基于机器学习的贫困地区识别算法对比——以陕西省为例

doi:10.13249/j.cnki.sgs.2022.08.010

摘要/Abstract

摘要：

以传统社会经济指标为主导的贫困识别依赖于详尽的普查抽查数据，收集和处理不同质量和数量的普查抽查数据来研究区域贫困需要耗费大量的人力物力和时间，难以快速动态地监测贫困状态。然而时间分辨率高且客观易获取的夜间灯光数据可以在一定程度上弥补统计数据的劣势，即时地反映地表社会经济现象。机器学习算法能够从这些数据中学习出规律和模式，从中挖掘出潜在信息来识别贫困地区。基于陕西省NPP-VIIRS夜间灯光数据，通过构造多维统计变量，利用逻辑回归、支持向量机、K近邻、随机森林、决策树和梯度提升树6种监督分类算法识别贫困地区。结果表明从夜间灯光数据提取的多维特征能够更好的应用于贫困地区的识别，6种算法都能够准确的识别贫困地区，分类结果在空间上具有相似性，且表现出一定的地域性，分类准确度达到76.82%~83.20%。根据混淆矩阵进一步对比各个算法的特点，认为随机森林算法在误差偏移和分类精度等方面综合表现最佳。

关键词: 贫困识别, 机器学习, 算法对比, 特征工程, 夜间灯光数据

Abstract:

Improving the platform of monitoring poverty-returning and making full use of advanced technical means to optimize the accuracy of monitoring is crucial to consolidate the existing achievements and build a solid foundation for rural revitalization. Poverty identification led by traditional socio-economic indicators depends on detailed census data. Collecting and processing census data with different qualities and quantities to research regional poverty requires a lot of expenditures, and it is difficult to monitor poverty quickly and dynamically. However, the night light data with high time-resolution and easy obtaining can to a certain extent make up for the disadvantage of statistical data and reflect the surface socio-economic phenomena in real-time. Machine learning can learn rules and patterns from these data and mine potential information to identify poor areas. Based on this, taking Shaanxi Province as an example, six supervised learning classification algorithms, such as Logical Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest Classifier (RFC), Decision Tree (DT) and Gradient Boosting Decision Tree (GBDT), are used to identify poor areas based on the rich features of nighttime light data. As a result, it shows that: 1) the multidimensional variables have about 5% higher simulation accuracy than single variables, and the accuracy can lift 1%-5% after downscaling, which can be better applied to identifying poverty-stricken regions. In addition, these variables perform similarly in feature importance, which indicate that these features are used reasonably in these algorithms. 2) All six machine learning algorithms can precisely classify poverty-stricken regions based on certain prior knowledge with an accuracy of about 80%, also have a stronger spatial consistency and regionalism. 3) The classification accuracy, F1-score, and Kappa coefficient are used to compare the characteristics between these selected algorithms, and the RFC algorithm ranks in first, the KNN, GBDT, and LR performed moderately, and the SVM and DT get poorer performance. Poverty-stricken regions act as a key region of rural revitalization, so establishing a dynamic system to instantly monitor its relative poverty level in real-time is crucial to enhance the existing poverty alleviation achievements, achieve long-term stable poverty eradication, and build a solid foundation for rural revitalization. The purpose of this paper is to find an auxiliary means to achieve the identification of poverty-stricken regions in an efficient and real-time manner, so as to provide some references for using machine learning algorithms for the dynamic monitoring of poverty-stricken regions.

Key words: poverty identify, machine learning, algorithm comparison, feature engineering, nighttime light data

中图分类号:

K901.2

赵雨, 白宇, 员学锋. 基于机器学习的贫困地区识别算法对比——以陕西省为例[J]. 地理科学, 2022, 42(8): 1421-1432.

Zhao Yu, Bai Yu, Yuan Xuefeng. Comparison of Machine Learning Algorithms for Identifying Poverty-stricken Regions: A Case of Shaanxi[J]. SCIENTIA GEOGRAPHICA SINICA, 2022, 42(8): 1421-1432.

