Poverty Mapping in the Dian-Gui-Qian Contiguous Extremely Poor Area of Southwest China Based on Multi-Source Geospatial Data

Abstract

Accurate information on the spatial distribution of poverty is of great significance to the formulation and implementation of the government’s targeted poverty alleviation policy. Traditional poverty mapping is mainly based on household survey data and statistical data, which cannot describe the spatial distribution of poverty well. This paper presents a study of mapping the integrated poverty index (IPI) in the Dian-Gui-Qian contiguous extremely poor area of southwest China. Based on multiple independent spatial variables extracted from NPP/VIIRS nighttime light (NTL) remote sensing data, digital elevation model (DEM), land cover information, open street map, and city accessibility data, eight algorithms were employed and compared to determine the optimal model for IPI estimation. Among these machine learning algorithms, traditional multiple linear regression had the lowest accuracy compared with the other seven machine learning algorithms and XGBoost showed the best performance. Feature selection was performed to reduce overfitting and five variables were finally selected. The final developed XGBoost model achieved an MAE of 0.0454 and an R2 of 0.68. The IPI map derived from the developed XGBoost model characterized the spatial pattern of poverty in the Dian-Gui-Qian contiguous extremely poor area well, which provided a good reference for the poverty alleviation work and public resources allocation in the study area. This study can also serve as a template for poverty mapping in other areas using remote sensing data.

1. Introduction

Poverty is a long-term concern of human society and leads to social instability and conflict [1]. Eradicating poverty is one of the largest social problems faced by humankind and is a significant challenge and indispensable requirement for sustainable development [2]. Since the reform and opening up of China, poverty alleviation work has achieved remarkable results. The number of poor people in China has been drastically reduced, but there were still 16.6 million poor people in 2018 [3]. Poverty-stricken areas in China are scattered with characteristics of large dispersion and small concentration and are mainly concentrated in remote mountainous areas, ethnic minority areas, and interprovincial border areas [4]. Constrained by natural conditions, regional transport, development history, and other factors, these areas have high poverty levels and large coverage areas [5]. To promote the development of these poor areas, the Chinese government launched a major project of poverty alleviation in 2011, known as the “National Program for Rural Poverty Alleviation (2011–2020)”, which defines 14 contiguous extremely poor areas and points out that these areas are the frontlines for poverty alleviation in the future.

Due to the differences in the natural environment and socioeconomic conditions, the contiguous extremely poor areas show significant spatial differences of poverty. Therefore, it is necessary to accurately reveal the spatial details of poverty. Poverty identification is an important component of poverty studies, and the basis for scientifically formulating poverty alleviation strategies [6]. Traditional poverty identification studies mainly rely on field survey data or officially released statistical data to measure poverty [7,8,9]. Field survey is time-consuming and labor-intensive and is also difficult to conduct at a large scale. Limited numbers of surveying samples leads to large uncertainties [10]. The official statistical data are generally collected by administrative districts and cannot accurately reflect the spatial differences within administrative units. Satellite remote sensing, especially night-time light (NTL) remote sensing, provides a new way for large-scale socioeconomic spatial monitoring and evaluation [11]. The Defense Meteorological Satellite Program (DMSP)/Operational Linescan System (OLS) data are the earliest NTL satellite data and have been widely used for socioeconomic activity monitoring. However, DMSP/OLS has the shortcomings of relatively coarse spatial resolution, low radiation resolution, lack of on-board calibration, and signal saturation in urban centers [12]. The new generation NTL satellite data, Suomi National Polar-orbiting Partnership (NPP)/Visible Infrared Imaging Radiometer Suite (VIIRS), was put into use in 2012. Compared with DMSP/OLS, NPP/VIIRS has significantly improved spatial resolution, quantization, dynamic range, and calibration [13]. Studies have shown that there is a close relationship between NTL satellite data and regional economic development [14,15]. NTL satellite data has been widely used in studies on urban expansion [16,17,18], economic evaluation [19,20], population estimation [21,22,23], and power consumption [24,25].

The NTL satellite data also has potential in poverty assessment. However, there are relatively few studies on poverty monitoring by NTL data. Ebener et al. [26] estimated the distribution of income per capita as the proxy for wealth at the country and sub-national level based on DMSP/OLS NTL data and population data, and analyzed its relationship with health indicators. Noor et al. [27] calculated the asset-based poverty index of 338 level-1 administrative districts in 37 African countries based on household survey data and analyzed their relationship with DMSP/OLS data, indicating that NTL is an effective alternative for the poverty survey data. Elvidge et al. [28] produced a global poverty map using a poverty index by dividing LandScan population count by the average values of DMSP/OLS NTL data, and also estimated the percentage of the population living in poverty using a univariate linear model. Wang et al. [29] calculated an integrated poverty index (IPI) at the provincial scale in China based on 17 socioeconomic indicators using principal component analysis and then used regression analysis to explore the relationship between IPI and the average light index (ALI) of DMPS/OLS data, achieving an R² of 0.854. Yu et al. [30] combined 10 socioeconomic indicators to calculate the IPI of the 38 counties of Chongqing, China, and verified the relationship between IPI and ALI using linear regression. The validation result showed an R² of 0.86, suggesting good correlation between them. Njuguna and McSharry [31] estimated the multi-dimensional poverty index (MPI) at the sector level in Rwanda based on VIIRS NTL data, Landscan population data, and detailed call records using multivariate regression, achieving a cross-validated correlation coefficient of 0.88. Pan et al. [32] developed a linear regression model to estimate MPI from NPP/VIIRS ALI in Chongqing Municipality, China, and tested the developed model in Shaanxi Province, showing an average relative error of 11.12%. The developed model was applied to map poverty in China at the county scale. The above studies mainly used correlation and univariate linear regression analysis to establish the quantitative relationship between poverty and NTL. However, due to the complexity and diversity of the causes and manifestations of poverty, the relatively simple one-factor linear model is insufficient to fit the relationship between poverty and light intensity. Li et al. [33] classified high-poverty counties in China based on the features extracted from DMSP/OLS NTL data using seven machine learning approaches (Gaussian process with radial basis function kernel, stochastic gradient boosting, partial least squares regression for generalized linear models, random forest, rotation forest, support vector machine, and neural network with feature extraction). The overall accuracies of all models were higher than 82%, suggesting that machine learning algorithms can effectively identify high-poverty counties. Zhao et al. [34] used the random forest method to estimate the household wealth index (HWI) of Bangladesh based on the spatial features extracted from NPP/VIIRS NTL data, Google satellite images, land cover, road maps, and division headquarter location data. The validation results showed an R² of 0.70 in Bangladesh and an R² of 0.61 in Nepal. Shi et al. [35] developed a comprehensive poverty index (CPI) in Chongqing, China by combining NPP/VIIRS NTL data, DEM, the normalized differential vegetation index (NDVI) and point of interest (POI) data. The developed CPI was validated by visual comparisons with Google Earth images and poverty-stricken villages, showing good fitness between CPI and them. CPI values were also compared with MPI values at the county level, achieving an R² of 0.931. Li et al. [36] calculated a farmer sustainable livelihood index (FSLI) from socioeconomic indicators to measure county-level poverty in Hubei province. First-order and second-order linear regression were performed to estimate FSLI from Luojia 1-01 ALI, achieving an R² of 0.85 and 0.88, respectively. Yin et al. [37] classified poverty areas in Guizhou, China based on the spatial features extracted from NPP/VIIRS NTL and geographical data using random forest, support vector machine, and artificial neural networks. Random forest outperformed the other two methods with an overall accuracy of 0.94. Niu et al. [38] estimated the multi-source data poverty index (MDPI) in Guangzhou, China based on the spatial variables derived from NPP/VIIRS NTL, Landsat8 image, POI, and housing rent data using the random forest algorithm, yielding a correlation coefficient of 0.954. The recent poverty studies based on remote sensing introduced more spatial variables and employed machine learning technology, showing better performances than the traditional univariate linear regression methods. However, various machine learning algorithms were mostly used for classifying poverty areas. For quantitatively identifying poverty index, only random forest algorithm was used, while other machine learning methods were still rarely used. In addition, relatively few studies were conducted in contiguous extremely poor areas, the most important regions for poverty alleviation in China.

