text
stringlengths
0
820
used, which is useful for feature discrimination.
In recent years, with the superiority of automatically extract-
ing deep features that reflect the inherent properties of objects,
DL has achieved an impressive breakthrough in the computer
vision (CV) field [11]–[17], as well as the RS field [18]–
[20]. In the aerial scene recognition domain, the most com-
monly used deep models are the convolutional neural networks
(CNN), which have a good ability for local perception and
global perception, where the former is achieved via applying
sliding window convolutions on the input image to extract the
local features while the latter is achieved by stacking multiple
convolutional layers to increase the receptive field. According
to the training scheme, the ways of using CNN in aerial scene
recognition methods can be divided into three categories, i.e.,
training from scratch, finetuning, and adoption as the feature
extractor. The first scheme does not involve any external data,
meaning there is no prior knowledge that can be leveraged.
To remedy this issue, the finetuning scheme uses the networks
pretrained on a large-scale dataset as the start point for further
training (i.e., finetuning). The last scheme directly extracts the
feature from the pretrained CNN without further finetuning,
therefore lacking the flexibility of adapting to aerial images
from different downstream tasks.
The existing literature [21] reveals that the finetuning strat-
egy performs better than the other ones. We attribute it to the
capacity of the used pretraining dataset, including the sample
size and the number of categories. In the current RS field
including the aerial scene recognition task, almost all finetuned
models are pretrained on the ImageNet-1K dataset [22], which
is the most famous image dataset in the CV field. Its millions
of real-world images from 1,000 different categories enable
the models to learn a powerful representation. Usually, off-
the-shelf deep models like VGG [11] and ResNet [12] are
used as backbone networks for aerial scene recognition, since
training a new network on the ImageNet from scratch is
time-consuming and requires a large number of computational
resources. To further improve the classification performance,
some methods [23], [24] adopt the ImageNet pretrained mod-
els as the backbone network and employ multi-level features
from it. In addition, many other components or strategies are
specially designed for the aerial recognition task, such as the
distillation [25] and feature partitioning [26].
Although the aforementioned methods have achieved re-
markable performance for aerial scene recognition, there are
still some issues needed to be further investigated. Intuitively,
when considering the characteristics of aerial images, there
exists a large domain gap compared with the natural images
in terms of view, color, texture, layout, object, etc. Previous
methods attempt to narrow this gap by further finetuning thepretrained model on the RS image dataset. Nevertheless, the
systematic bias introduced by ImageNet Pretraining (IMP)
has a non-negligible side impact on the performance [27].
On the other hand, we notice that there are abundant aerial
images captured by diverse multiple sensors with the progress
of RS technology, which can be used for pretraining. As a
representative example, MillionAID [28] is so far the largest
aerial image dataset and has a million-level volume similar to
ImageNet-1K, making the Remote Sensing Pretraining (RSP)
become possible.
RSP enables training a deep model from scratch, implying
that the candidate model is not necessary to be limited to
the off-the-shelf CNN. In this paper, we also investigate the
impact of RSP together with vision transformers, which have
shown surprisingly good performance in the CV domain.
Compared with convolutions in CNN which are skilled in
modeling locality, the multi-head self-attention (MHSA) in
transformers, e.g., Swin transformer [13], can flexibly capture
diverse global contexts. Recently, ViTAE [14], [29] explores
both convolutions and MHSA for modelling locality and
long-range dependency simultaneously, achieving state-of-the-
art (SOTA) performance on the ImageNet classification task
and downstream vision tasks. In addition, it also extracts
multi-scale features through a dilated convolution module and
the stage-wise design, which have been shown effective in
previous works, especially for aerial image interpretation [30].
Since both the CNN and aforementioned vision transformers
can also produce intermediate features in different stages,
which are useful for many downstream tasks, we also in-
vestigate their finetuning performance of them after RSP on
semantic segmentation, object detection, and change detection.
To achieve these goals, we conduct extensive experiments on
nine popular datasets and have some findings. The RSP is
an emerging research direction in aerial image understanding,
which however is still underexplored, especially in the context
of vision transformers. We hope this study could fill this gap
and provide useful insights for future research.
The main contribution of this paper is three-fold.
(1) We investigate the impact of remote sensing pretraining
by training on a large-scale remote sensing dataset using
three types of backbone networks, including traditional
CNN, competitive visual transformer models, and the
advanced ViTAE transformers.
(2) We further finetune the above models that are initialized
with the remote sensing or ImageNet pretraining weights
on four kinds of tasks including scene recognition, se-
mantic segmentation, object detection, and change detec-
tion using a total of nine datasets, and compare them with
other methods.
(3) Experimental results show that typical vision transformer
models can obtain competitive performance or perform
better than CNN. Especially, the ViTAE achieves the