-off streetlights) decreases NTL brightness, indicating land-use changes or reduced human activity.

). First, with the addition and removal of light sources, NTL values change as a reflection of real changes—the fundamental assumption of NTL data. However, real changes can also occur with the replacement of light sources; for example, switching from incandescent bulbs to LEDs. This may represent a real brightness change or a change in spectrum [49,62,63,64]. It remains an open question what the gold standard in brightness measurement should be: the value from a lux meter, the value closest to human perception, or something else? Most satellite sensors have a spectral response different from that of human eyes, and as of today, the debate continues. A standard, however, should be reached within the research community in the era of color remote sensing of NTLs [19,36,63].

Observation angle is another known yet complex factor. The directional effects of NTL call for additional angular correction [30,31]. But unlike daytime remote sensing, NTL exhibits more complicated directional effects, especially in cities [65]. In our study, we have detailed how the observation angle has changed significantly along satellite trajectory. Yet, due to the medium-resolution nature, and the lack of 3-D building data to properly model the urban canopy, we did not provide quantitative analysis. In the current literature, so far, although specific corrections and investigations have been discussed and specially handled in several studies [66], a universal solution for large-scale application is still pending in NTL research due to the complexity of this factor.

Some other known factors, often too complex to discuss in detail, include moonlight and atmospheric conditions such as PM2.5 and humidity that do not block NTL transmission [5,67,68]. These factors, in addition to the changing observation angle across snapshots taken within seconds, contribute to the imperfect correlations between such snapshots. As we observe better agreement in urban environments ($R=0.96$) than in natural environments ($R=0.88$), natural factors likely dominate.

5.3. ISS Imagery for Artificial Light Research

The ISS image repository has long been considered a valuable source for artificial light research [69], but it has not been used in practice until recent years [19,20,53,70]. One of the major concerns is its lack of radiometric calibration. In this study, we show that even without radiometric calibration, the ISS images still present high potential for artificial light research in a relative comparison fashion. Within an ISS image, the color comparison can be done relatively, and we saw a consistent, reasonable trend from land use lighting patterns in Hong Kong. Even so, in a cross-scene validation, the agreement between NSB and ISS is on par with the agreement between NSB and Luojia-1 (a calibrated data source). When calibrated data are not available, directly using the uncalibrated ISS images can be a starting solution to further our understanding of artificial light in cities.

We should note the limitations coming from using uncalibrated ISS images. First, as the ISS images were uncalibrated, their direct cross-comparison should be treated carefully. Second, there are also some other additional considerations, including flat-field correction and spectral response. In our analysis, as the study area is relatively small compared to the entire image, and we have georeferenced the image data carefully, factors related geometric correction remain but should be largely handled. The ISS images we used were captured by Nikon D4 and D5, and it is reasonable to assume that their spectral response should be similar [19]. But the spectral response is different with SQM [71] and Luojia-1 [51]. Yet, as the ISS-NSB agreement is on par with and even better than the agreement between Luojia-1 and NSB, the limitations mentioned here may not be as critical as we thought. Nonetheless, using uncalibrated ISS images is sufficient for our study that is large scale and across multiple years, but for a strictly rigorous and granular study, these limitations should be handled properly.

5.4. Conclusions and Future Directions

In conclusion, our study provides a comprehensive analysis of multi-temporal NTL changes over the night as observed from medium-resolution satellite data. We emphasize the importance of using multi-temporal data to analyze changing NTL patterns in cities, particularly in Hong Kong, one of the most light-polluted cities globally. By analyzing NTL data across different times of the night, we identified significant variations in NTL across land use categories, especially commercial areas. As NTL data are influenced by socioeconomic factors, physical characteristics, sensor parameters (e.g., spectrum and spatial resolution), and the changing position of satellites, conclusions drawn from these analyses should be interpreted with caution. Future studies exploring the scale, color, and potential multi-angle analysis on the same night would contribute significantly to understanding NTL sources and their broader socioeconomic implications.