Table 1.
Datasets and number of data points.
Fig 1.
Visual representation of the LSTM model graph.
Fig 2.
Visual representation of the GAN model graph for prediction.
Fig 3.
Visual representation of the Generator and Discriminator models of the GAN.
Regularization (Dropout and Batch Normalization) and activation (LeakyReLU) layers are not shown for simplicity.
Table 2.
The values in bold are the lowest on each column.
Table 3.
Results summary (N-ADE/N-FDE).
The values in bold are the lowest on each column.
Table 4.
GAN-1 / GAN-3 ratio: The higher the ratio, the higher is the improvement of GAN-3 over GAN-1, meaning that better solutions are found by the former (a sign of multimodality).
Table 5.
N-FDE / N-ADE ratio: The higher the ratio, the faster the error grows (a sign of multimodality).
The values in bold are the lowest on each column, the ones where the error is the least affected by multimodality.
Fig 4.
Legends for trend on number of positions (a) and for hot spot analysis (b).
Fig 5.
Visual comparison between real and predicted (LSTM) positions in Milan (top left), Naples (top right), Rome (bottom left) and Turin (bottom right).
Republished from ESRI and HERE under a CC BY license, with permission from ESRI, original copyright 2021.
Table 6.
Position distance statistics (m) measured in the 4 cities.
Fig 6.
Visualization of space-time data cube with 500 m resolution using the 12 points prediction over Milan area.
Republished from ESRI and HERE under a CC BY license, with permission from ESRI, original copyright 2021.
Fig 7.
Visualization of space-time data cube with 500 m resolution using the 12 points prediction over Naples area.
Republished from ESRI and HERE under a CC BY license, with permission from ESRI, original copyright 2021.
Fig 8.
Visualization of space-time data cube with 500 m resolution using the 12 points prediction over Rome area.
Republished from ESRI and HERE under a CC BY license, with permission from ESRI, original copyright 2021.
Fig 9.
Visualization of space-time data cube with 500 m resolution using the 12 points prediction over Turin area.
Republished from ESRI and HERE under a CC BY license, with permission from ESRI, original copyright 2021.
Table 7.
Variation in % of the cells in each trend class in the 4 cities computed using the 500 m aggregation: Positive values are classes where real values have more occurrences than predicted ones.
In brackets the real and the predicted percentage values: for the total number of cells with at least one value, the real and the predicted total numbers are reported in brackets.
Table 8.
Variation in % of the cells in each trend class in the 4 cities computed using the 100 m aggregation: Positive values are classes where real values have more occurrences than predicted ones.
In brackets the real and the predicted percentage values: for the total number of cells with at least one value, the real and the predicted total numbers are reported in brackets.
Table 9.
Overall accuracy (%) measured on the count trends in the 4 cities.