In this study, we aim to map poverty in the Dian-Gui-Qian contiguous extremely poor area in southwest China based on multi-source spatial datasets using several machine learning algorithms. The Dian-Gui-Qian contiguous extremely poor area is one of the most important extremely contiguous poor areas in China and is characterized by complicated social and natural conditions, making it a key area for poverty identification. The IPI values were calculated at the county scale by integrating 13 socioeconomic indicators as the dependent variables and various spatial features derived from multi-source geospatial data were used as independent variables. To develop the best model for poverty identification, eight machine learning algorithms were employed and compared using cross validation. The model with best performance was applied to spatial input variables to map poverty in the study area. The remotely sensed poverty map measures poverty on a fine spatial scale and provides a reference for the formulation and implementation of the strategies for targeted poverty alleviation in the study area, and the study provides a practical way for identifying and evaluating the spatial pattern of poverty.

2. Study area and data

2.1. Study Area

The Dian-Gui-Qian contiguous extremely poor area is located in southwest China (Figure 1). It covers a total area of 227,544 km² and has a permanent population of 29.15 million. There are 91 county-level administrative districts in this region, including six municipal districts, five county-level cities, 66 counties, and 14 minority autonomous counties. In China, administrative districts are divided into four levels: provincial (municipality, province, autonomous region, and special administrative region), prefectural (prefecture-level city, autonomous prefecture, prefecture, and league), county (district, county-level city, county, autonomous county, banner, autonomous banner, special district, forest district), and township (subdistrict, town, township, ethnic township, sum, ethnic sum, and county-controlled district).

The Dian-Gui-Qian contiguous extremely poor area is one of the most typical karst areas in the world, whose landforms are mainly plateaus and mountains. Due to the remote geographical location, harsh natural conditions, frequent natural disasters, and weak infrastructure, the poverty problem in this region is serious. Among the 14 extremely contiguous poor areas in China, this region has the largest number of counties, the most poverty alleviation targets, and the largest ethnic minority population, making it a key area of poverty alleviation.

2.2. Data

The data used in this study mainly include county-level socioeconomic statistical data and remote sensing data.

The county-level socioeconomic statistical data was mainly collected from the China County Statistical Yearbook of 2017, which provides socioeconomic statistical data in 2016. The supplemented socioeconomic statistical material included the statistical yearbooks of the cities and the statistical bulletins on national economic and social development in the study area.

The remote sensing data include NPP/VIIRS NTL data, SRTM DEM data, Finer Resolution Observation and Monitoring Global land cover (FROM-GLC) data, open street map (OSM) data, accessibility to cities, and natural disaster data.

NPP/VIIRS NTL data were the NPP/VIIRS DNB annual cloud-free composite data in 2016 provided by the National Oceanic and Atmospheric Administration (NOAA) Earth Observation Group. This data provides annual average radiance by excluding the influence of stray light, lightning, lunar illumination, cloud-cover and temporal lights, with a spatial resolution of 15 arc-seconds (~500 m).

The FROM-GLC land cover data in 2015 were provided by the Department of Earth Sciences at Tsinghua University. FROM-GLC is a 30 m resolution global land cover product developed by a random forest-based mapping framework based on Landsat remote sensing data. The classification scheme contains 11 land cover categories: cropland, forest, grassland, shrubland, wetland, water, tundra, impervious surface, bareland, snow/ice, and cloud.

OSM is a free editable map amended by volunteers from all over the world, providing geospatial information such as roads, buildings, and water bodies. The road network in the study area was extracted from OSM database.

The DEM data used in this study were the SRTM DEM data provided by the United States Geological Survey (USGS) with a spatial resolution of 1 arc-second (~30 m).

Accessibility to cities data, provided by the Malaria Atlas Project, is a global map of travel time to cities in 2015. It quantifies travel time to cities at a spatial resolution of 30 arc-seconds (~1 km) by integrating spatial variables that affect human movement rates such as roads, railways, land covers, topographical conditions and national borders [39].