图/表 12

图1

图2

表1

表2

图3

表3

图4

表4

图5

图6

图7

表5

参考文献 41

[1]	李玉恒, 武文豪, 宋传垚, 等. 世界贫困的时空演化格局及关键问题研究[J]. 中国科学院院刊, 2019, 34(1): 42-50.
	Li Yuheng, Wu Wenhao, Song Chuanyao et al. Spatial-temporal pattern of world poverty reduction and key problems analysis. Bulletin of Chinese Academy of Sciences, 2019, 34(1): 42-50.
[2]	Undp O. Global multidimensional poverty index 2020[R/OL]. https://ophi.org.uk/multidimensional-poverty-index/.
[3]	Zhou Y, Liu Y S. The geography of poverty: Review and research prospects[J/OL]. Journal of Rural Studies, 2019. doi: 10.1016/j.jrurstud.2019.01.008
[4]	王小林, 冯贺霞. 2020年后中国多维相对贫困标准: 国际经验与政策取向[J]. 中国农村经济, 2020(3): 2-21.
	Wang Xiaolin, Feng Hexia. China’s multidimensional relative poverty standards in the post-2020 era: International experience and policy orientation. Chinese Rural Economy, 2020(3): 2-21.
[5]	Liu Y, Liu J, Zhou Y. Spatio-temporal patterns of rural poverty in China and targeted poverty alleviation strategies[J]. Journal of Rural Studies, 2017, 52: 66-75. doi: 10.1016/j.jrurstud.2017.04.002
[6]	周扬, 郭远智, 刘彦随. 中国县域贫困综合测度及2020年后减贫瞄准[J]. 地理学报, 2018, 73(8): 1478-1493. doi: 10.11821/dlxb201808007
	Zhou Yang, Guo Yuanzhi, Liu Yansui. Comprehensive measurement of county poverty and anti-poverty targeting after 2020 in China. Acta Geographica Sinica, 2018, 73(8): 1478-1493. doi: 10.11821/dlxb201808007
[7]	丁建军, 冷志明. 区域贫困的地理学分析[J]. 地理学报, 2018, 73(2): 232-247. doi: 10.11821/dlxb201802003
	Ding Jianjun, Leng Zhiming. Regional poverty analysis in a view of geography science. Acta Geographica Sinica, 2018, 73(2): 232-247. doi: 10.11821/dlxb201802003
[8]	王艳慧, 钱乐毅, 段福洲. 县级多维贫困度量及其空间分布格局研究——以连片特困区扶贫重点县为例[J]. 地理科学, 2013, 33(12): 1489-1497.
	Wang Yanhui, Qian Leyi, Duan Fuzhou. Multidimensional poverty measurement and spatial distribution pattern at the country scale: A case study on key country from national contiguous special poverty-stricken areas. Scientia Geographica Sinica, 2013, 33(12): 1489-1497.
[9]	刘小鹏, 李永红, 王亚娟, 等. 县域空间贫困的地理识别研究——以宁夏泾源县为例[J]. 地理学报, 2017, 72(3): 545-557. doi: 10.11821/dlxb201703014
	Liu Xiaopeng, Li Yonghong, Wang Yajuan et al. Geographical identification of spatial poverty at county scale. Acta Geographica Sinica, 2017, 72(3): 545-557. doi: 10.11821/dlxb201703014
[10]	丁建军. 中国11个集中连片特困区贫困程度比较研究——基于综合发展指数计算的视角[J]. 地理科学, 2014, 34(12): 1418-1427.
	Ding Jianjun. Comparative analysis on poverty degree of China’s 11 contiguous destitute areas: With view of comprehensive development index. Scientia Geographica Sinica, 2014, 34(12): 1418-1427.
[11]	武鹏, 李同昇, 李卫民. 县域农村贫困化空间分异及其影响因素——以陕西山阳县为例[J]. 地理研究, 2018, 37(3): 593-606.
	Wu Peng, Li Tongsheng, Li Weimin. Spatial differentiation and influencing factors analysis of rural poverty at county scale: A case study of Shanyang County in Shaanxi Province, China. Geographical Research, 2018, 37(3): 593-606.
[12]	施琳娜, 文琦. 相对贫困视角下的精准扶贫多维减贫效应研究——以宁夏彭阳县为例[J]. 地理研究, 2020, 39(5): 1139-1151. doi: 10.11821/dlyj020181292
	Shi Linna, Wen Qi. Effects of the targeted poverty alleviation policy on poverty reduction from the perspective of relative poverty in Loess Plateau. Geographical Research, 2020, 39(5): 1139-1151. doi: 10.11821/dlyj020181292
[13]	刘小鹏, 苏胜亮, 王亚娟, 等. 集中连片特殊困难地区村域空间贫困测度指标体系研究[J]. 地理科学, 2014, 34(4): 447-453.
	Liu Xiaopeng, Su Shengliang, Wang Yajuan et al. The index system of spatial poverty of village level to monitor in concentrated contiguous areas with particular difficulties. Scientia Geographica Sinica, 2014, 34(4): 447-453.
