Population Estimation At The Building Level Based On Multi-source Data And Random Forest Algorithm

Building-level population data are of vital importance in disaster management,homeland security,and public health.One traditional way to acquire population distribution data is through population census.However,the population census is labor-intensive and time-consuming and thus constrains population investigations to a coarse temporal scale.Studies have thus turned to remote sensing techniques which conduct simultaneous observation over large areas in a cost-effective way to estimate population.Remotely sensed data,especially LiDAR data,which allows measures of three-dimensional morphological information,have shown to be capable for fine-scale population estimations.However,studies using LiDAR data for population estimation have noted a non-stationary relationship between LiDAR-derived morphological indicators and populations due to the unbalanced characteristic of population distribution.In this research,we proposed a framework to estimate population at the building-level by integrating both the three-dimensional morphological information derived from LiDAR data and the human activity information extracted from POI and NTL data.The main contents of this study include the following three points:(1)The LiDAR data and aerial photographs were used to extract individual building objects.A progressive morphological filter was used to extract the nDSM from the airborne LiDAR data.To eliminate the influence of vegetation and non-building objects,vegetation and non-building pixels were masked from the original nDSM using vegetation index and a height threshold value.Finally,the polyline compression algorithm was applied to regularize building boundaries.The overall accuracy of detection rate and matched overlay of building objects extracted was 92.58%and 94.28%,respectively.The results indicated a high accuracy of the building detection method in our study.(2)We proposed a framework for population estimation at the building-level by integrating both the three-dimensional morphological information derived from LiDAR data and the human activity information extracted from POI and NTL data.Based on the Luojia1-01 NTL data,POI data,and LiDAR data,33 features were preliminarily selected,including six NTL intensity features,two geometric features,and 25 POI features.Through a rigorous feature selection process,20 of the initial 33features were finally selected for the well-trained RF model.The trained RF model was applied to map the spatial distribution of population at the building-level.Subsequently,the accuracy of the model estimation results was evaluated and analyzed.The R~2 between the reference and estimated population was 0.65,indicating a satisfying overall accuracy.(3)The feature contributions are calculated separately for each feature to provide the detailed information of the relationships between features and the estimate population.Building volume was the most important feature to estimate population with an importance of 20.7%.The NTL radiance intensity was found to be less important than two POI-related features(i.e.,Sum＿POI＿a(residential)and Sum＿POI＿d(institutional)).From the perspective of feature contributions,we found that when using the random forest model to estimate population for each individual building,NTL radiance intensity has a positive effect on population estimation in residential areas,while shows a significant negative effect in office and commercial areas.This study proposes a new building-level population estimation framework,which improves the accuracy of building-level population estimation by fusing three-dimensional shape information extracted from LiDAR data with human activity information extracted from POI and night light data.