Natural disaster data was derived from the global estimated risk index provided by the Global Risk Data Platform. The index at a spatial resolution of 5 arc-minutes (~10 km) was estimated from multiple hazards including tropical cyclone, flood, and landslide induced by precipitations, with the index value ranges from 1 (low) to 5 (extreme).

3. Methods

3.1. IPI Calculation from Census Data

Poverty is a comprehensive social phenomenon that is not only related to income and assets but is also closely related to factors such as health and education. Simply using a single indicator cannot fully reflect the poverty situation. Some studies have proposed to quantify poverty by combing a collection of socioeconomic indicators. In this study, IPI was used to indicate poverty conditions. According to previous research, IPI was calculated from multiple socioeconomic indicators, which cover economic development, living conditions, health, and education. In total, 13 socioeconomic indicators were employed to calculate IPI (

Table 1. Measurement indicators of IPI.

Dimension	Indicator	Attribute	Weight
Economic development	Per capita GDP	Positive	0.0645
	Per capita fiscal expenditure	Positive	0.0457
	Per capita savings	Positive	0.0881
Living conditions	Per capita income	Positive	0.0627
	Per capita fixed assets	Positive	0.1727
	Endowment insurance coverage	Positive	0.0361
	Per capita telephone ratio	Positive	0.1006
Health	Per capita beds in health	Positive	0.1151
	Per capita doctors	Positive	0.0481
	Medical insurance coverage	Positive	0.0135
Education	Per capita students present in primary and high schools	Positive	0.1220
	Per capita teachers at primary and high schools	Positive	0.0589
	Per capita education expenditure	Positive	0.0722

).

To eliminate the dimension differences of different indicators, each indicator was min–max normalized:

$$X_{ij}=\frac{x_{ij}-\min \left(x_i\right)}{\max \left(x_i\right)-\min \left(x_i\right)}$$

(1)

where X_ij is the normalized value of the jth sample of indicator i, x_ij is the original value of the jth sample of indicator i, and min(x_i) and max(x_i) are the minimum and maximum values of the ith indicator, respectively.

The weight of each indicator was determined using the entropy weight method to avoid the bias caused by subjective factors. For an indicator, higher information entropy indicates greater role of this indicator in the comprehensive evaluation.

The entropy of indicator i (e_i) was calculated as follows:

$$e_i=-\ln {\left(n\right)}^{-1}·{\displaystyle \sum }_{i=1}^n{P}_{ij}·\ln \left(P_{ij}\right)$$

(2)

where n is the number of samples, and:

$$P_{ij}=\frac{X_{ij}}{{\sum }_{j=1}^n{X}_{ij}}$$

(3)

The deviation degree of indicator i (d_i) was calculated as follows:

$$d_i=1-{\mathrm{e}}_i$$

(4)

The weight of indicator i (w_i) was calculated as follows:

$$w_i=\frac{d_i}{{\sum }_{i=1}^m{d}_i}$$

(5)

where m is the number of predictors.

The calculated weights of the indicators are shown in Table 1. Then, the IPI of each county-level administrative district was calculated as follows:

$$IP{I}_j={\displaystyle \sum }_{i=1}^m{X}_{ij}·w_i$$

(6)

where IPI_j is the IPI value of the jth sample. A high IPI value indicates a low level of poverty, while a low IPI value refers to a high level of poverty.

3.2. IPI Identification from Spatial Data

Poverty is a complex phenomenon, which should be understood from multi-source data to improve the accuracy of poverty mapping [40]. In this study, a number of spatial independent variables were extracted from multi-source geospatial data for the spatial identification of poverty. The average NTL was extracted from the NPP/VIIRS data, the traffic accessibility (TA) was extracted from the city accessibility data, the altitude (H) was extracted from the SRTM/DEM data, the flat area (slope < 5°) coverage (FAC) was derived from the SRTM/DEM data, the impervious surface coverage (ISC) and cropland coverage (CC) were calculated from the FROM-GLC data, the road density (RD) was calculated based on the OSM road network data, and the risk index (RI) was extracted from the globally estimated risk index data. A total of eight independent variables are used for the spatial identification of poverty: NTL, H, FAC, ISC, CC, TA, RD, and RI.

