Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
gressively increased for various applications worldwide.
Progress in this domain can substantially improve our abil-
ity to understand the earth and how we interact with it. With
the rising popularity of foundation models in vision and nat-
ural language, researchers have begun to investigate apply-
*Work done as an intern at Amazon Web Services
92.5
94.1
95.8
97.4
99.0
30.537.945.252.660.0
43.050.257.564.872.0
81.083.586.088.591.0
68.0
69.9
71.8
73.6
75.50.59
0.6
0.62
0.63
0.6440.0
50.5
61.0
71.5
82.0DSFIN
(Change Detection)
WHU Aerial
(Segmentation)
Vaihingen
(Segmentation)
SpaceNet2
(Super-resolution)BigEarthNet
(Multi-label
Calssification)UC Merced
(Classification)OSCD
(Change Detection)
N/AFigure 1. Our geospatial foundation model (GFM) achieves favor-
able performance on a broad set of tasks in comparison to other
state-of-the-art geospatial pretraining methods (SeCo [28], Sat-
MAE [9]) and ImageNet supervised pretraining baselines. Leg-
end is as follows. Cyan: ImageNet-1k Supervised (ResNet50),
Blue: SeCo [28], Purple: ImageNet-22k Supervised (ViT), Or-
ange: SatMAE [9], Gray: ImageNet-22k Supervised (Swin),
Green: GFM (ours).
ing such principles to the geospatial domain in order to en-
hance the suitability of deep learning models in downstream
tasks [29, 28, 9, 2]. In the literature, various works have ex-
plored two prominent approaches for introducing pretrained
foundation models in geospatial applications. The first ob-
vious approach is to leverage existing foundation models
from the natural image domain, like those trained on the
large-scale ImageNet-22k dataset [11]. In practice, this
is done by directly finetuning publicly-available ImageNet
pretrained models on the downstream tasks . This approach
has the advantage of being straight-forward, as ImageNet
models can be simply downloaded from many open-source
model zoos, and has been shown to be effective [29, 30].
However, due to the domain gap between natural images
and remote sensing, this approach is not optimal for geospa-
tial data, and still leaves performance gains on the table.
In recent years, a second approach has gained significant
traction, where researchers aim to pretrained models spe-
cific to the geospatial domain [28, 2, 9, 38]. These methods
typically train a network from scratch on a large corpus
of remote sensing imagery to learn in-domain representa-
tions transferable to downstream tasks. Unfortunately, this
can require a significant amount of data and training time
to achieve good performance, especially when employing
large state-of-the-art (SOTA) transformer models. For in-
stance, the current SOTA in geospatial foundation models,
SatMAE [9], requires 768 hours on a V100 GPU for train-
ing a vision transformer [14]. This has substantial cost
associated with producing the model, not just in terms of
time and computation but also environmentally, with a to-
tal estimated carbon footprint of 109.44 kg CO 2equivalent.
Additionally, the final performance of such models are not
consistently better across various tasks than simply utilizing
publicly-available ImageNet pretrained models (Section 4),
despite the high resource expense.
In this work, we propose to investigate a different
paradigm for producing more effective geospatial founda-
tion models with substantially less resource costs. First, we
begin with a discussion on pretraining data selection, and
ultimately construct a concise yet diverse collection of data
from various sources to promote feature diversity and ef-
fective pretraining. Second, rather than following the afore-
mentioned typical approaches, we investigate the potential
ofcontinual pretraining for the geospatial domain from
readily-available ImageNet models. Continual pretraining
has been practiced in the NLP domain with success in var-
ious works [16, 17, 26]. In this paradigm, existing founda-
tion models are further improved for a specific domain or
task through a secondary pretraining stage. This new single
model can now be fine-tuned on the various downstream
tasks in that domain. In principle, we reason that such a
paradigm has the potential to boost performance by utiliz-
ing large-scale ImageNet representations as a base on which
stronger geospatial foundation models can be built. Further-