THE EXTRACTION OF URBANIZED AREAS THROUGH IMAGES OF HIGH RESOLUTION NIGHTTIME LIGHTS

Satellite nocturnal images of the earth are a useful way to identify urbanisation. Nighttime lights have been used in a variety of scientific contributions, including studies on the identification of metropolitan areas as well as landscapes impacted by urbanization. However, the study of urban systems by nighttime light imagery has had a fundamental limitation to date: the low spatial resolution of satellite sensors. Although the DMSP Operational Linescan System (OLS) has been gathering global low-light imaging data for over 40 years, its 2.7 km/pixel footprint has limited its use for in-depth studies of urban development. The 2011 launch by NASA and the NOAA of the Suomi National Polar Partnership (SNPP) satellite, with the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor on board, has led to a significant improvement. This instrument has better spatial resolution (742 m/pixel), on-board calibration, a greater radiometric range, and fewer saturation and blooming problems than DMSP-OLS data. However, it still has considerable limitations for the in-depth study of the area and internal structure of urban systems. The launch of Luojia 1-01 in June 2018 has increased expectations. LJ1-01 is a nano satellite that can obtain high-resolution nocturnal images (130 metres/pixel). The aim of this paper is to analyse, and compare with previous satellites, the new instrument’s capacity to delimit the urbanised area and its efficiency in identifying types of urban landscape (compact, dispersed and rurban). The study cases are Barcelona Metropolitan Region (Spain) and Shenzhen City (China). * Corresponding author


INTRODUCTION
Since the mid-twentieth century there has been a true "explosion" of urbanization on a global scale. The urban population has grown from 750 million people in 1950 to 2,860 in 2000, more than 50% of the world population. In the Developed World, the urban model has suffered significant changes in recent decades, transforming from a model of urban continuum of medium and high densities to a model of an endless diffuse and sprawled city, driven by technological innovation processes, separation of functions and seeking proximity to nature. Therefore, since 1950 there has been a real reversal in the topology of the landscape. The process of urban sprawl has relegated to open spaces the role of auxiliary elements within the spatial structure. The sprawl in residential areas is linked to the gradual decentralisation of economic activity, first the industry, and then the services and even the most qualified tertiary activities. Urban sprawl, the massive consumption of land, can be found worldwide, although it takes many different forms in different regions and continents.
Satellite nocturnal images of the earth are a useful way to identify urbanisation (Elvidge et alt, 2001). Nighttime lights have been used in a range of scientific contributions, including studies the identification of megalopolises (Florida et alt., 2008) and urban landscapes (Arellano & Roca, 2016). This paper assumes that night-time lights satellite imagery provides valuable information for the identification of human landscapes, such as rural and urbanized areas. The "dark" landscapes are certainly related to rural settlements. The landscapes of light and darkness detect more clearly than traditional statistics based on the percentage of urban/rural forms of human settlement on the world population, with the advantage, in turn, of allow it to be studied on a subnational level, which is not possible when simply using official statistics. Concerning the "lit" landscapes, it clearly identifies areas of the world characterized by high human artificialisation. The electricity supply, along with the division of land into plots and the "lines" of the streets, represent the first steps in the process of urbanisation. The almost universal access to electrical energy as well as the diverse intensity of its use makes the analysis of night-time images an exceptional tool for studying the urbanisation gradient on a world scale.
However, the study of the extent and internal structure of urban systems by nighttime light imagery has had a fundamental limitation to date: the low spatial resolution of satellite sensors. Although the DMSP Operational Linescan System (OLS) has been gathering global low-light imaging data for over 40 years, its 2.7 km/pixel footprint has limited its use for in-depth studies of urban development. The 2011 launch by NASA and the NOAA of the Suomi National Polar Partnership (SNPP) satellite, with the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor on board, has led to a significant improvement. This instrument has better spatial resolution (742 m/pixel), onboard calibration, a greater radiometric range, and fewer saturation and blooming problems than DMSP-OLS data (Elvidge et alt., 2013). However, it still has considerable limitations for the in-depth study of the extent and internal structure of urban systems. The launch of Luojia 1-01 in June 2018 has increased expectations (Jiang et alt., 2018). LJ1-01 can obtain high-resolution nocturnal images (130 meters/pixel), with a dynamic range above 14 bits at night and a spectrum rate of 0.46-0.98 µm.
The aim of this paper is to analyse, and compare with previous satellites, the new instrument's capacity to delimit the urbanized area and its efficiency in identifying types of urbanized landscapes (compact, dispersed and rurban).
The study cases are Barcelona Metropolitan Area (MRB, 3,200 km 2 , 4.7 million inhabitants) and Shenzhen City (1,997.47 km 2 , 13.03 million inhabitants). The methodology used consisted of developing a series of models showing whether land has been urbanized according to information provided by Corine Land Cover (  Where r is the radiance value, DN is the digital number obtained by Luojia1-01, and w is the bandwidth, whose unit is Wm -2 sr -1 . The radiometric range of Luojia1-01 is 0.46-0.98 µm, so that w is equal to 0.52 µm.


