text
stringlengths
0
820
ViT (ImageNet-22k) [14] 48.34 22.52 30.73
SatMAE [9] 48.19 42.24 45.02
Swin (random)[25] 51.80 47.69 49.66
Swin (ImageNet-22k)[25] 46.88 59.28 52.35
GFM 58.07 61.67 59.82
Figure 4. Qualitative results of downstream performance on OSCD
comparing our GFM with ImageNet-22k and randomly initialized
baselines. White, green, red colors show true positive, false posi-
tive, and false negative respectively.
Table 4. DSFIN Change Detection Results
Method Precision ↑Recall ↑F1↑
ResNet50 (ImageNet-1k) [19] 28.74 92.07 43.80
SeCo [28] 39.68 81.02 53.27
ViT (ImageNet-22k) [14] 70.77 66.34 68.49
SatMAE [9] 70.45 60.29 64.98
Swin (random)[25] 57.97 62.06 59.94
Swin (ImageNet-22k)[25] 67.11 72.33 69.62
GFM 74.83 67.98 71.24
will be an intriguing avenue for future research. Additional
training details for these tasks are provided in the supple-
mentary material .
4.1. Change Detection
Change detection is a particularly important remote sens-
ing task, helping us understand how humans interact with
our planet over time, and natural phenomena that change
our planet’s landscape. We conduct experiments on both
the Onera Satellite Change Detection (OSCD [5]) in Table
3 and DSIFN [43] in Table 4.
OSCD consists of 14 image pairs extracted from various
regions around the world within a three year period of 2015
to 2018. The images are taken from Sentinel-2 with GSDs
ranging from 10m to 60m, and split into 14 images for train-
ing and 10 for evaluation. The annotations indicate whether
the change has occurred on a pixel level, and focus primarily
on urban developments. Similarly, we also test our method
on DSIFN dataset. This dataset contains high-resolution
imagery, such as WorldView-3 and GeoEys-1 [43]. This
dataset contains 3490 high resolution samples for training
and 48 images for evaluation respectively. Every pair of im-
ages from a given location at two different timestamps will
be fed into the swin encoder [25] for feature extraction. The
difference between the features from each pair is computed
and fed into an UPerNet [39] to generate the final binary
segmentation masks [28, 4]. The encoder is initialized with
the pretrained weights.
For both datasets, we report the precision, recall, and
F1 score on the “change” class. As the results presented
from OSCD (Table 3 and Figure 4) and DSIFN (Table 4),
GFM shows a consistent improvement over the ImageNet-
22k baseline across both datasets. Notably, SatMAE is able
to improve over its ImageNet-22k baseline on OSCD, but
lags behind on DSIFN. This further highlights the difficulty
of training large vision transformers from scratch that can
perform consistently across different GSDs.
4.2. Classification
Another common remote sensing application is that of
classification. We evaluate two datasets common in the
literature [28, 1]: UC Merced Land Use Dataset [42] and
BigEarthNet [36]. The UC Merced Land Use Dataset is a
classic dataset in the remote sensing field. It contains 21
classes, each with 100 images at 256x256 pixels and an ap-
proximate GSD of 1 foot. We split the data into train and
validation according to [13]. BigEarthNet [36] (BEN) is a
large-scale remote sensing dataset for multi-label classifi-
cation. The data consist of 12-band Sentinel-2 images with
sizes of 120x120, 60x60, and 20x20 pixels for the bands at
10m, 20m, and 60m GSDs, respectively. We employ the
data split and 19 class evaluation as common in the litera-
ture [29, 28, 9].
In Table 5, we report the classification accuracy on
UC Merced (UCM) and mean average precision results
on BigEarthNet (BEN) for all methods. On UC Merced,
we note the SeCo [28] pretrained model performs signif-
icantly worse than its ImageNet-1k pretrained counterpart
with ResNet-50. These two datasets are very different in
both classes, satellite source, and GSDs, and therefore hav-
ing a diverse feature knowledge is imperative to maintain-
ing performance despite these distinctions. Our model can
provide robust performance in both cases by leveraging Im-
ageNet representations and remote sensing data in its learn-
ing. Furthermore, one key motivation for training a geospa-
tial foundation model is to improve the sample efficiency forTable 5. UC Merced classification accuracy and BigEarthNet
multi-label classification mean average precision results.
Method UCM BEN 10% BEN 1%
ResNet50 (ImageNet-1k) [19] 98.8 80.0 41.3
SeCo [28] 97.1 82.6 63.6
ViT (ImageNet-22k)[14] 93.1 84.7 73.6
SatMAE [9] 92.6 81.8 68.9
Swin (random)[25] 66.9 80.6 65.7
Swin (ImageNet-22k) [25] 99.0 85.7 79.5
GFM 99.0 86.3 80.7
downstream tasks. Notably, we find that our model main-
tains strong performance on BigEarthNet, even when only
given 1% of the training data.
4.3. Segmentation
Segmentation is a popular remote sensing application for
enabling automated extraction of building footprints or land
cover mappings over wide regions. We therefore conduct
experiments on this task on two different datasets. Vai-
hingen [35] is an urban semantic segmentation dataset col-