text
stringlengths
0
820
merely takes ViT as a branch parallel to CNN [52], implying
an urgent need to further explore the applications of vision
transformers on aerial scene tasks.
C. Remote Sensing Pretraining
Pretraining using RS dataset for aerial scene recognition is
a very intuitive idea. However, to our best knowledge, there
are few explorations in this direction since the insufficiency
of large-scale RS datasets like ImageNet. Nevertheless, re-
searchers have attempted to obtain the RS representations from
other resources. For example, GeoKR [53] leverages the global
land cover product as the labels, and they use the mean-teacher
framework to alleviate the influences of imaging time and res-
olution differences between RS images and geographical ones.
However, forcing alignment of different datasets inevitably
brings errors due to the intrinsic different data distributions.
The scarcity of large capacity RS dataset is mainly in the
aspect of category labels instead of images. In this case, it
4 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Reduction CellReduction CellReduction CellNormal CellNormal CellP
1/4 1/8 1/16
Class tokenConcatLinear
AdditionSchool
Reduction Cell
1/4 1/8 1/16Linear
School
Normal CellReduction CellNormal CellReduction CellNormal CellReduction CellNormal Cell
1/32GAP(a)
(b)×𝑵𝟏 ×𝑵𝟐 ×𝑵𝟑 ×𝑵𝟒𝑵
Fig. 2. The diagram of the adopted ViTAE models. (a) Original ViTAE [14]. (b) ViTAEv2 [29].
is promising to develop self-supervised pretraining methods
[54]–[57] and some related methods have been developed in
the RS area [27], [58]–[60]. For instance, SeCo [58] leverages
the seasonal changes to enforce consistency between positive
samples, which are the unique characteristics of aerial scenes,
while [59] simultaneously fuses the temporal information and
geographical location into the MoCo-V2 [55] framework.
Moreover, the channel properties [60] and spatial variance
[27] are also explored in some approaches. In this study,
since the adopted MillionAID dataset has the ground truth
labels annotated by experts and does not contain any temporal
information, we directly conduct the supervised pretraining
like the conventional IMP
III. R EMOTE SENSING PRETRAINING
In this section, we first provide a brief introduction of
the adopted large-scale RS dataset—MillionAID. Then, we
describe the details of the utilized ViTAE transformer. The
whole RSP procedure will be presented finally.
A. MillionAID
To our best knowledge, the MillionAID is by far the largest
dataset in the RS area. It contains 100,0848 non-overlapping
scenes, exceeding the competitive counterpart fMoW [61]
and BigEarthNet [62], which separately includes 132,716
and 590,326 scenes. Note that the fMoW contains 1,047,691
images since they provide multiple temporal views for each
scene. In addition, it should be noted that the fMoW and
BigEarthNet are multispectral datasets, while the MillionAID
is an RGB dataset, which is more suitable for existing deep
vision models. The categories of MillionAID consist of a
hierarchical tree that has 51 leaves, which locate on 28 parent
nodes on the second level, while the 28 groups are belonging
to 8 base classes: agriculture land, commercial land, industrial
land, public service land, residential land, transportation land,
unutilized land, and water area, and each leaf category has
about 2,000 ∼45,000 images. This dataset is obtained from the
Google Earth that is made up of diverse sensors including but
not limited to SPOT, IKONOS, WorldView, and Landsat series,resulting in different resolutions. The maximum resolution can
reach 0.5m, while the smallest is 153m. The image size ranges
from 110 ×110 to 31,672 ×31,672.
B. ViTAE
The original ViTAE [14] follows the deep-narrow design
of T2T-ViT [63], which found that simply decreasing channel
dimensions and increasing layer depth can improve the feature
richness of ViT, boosting the performance while reducing the
model size and computational cost. Thus, the original ViTAE
firstly downsamples the input image to 1/16 size by three
reduction cells. Similar to ViT, a class token is concatenated
with the output of the third reduction cell before adding
an element-wise sinusoid position encoding. Then, multiple
normal cells are stacked, and the feature size is always kept
till the end. The class token feature of the last normal cell is
used for classification through a linear layer.
Although the original ViTAE performs well on ImageNet
classification, it is unsuitable for transferring to other tasks
like segmentation, detection, pose estimation, and so on, since
it cannot generate abundant intermediate features at different
scales. Thus, the authors propose the ViTAEv2 variant [29],
which adopts the classic stage-wise design of popular back-
bone networks such as ResNet and Swin. Figure 2 shows
the comparison between the original ViTAE and ViTAEv2.
In ViTAEv2, the network is split into multiple stages, usually,
the number is 4. In each stage, the first cell is a reduction cell
for downsampling, which is followed by the stacked normal
cells. An average pooling layer is used after the last normal cell
to replace the class tokens. When finetuning on downstream
tasks, this pooling layer is removed and the remained network
is connected with corresponding task decoders.
In this paper, we employ the ViTAEv2 model for RSP.
Concretely, inspired by Swin [13], some MHSAs in ViTAE are
replaced by the WMHSA to reduce computational cost. Specif-
ically, considering the feature size becomes smaller in later