1. Introduction
Poverty is a long-term worldwide predicament and a main cause of instability for society [
1]. It is estimated that 735.9 million people remained in poverty (
$1.90 per day) during 2015. Poverty eradication is the primary goal of the Sustainable Development Goals (SDGs) set forth by the United Nations. Therefore, poverty reduction has become a vital task faced by many countries [
2]. China has implemented huge amounts of work for poverty alleviation and has achieved remarkable results. By the end of 2020, all the 770 million Chinese rural population living below the poverty line were lifted out of poverty, and region-wide absolute poverty was resolved. However, imbalanced and inadequate development are still urgent issues for China, which are barriers to China’s sustainable development. Meanwhile, consolidating achievements of poverty alleviation also will be the next priority. There is still a long and tough road to completely eliminate poverty in China [
3]. Hence, accurately and objectively evaluating and monitoring poverty level and development situations are crucial for governments to continue the strategy of rural vitalization and promote balanced development.
Traditionally, statistics data provides the main basis of poverty level evaluation and analysis. In the beginning, poverty is considered as an economic phenomenon [
4], thus poverty measurement and determination is mainly based on the single economic dimension [
2], and census data such as gross domestic product (GDP) is the most prominent indicator [
5]. However, GDP cannot be treated as a sole measure of a country’s well-being because it is unable to express many components related to individual and social well-being [
6]. With the development of the poverty concept, poverty studies have turned to a multidimensional assessment that involves not only economic but also natural, human and other aspects [
7]. Multidimensional indicators are often more accurate than using a single economic dimensional [
8], therefore many studies use this way to identify regional poverty levels and have obtained quite reasonable results [
8]. However, more dimensions also mean higher requirements for data integrity. Besides, the time-sensitivity of data limits these traditional methods based on statistic or survey data. Because of the high cost, most existing statistics are updated by census or survey at long intervals [
9], and detailed poverty-related data still needs to be provided between the intervals [
10].
Satellite remote sensing data, especially nighttime light (NTL) imagery, offers a timely, objective and consistent way for direct observations about human activities [
11]. Compared to other satellite products, NTL data has a distinct advantage in quantifying human activities [
12], and it has been widely used to monitor socioeconomic dynamics [
13], such as urbanization [
14], population [
15], GDP [
16], electricity consumption [
17] and so on. Therefore, NTL data correlates closely with poverty distribution as well. The most used NTL data sources include the Defense Meteorological Satellite Program’s Operational Line Scan System (DMSP-OLS) and the Suomi National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) Day-Night Band. Both datasets have been utilized in poverty analysis. For the sake of simplicity and convenience, some studies accomplish poverty level evaluation by exploring the correlation between NTL data and poverty indices [
18]. However, most of these studies tended to focus on single NTL feature, which may cause the omission of important information that can be useful for exploring poverty [
19]. Meanwhile, the commonly regression models that relate NTL to poverty indices might not work well without a huge prior knowledge of the potential relationship between inputs and outputs [
11]. Therefore, some other scholars utilized classification features of NTL data, and adopted machine learning approaches to identify the poor counties [
11]. These previous studies provide useful thoughts for poverty research based on NTL data, but there are some uncertainties that poverty probability obtained from these classification approaches. The previous training model is based on the output of “poor counties” as “1” and “non-poor counties” as “0”, without linking NTL to the real socioeconomic information [
19], but only relying on the night light itself to interpolate between 0 and 1, which can hardly reflect the multidimensional characteristics of poverty. In addition, poverty is a long-term dilemma, but the applications of NTL data in poverty analysis are hindered by the sensor differences and temporal coverage differences of different NTL products [
12].
The DMSP-OLS NTL dataset provides a continuous time series during the period 1992–2013, while the NPP-VIIRS NTL product runs from 2012 to the present. As the result of the evident inconsistency between the two data sources, however, it is difficult to directly apply multi-source NTL data for long-term studies [
20]. Therefore, it is essential to connect the DMSP-OLS and NPP-VIIRS NTL to construct a continuous time series NTL dataset for facilitating long-term research. There have been some studies attempting to integrate NTL from both sources for extending NTL data. For example, Zhu et al. [
21] attempted to integrate the DMSP-OLS and NPP-VIIRS NTL datasets at the provincial level using the relationship of power function, and Zhao et al. [
22] achieved this with a quadratic polynomial model at the level of county. However, these datasets generated at the administrative level have limited application at a finer scale. Meanwhile, the applicability of current approaches still suffers from some limitations, especially the limited accessibility of datasets [
23], regional limitations of methods et al. [
24] Moreover, the widely-used fixed regression model [
25] may fail to explain the undefined mechanisms correlating the two NTL datasets and lead to mismatch. For the well-acknowledged NTL data [
26], whose accuracy is still limited by saturation, this may affect the final results obtained in this study. On the contrast, back propagation (BP) artificial neural network (ANN) is an efficient multilayer feedforward neural network with wide applicability. However, there are some inherent problems in BP algorithm, such as trapping in local minima easily and the low convergent speed [
27]. Particle swarm optimization (PSO) algorithm has a great ability to obtain the global optimistic results [
28], which is also adopted to improve the performance of BP algorithm. Hence, in the study, by combining the PSO with the BP, a PSO-BP hybrid algorithm was used to construct the ANN for exploring the potential relationship between the DMSP-OLS and NPP-VIIRS data.
Hence, the major objectives of this study are: (1) to integrate DMSP-OLS and NPP-VIIRS NTL data at the pixel level based on the PSO-BP hybrid algorithm, and establish a consistent time-series of NTL dataset from 2000 to 2019; (2) to establish an efficient method for poverty evaluation based on NTL feature variables by exploring the relationship between them and an actual comprehensive poverty index (ACPI), and implement poverty evaluation in southwestern China from 2000 to 2019 based on the consistent NTL dataset; (3) to analyze the spatiotemporal variations of regional poverty in southwestern China from 2000 to 2019.