The man-land areal system comprises a complex arrangement with certain structures and functions that form through interactions between the two subsystems of human and environment over a particular time period and spatial area (Wu, 1991; Wang, 1997; Lu and Guo, 1998). Within such a giant open system, man is the main subject for activities while environment refers to the geographical setting formed by the interweaving of natural and human elements on the surface of the earth (Li et al., 2016). These associations mean that human, natural, and socioeconomic factors all make up the man-land areal system, while interactions between these three components and their internal parts promote system evolution (Figure 1). The living and development needs of humans are met through the use of natural resources (Zhou and Wang, 2016). Groups of humans with higher level abilities are able to obtain and utilize more resources. In contrast, natural resources provide the material basis and spatial carrier for socioeconomic development, the social economy in this context is therefore the result of interactions between human and environment in the process of systems evolution. Natural resources are then incorporated into the production system via an input-output relationship and provide the necessities required by humans. In other words, if a system is well structured and its internal components are arranged properly, then it will develop in a benign direction, but if conflicts are presented amongst internal structures and elements within a system, it will gradually decline (Wang et al., 1999).

Although China is rich in natural resources, these exhibit small per capita availability levels compared with other countries globally. At the same time, Chinese natural resources are also uneven in spatial distribution and poorly matched. For example, the cultivated land and water resources in north China account for 58% and 19% of the national total, respectively, while these resources in south China comprise 42% and 81% (Song et al., 2005; Zhou and Wang, 2016). Increased attention has also been paid to urban areas because of the national urban-rural dual structure and this has caused infrastructure construction in rural areas lagging behind cities. These disadvantages have therefore promoted the rural outflow of young and middle-aged labor, capital, and other development elements, all further aggravating the backward situation in the countryside. Indeed, if problems relating to food and clothing have not been solved, people will be unable to afford higher levels of demands, including education and culture, which results in decreased population quality. In these circumstances, poverty will enter a vicious cycle leading to an intergenerational transmission dilemma. Rural poverty is therefore an external manifestation of the imbalance between human and environment in rural areas and the result from inconsistencies in individual abilities, natural resource endowments, and socioeconomic development levels (Figure 1).

Accurate measurement is a key issue in poverty research. As studies in this field were initiated, poverty was always measured on the basis of per capita household income. This single indicator is, however, unable to synthetically reflect the vulnerabilities of the poor and so the concept of multidimensional poverty was developed which includes health, education, and living standards (Wang, 2017). As a result of cognitive deviations, however, these methods do not take geographical environment into consideration (Liu et al., 2017). Thus, based on previous theoretical analyses and existing related studies (Liu and Xu, 2016; Liu and Li, 2017; Liu et al., 2017), a multidimensional indicator system is developed here to measure comprehensive poverty level (CPL) in rural areas at county level. This approach encapsulates human development ability (HDA), socio-economic development (SED), and natural resources endowment (NRE) (Table 1).

In general, HDA refers to the abilities of people utilizing natural resources to meet their needs and is therefore closely related to population quantity and quality, represented by education level and the size of work-aged population, respectively. However, due to the socioeconomic gap between ethnic minorities and the dominant Han nationality in China, the proportion of ethnic minority population versus total population was also included to measure HDA. The concept of SED is more comprehensive, however, and includes urbanization, economy, public service, transportation and housing conditions. As a result, per capita GDP, public budget revenue, and disposable income of rural residents were all chosen to represent economic development level, while the number of beds in health institutions and child survival were selected to characterize medical service supply. Similarly, per capita living space and road density were respectively used to reflect housing and transportation conditions. NRE reflects development foundations. Altitude, slope, and degree of fragmentation are all major topographic features. Annual rainfall, per capita arable land, NPP, and FPP all describe the basic conditions of agricultural production within a region which plays a significant role in poverty reduction and development (Christiaensen et al., 2011).

The incidence of poverty refers to the proportion of population living below a certain poverty standard relative to the total within a region. The relationship between the indicators selected in this study and the incidence of poverty was therefore investigated using a Pearson correlation coefficient analysis. Results show that minority nationality (A3), altitude (C1), slope (C2), and degree of fragmentation (C3) are all positively correlated to the incidence of poverty, which means that they all contribute to poverty (Table 2). The remainder of indicators are negatively correlated, which means that they contribute to anti-poverty. When significant levels of per capita public budget revenue (B2) and annual rainfall (C4) are both greater than 0.01, 0.069 and 0.150, respectively, all the other indicators considered here are less than 0.01. This result means that the indicators utilized in this study are all closely related to the incidence of poverty.

