地球信息科学学报
地毬信息科學學報
지구신식과학학보
GEO-INFORMATION SCIENCE
2014年
3期
435-442
,共8页
村级人口%空间分布%GWR模型%格网方法%BP神经网络%精度分析
村級人口%空間分佈%GWR模型%格網方法%BP神經網絡%精度分析
촌급인구%공간분포%GWR모형%격망방법%BP신경망락%정도분석
village-level census%spatial distribution%geographically weighted regression model%grid method%BP neural network%accuracy analysis
在人口分布及其相关研究中，常常会遇到小尺度人口数据部分缺失的问题。本文以湖北省鹤峰县为例，在分析土地利用与人口分布关系的基础上，从全局与局部、线性回归与非线性回归考虑，基于土地利用类型，分别利用地理加权回归（GWR）方法、格网方法、BP神经网络方法对缺失数据的行政村人口数据进行模拟，并进行了多角度精度对比验证。研究结果表明：（1）各种土地利用类型中，耕地、林地、城镇村及工矿用地、交通用地是影响研究区村级人口分布的主要因素；（2）30个调查村中，3种方法模拟的人口总数误差小于3%，通过每个村的模拟值与实际值相比，BP神经网络方法能更好地模拟研究区村级人口的分布,格网方法次之,GWR方法最差；（3）研究区各村人口分布呈现较高的空间正相关性，各乡镇的人口密度在空间上并不独立，而是呈现紧密的集聚特征。
在人口分佈及其相關研究中，常常會遇到小呎度人口數據部分缺失的問題。本文以湖北省鶴峰縣為例，在分析土地利用與人口分佈關繫的基礎上，從全跼與跼部、線性迴歸與非線性迴歸攷慮，基于土地利用類型，分彆利用地理加權迴歸（GWR）方法、格網方法、BP神經網絡方法對缺失數據的行政村人口數據進行模擬，併進行瞭多角度精度對比驗證。研究結果錶明：（1）各種土地利用類型中，耕地、林地、城鎮村及工礦用地、交通用地是影響研究區村級人口分佈的主要因素；（2）30箇調查村中，3種方法模擬的人口總數誤差小于3%，通過每箇村的模擬值與實際值相比，BP神經網絡方法能更好地模擬研究區村級人口的分佈,格網方法次之,GWR方法最差；（3）研究區各村人口分佈呈現較高的空間正相關性，各鄉鎮的人口密度在空間上併不獨立，而是呈現緊密的集聚特徵。
재인구분포급기상관연구중，상상회우도소척도인구수거부분결실적문제。본문이호북성학봉현위례，재분석토지이용여인구분포관계적기출상，종전국여국부、선성회귀여비선성회귀고필，기우토지이용류형，분별이용지리가권회귀（GWR）방법、격망방법、BP신경망락방법대결실수거적행정촌인구수거진행모의，병진행료다각도정도대비험증。연구결과표명：（1）각충토지이용류형중，경지、임지、성진촌급공광용지、교통용지시영향연구구촌급인구분포적주요인소；（2）30개조사촌중，3충방법모의적인구총수오차소우3%，통과매개촌적모의치여실제치상비，BP신경망락방법능경호지모의연구구촌급인구적분포,격망방법차지,GWR방법최차；（3）연구구각촌인구분포정현교고적공간정상관성，각향진적인구밀도재공간상병불독립，이시정현긴밀적집취특정。
The problem that population data is usually missing in small scale areas such as administrative villag-es which are always mentioned in population distribution studies and related researches. In this context, we took the Hefeng County in Hubei Province as the study area and analyzed the correlation between land use type index and population density. The simulation of the village-level population distribution is performed using Geographi-cally Weighted Regression (GWR) method, grid method and BP neural network method respectively. Then, from the perspective of global-local and linear-nonlinear, the comparative precision validation was taken to verify the simulation accuracy of the population in villages with missing population data, which utilizes cross-validation method between the simulated population and the actual population. Results show that:(1) in all kinds of land use types, the main factors affecting population distribution are farmland, woodland, urban industrial land, and transportation land;(2) with regard to the three simulation methods we concerned, the errors of the simulated to-tal population using these methods are all less than 3%for the 30 invested villages. By comparing the ratios of estimated values to the actual values of population in each village, and taking 10%as the tolerance, the reliability of GWR method is 43.33%, while grid method is 76.67%and BP neural network is 86.67%. It shows that the BP neural network method is the optimal method among the three methods for the study area, and grid method is better than GWR method. In addition, the prediction accuracy of nonlinear regression is higher than that of linear regression; (3) population spatial distribution in the study area shows a high spatial positive correlation and a“high-high”agglomeration type which is also the main type in the study area;moreover, it shows that the popu-lation densities of the county are not spatially independent but intensively agglomerated.
