Population Estimation Of Different Administrative Scales Based On DMSP/OLS Lighting Data

Population has always been a key and hot issue which the academic circle and decisionmaking departments pay attention to.Quickly and accurately obtaining basic population information and understanding the status quo and development trend of population not only play a decisive role in the formulation and improvement of the national economy and people's livelihood policy,but also have a significant and far-reaching impact on regional economic and social development and production.Therefore,establishing a more scientific,proper and dynamic population estimation model show important application value.With the rapid development of the remote sensing technology,remote sensing data sources have become an important auxiliary means of population estimation.Because of the unique ability of DMSP/OLS night lighting images in reflecting human social activities,they are widely used in human activity analysis and estimation of related socioeconomic indicators.The research content of this paper is based on the DMSP/OLS night lighting data at the provincial,municipal,and district(county)scales to establish a linear regression model for population estimation and a BP-neural-network-based demographic time series estimation model.In response to population overestimation and underestimation of in linear models,an improved model,linear regression overestimation/underestimation,is proposed.The results show that the BP-neural-network-based demographic time series estimation model is ideal for estimation on the three scales.On the provincial and municipal scales,the accuracy of population estimation is high and the accuracy fluctuation is small.Specifically,the relative error on the provincial scale is 2% to 3%,and that on the municipal scale is 1.5% to 3%.The relative error at the county level is 2% to 11%.Although the relative error of population estimation in 2012 fluctuates considerably,the estimation effect is still within the proper range.The improved overestimation/underestimation-targeted linear regression model only has a relatively stable estimation accuracy at the provincial scale.The average relative error is about 20%.However,the estimation results on the municipal scale in Sichuan Province are not ideal,but some scholars have used the same model to estimate the population on the municipal scale of Hubei Province and achieved ideal results,indicating that the location and economic development level are at the municipal scale.The effect on population distribution is more significant than that on the provincial level,and spatial heterogeneity is more prominent.To sum up,on the provincial,municipal and district scales,the estimation results of the linear regression model and the BP-neural-network-based demographic time series estimation model show that the smaller the scale,the more prominent the spatial heterogeneity is,and the greater the impact on the population estimation accuracy.However,compared with traditional statistical models,the dynamic learning ability of the BP neural network can be used to identify and fit the data of population fluctuation on different scales.Therefore,the BP-neural-network-based demographic time series estimation model is more suitable for multiscale population estimation.