Prediction of People Density Distribution

  • Objectives: using deep learning method to predict spatial-temporal distribution of people based on the Call Detail Record (CDR) dataset
  • Method: Use CDR to map the dynamic people density distribution by kernel density estimation (KDE) method, and then input to a convolution long short-term memory (ConvLSTM) model
  • Results: The mean absolute error of the predicted results of ConvLSTM ranged from 0.6 to 1.8 over 17 February 2015, which means that the model was much more stable and accurate than the other two baseline methods. Moran’s I index for the error distribution was still lower than that of the other baseline methods in space
  • Conclusions: the predicted density correlated much better with the original data at the temporal and spatial scales used when using ConvLSTM as compared to the other two methods, which do not consider the spatial autocorrelation.
Posted in Research interests, Spatial Intelligence Tagged with: