Multi-source Data Based Spatial Model Of Population Distribution Using Random Forests

It has been a severe challenge for urban management that population are of excessively rapid growth and excessive agglomeration.To perform refined monitoring system of spatial distribution of population cannot be ignored.Shanghai is one of the largest cities in China,which is typical on population control and urban management.Therefore,it is very necessary to study on a refined monitoring system of the spatial distribution of population for this megacity.It is of great importance to establish a fine system for monitoring spatial distribution of population and understanding the rules of population spatial distribution,and it is of great theoretical and practical significance for urban scientific governance and optimum allocation of resources.Previous Studies have been establishing models at regional scales using statistics data,which are unable to explore the relationships between factors and population.What's more,studies have paid less attention to internal running mechanism and interpretation of models.In this study,on the basis of the fundamental results of previous research,we introduced a great volume of spatial data from nighttime light data,land use and POI(Points of Interest)data,and then extract many features related to spatial distribution of population.Next,a useful 500m gridded population model of spatial distribution is created and trained based on the random forests algorithm.Then using this model,we made an estimation for shanghai population,and got a population distribution grid at500m spatial resolution.In detail,this study has results in two aspects:(1)A refined population spatial model based on random forests using multi-source spatial data.The model worked well,for it gets R~2 at 77.4%,and the mean absolute error is about 217.40/grid.In addition,the examining results show that the model has good capability of generalization.(2)Using a random forests model interpreting method based on decision routine estimation,calculated feature contributions,and discussed the relation between the model and feature contributions.We found that some land use classes of residence,whose contributions are increasing with feature values increasing,are of great impact on model estimation results.Some POIs of functional services are playing an importance role on model estimation,such like POIs of medical care services and daily life services.Using the results of this thesis,people can get spatial distribution of a city conveniently and quickly,and we can provide powerful support for urban delicacy management and decision-making.Besides,the population model based on random forests can be generalized to other cities,and we can build up different scenes of urban patterns for future city population simulations.