[14]	周扬, 李寻欢. 平原农区贫困地理格局及其分异机制——以安徽省利辛县为例[J]. 地理科学, 2019, 39(10): 1592-1601.
	Zhou Yang, Li Xunhuan. Geographical pattern and mechanism of poverty differentiation in plain areas: A case study of Lixin County, Anhui Province. Scientia Geographica Sinica, 2019, 39(10): 1592-1601.
[15]	刘艳华, 徐勇. 中国农村多维贫困地理识别及类型划分[J]. 地理学报, 2015, 70(6): 993-1007. doi: 10.11821/dlxb201506012
	Liu Yanhua, Xu Yong. Geographical identification and classification of multi-dimensional poverty in rural China. Acta Geographica Sinica, 2015, 70(6): 993-1007. doi: 10.11821/dlxb201506012
[16]	丁建军, 金宁波, 王璋, 等. 湘渝川滇黔桂地市城镇化减贫效应的异质性及演变特征分析[J]. 地理科学, 2021, 41(3): 522-533.
	Ding Jianjun, Jin Ningbo, Wang Zhang et al. Heterogeneity and evolution characteristics of poverty reduction effect of prefecture-level cities’ urbanization in Hunan, Chongqing, Sichuan, Yunnan, Guizhou and Guangxi. Scientia Geographica Sinica, 2021, 41(3): 522-533.
[17]	陈烨烽, 王艳慧, 王小林. 中国贫困村测度与空间分布特征分析[J]. 地理研究, 2016, 35(12): 2298-2308.
	Chen Yefeng, Wang Yanhui, Wang Xiaolin. Measurement and spatial analysis of poverty-stricken villages in China. Geographical Research, 2016, 35(12): 2298-2308.
[18]	刘彦随, 曹智. 精准扶贫供给侧结构及其改革策略[J]. 中国科学院院刊, 2017, 32(10): 1066-1073.
	Liu Yansui, Cao Zhi. Supply-side structural reform and its strategy for targeted poverty alleviation in China. Bulletin of Chinese Academy of Sciences, 2017, 32(10): 1066-1073.
[19]	Yu B, Shi K, Hu Y et al. Poverty evaluation using NPP-VIIRS nighttime light composite data at the county level in China[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2015, 8(3): 1217-1229. doi: 10.1109/JSTARS.2015.2399416
[20]	Hall G B, Malcolm N W, Piwowar J M. Integration of remote sensing and GIS to detect pockets of urban poverty: The case of Rosario, Argentina[J]. Transactions in GIS, 2001, 5(3): 235-253. doi: 10.1111/1467-9671.00080
[21]	Ghosh T, Anderson S, Elvidge C et al. Using nighttime satellite imagery as a proxy measure of human well-being[J]. Sustainability, 2013, 5(12): 4988-5019. doi: 10.3390/su5124988
[22]	Elvidge C D, Sutton P C, Ghosh T et al. A global poverty map derived from satellite data[J]. Computers and Geosciences, 2009, 35(8): 1652-1660. doi: 10.1016/j.cageo.2009.01.009
[23]	Wang W, Cheng H, Zhang L. Poverty assessment using DMSP/OLS night-time light satellite imagery at a provincial scale in China[J]. Advances in Space Research, 2012, 49(8): 1253-1264. doi: 10.1016/j.asr.2012.01.025
[24]	潘竟虎, 赵宏宇, 董磊磊. 基于DMSP-OLS数据和可持续生计的中国农村多维贫困空间识别[J]. 生态学报, 2018, 38(17): 6180-6193.
	Pan Jinghu, Zhao Hongyu, Dong Leilei. Spatial identification of multi-dimensional poverty in rural China by using nighttime light and sustainable livelihoods. Acta ecologica Sinica, 2018, 38(17): 6180-6193.
[25]	潘竟虎, 胡艳兴. 基于夜间灯光数据的中国多维贫困空间识别[J]. 经济地理, 2016, 36(11): 124-131.
	Pan Jinghu, Hu Yanxing. Spatial identification of multidimensional poverty in China based on nighttime light remote sensing data. Economic Geography, 2016, 36(11): 124-131.
[26]	Jean N, Burke M, Xie M et al. Combining satellite imagery and machine learning to predict poverty[J]. Science, 2016, 353(6301).
[27]	Zhao X, Yu B, Liu Y et al. Estimation of poverty using random forest regression with multi-source data: A case study in Bangladesh[J/OL]. Remote Sensing, 2019, 11(4). https://www.mdpi.com/ 2072-4292/11/4/375/htm.
[28]	Li G, Chang L, Liu X et al. Monitoring the spatiotemporal dynamics of poor counties in China: Implications for global sustainable development goals[J]. Journal of Cleaner Production, 2019, 227: 392-404. doi: 10.1016/j.jclepro.2019.04.135
