text
stringlengths
0
820
as training. For the network trained without any adapta-
tion, the mean IoU accuracy difference between rural and
urban images on the target domain is around 64% higher
than the difference reported in the single-domain task. Sim-
ilarly, all other fairness metrics show much higher differ-
ences between groups on transferring to the target domain.
Notably, when applying UDA methods, CBST and IAST,
transfer accuracy was improved, but at the cost of fairness
damage. These methods further enlarged the performance
gap between rural and urban groups measured in all four
metrics. These findings indicate a need for new domain
adaptation methods that tackle the challenge of fair trans-
fer learning.
One of the possible reasons why the network can bet-
ter adapt to urban images than rural images is the unequal
domain discrepancy. To estimate how similar the source
and target images are in the rural and urban groups, we use
two metrics – Proxy-A-distance (PAD) [12] and Maximum
mean discrepancy (MMD) [15]. Both measure the dissim-
ilarity between data distributions of different domains. We
randomly sample 100 images from each domain at a time
and ran 30 trials to compute the two measures. The mean
and standard deviation of distance across the trials are re-
ported in Table 4. We observe that the raw satellite imagesGroup Source Target PAD MMD
Urban Gulou Yuhuatai 0.26 0.207
Qinhuai Jintan (±0.12) (±0.0235)
Qixia
Jinghan
Rural Pukou Liuhe 0.64 0.262
Gaochun Huangpi (±0.21) (±0.0437)
Lishui
Jingxia
Table 4. Shift in raw image distribution. Measurement of
source-target domain distance using two metrics – Proxy-A-
distance (PAD) and Maximum mean discrepancy (MMD). The
implementation is based on the online codebase in [51]. Rural
images shift more than urban images.
Figure 4. Shifts in class distribution. Class distribution in terms
of proportion of pixels per class is plotted for urban images from
source domain (S-urban), urban images from target domain (T-
urban), rural images from source domain (S-rural), and rural im-
ages from target domain (T-rural). Class distribution is substan-
tially different for all subsets.
are overall more dissimilar between source and target dis-
tricts for rural than for urban, which is a likely cause of
the unequal transfer learning performance between the two
groups. Figure 1 illustrates example images from both do-
mains and the segmentation predictions our network made
on the target. We see that Buildings across source and tar-
get districts are of similar shape and arrangement for urban,
but they are disordered and dissimilar for rural. Accord-
ingly, the model segmentation map shows that the model
segments urban buildings well but fails to detect most of
the rural buildings. This observation indicates the impor-
tance of checking class differences besides overall differ-
ences across the whole image.
Along these lines, we define pixel-wise class distribu-
tion as the proportion of pixels belonging to each class. We
assess shifts in the class distribution between source and
target for both rural and urban images. Class distributions
are shown in Figure 4. For urban locations, the Water and
2922
Agriculture classes have the largest shifts from source to
target. For rural, the Background and Forest classes have
the largest shifts. Indeed there are large class shifts between
source and target for both the urban and rural data. For
example, the class distribution of urban target data seems
more rural-like with an increased proportion in the Agricul-
ture class. This emphasizes the internal variation in rural
and urban categories. Moreover, as our data consists of im-
ages from just one set of locations, data from different loca-
tions are needed for more generalizable conclusions. How-
ever, based on the selection of classes examined in our data
which are common land-use classes globally, some results
(such as in Table 4 indicate common threads that can be
applicable to rural-urban disparities in general.
In the second transfer learning task, the networks show
unfairness on rural to urban domain transfer. Differences
between rural and urban scenes provide explanation for
why almost all classes lost accuracy on the target do-
main. Some classes show opposite transfer performance
in the two sub-tasks of Task C. The Road class lost only
4% accuracy on Rural →Urban transfer but lost 41% ac-
curacy on Urban →Rural transfer. The Water class shows
similar patterns, whereas the Forest class lost 53% accu-
racy on Rural →Urban transfer but gains 33% accuracy on
Urban→Rural transfer. These observations indicate that for
some classes, the features learnt by the networks from rural
scenes can be easily adapted to interpret urban scenes, and
some classes have the opposite case. From this perspective,
features of different classes can have very different general-
ization ability, which will cause the transfer performance to
be highly unequal across classes. This feature-specific char-
acteristic may be leveraged in the design of future transfer
learning methods.
7. Conclusion
Transfer learning models for semantic segmentation are
often evaluated based on overall accuracy metrics. Here,
we expand the scope of their evaluation by conducting a