Considering the different natural and socioeconomic conditions in different regions and the complex relationships between various factors, the relationships between the poverty level and independent spatial variables in different regions may be different. In addition to the traditional multiple linear regression (MLR), machine learning methods such as the bidirectional recurrent neural network (BRNN), generalized additive model (GAM), support vector machine (SVM), MARS, random forest (RF), XGBoost, and Cubist, were also employed to estimate the IPI from independent spatial variables. Compared with the MLR, machine learning models have the advantage of being able to fit complex, high-dimensional nonlinear relationships [41].

Ten-fold cross-validation (CV) was used to verify model performance. The whole dataset was randomly divided into ten folds, one of which was used as the test dataset and the other nine folds were used as the training dataset. This process was repeated 10 times so that all samples were used once as validation samples. In the CV process, the parameters of each model are tested to optimize the model, and the parameters that obtain the minimum error are used to fit the final model. The mean absolute error (MAE) and coefficient of determination (R²) are calculated and used as accuracy evaluation indicators. The method with the highest accuracy is used to establish the final IPI estimation model.

To reduce the overfitting problem of machine learning models, feature selection was performed to control the complexity of the model. The eight variables were screened to determine the optimal variable combination. First, all the combinations of two variables were tested to determine the combination with the highest accuracy. Then, a new independent variable was added to fit the model. The new independent variable with the best performance was retained. By analyzing the influence of adding new variables to the model performance, the optimal robust combination of variables was determined.

When the final model was developed, the importance of each variable was quantified to evaluate the contributions of each independent variable to the IPI estimation. For each variable, the percentage increase in MAE (% IncMAE) was calculated as the measure of its importance when the variable was randomly permuted and the other variables remained unchanged.

4. Results

4.1. Relationship between IPI and NTL

Based on the values and corresponding weights of the 13 socioeconomic indicators of the 91 county-level administrative districts in the study area, the IPI values of these administrative districts were calculated. The Luocheng Mulao Autonomous County has the lowest IPI of 0.1436, and the Kaili County-level City has the highest IPI of 0.7057. For all administrative districts, the average IPI is 0.3045, the median IPI is 0.2792, and the skewness is 1.357. The majority of administrative districts show lower IPI values than the average value.

Previous studies have shown that NTL data is an effective spatial indicator of poverty. Figure 2 shows the scatterplots between the NTL and IPI values of the 91 county-level administrative districts. A correlation coefficient (R) of 0.63 indicates that there was a positive but not strong correlation between them. The NTL values of most administrative districts were relatively low (<1 nanoWatt·cm⁻²·sr⁻¹), and some administrative districts with similar NTL values had obviously different IPI values. Therefore, it is not sufficient to use only the NTL to indicate poverty.

4.2. Model Performance

In the modeling process, the parameters of the eight methods were tuned to obtain the optimal settings. The machine learning models based on the optimal parameter settings were used to estimate IPI and evaluated using 10-fold CV. During the CV process, the total 91 county-level districts were randomly divided into 10 folds, with one fold used for validation and the remaining nine folds used for fitting models; this was repeated 10 times until all folds had been used as validation datasets once. CV-MAE and CV-R² were calculated based on the comparison of validation and training data. The CV-MAE and CV-R² of each model are shown in

Table 2. Accuracies of the nine models.

Model	MAE	R2
MLR	0.0767	0.23
BRNN	0.0593	0.47
GAM	0.0498	0.59
SVM	0.0583	0.49
MARS	0.0581	0.46
RF	0.0527	0.57
XGBoost	0.0479	0.61
Cubist	0.054	0.53

. The performances of these models were quite different. The MLR had a significantly lower accuracy than the other models, with an MAE of 0.0767 and R² of 0.23, indicating that a simple linear regression cannot fit the relationship between the IPI and independent variables very well. The XGBoost model had the highest accuracy (MAE = 0.0479, R² = 0.61) and therefore was selected as the final model to map IPI from spatial variables.

