text
stringlengths
0
820
RSP-ViTAEv2-S-E40 Ours 99.71±0.10 96.72±0.06 97.92±0.06 94.12±0.07 95.35±0.03
RSP-ViTAEv2-S-E100 Ours 99.90±0.13* 96.91±0.06* 98.22±0.09 94.41±0.11* 95.60±0.06*
the traditional CNN, the recent ViT has also been applied in
some works. Compared with the IMP-ViT-B, the RSP-Swin-
T-E300 performs better, although the former model has more
trainable parameters. It can be observed that the backbones are
changing over time. The VGG-16 used in the early years is
gradually replaced by the deeper networks such as ResNet-101
or DenseNet-121 due to their better representation ability.
In the implemented networks, the SeCo-ResNet-50 performs
the worst compared with its counterparts, it may be because
there still exists a gap between the Sentinel-2 multispectral
images where the SeCo trained on with the RGB images
for aerial scenes recognition. Compared with the ImageNet
pretrained ResNet-50, our RS pretrained ResNet-50 improves
the accuracy on all settings. These results imply that RSP
brings a better starting point for the optimization of the
subsequent finetuning process, attributing to the aerial im-
ages used for pretraining compared with the natural images
in ImageNet. Similarly, the RSP-Swin-T outperforms IMP-
Swin-T on three settings and achieves comparable results
on the other two settings. In addition, the ResNet-50 and
Swin-T can perform to be competitive compared to other
complicated methods by only using the RSP weights without
changing the network structures. Besides, when comparing
the ImageNet pretrained ResNet-50 and Swin-T, we can find
that the IMP-Swin-T performs better in all settings since the
vision transformers have stronger context modeling capability.
While being initialized by RSP weights, the ResNet becomes
to be more competitive and surpasses the IMP-Swin-T on
the AID (2:8), NWPU-RESISC (1:9), and NWPU-RESISC(2:8) settings, showing the benefit of RSP again. Owing to
the excellent representation ability of ViTAEv2-S, which has
both the locality modeling ability and long-range dependency
modeling ability, it outperforms both ResNet-50 and Swin-
T on almost all the settings, regardless of IMP and RSP.
Moreover, the RSP-ViTAEv2-S achieves the best performance
compared with all other methods on almost all settings except
for the AID (5:5), though on which it also delivers comparable
performance with the best one, i.e., RSP-Swin-T-E300.
In our experiments, RSP helps the networks obtain bet-
ter performance on small datasets, it may be because the
models are easier to converge when adopting the RS pre-
trained weights. While for the case where training samples
are abundant, like AID (5:5), the representation ability of
deeper models can be fully exploited. For example, the
DenseNet-121 based ESD-MBENeT obtain the best accuracy.
Nevertheless, it should be noted that only the feature output
from the last layer of RSP-ResNet-50, RSP-Swin-T, or RSP-
ViTAEv2-S is used for classification, and it is expected that
their performance can be further improved when employing
the multilayer intermediate features. In this sense, these RS
pretrained models can serve as effective backbones for future
research in the aerial recognition field. Furthermore, Table
V also shows that the models pretraining with more epochs
will probably have stronger representation abilities. Since
RSP-ResNet-50-E40 and RSP-Swin-T-E40 fall behind their
counterparts with more epochs, we only evaluate the “E120”
and “E300” pretrained weights for these two types of networks
in the rest experiments, while for ViTAEv2-S, both the “E40”
WANG et al. : EMPIRICAL STUDY OF REMOTE SENSING PRETRAINING 9
Terrace
Mountain
River
Bridge
Church
Airplane
Stadium
Airport
Thermal
power
station
Sparse
residential
Medium
residential
Dense
residential
School-1
School-2
(b) (c) (d) (e) (f) (g) (h)
(a)
Fig. 4. Response maps of the evaluated models on different scenes. (a)
Original image. (b) IMP-ResNet-50. (c) SeCo-ResNet-50. (d) RSP-ResNet-50.
(e) IMP-Swin-T. (f) RSP-Swin-T. (g) IMP-ViTAEv2-S. (h) RSP-ViTAEv2-S.
and “E100” weights are still used.
Qualitative Results and Analyses: Figure 4 shows the
response maps of the above evaluated models using Grad-
CAM++ [80] on images from various scenes. The warmer the
color is, the higher the response is. To better show the impact
of RSP, we use the pretrained weights of “E300” for ResNet-
50 and Swin-T, and the weights of “E100” for ViTAEv2-S.
The first three rows are the natural landscapes, and the scenes
in 4-8 rows mainly contain specific foreground objects, while
the next six rows present some scenes with different artificial
constructions. For example, the “Thermal power station” scene
includes not only chimneys but also cooling towers.
Corresponding to the quantitative results in Table V, the
response maps of SeCo-ResNet-50 are scattered and they can
not precisely capture the semantic-relevant areas, especially
in the natural landscapes or complex scenes with artificial