The data on Chinese county boundaries and road network utilized here were derived from the National Geomatics Center of China (http://www.ngcc.cn/). A total of 2731 county-level units were therefore obtained for this analysis, excluding those that lack rural population as well as Hong Kong, Macao, and Taiwan. A digital elevation model (DEM) at a resolution of 90 by 90 m was downloaded from WebGIS (http://www.webgis.com/index.html), and data about rainfall, land use, NPP, and FPP were obtained from the Resource and Environment Data Cloud Platform (http://www.resdc.cn). Data on social and economic development level were sourced from the Statistical Yearbook of Provinces (2016) and the Statistics Bulletin of National Economy and Social Development of Counties (2015). Data on urbanization, educational attainment, labor force, minority population, and per capita housing area were extracted from tables published in 2010 population census of the PRC by county. Missing values were replaced by substituting variable mean values for upper administrative units.

Rural poverty is the undesirable manifestation of interactions between a range of factors, including individual abilities, natural resources, and social economy. Due to the interactions within each dimension and its components, a complex nonlinear relationship is therefore presented between dependent and independent variables. Because of the unique advantages inherent to arbitrary complex pattern classification, excellent multidimensional function mapping, strong self-learning, and searches for optimal solutions at high-speeds, BP neural networks can be used to reveal rural poverty (Xu, 2002). This method is a multilayer feedforward network which is trained by an error back propagation algorithm, including input, hidden, and output layers (Figure 2). And the training procedure is comprised of both forward propagation and backward feedback (Du et al., 2014). In the former, information is developed from the input to the hidden layer and then to the output layer. As part of this process, if expected results are not attained, then error singles will return along their original paths and are minimized by modifying the weights of neurons in each layer. Theoretically, the three-layer BP neural network model used in this study is able to achieve arbitrary multidimensional continuous mapping (Kitahara et al., 1992).

In BP neural network model, the indicators that characterize HDA, SED, and NRE are respectively used as input neurons, while corresponding outputs are human development index (HDI), socioeconomic index (SEI), and resource endowment index (REI). On this basis, a model for CPL can be constructed as follows:

$CPL=HDI+SEI+REI$ (1)

This expression describes a relationship such that the smaller the CPL, the deeper the poverty, and vice versa. Natural breaks (Jenks) were then applied to divide the standardized values of each indicator into 20 categories. Comparing the results output from different functions enabled us to select TRAINLM, LEARNGDM, and TANSIG as training, adaption learning, and transfer functions, respectively. The numbers of nodes within hidden layers in the training model were three, six, and seven, respectively, which were obtained by cross validation. The training batch was set as 10,000, and system defaults were used for all other parameters.

The data collected in this study were input into the BP neural network model based on previous analysis to measure HDI, SEI, and REI values. Formula (1) was then used to calculate CPL, and spatial patterns in these variables were obtained via an ArcGIS spatial analysis.

Figure 3 shows that there is an obvious regional difference in CPL, which decreases from the southeast coast towards the inland northwest of China. These values can be divided into three areas based on three terrain ladders, where the first ladder and its peripheries are dominated by low values, the second is dominated by medium-low values, and the third is dominated by medium-high and high values. Spatial patterns of REI and SEI are similar to that of CPL, while the HDI exhibits its own characteristics. Figure 3 also shows that HDI values are subdivided into two regions by the Xiamen-Karamay Line; values tend to be low to the southwest of this line and are dominated by low and medium-low figures while high and medium-high values characterize the northeast side.

The data presented in this study show that a total of 288 counties can be characterized by low CPL values. Most of these regions are on the Tibetan Plateau as well as three prefectures in southern Xinjiang (TPSX), Wumeng mountain area (WMMA), and rocky desertified areas of Yunnan-Guangxi-Guizhou (RDA). These regions, which are characterized by formidable natural conditions, are short of resources and tend to be gathering points for minorities. In contrast, 613 counties are characterized by medium-low CPL values. These regions are concentrated within central Xinjiang and contiguous poor areas with difficulties (CPADs), including Liupan mountain area (LPMA), border mountain area in western Yunnan (BMA), Wuling mountain area (WLMA), Qinba mountain area (QBMA), Dabie mountain area (DBMA), Luoxiao mountain area (LXMA), and Southern Greater Khingan Range area (SGKR). Although extremely restrictive conditions for development are seen within these areas, they nevertheless do possess some advantages in terms of water and land resources or transportation condition. Data show that 902 counties are characterized by medium CPL values; these areas are mainly located within traditional agricultural regions and include zones with obvious abundances of mineral and land resources, such as the Loess Plateau, central Inner Mongolia, and northern Xinjiang. Some of this kind of areas are also scattered in the mountainous and hilly areas of south China. A total of 645 counties are characterized by medium-high CPL values. Most of these are located in the western part of Inner Mongolia and Ordos Plateau where resources are abundant, as well as in the Chengdu Plain, the middle-to-lower reaches of the Yangtze River, and eastern coastal areas, where development foundations are good or counties are close to developed areas. The remainder of the counties assessed here, 283, are dominated by high CPL values. These characterize the developed areas in China, including the Yangtze River and Pearl River deltas, the Beijing-Tianjin and Changsha-Zhuzhou-Xiangtan urban agglomerations, and the Wuhan metropolitan area.

