Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
vise three additional metrics that quantify how accuracy is
distributed across the classes. The metrics have been used
in existing fairness analysis of segmentation models [19].
These are:1.Class-std . Standard deviation of IoUs across the 7
classes;
2.Worst . IoU of the worst-performing class (Worst);
3.Sorted 30% (bottom, top) . Mean of the bottom 30%
and top 30% classes of the sorted class IoUs. In our
case, 30% is 2 classes out of 7.
Next, we describe the setup for the two transfer tasks.
4.2. Transfer learning
As mentioned earlier, we consider the setting of unsu-
pervised domain adaption (UDA) that is we have a single
task (image segmentation) on the two domains. We assume
access to images and labels for the source (train) domain
and only the images for the target (test) domain. This is
a practical setting in satellite imagery since collecting im-
ages is inexpensive due to advancements in remote sensing,
however, annotating the segmentation labels is expensive.
Thus, we want to be able to use the labelled source images
to segment known but as yet unlabelled target domain.
We consider two UDA methods which performed the
best on the LoveDA benchmark [52] – class-balanced self-
training ( CBST ) [59] and instance adaptive self-training
(IAST ) [33]. CBST optimizes the generation of pseudo-
labels used during self-training to be more balanced among
the classes by using class-wise confidence scores. IAST
adaptively adjusts the pseudo-label generation to improve
the diversity of pseudo-labels and saves useful information
from hard instances. We also use a natural method which
ignores transfer learning and trains only with data from the
source domain ( No adaptation ).
For all transfer learning experiments, we use a
DeepLabV3+ [7] network with ResNet50 encoder. An
Adam optimizer is used with a momentum of 0.9. The batch
size is 8 and the total training iterations are 15000. Other ex-
perimental setup parameters are the same as in the semantic
segmentation task.
4.2.1 Task B - Transfer across districts
First, we consider the scenario where a model is transferred
across different geographical locations, which in our case
are administrative districts. Source domain comprises of
8 districts: Gulou, Qinhuai, Qixia, Jinghan (urban), and
Pukou, Gaochun, Lishui, Jingxia (rural); and Target do-
main has 4 districts: Yuhuatai, Jintan (urban), and Liuhe,
Huangpi (rural). The ”No adaptation” and two UDA meth-
ods (CBST, IAST) are applied to train the network on the
source domain, and are tested on urban and rural images
from the unseen target domain, separately. The same accu-
racy and fairness metrics listed in Section 4.1.2 are used for
evaluation.
2919
Mean Class-std
Model rural urban rural−urban (%) rural urban rural−urban (%)
UNet 0.639 0.595 0.044 (6.9%) 0.106 0.0946 0.0114 (10.8%)
DeepLabV3+ 0.632 0.597 0.035 (5.5%) 0.0982 0.0896 0.0086 (8.8%)
Worst Sorted 30% (bottom, top)
Model rural urban rural−urban (%) rural urban rural−urban (%)
UNet 0.453 0.474 −0.021 (-4.4%) (0.491, 0.742) (0.480, 0.705) (0.011 , 0.037)(2.2%, 5.0%)
DeepLabV3+ 0.473 0.473 0 (n/a) (0.504, 0.740) (0.489, 0.706) (0.015, 0.034)(3.0%, 4.6%)
Table 1. Task A: Evaluation on single-domain semantic segmentation. Two networks, UNet and DeepLabV3+, are tested on rural and
urban districts from the same domain as training set. For metrics, Mean, Class-std, and Worst, the better performing group (between rural
and urban) is in bold. The difference in performance between rural and urban is shaded. Typically, performance is better for rural than
urban.
Mean Class-std
Method rural urban rural−urban (%) rural urban rural−urban (%)
No adaptation 0.364 0.486 −0.122 ( −25%) 0.200 0.135 0.065 (33%)
CBST 0.374 0.523 −0.149 ( −28%) 0.215 0.105 0.110 (51%)
IAST 0.376 0.493 −0.117(−24%) 0.223 0.135 0.088 (39%)
Worst Sorted 30% (bottom, top)
Method rural urban rural−urban (%) rural urban rural−urban (%)
No adaptation 0.0609 0.244 −0.183 ( −75%) (0.098, 0.581) (0.317, 0.630) (−0.219,−0.049) ( −69%,−7.7%)
CBST 0.0172 0.362 −0.345 ( −95%) (0.0943, 0.609) (0.398, 0.647) (−0.304,−0.038) ( −76%,−5.9%)
IAST 0.0304 0.232 −0.202 ( −87%) (0.0772, 0.598) (0.327, 0.640) (−0.250,−0.042) ( −76%,−6.6%)
Table 2. Task B: Evaluation of transfer across districts. Three methods (No adaptation, CBST, IAST) are trained source districts, and
evaluated on target rural and target urban districts. For the metrics, Mean, Class-std, and Worst, the better performing group (between rural
and urban) is in bold. The differences in performance between rural and urban are shaded. Models have high unfairness upon transfer.
4.2.2 Task C - Transfer across urban and rural areas
Second, we consider the scenario where the segmentation
model is transferred either from urban to rural areas or from
rural to urban areas. The source and target domain consists
of data from the same set of districts. So for this task, the
only source for domain discrepancy is rural and urban dis-
crepancy. The no- adaptation method and two UDA meth-
ods are trained on the source domain, and tested on the tar-
get domain. Evaluation metrics are the same as earlier.
5. Results
We summarize results for the single-domain in Table 1.
Both the networks (UNet, DeepLabV3+) have better over-
all accuracy, shown with Mean IoU over the 7 classes, for
the rural districts compared to the urban districts. Fairness
metrics such as IoU of the worst class and mean IoU of
30% bottom classes are comparable between rural and ur-
ban. The worst class is Barren for both rural and urban. The
30% bottom classes include Barren and Road for rural, and
Barren and Forest for urban (see Table A.1 in Appendix
for class-wise results). Rural results show higher mean IoU
Figure 2. Task B: Mean IoU upon transfer across districts.
Mean IoU on the union of rural and urban data from the source