Forward feature selection was applied in the XGBoost modelling to develop a simple model with fewer variables to reduce overfitting. Figure 3 shows the variations of minimum MAE with the number of variables. As the number of variables increased from 1 to 5, the model error significantly decreased. Then, as the number of variables continued to increase, the error tended to increase. Therefore, the combination of five variables (NTL, TA, RI, RD, and FAC) was used to develop the XGBoost model for IPI estimation.

Figure 4 shows the scatterplot between the actual IPI and the IPI estimated by the XGBoost model, which is constructed based on the selected five variables. The MAE is 0.0454, and R² is 0.68. Most samples were clustered near the 1:1 line, indicating that the predicted IPI values were in good agreement with the actual IPI values. Wealthy county-level administrative districts (IPI > 0.5) were slightly underestimated, and other samples were not obviously overestimated or underestimated. In addition, the fitting error did not vary with the IPI value, indicating no heteroscedasticity problem.

Figure 5 shows the influence of the five variables on the XGBoost model. NTL had the highest importance (% IncMAE = 38.51%), indicating that NTL was the most important variable for capturing the spatial variance in poverty. NTL can effectively reflect the intensity of human activities and is closely related to the social economy, thus resulting in a very high importance index. The importance of TA (% IncMAE = 13.33%) was secondary to that of NTL, suggesting that the connection to cities in the study area restricted the economic development to a great extent. RI also had a high importance (% IncMAE = 10.03%), indicating that natural disasters in this region were also an important factor restricting residential poverty levels. The % IncRMSE values of RD and FAC were 5.38% and 4.01%, respectively, indicating that local traffic and terrain conditions had a certain influence on the level of poverty, but such influence was less than that of the other three variables.

4.3. Spatial Distribution of IPI

The validation results suggested that the XGBoost model developed based on five variables (NTL, TA, RI, RD, and FAC) was the best model for poverty mapping in the Dian-Gui-Qian contiguous extremely poor area. Figure 6 shows the spatial distribution of the estimated county-level IPI derived from the XGBoost model. Generally, the overall IPI values of northern parts were obviously higher than those of southern regions. The wealthy regions with high IPI values were mainly prefectural districts and county-level cities, such as Zhongshan District, Kaili County-level City, Xingyi County-level City, Xixiu District, and Duyun County-level City (labeled in the figure). These regions had relatively good socioeconomic foundations, decent infrastructure, and well developed industry and commerce, making the poverty level relatively low in these county-level districts. The poor regions with low IPI values were mainly located in the southwest and east of the study area. These regions had poor transportation, ethnic minority settlements, irrational industrial structures, and weak social and economic foundations, leading to deep poverty.

To assess the reliability of the spatial distribution of the remotely sensed county-level IPI map, it was compared with the actual county-level IPI map (Figure 7a). Similar to Figure 6, the actual IPI map also shows a high pattern in the north and low pattern in the south, suggesting good spatial prediction of the remotely sensed poverty. The relative errors of all county-level districts were also calculated (Figure 7b). Most of the county-level districts showed relative errors within ±15%, and only a few county-level districts had relative errors higher than 30% or lower than −30%. The spatial assessment indicates that the remotely sensed IPI map can effectively characterize the spatial distribution of poverty. It was also noted that the county-level districts with high IPI values generally showed negative relative errors, suggesting that the remotely sensed IPI values were less than the actual IPI values. This phenomenon indicated the slight underestimation of the wealthy regions in the remotely sensed result, as pointed out in the aforementioned analysis.

Based on the developed XGBoost model and the spatial variables, the distribution of IPI at town level in the study area was also mapped (Figure 8). Because of lacking town-level socioeconomic statistical data, the remotely sensed town-level IPI map cannot be quantitatively validated, but can be compared with the county-level IPI map to assess the spatial consistency between them. Compared to the county-level IPI map, the town-level IPI map showed a similar overall distribution pattern that most wealthy units were located in the northern part. However, the township-level IPI map reflected more detailed spatial differences. In the northern part of the study area, there were many wealthy town-level administrative districts (IPI > 0.45), but there were also some very poor town-level districts (IPI < 0.25), indicating strong inequality in this area. The town-level administrative districts in the southern part generally showed low IPI values, and the gap in wealth was relatively small. These detailed spatial patterns cannot be reflected by the county-level IPI map but can be well identified by the town-level IPI map. The developed remote sensing model for IPI estimation can also be applied at a finer scale, such as the village-level administrative districts or pixels, to reveal more detailed information about the spatial distribution of poverty.

