text
stringlengths
0
820
In this work we study the fairness impacts of transfer
learning with satellite imagery . To accomplish this goal,
we test multiple semantic segmentation models across dif-
ferent geographies. We then assess if such models made
fair predictions on both the source and the target data. We
focus on differences between urban and rural areas (i.e. ur-
ban/rural categorization is the sensitive attribute) due to per-
sistent and striking disparities between urban and rural ar-
eas, especially for poor populations [2, 3]. The unfairness
criterion in this work is based on differences in error rates
across protected groups where the error rate is computed
using Intersection-over-Union (IoU), a standard segmenta-
tion metric. We also examine model performance disparity
across different land-cover classes. Results show that ex-
isting domain adaptation methods do not maintain fairness
properties on transfer, either across protected groups or fea-
ture classes. This work serves as a valuable demonstration
of fairness being an critical issue in transfer learning using
a large freely-available satellite imagery dataset.
Important takeaways are as follows.
• Studied models have better overall accuracy (via mean
IoU over the 7 classes) on rural districts as compared to
urban districts.
• For common unsupervised domain adaptation methods,
transfer accuracy is improved, but at the cost of fairness;
the performance gap between rural and urban group is
enlarged indicating the need to design new methods thattransfer well for both the groups.
• Investigating reasons for the above findings, we find that
images from rural districts differ more across locations
than those of urban districts.
2. Related Work
Before discussing prior work on transfer learning for
satellite images, we describe some of the alternative ways
to address label scarcity and their shortcomings. Lastly, we
summarize work in the nascent area of fair transfer learning.
Approaches to tackle annotation burden for satellite
images Given the difficulty in labelling data for semantic
segmentation of satellite images, Schmitt, et al. [45] de-
veloped weakly-supervised learning methods, where noisy,
limited, or imprecise data sources are used to provide su-
pervision signal. Previous work has leveraged the spatial
context to develop unsupervised losses which, for exam-
ple, penalize nearby pixels with different predicted labels
[35, 50]. In another approach, Castillo-Navarro, et al. [6]
proposed auxiliary losses based on self-supervised image
reconstruction to improve the performance on the main task
of image segmentation. To improve efficiency of label-use,
Wang et al. [53] transferred classification models trained
with image-level labels to image segmentation tasks and
achieved high accuracy. While these approaches demon-
strate successful combination of labeled and unlabeled im-
ages, they assume that the images are from the same domain
(or distribution). However, the assumption of a consistent
domain is not realistic for problems involving satellite im-
ages which are often from different geographies. Thus, such
approaches are not straight-forwardly applicable in our set-
ting.
Transfer learning for satellite images Transferability
of satellite image segmentation models across different ge-
ographic locations has been studied in Ghorbanzadeh et
al.[14]. Using train and test sets across 3 different ge-
ographies (Taiwan, China, and Japan), they show consis-
tent decrease in evaluation scores upon transfer. Previous
work has incorporated domain adaptation methods to deal
with the challenge. For instance, Tran et al. [49] pro-
posed a two-stage transfer learning structure which gen-
erated pseudo-ground truth segmentation labels for target
data. Algorithms to improve the quality of such pseudo la-
bels were studied in [33,59]. Data augmentation is another
strategy for domain adaptation. Ji et al. augments images
to simulate perturbations due to atmospheric radiation and
demonstrate improved generalization of CNN-based mod-
els [23]. These studies show the promise of adapting mod-
els to data from different locations. But, the transfer is only
evaluated based on overall accuracy for the domain, such as
using Intersection-over-Union (IoU) to measure the over-
lap between predicted segmentation maps and ground-truth
2917
masks. Past work does not study fine-grained measures of
model performance on transfer, like how is the performance
for different subgroups in the domain (based on sensitive
attributes or land-use types) impacted. The risk that dis-
crepancy between domains in transfer learning may impact
subgroups unfairly remains unexplored.
Fairness in transfer learning Following the work in fair
machine learning literature [32], we will narrowly classify
the study of performance differences between subgroups
asmodel fairness analysis. Compared to fairness analysis
within the same domain, little work has studied transfer of
fair models across domains. The two objectives–improving
transfer accuracy and maintaining fairness–can be at odds
with each other [47, 56]. Schumann et al. [46] formalized
the problem of fair transfer learning which sets the learn-
ing objective to improve accuracy as well as fairness in the
target domain. Multiple approaches to fair transfer learning
have been proposed [8, 27, 30, 37, 41, 42, 47] that make var-
ious assumptions on how the domains differ and what data
is available. Even when labels are not available for the tar-
get domain, like in our setting, methods typically make the