We also analyzed CPL values within CPADs²(2 Lenghu, Da Qaidam, Mangya and Shuanghu are not included, so the total number is 676, similarly hereinafter.) as part of this study (Table 3). Resultant statistics show that both CPLs and their sub-dimensions are all lower than the national average, especially HDIs and SEIs. The lowest CPL value recorded here is in Tibet, with TPSX and WMMA values being very close to this level, while the top three areas are DBMA, SGKR, and LLMA. We also show that SEI values for each area are lower than the national average; only HDIs for SGKR, LLMA and QBMA, and REIs for DBMA, RDA, LXMA, and WLMA are higher than the national average.

Because of the continuity of rural poverty and limitations on resources for poverty alleviation, identifying national key counties for poverty alleviation and development (NKCs) is a key initial step (Park et al., 2002; Wang et al., 2007). As a result, the Chinese government firstly designated 331 NKCs in 1986, and increased this number to 592 in 1994. Subsequently, in spite of two revisions in 2001 and 2011, the total number of NKCs has remained unchanged at 592 (Figure 4). Targeting at these counties has allowed the Chinese government to actively promote anti-poverty policy innovations and infrastructural construction, greatly alleviating poverty in rural areas. Indeed, the aim of the PRC central government is to lift all rural residents living below the current poverty standard above this level and to remove caps from all counties designated as poverty-stricken. This process is, however, limited by both natural and social conditions, as some counties do not have long-term, stable development abilities. This means that a certain number of poor people will always remain with the improvements of poverty standard. It is therefore necessary to study anti-poverty targeting in rural China after 2020 because this will provide benchmarks for the formulation of subsequent anti-poverty policies.

SD classification is often used to illustrate the differences between the attribute and average values of elements. In order to meet the need of this research, we applied an SD level to divide Chinese counties into seven categories (Figure 5). We therefore recognize counties with CPL values of VII, VI, and V as CRNPSs, and identify 31, 213, and 472 counties in these classes, respectively. Due to disadvantages in personal abilities, socioeconomic development, and natural endowments, development in these regions remains difficult and unsustainable, facing different degrees of pressure to return to poverty.

We compare the 832 national poverty-stricken counties (Figure 6), which contain NKCs and all counties in contiguous poor areas with difficulties (CCPADs). It is clear that 149 counties are newly identified by this analysis, accounting for 20.81% of total number of CRNPSs. The bulks of these newly identified areas are mainly distributed in central Xinjiang, the peripheries of CPADs in southwest China, and the border area of Shandong-Henan-Anhui. It is also clear that 567 counties will still require policy-support after 2020. Of these, 387 counties are NKCs and 509 counties are CCPADs. Specifically, 329 counties belonging to both NKCs and CCPADs are distributed within CPADs with the exception of Tibet and SGKR. A total of 58 are just NKCs, scattered in the peripheries of CPADs and northern Xinjiang, while 180 are just CCPADs, distributed across the Tibetan Plateau. Besides, we show that 261 counties have managed to overcome poverty, which are mainly distributed in transitional zones of the three-gradient terrain, central China, and the east of northeast China. A total of 111 counties are classified as both NKCs and CCPADs, while 94 counties are just NKCs, and 56 counties are just CCPADs. Our data suggest that post-2020 CRNPSs are mainly distributed in high-arid areas of the Tibetan Plateau, transitional zones of the three-gradient terrain, and karst areas of southwest China, where ecosystems are extremely fragile and farmland is exceedingly limited. The bulks of these regions are deeply poverty-stricken at present and have been in poverty for a long time. Our CPL data suggest that only the SEI of CRNPSs is higher than that of CCPADs, and that all other indexes are lower than the case for NKCs and CCPADs and large gaps remain in both HDIs and CPLs (Table 4).

We applied Ward’s cluster analysis to divide the CRNPSs identified here into different categories. However, as we still have no clear understanding of the similarities between these groups, the number of clusters was set according to the hierarchical clustering diagram and four were eventually selected. The four types of aiding counties were then classified following the principle of leading factors consistency as key aiding counties restricted by multidimensional factors (Type I), aiding counties restricted by human development ability (Type II), aiding counties restricted by both natural resource endowments and socioeconomic development level (Type III), and aiding counties restricted by both human development abilities and socioeconomic development level (Type IV) (Figure 7). These categories respectively encompass 29.05%, 12.15%, 19.83%, and 38.97% of the total CRNPSs (Table 5).