text
stringlengths
0
820
covariate shift assumption which says that the labeling rule
remains the same between the domains and only the fea-
ture distribution changes. In this setting, Coston et al. [8]
propose a method for fair transfer even in cases where sen-
sitive attributes are absent from one of the domains. Other
approaches do not require access to target domain data al-
together and instead either make causal assumptions on the
discrepancy [47] or hypothesize a set of target domains and
optimize against them [10, 30]. Finally, Szab ´oet al. [19]
conduct a fairness evaluation of segmentation methods as-
suming a single domain.
However, none of the existing works study fair transfer
learning for semantic segmentation models. The task differs
considerably from the above settings as the input data is
high-dimensional, and the model output and loss function
for segmentation are different. We take the first step in this
direction by demonstrating the need for such methods via a
thorough empirical study on a relevant application.
3. Dataset
We use the publicly available, high spatial resolution
land-cover dataset called LoveDA [52] in this study. Com-
pared to other popular satellite image datasets, such as
Zurich Summer [50], DeepGlobe [9], and DSTL [20], the
recently released LoveDA has more annotated images and
includes images from diverse locations. The dataset con-
sists of 5987 images of size 1024 ×1024 and spatial reso-
lution 0.3m. The images are collected from 18 adminis-
trative districts from three cities in China, namely Nanjing,
Changzhou, and Wuhan. Out of these, there are 9 urban dis-
tricts and 9 rural districts, categorized based on their popu-
lation density and level of economic development. We use
Figure 1. Sample images from urban and rural scenes. For each
scene, one image from source domain districts, one from target
domain districts, and the network’s segmentation predictions (Seg-
map) for the 7 land-cover classes on that target image are shown.
satellite images from the 12 districts for which ground-truth
masks are available: Gulou, Qinhuai, Qixia, Yuhuatai, Jin-
tan, and Jianghan (urban); Pukou, Lishui, Liuhe, Huangpi,
Gaochun, and Jiangxia (rural). The remaining 6 districts
are not availanble as they are held out for the benchmark
challenge. The dataset contains segmentation masks, which
are pixel-level labels, for 7 land-cover classes: Background,
Building, Road, Water, Barren, Forest, and Agricultural.
The Background class consists of any pixel not belonging
to the other classes. Statistics of the dataset, given in Figure
3 in [52], show that the pixel counts across the 7 classes are
imbalanced. Further, distribution of classes and of building
scales differ between urban and rural scenes. Thus the rural
and urban groups of images, which we use in our fairness
analysis, have different characteristics.
4. Methods
We study three tasks, namely semantic segmentation
within the same domain, across districts, and across rural-
urban areas. Next, we describe the setup for each task.
2918
4.1. Task A - Semantic segmentation
Our task is to train multi-class semantic segmentation
models for detecting the 7 land-cover classes from a given
image. Sample images and predicted segmentation maps
are shown in Figure 1. Same as the models studied in
the LoveDA study [52], we use two commonly used deep
learning-based segmentation methods – U-Net [44] and
DeepLabV3+ [7] network, both with pre-trained ResNet50
[17] as the backbone model for the encoder [13].
4.1.1 Training-testing details
Images from the 12 districts with labeled data are shuffled
and split into training ( ≈80%) and testing sets ( ≈20%).
Training set has 3148 images with a mix of 1377 urban and
1771 rural images, and the rest of the images comprise the
testing set with a mix of 368 urban images and 473 rural
images. Images are augmented during training by mirroring
and rotation. Dimension of the input image to the network is
512×512×3 where 3 indicates the RGB bands. The output
dimension of the network is 512 ×512×7, where 7 repre-
sents the probability of each pixel belonging to each land-
cover class. We use cross-entropy (CE) loss, and stochastic
gradient descent (SGD) as the optimizer with a momentum
of 0.9 and a weight decay of 10−4. The batch size is set to
16 and the total training iterations are 15000, during which
the learning rate is decayed using a polynomial learning rate
scheduler implemented in PyTorch [40].
4.1.2 Evaluation metrics
We test the models on either the whole test set or the ur-
ban and rural subsets in the test set separately, and evaluate
model performance using the following metrics:
Accuracy metrics: Intersection-over-Union (IoU), also
called Jaccard index, is used to measure segmentation ac-
curacy which is a common method to evaluate the quality
of image segmentation [11, 54]. IoU for a class is defined
as the intersection of class-wise ground-truth masks and the
predicted segmentation divided by their union,
IoU :=TP
TP+FP+FN,
where TP,FP, andFN are pixel-wise true positives, false
positives, and false negatives. We report IoU score of the
model on each land-cover class as well as mean over class-
wise IoU (referred to as Mean ) over the 7 classes.
Fairness metrics: Besides looking at IoU on the rural
and urban subsets and comparing the two values, we de-