In the rest of the sensors, the logarithmic transformation of the digital numbers (Yu et al, 2018) has been used as independent variables in the logistic regression models.  Identify urbanized areas through CLC and FROM-GLC (rural = 0; urban = 1).  Below is a set of logistic regressions (a model for each of the images analyzed) with urbanization (1-0) as a dependent variable and the intensity of the night light supplied by the different sensors as independent variables.  Once the logistic regression is carried out, analyze the % of successes, especially in the urbanized land, as well as other indicators of the goodness of the adjustment achieved.  Select all the urbanized land pixels as well as the rural ones correctly classified in the previous logistic regression, and, with that selection perform another logistic regression, repeating iteratively, until the % of urbanized land successes do not improve meaningfully.  Once the process is finished, it is necessary to "recover" the rural pixels discarded in the previous stages (due to being poorly classified) as well as calculate the cross table with all the pixels, definitively establishing the % of successes in the urbanized land and the rural land (overall accuracy). Check the statistical significance, as well as the "measure of agreement" Kappa, estimating the validity of each model developed.    SNPP-VIIRS, with a pixel depth of 32 bits, and a spatial resolution of 0.0041666667 degrees, obtains medium-high results to identify the extent of artificialized covers of the MRB, as reflected by the Kappa coefficient (0.578).

Figure 7. SNPP-VIIRS (MRB)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2020, 2020 XXIV ISPRS Congress (2020 edition) In the case of Luojia 1-01 (figure 8 and table 3), logistic regressions converge after 12 iterations, obtaining 87.8% of successes in rural land and 89.2% in urbanized land, with an overall accuracy of 88.08%. The degree of validity of the night light intensity obtained by means of the Luojia 1-01 sensor (with a 32-bit pixel depth and a spatial resolution of 130 meters/pixel) can be considered high, as demonstrated by Kappa coefficient (0.691) achieved as well as the rest of the indicators used.

Shenzhen City
Figures 9 to 10 show SNPP-VIIRS (2019) and Luojia 1-01 (2019) nighttime lights and the urbanized land according to FROM-GLC (2017) in Shenzhen City (Zheng, 2019). Tables 4  and 5 show the synthetic results (overall accuracy) of the iterative logistic regressions obtained for these sensors, since it has not been possible to obtain stable results with DMSP-OLS and Black Marble.  The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2020, 2020 XXIV ISPRS Congress (2020 edition)   Table 7. Adjustment accuracy indicators (Shenzhen) As in the case of the Metropolitan Region of Barcelona, in Shenzhen City Luojia 1-01 performed better in overall accuracy, and urbanized and rural land prediction and in the determination of land cover types. However, the results are worse than those obtained in the Metropolitan Region of Barcelona.
The lower goodness of fit of the regression models achieved in Shenzhen in relation with Barcelona is due to the lower accuracy of the FROM-GLC database comparing to Corine Land Cover (see Appendix). Particularly FROM-GLC does not accurately identify the urbanized land, considering only impervious surfaces. Urban green spaces are classified as rural covers, such as forest, grass or shrubs.