text
stringlengths
0
820
best performance on almost all settings even if compared
with the existing state-of-the-art methods. In addition, a
series of findings of remote sensing pretraining will be
presented, including the comparison with the traditional
ImageNet pretraining and the performances on different
downstream tasks. These findings provide useful insights
WANG et al. : EMPIRICAL STUDY OF REMOTE SENSING PRETRAINING 3
for future research.
The remainder of this paper is organized as follows. Section
II introduces the related works, including the aerial scene
recognition methods, especially CNN and vision transformer
related ones, and the existing works of RSP. Section III
describes the implementations of RSP, as well as the employed
large capacity MillionAID dataset and the adopted ViTAE
network. The experiment results on the four tasks and the
related comprehensive analyses are presented in section IV .
Finally, Section V concludes this paper.
II. R ELATED WORK
A. Aerial Scene Recognition
There are a large number of CNN-based approaches to
aerial scene recognition. Many off-the-shelf CNN classifica-
tion models that are pretrained on ImageNet, such as the
VGG [11], ResNet [12], and DenseNet [31], have been used
and further finetuned on aerial images. Nonetheless, the chal-
lenging aerial scene that possesses inter-class similarity and
intra-class diversity can not be easily interpreted by only
using features of the last layer, which are also considered as
the “global features” compared with the features in previous
layers, since it is also useful to highlight the important scene-
relevant local regions for scene understanding. To address this
issue, [23], [32] jointly exploits the multi-level CNN features,
where the high-level features from deep layers usually have
abundant semantic information, while the low-level features
of the shallow layers tend to provide visual structures. For
example, [32] conducts varied dilated convolutions on multiple
features of VGG to obtain the more effective multiscale
features. In addition, they optimize the category probabilities
by preserving the local maximum and revising others with
the two-dimensional Gaussian-like distribution in a window
to strengthen the local regions. [23] separately applies graph
convolutions on multilayer VGG features, where each pixel
representation is regarded as a node, and the extracted graph
features are concatenated with the last global visual features.
Apart from feature fusion, the attention mechanisms have
also been commonly used in aerial scene recognition since
they can enhance the local features by simulating the human
vision that directly assigns different weights to various areas of
the current scene. The attention modules can be easily inserted
into the CNN [33]. For example, [34] adopts channel attention
and spatial attention modules in parallel like [35] to form a
complementary attention. [32] also employs spatial attention
to further adjust the optimized category probabilities. Another
point for aerial scene recognition is to model the relationships
of different regions. For example, [23] captures the topological
relations of different objects, where the adjacency matrices are
carefully designed. Besides, some interesting topics such as
the multiple instance learning [36], self-distillation [25] and
feature partition [26] have also been explored in aerial scene
recognition research.
Networks before linear layers in scene recognition task can
be used as feature encoders for many downstream tasks, where
the most representative ones for aerial images are semantic
segmentation, object detection, and change detection. Whilesemantic segmentation and object detection are common tasks
in CV , change detection is a specific task in RS. So far, a large
number of related approaches in the aforementioned fields
have been developed. Please refer to [37]–[40] for details.
B. Vision Transformer
Transformer is first proposed in [41], and has been widely
used in the natural language processing (NLP) field [42],
[43]. Besides NLP, the recently proposed vision transformers
inspire a wave of researches [13], [14], [29], [44]–[51] in
the CV field. The core component of the vision transformer
is the MHSA, which is the extension of self-attention (SA).
Compared with the convolution operation, the SA can capture
long-range context and the relationships between any different
positions. On the foundation of SA, the MHSA separately
conducts the SAs in different projected subspaces, possess-
ing more powerful representative abilities. ViT [44] is the
pioneer vision transformer, where the input image is split
into fixed-size patches to form tokens, which are then fed
into the MHSA. However, the fixed receptive field restricts
its applications on downstream tasks, and the global MHSA
brings high computational complexities. To address the former
issue, PVT [46] adopts the classical pyramid structure, improv-
ing the model transferability with the generated hierarchical
multiscale features. Swin [13] further substitutes the global
MHSA to the shiftable window MHSA (WMHSA), it reduces
the computational overhead significantly, achieving excellent
performance on many CV tasks. Nonetheless, it still suffers
from the common issues of vision transformers, such as being
inefficient in modeling locality and scale invariance, which are
exactly the advantages of CNN. Thus, besides Swin, we also
employ another advanced vision transformer named ViTAE
[14], [29], where the intrinsic biases of CNN are introduced
into the transformer. It also adopts a pyramid structure to
generate hierarchical features and local window attention,
achieving better performance on many CV tasks while having
a reasonable computational overhead and memory footprint.
Using vision transformers as the backbone for aerial scene
recognition is still under-explored, while the existing method