text
stringlengths
0
820
constructions. Compared with the IMP-ResNet-50, the RSP-
ResNet-50 pays more attention to the important targets. It
(a)
(b)
(c)
(d)
Fig. 5. (a), (b) and (c) are separately the training loss curves of ResNet-50,
Swin-T and ViTAEv2-S, where the loss is recorded every 10 iterations. (d)
is the testing accuracy curves of these models. All curves are obtained by
training in the setting of UCM (8:2). The curves in (a), (b), and (c) have been
smoothed by moving average.
implies that RSP facilitates ResNet-50 to learn better semantic
representations, probably by seeing semantic-similar images
in the MillionAID dataset. Compared to the ResNet-50, the
Swin-T has a better context modeling ability by attending
faraway regions with the help of MHSA. Thus, their range
of high response areas is wider. Surprisingly, the IMP-Swin-T
mainly concentrates on background context, but the foreground
responses have been enhanced when adopting the RSP. By
combining the advantages of CNN and vision transformers, the
ViTAEv2-S achieves a comprehensive perception of the whole
scene. Especially, the RSP-ViTAEv2-S can better recognize
the typical natural RS scenes, such as terrace, mountain and
river. In the foreground object based scenes, compared with
RSP-ResNet-50, the RSP-ViTAEv2-S not only focuses on the
primary objects, but it also considers the related regions in
the background. While on the objects, the RSP-ViTAEv2-S
assigns higher attention, such as the airplane with a warmer
color compared with IMP-ViTAEv2-S. In the residential areas
with complex object distributions, the RSP-ViTAEv2-S can
correctly capture the sparse buildings and connect these re-
gions to form a holistic representation, effectively perceiving
the overall information of the scene. In the first image of the
school scene, the RSP-ViTAEv2-S simultaneously focuses on
the playground and the surroundings, surpassing the SeCo-
ResNet-50. As for the “school-2” image that is even difficult
to be recognized by humans, these models show different
recognition priorities. For example, the RSP-ViTAEv2-S not
only considers the campus (it can be possibly distinguished by
the irregular shape of buildings) like the IMP-ResNet-50, but
also notices the surrounding roads. The results in Table V and
Figure 4 validate the effectiveness of RSP and the superiority
of vision transformers in the aerial scene recognition task.
We also provide training loss curves and testing accuracy
curves to investigate the performances of different pretraining
10 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
methods during the training phase. Here, the setting of UCM
(8:2) is chosen as our example. The corresponding results
have been displayed in Figure 5, where the loss curves of
three kinds of networks are separately plotted. It can be seen
that SeCo-ResNet-50 performs the worst and it has the largest
initial loss, further confirming our aforementioned hypothesis
that there is a large gap between Sentinel-2 multispectral
images and the used RGB aerial images, although they are both
RS images. Compared with the IMP, it can be observed that the
RS pretrained models have better starting points, proving our
intuition that the RS pretrained weights are easier to transfer
between aerial scenes. It is also noteworthy that the shapes
of these curves are similar for the same network, implying
the unique characteristics of different network structures. We
can also find the advanced structures enable ViTAEv2-S to
reduce the performance gap indicated by the different starting
points between IMP and RSP, while other networks failed.
In addition, we can also find that the RSP accelerates the
learning of ResNet-50 and makes the corresponding accuracy
curve similar to Swin-T compared with IMP-ResNet-50. For
Swin-T, the RSP also helps it converge fast. When adopting
RSP, among all models, the advanced transformer network
ViTAEv2-S can simultaneously achieve the best accuracy and
the fastest convergence speed.
B. Aerial Semantic Segmentation
Aerial semantic segmentation is also a classification task
like aerial scene recognition but at the pixel-level rather than
the scene-level. We then evaluate the above three models on
the aerial semantic segmentation task, including the scene
parsing and object segmentation subtasks, where the former
focuses on labeling each pixel of the whole scene, while the
latter emphasizes the segmentation of foreground objects.
1) Dataset: We use the ISPRS Potsdam dataset1and the
large-scale segmentation benchmark — iSAID [81], to serve
as the testbed of the corresponded subtasks, respectively.
•Potsdam: This dataset is released by ISPRS Commission
WG II/4. It covers a large scene that is collected over
3.42 km2area of the Potsdam city. It contains 38 images,
whose average size is 6,000 ×6,000 pixels, and the
resolution is 0.5m. Among them, the training and testing
sets separately have 24 and 14 images. There are 6
categories included in these scenes, namely impervious
surface, building, low vegetation, tree, car, and clutter.
•iSAID: This is a large-scale dataset that is mainly for in-
stance segmentation. Also, it provides the semantic mask
including 15 foregrounds and 1 background category over
the aerial objects. It consists of 2,806 high-resolution
images that range from 800 ×800 to 4,000 ×13,000
pixels. The training, validation, and test set separately
have 1,411, 458, and 937 images. In this paper, only the
validation set is used for evaluation since the testing set
is unavailable.
1https://www.isprs.org/education/benchmarks/UrbanSemLab/2d-sem-label-
potsdam.aspx