Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
compute power to train. Instead, we reason that the valu-
able knowledge available in models like those trained on
ImageNet-22k should be leveraged to produce strong per-formance with minimized overhead. To this end, we pro-
pose an unsupervised multi-objective training paradigm for
effective and efficient pretraining of geospatial models, il-
lustrated in Figure 3.
There are two main components in our framework. First,
we randomly initialize an encoder FSand decoder Dset up
for MIM as in [41]. During training, the input is randomly
masked, and the network attempts to reconstruct the image
at the output. This MIM objective is enforced with an L1
loss [41]:
LMIM =∥Oκ−Gκ∥1
N, (2)
whereOκare the original pixel values from κmasked re-
gions,Gκare the generated reconstructions for those re-
gions, and Nis the total number of masked pixels.
For the continual pretraining of our framework, we ini-
tialize a second encoder branch FTup to a chosen stage L
and load the ImageNet-22k pretrained weights. This branch
behaves as a form of teacher during the training process
to the student branch ( FS), which will serve as our final
model. For the ImageNet teacher, we freeze the weights,
to both ensure that the structured representations are main-
tained during the training process, and also reduce the com-
putation required during optimization.
Rather than using the masked input as in the student
branch, the teacher receives the unmasked image as input,
and provides a feature output fT
Lat stage L. This feature
has access to the full context of the input, enabling it to
capture informative representations. We utilize this feature
to guide the representations of the student, and form a sec-
ondary objective with the cosine similarity between branch
features:
Lfeat =−P(fS
L)
P(fS
L)
2·fT
L
fT
L
2, (3)
where fS
LandfT
Lare the intermediate features of the stu-
dent and teacher branches at stage L, andPis an linear
projection layer. Therefore, the final loss during training is
simply the summation of these objectives:
L=LMIM +Lfeat. (4)
This training paradigm enables an ideal two-fold optimiza-
tion. Distillation from the intermediate features of the
teacher ensure that the student can benefit from the teacher’s
diverse knowledge, learning more in less time. Further-
more, the student is simultaneously given freedom to adapt
to in-domain data through its own pretraining objective,
gathering new features to improve performance.
We analyze the ARP and resource cost of this approach
in Table 1. Notably, our GFM is able to achieve better over-
all performance with substantially less computation and
emissions impact compared to vanilla continual pretraining
with the same dataset, illustrating that our multi-objective
continual pretraining paradigm is an effective method for
training these models. Comparatively, the SOTA geospa-
tial pretrained method SatMAE [9] requires 768 hours on a
V100 GPU and 109.44 kg equivalent CO 2according to their
reported results. Therefore, GFM enables more than 8×
reduction in total training time and carbon impact. More-
over, we find that the performance of SatMAE is often not
superior to the off-the-shelf ImageNet-22k pretrained ViT
(Section 4). This implies that building powerful geospatial
pretrained models from scratch is challenging and further
underscores the benefits of utilizing continual pretraining
instead. We show these results in the following section.
4. Experiments
To verify the effectiveness of our model in detail, we
conduct experiments on seven geospatial datasets of vari-
ous tasks including change detection (Section 4.1), classifi-
cation (Section 4.2), segmentation (Section 4.3), and super-
resolution (Section 4.4).
For pretraining, we employ 8 NVIDIA V100 GPUs with
a batch size of 2048 (128 per GPU) and the image size
of 192×192. All pretraining settings are the same as in
[41]. For downstream tasks, 4 NVIDIA A10G GPUs are
employed. During the pretraining stage, we utilize RGB
bands as they are most commonly available among data
sources and tasks. For downstream tasks with additional
band inputs, we initialize the RGB patch embeddings with
the pretrained weights and randomly initialize the remain-
ing channels. Potentially improving performance even fur-
ther though the employment of additional data modalitiesTable 3. Onera Satellite Change Detection Results
Method Precision ↑Recall ↑F1↑
ResNet50 (ImageNet-1k) [19] 70.42 25.12 36.20
SeCo [28] 65.47 38.06 46.94
MATTER [1] 61.80 57.13 59.37