[29]	樊杰, 周侃, 伍健雄. 中国相对贫困地区可持续发展问题典型研究与政策前瞻[J]. 中国科学院院刊, 2020, 35(10): 1249-1263.
	Fan Jie, Zhou Kan, Wu Jianxiong. Typical study on sustainable development in relative poverty areas and policy outlook of China. Bulletin of Chinese Academy of Sciences, 2020, 35(10): 1249-1263.
[30]	王永生, 文琦, 刘彦随. 贫困地区乡村振兴与精准扶贫有效衔接研究[J]. 地理科学, 2020, 40(11): 1840-1847.
	Wang Yongsheng, Wen Qi, Liu Yansui. Achieving effective connection between rural revitalization and targeted poverty alleviation in poverty-stricken regions. Scientia Geographica Sinica, 2020, 40(11): 1840-1847.
[31]	刘彦随. 中国新时代城乡融合与乡村振兴[J]. 地理学报, 2018, 73(4): 637-650. doi: 10.11821/dlxb201804004
	Liu Yansui. Research on the urban-rural integration and rural revitalization in the new era in China. Acta Geographica Sinica, 2018, 73(4): 637-650. doi: 10.11821/dlxb201804004
[32]	Kraay A, Mckenzie D. Do poverty traps exist?[J]. Social Science Electronic Publishing, 2014, 28(3): 127-48.
[33]	刘彦随, 周扬, 刘继来. 中国农村贫困化地域分异特征及其精准扶贫策略[J]. 中国科学院院刊, 2016, 31(3): 269-278.
	Liu Yansui, Zhou Yang, Liu Jilai. Regional differentiation characteristics of rural poverty and targeted poverty alleviation strategy in China. Bulletin of Chinese Academy of Sciences, 2016, 31(3): 269-278.
[34]	周扬, 李寻欢. 贫困地理学的基础理论与学科前沿[J]. 地理学报, 2021, 76(10): 2407-2424.
	Zhou Yang, Li Xunhuan. Basic theories and frontiers of poverty geography[J]. Acta Geographica Sinica, 2021, 76(10): 2407-2424.
[35]	胡为安, 刘传立, 詹淇雯. 中国区域NPP-VIIRS年度夜间灯光数据的合成方法与对比验证[J]. 桂林理工大学学报, 2021, 41(1): 141-148. doi: 10.3969/j.issn.1674-9057.2021.01.018
	Hu Weian, Liu Chuanli, Zhan Qiwen. Synthetic method and comparison verification of NPP-VIIRS annual nighttime lighting data in China. Journal of Guilin University of Technology, 2021, 41(1): 141-148. doi: 10.3969/j.issn.1674-9057.2021.01.018
[36]	杨剑锋, 乔佩蕊, 李永梅, 等. 机器学习分类问题及算法研究综述[J]. 统计与决策, 2019, 35(6): 36-40.
	Yang Jianfeng, Qiao Peirui, Li Yongmei et al. A review of machine-learning classification and algorithms. Statistics & Decision, 2019, 35(6): 36-40.
[37]	Eray Sevgen, Sultan Kocaman, A Nefeslioglu H et al. A novel performance assessment approach using photogrammetric techniques for landslide susceptibility mapping with logistic regression, ANN and random forest[J]. Sensors (Basel, Switzerland), 2019, 19(18): 3940 doi: 10.3390/s19183940
[38]	Garosi Y, Sheklabadi M, Conoscenti C et al. Assessing the performance of GIS-based machine learning models with different accuracy measures for determining susceptibility to gully erosion[J]. Science of the Total Environment, 2019, 664: 1117-1132. doi: 10.1016/j.scitotenv.2019.02.093
[39]	吕倩, 刘海滨. 基于夜间灯光数据的黄河流域能源消费碳排放时空演变多尺度分析[J]. 经济地理, 2020, 40(12): 12-21.
	Lyu Qian, Liu Haibin. Multiscale spatio-temporal characteristics of carbon emission of energy consumption in Yellow River Basin based on the nighttime light datasets. Economic Geography, 2020, 40(12): 12-21.
[40]	曾冰. 基于NPP/VIIRS夜间灯光数据的湘鄂赣省际交界区县域经济空间格局及影响因素[J]. 地理科学, 2020, 40(6): 900-907.
	Zeng Bing. Spatial pattern evolution and influencing factors of county-level economy of border regions in Hunan-Hubei-Jiangxi based on nighttime light data. Scientia Geographica Sinica, 2020, 40(6): 900-907.
[41]	Joharestani M Z, Cao C X, Ni X L et al. PM_2.5 prediction based on random forest, XGBoost, and deep learning using multisource remote sensing data[J]. Atmosphere 2019, 10(7): 19

