text
stringlengths
0
820
(S-overall) and target (T-overall), urban data from the source (S-
urban) and target (T-urban), rural data from the source (S-rural)
and target (T-rural) is plotted when transferring models across dis-
tricts. No adaptation is the source-only method, CBST and IAST
are UDA methods. While overall accuracy drops on transfer, UDA
methods have smaller accuracy drop.
2920
Sub-task Method Mean Class-std Worst Sorted 30% (bottom, top)
No adaptation 0.437 0.108 0.301 (0.322, 0.566)
Rural→Urban CBST 0.469 0.123 0.326 (0.332 , 0.617)
IAST 0.443 0.175 0.205 (0.211, 0.638 )
No adaptation 0.426 0.108 0.226 (0.271, 0.531)
Urban→Rural CBST 0.467 0.129 0.228 (0.283, 0.599 )
IAST 0.454 0.120 0.229 (0.307 , 0.592)
Table 3. Task C: Evaluation of urban-rural transfer. Three methods (No adaptation, CBST, IAST) are trained on rural districts and
evaluated on urban districts, and vice versa. Results with the most improvements are marked in bold. UDA methods improve Mean IoU
compared to No adaptation but increase standard deviation of IoUs across classes.
from top 30% classes than urban results, but higher class-
wise standard deviation. Overall, we observe rural-urban
disparities in all four metrics.
For Task B on transfer learning across districts, we sum-
marize the results in Figure 2 and Table 2. Figure 2 visual-
izes the mean IoU metric for both source and target districts.
Based on the first two bar plots (dark and light orange) for
No adapation, CBST, and IAST, we conclude that UDA
methods improve overall segmentation accuracy on the tar-
get domain (T-overall) compared to No adaptation (a de-
crease of 22% and 24% vs that of 26%). Similar trend is ob-
served for each of the source-target pairs for rural and urban
separately. However, the performance gap between the ru-
ral and urban data from target (T-rural and T-urban) remains
large. For instance, from Table 2 we observe that CBST ob-
tains mean IoU of 0.523 on urban area which is better than
the ”No adaptation” 0.486, and IAST obtains 0.376 on rural
area better than the ”No adaptation” 0.364. However, the
networks remain unfair across rural-urban groups after the
transfer (large values in the rural −urban columns). UDA
methods further lower fairness: CBST increases the differ-
ence of mean IoU between urban and rural by 22% (-0.122
to -0.149), increases the difference of standard deviation by
69% (0.065 to 0.11), and increases the difference of worst-
performing class’ IoU by 89% (-0.183 to -0.345).
Next, we examine Task C on transfer learning from ur-
ban domain to rural domain and vice versa. Results are
summarized in Table 3. For both the transfer directions,
the two UDA methods improve the overall accuracy, shown
as higher mean IoU, higher IoU on the worst-preforming
class, and higher mean IoU on bottom and top 30% classes.
However, compared to ”No adaptation”, UDA methods dis-
perse model performance across the classes, measured by
higher standard deviation. For example, CBST increases
Class-std from 0.108 to 0.123 on rural to urban transfer and
from 0.108 to 0.129 on urban to rural transfer. Looking
more closely into the CBST method which obtains the best
overall transfer accuracy (0.469 and 0.467), its performance
on each class is visualized in Figure 3. We observe that
the IoU changes for each class upon transfer are highly un-
Figure 3. Task C: Mean IoU upon transfer across rural-urban.
Mean IoU for 7 landscape classes on source and target domain
when transferring from rural area to urban area, and from ur-
ban area to rural area, with the UDA method CBST. Performance
changes vary substantially by class.
equal. For example, in transferring from Rural →Urban, the
network retains accuracy on the Water and Road classes, but
lost significant accuracy ( 53%) on the Forest class. In trans-
ferring from Urban →Rural, accuracy drops significantly on
Road, Water, and Barren classes, but increases by 33% on
the Forest class.
2921
6. Discussion
For the locations included in this study, segmentation re-
sults showed a disparity in performance between rural and
urban areas. Though the two groups obtain similar accu-
racy on the respective bottom 30% performing land-cover
classes, rural areas obtain better accuracy on the top 30%
performing classes. Moreover, performance distribution
across classes are different between rural and urban im-
ages. Specifically, the segmentation model detects Forest
and Agriculture classes well in their rural form, and detects
Road and Water classes well in their urban form (detailed
results are reported in Table 5 in the Appendix). Due to ur-
banization, rural and urban areas have clear landscape dif-
ferences. For example, roads are typically wider in the ur-
ban scenes and narrower in rural scenes and water takes on
larger shapes like lakes in urban scenes, and smaller shapes
like ditches in rural scenes [52]. This may explain why the
networks show advantages in urban images on Road and
Water classes. Moreover, agricultural land covers large area
and is continuously distributed in rural scenes. The per-
centage of pixels with Agriculture and Forest elements is
also higher [52]. This can facilitate learning on these two
classes in rural areas as compared to urban areas.
We considered two practical transfer learning tasks with
satellite images and assess network fairness while trans-
ferring across geographical locations, and across rural and
urban areas. Broadly, we observed that when transferring
across districts, networks made more unfair predictions on
data from the new domain than data from the same domain