Small area population estimation with high-resolution remote sensing and lidar - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 181

buildings attributes (building count, area, and volume) generally
yielded lower values than those derived from the reference
building footprint layer. This trend was most obvious for blocks
dominated by SF dwelling units. Therefore, population counts
were also underestimated from all the models. The sensitivity
of each model to the underestimation errors of explanatory
variables was then assessed through the MAE of population
estimates.
Figure 13.6 compares the MAE yielded by each model
when using the reference building layer and when using the
extracted buildings layer. The two curves are negatively correlated
( R 2

13.4.3.4 Population estimation
uncertainty analysis

The calibrated models were applied to the entire study area using
the extracted buildings layer and the classified land-use layers.
The goal here was to examine the effect of errors from building
detection and land-use classification to population estimation.
The reference building footprint layer was first replaced by
the automatically detected buildings layer, while the reference
land use layer was still used for assigning the type of dwelling.
Geometric attributes were calculated fromthe extracted buildings
layer and aggregated at census block level. Figure 13.5 shows the
scatter plot of estimated versus measured building attributes at
census block level. These scatter plots indicated that three detected

= 0 . 86, n = 7, p< 0.05), showing that the trend observed in
the model fitting stage (Fig. 13.3) is inverted due to building
detection errors alone. Although building volume fits block pop-
ulation better than building area, and building area fits better

256
128
64
32
16
8
4
2
1

32768

8192

Single-family

Bldg.

2048

Multi-family

512

Bldg.

128

45-deg Line

32

32 128 512 2048 8192

Actual Building Area [m 2 ]

32768

1248 6
Actual Building Count

32 64 128

131072

Legend

32768

Single-family Bldg.

8192

2048
512

Multi-family Bldg.

128 45-deg Line

128 512 2048 8192 32768 131072

Actual Building Volume [m 3 ]

FIGURE 13.5 Scatter plots of estimated versus reference building count, building area and building volume at census block
level. Axes are in a logarithmic scale.

Effect of Building Detection

30

25

20

15

True Bldg
DS Bldg

10

5

0

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7

FIGURE 13.6 Effect of building detection accuracy on population estimation.