分类器名称	简称	优点	缺点
逻辑回归	LR	线性关系拟合效果好	易发生欠拟合现象
随机森林	RFC	对特征要求低	噪声较大的数据上会过拟合
支持向量机	SVM	可用于小样本，非线性分类	核函数的选择较复杂，对缺失数据敏感
K近邻	KNN	对数据的分布情况无要求	样本不平衡时，预测偏差比较大
决策树	DT	结构简单，分类规则易理解	易发生过拟合现象
梯度提升树	GBDT	灵活处理各种类型的数据	难以并行训练数据

序号	指标	计算方法
注：DN_i为第i个像元灯光亮度(DN)值，N为县域内所有像元数， $ {DN}_{i}^{\ne 0} $ 为第i个非零像元DN值， $ {N}^{\ne 0} $ 为县域内所有非零像元数， ${ {DN}_\text{max} }^{\ne 0}$ 为县域内非零像元的最大DN值， ${ {DN}_\text{min} }^{\ne 0}$ 为非零像元最小DN值。 $ {Z}_{i}={y}_{i}-\bar{y} $ ， $ {Z}_{j}={y}_{j}-\bar{y} $ ，Z_i 和Z_j 为样本i和j 与平均水平的偏差， y_i 和 y_j 为县域i和j内所有像元平均DN值， $ \bar{y} $ 为样品总平均值， $ {S}^{2} $ 为所有县DN值的标准差， $ {w}_{ij} $ 为空间权重值。
F₁	县域内所有像元平均DN值	$ {\displaystyle\sum }_{i=1}^{N}{DN}_{i}/N $
F₂	县域内所有像元DN值总和	$ {\displaystyle\sum }_{i=1}^{N}{DN}_{i} $
F₃	县域内自然对数变换的平均DN值	$ {\displaystyle\sum }_{i=1}^{N}\mathrm{l}\mathrm{n}\left({DN}_{i}\right)/N $
F₄	县域内非零像元平均DN值	$ {\displaystyle\sum }_{i=1}^{N}{DN}_{i}^{\ne 0}/{N}^{\ne 0} $
F₅	县域内非零像元个数/所有像元个数	$ {N}^{\ne 0}/N $
F₆	县域内非零像元最大DN值	${ {DN}_\text{max} }^{\ne 0}$
F₇	县域内非零像元最小DN值	${ {DN}_\text{min} }^{\ne 0}$
F₈	县域内非零像元DN值极差	${ {DN}_\text{max} }^{\ne 0}-{ {DN}_\text{min} }^{\ne 0}$
F₉	县域内非零像元DN值标准差	$\surd {\displaystyle\sum }_{i=1}^{ {N}^{\ne 0} }{({DN}_{i}^{\ne 0}-{DN}_\text{mean}^{\ne 0})}^{2}/{(N}^{\ne 0}-1)$
F₁₀	县域内平均DN值局部莫兰指数	$\dfrac{ {Z}_{i} }{ {S}^{2} }\displaystyle\sum\nolimits _{j\ne i}^{n} {w}_{ij}{Z}_{j}$

	F1	F2	F3	F4	F5	F6	F7	F8	F9	F10
注：^**表示在5%水平下显著，黑体字为最大值、最小值标注；F1~F10解释见表2。
F1	1
F2	0.459^**	1
F3	0.855^**	0.587^**	1
F4	0.983^**	0.513^**	0.915^**	1
F5	0.834^**	0.593^**	0.895^**	0.838^**	1
F6	0.461^**	0.843^**	0.540^**	0.527^**	0.481^**	1
F7	0.842^**	0.142	0.655^**	0.818^**	0.612^**	0.224^**	1
F8	0.414^**	0.844^**	0.506^**	0.483^**	0.449^**	0.998^**	0.165	1
F9	0.710^**	0.768^**	0.826^**	0.791^**	0.703^**	0.830^**	0.377^**	0.816^**	1
F10	0.951^**	0.276^**	0.729^**	0.922^**	0.685^**	0.345^**	0.940^**	0.290^**	0.542^**	1

真实情况	预测结果
真实情况	正例（贫困县）	反例（非贫困县）
正例（贫困县）	TP（真正例，预测准确）	FN（假反例，预测偏高）
反例（非贫困县）	FP（假正例，预测偏低）	TN（真反例，预测准确）

ID	P/%	R/%	F_score/%	Kappa系数	运行时间/s
注：P为查准率，R为召回率，F_score为调和平均；分类器解释见表1。
RFC	82.69	82.69	82.69	0.66	0.27
LR	74.60	90.38	81.74	0.61	0.13
SVM	71.64	92.31	80.67	0.57	0.35
KNN	83.33	76.92	80.00	0.63	0.11
GBDT	79.25	80.77	80.00	0.61	0.25
DT	72.88	82.69	77.48	0.53	0.16