Introduction
Satellite imagery of nighttime luminosity has been used as a reliable proxy for national and sub-national economic activity since Henderson et al.’s seminal 2012 study. Luminosity is particularly useful as it is frequent (it is available monthly), timely (it is made available within a few weeks following the end of the month), and is granular (it can be used to measure economic activity in small regions).
We implement a new method of pre-processing nighttime satellite imagery using VIIRS data from NASA’s Black Marble project. This method involves removing ephemeral light, light reflected on water, and non-electric sources of light. As part of the World Bank’s Big Data Observatory program, we examine luminosity for Vietnam at the national and provincial level. We find that, compared to off the shelf VIIRS data from the Colorado School of Mines, our measures of luminosity are less volatile, more correlated with sub-national GDP, and show an explicit increase in luminosity over time.
Methodology
1. Data
The primary dataset used is monthly night time satellite imagery from NASA’s Black Marble project (product code: VNP46A3). We query this data via API using the NASA LAADS DAAC portal, and subsequently convert it from HDF to GeoTIFF format.
The data is available monthly from January 2012, and has a resolution of 15 arc seconds (approximately 500m × 500m at the equator).
We compare this data to monthly VIIRS data from the Colorado School of Mines, which has been processed and uploaded to the Big Data Observatory’s night lights portal: https://lights.nltglobal.com.
We also use provincial-level GDP figures for 2018 from the Ministry of Planning and Investment, the General Statistical Office, and some of the provinces’ information portals, as well as national GDP data (PPP, 2017 constant international $) from the World Bank.
2. Method
In order to process the NASA Black Marble (BM) VIIRS images, we conduct pixel-level pre-processing. This involves discarding pixels that are the most affected by cloud-cover and off-nadir viewing observations. We then impute values for missing pixels using a linear model based on the previous 12 months of luminosity values for each missing pixel.
In addition, we discard pixels that fall on water bodies (as water reflects light and can provide false luminosity observations), as well as luminosity caused by ephemeral sources (e.g. lightning). We detect non-ephemeral light using Colorado School of Mines’ (CSM) Annual Lit Masks, which provide pixel-level information on non-ephemeral sources of light.
We subsequently compute monthly luminosity data (by summing up all luminous pixels) for each province in Vietnam from January 2014 to August 2022.
Results
1. Mapping Black Marble and Colorado School of Mines data
Figure 1: Normalised log luminosity from Black Marble (left) and Colorado School of Mines (right) for provinces in Vietnam from January 2014 to December 2020
We can see that the CSM data has higher levels of log luminosity for provinces such as Lai Chau, Lao Cai and Nghe An in the North, and Quang Nam in the centre. However, on the whole, there appears to be broad similarities between BM and CSM luminosity patterns.
2. Graphing luminosity for Black Marble and Colorado School of Mines
We subsequently graph luminosity for Vietnam, as well as for the provinces of Hanoi and Ho Chi Minh City.
Figure 2: Monthly line graph of (1) log luminosity, and (2) monthly percentage change for Vietnam (January 2014-December 2020).
Figure 3: Monthly line graph of (1) log luminosity, and (2) monthly percentage change for Ha Noi (January 2014-December 2020).
Figure 4: Monthly line graph of (1) log luminosity, and (2) monthly percentage change for Ho Chi Minh City (January 2014-December 2020).
For all three regions: Vietnam, Hanoi and Ho Chi Minh City, we see that our pre-processed BM data is significantly less volatile than the CSM data. This is particularly pronounced in the monthly percentage change graphs, whereby CSM experiences absolute growth rates of over 200%, whereas BM experiences absolute growth rates of 50% at most. In addition to this, the BM data for Hanoi and Ho Chi Minh City exhibits a clear upwards trend in luminosity over time, which aligns with the upwards trend in GDP that Vietnam has been experiencing since 2014.
When comparing BM to CSM, we see a strong correlation between log luminosity from BM and log luminosity from CSM, however this correlation disappears when we look at monthly growth rates.
Figure 5: The first panel illustrates the correlation between log luminosity from Black Marble and log luminosity from Colorado School of Mines. The second panel shows the lack of correlation between annual growth rates in luminosity from Black Marble and that from Colorado School of Mines.
In the regression output below, we model the relationship between log luminosity from Black Marble and log luminosity from CSM (column 1). We find that the relationship is statistically significant at the 1% level. However, when examining the relationship between monthly growth rates, we find that the relationship becomes negative and is no longer significant.
Figure 5: The first panel illustrates the correlation between log luminosity from Black Marble and log luminosity from Colorado School of Mines. The second panel shows the lack of correlation between annual growth rates in luminosity from Black Marble and that from Colorado School of Mines.
In the regression output below, we model the relationship between log luminosity from Black Marble and log luminosity from CSM (column 1). We find that the relationship is statistically significant at the 1% level. However, when examining the relationship between monthly growth rates, we find that the relationship becomes negative and is no longer significant.
Figure 6: The first panel maps log luminosity for 2018 from our processed BM data, the second panel maps log luminosity for 2018 from CSM, and the third panel maps log GDP in 2018 for each province in Hanoi.
When we conduct a regression analysis between each source of luminosity and log GDP, we find that our processed BM data has a higher level of statistical significance (significant at the 1% level) and a higher R² than the CSM log luminosity data (significant at the 5% level). This indicates that the pre-processed BM data is a stronger indicator of economic activity than the CSM data.
Table 2: Column 1 measures the relationship between log luminosity from CSM and log GDP for 2018. Column 2 measures the relationship between log luminosity from BM and log GDP for 2018.
Conclusion
To summarise, we find that our processed BM luminosity data provides several improvements to the off the shelf luminosity data from the CSM in measuring economic activity in Vietnam. We see that the BM data is significantly less volatile than CSM data, has a stronger relationship with provincial-level GDP, and exhibits a clear upwards trend from 2014-2021 which matches the upwards trend in GDP that Vietnam as a whole has experienced over the same period.
This suggests that, while luminosity may be a promising proxy for economic activity, several improvements can be made to it by conducting pixel-level processing. This has the effect of making the results less noisy, and a better measure of economic activity at a sub-national level.
