Multi-scale Modeling Of Population Spatialization In Changsha Based On Multi-Source Data

Accurate and high-resolution population spatial distribution information has an important reference value for urban planning,disaster assessing,disease preventing,and precision poverty alleviating.However,the traditional researches on population spatialization have the disadvantages including of single model,single modeling cofactor,single grid scale and low accuracy of results.In particularly,the lack of both comparison studies of multi-scale models of population spatialization results comparison studies and correlation analysis between multi-source co-factor data,is likely to cause problems such as multi-collinearity of the models and application limitations of population spatialization products.In view of the above problems,the author in this paper selects four types of data including land-use,nighttime light(NPP-VIIRS DNB),digital-elevation-model(DEM)and point-of-interest(POI)data as modeling co-factors.By correlation analysing and least squares analysing,the factors and optimal factor combinations for modeling that have significant effects on population spatial distribution are gathered.Based on the previous experiences in selection of population spatial grid scale in the city areas,the author used three methods such as Multiple Linear Regression(MLR),Geographic Weighted Regression(GWR)and Spatial Lag Model(SLM)to establish the population spatialization models of Changsha in Hunan province,P.R.C.at fourgrid scales including 1km×1km,500m×500m,250m×250m and 100m×100m(referred to as 1km,500 m,250m,100m)grid cells.According to the accuracy testing and comparative analysing,the pros and cons of the three models were obtained in this paper.At the same time,according to the results of accuracy analysis,the 250 m grid scale of population simulated data by the GWR model was selected as the sample data for interpolation by both Ordinary Kriging(O-Kriging)and Regularized Spline(R-Spline)methods to refine the100 m grid scale of population spatialization scale,that is,the sample date was interpolated to 100 m scale grid population,and accuracy testing and comparative analysing were performed.The results showed that:(1)The overall pattern of population aggregation in Changsha was presented as “one large aggregation in the center,and two small aggregations on the left and right sides of the center”.That is that population were tended to high-high aggregation state in economically developed urban areas and there were also small areas with high-high aggregation state in counties(cities)close to the main urban areas.But population were population tended to low-low aggregation state in rural areas with complex topography and far away from urban areas.At the same grid scale,the simulation results from both MLR and GWR models were slightly larger than that from SLM model in the high-density area of urban population,and the simulation results from both GWR and SLM models were more detailed than that from MLR model in the areas with moredeveloped river systems and higher elevations.The detailed level of population spatial information display was from GWR model,SLM model,and MLR model in order from high to low.(2)In the population spatialization results from the three models at different scales,the average relative error variation trend in the township-level tests was different.The accuracy of the MLR model did not increase with the decrease of the scale and the error at the 250 m scale was the smallest.But the errors of both GWR and SLM models decreased as the decrease of the scales and the two models had the smallest error at the 100 m scale.Among them,the accuracy of SLM model at 250 m and 100 m scales ware similar.The average relative errors of the community-level tests of the spatialization results from the three models at different scales were the same as that of the township-level tests.At the same scale,the average errors of the community-level tests for the three models were higher than that of the township-level tests.On the whole,the GWR model had the highest accuracy,followed by the SLM model,and the MLR model had the lowest accuracy.(3)In the two kinds of spatial interpolation simulation results from both R-Spline and O-Kriging methods,the average relative errors of the 100 m scale of R-Spline simulation population in both township-level tests and community-level tests were higher than the 250 m scale of GWR model simulation result and were lower than that of the 100 m scale of GWR modelsimulation result.However,the average relative errors of O-Kriging interpolation in the township-level and community-level tests were higher than that of both 250 m and 100 m scale of the GWR model simulation results.That is that the accuracy of the R-Spline interpolation result was lower than the simulation results of the same-scale GWR model,but to a certain extent,it improved both of the accuracy and resolution of the large-scale of grid population data from the GWR model and could be used to obtain more accurate population data quickly and easily.