5. Discussion

Compared with previous studies, the correlation coefficient between poverty and NTL in the Dian-Gui-Qian contiguous extremely poor area is quite low (R = 0.63) in this study. This may be attributed to the complex terrain in the study area and the large differences between the natural and economic conditions of the administrative districts. Due to the low correlation between IPI and NTL, it is difficult to achieve ideal results by simple regression models with NTL as the only independent variable. Therefore, auxiliary spatial variables shall be introduced to improve the accuracy. Under this consideration, terrain, land cover, traffic, and raster variables were also included. Given the complicated relationships between poverty and these spatial variables, several machine learning algorithms were employed to fit the relationships between poverty and spatial variables. Considering that linear regression was widely used in previous studies, a MLR model was also developed as a comparison. The validation results indicated that machine learning models, especially the ensemble learning models such as XGBoost, achieved much higher accuracies than the MLR model. This study indicates that machine learning algorithms based on multi-source spatial data can effectively identify the spatial distribution of poverty, providing a reliable way to map poverty in regions with complex natural and socioeconomic conditions. All the spatial datasets used in this study can be timely and freely obtained. Therefore, the proposed method can be conveniently applied in other areas for poverty mapping without being constrained by data acquisition.

Although the method in this study shows efficiency for poverty mapping, there are still some limitations. Poverty is a complicated issue related to various socioeconomic conditions. This paper introduced multi-source spatial variables to identify poverty. However, most of the spatial variables are environmental indicators that can just indirectly reflect poverty. In addition, it should be noted that with the development of information technology, new economic forms such as the internet economy have broken through the limitation of geographic space. The associated factors of wealth have become increasingly complex and diverse, and their spatial identifications are therefore facing greater challenges. Nighttime light remote sensing data provide a unique way to monitor human activities from space at night, making it a useful indicator to identify poverty. In this study, NTL derived from NPP/VIIRS showed the highest importance for poverty mapping. However, the spatial resolution of NPP/VIIRS is not very high, which may not effectively reflect spatial details of human activity patterns. Moreover, due to the lack of household survey data, further validation cannot be performed at a finer scale in this study. If sufficient household survey data can be collected, the method proposed in this paper will be more comprehensively validated on a small scale.

6. Conclusions

The capability of machine learning technology on poverty mapping from multiple remote sensing data has been evaluated in this paper. The IPI was calculated from the socioeconomic indicators of county-level administrative districts in the Dian-Gui-Qian contiguous extremely poor area of southwest China. Then, multiple spatial indicators were extracted from NTL, DEM, land cover, OSM, city accessibility, and natural disaster data as the spatial variables for poverty mapping. Several machine learning algorithms were employed and compared to develop models for estimating county-level IPI over the study area. Cross validation results suggested that XGBoost outperformed other methods and exhibited the best performance. By feature selection, the final XGBoost model was developed, which achieved an MAE of 0.0454 and an R² of 0.68.

The spatial pattern of the remotely sensed county-level IPI map is similar to that of the actual IPI map, indicating that it can effectively characterize the spatial pattern of poverty. By applying the developed model to spatially independent variables, an IPI map at a finer resolution could be produced, such as town-level or even pixel-level. The remotely sensed IPI provides a valuable reference for a more detailed understanding of the poverty situation in the Dian-Gui-Qian contiguous poor area, which is helpful for the formulation and implementation of the government’s targeted poverty alleviation strategy.

In the future, some efforts could be considered to improve the spatial identification of poverty: Additional spatial datasets, especially the datasets that have direct relationships with poverty should be employed to quantify poverty from a more comprehensive perspective, including POI, location-based social media data, and mobile signaling data; nighttime light remote sensing data with higher spatial resolution, such as Luojia 1-01, can be used to provide more details of human activities for poverty mapping. Deep learning technology is expected to extract detailed building information from high resolution satellite images, which is an important and complementary spatial variable for poverty identification.