text
stringlengths
0
820
ration of the masked image modeling framework in geospa-
tial applications is still in its early stages, and could help
alleviate some concerns with contrastive approaches in this
domain. Particularly, the choice of augmentations with con-
trastive methods can be quite difficult, as common selec-
tions such as greyscale, color jitter and others that heav-
ily affect the intensity of the image can instill undesirable
invariances [29]. On the other hand, MIM objectives like
[41, 18] rely only on simple spatial augmentations such
as flipping and cropping. Furthermore, a common remote
sensing application is that of change detection, which re-
quires a model to detect changes in two images from the
same location but at different times. In order to still be ef-
fective on this task, works that use contrastive approacheson temporal positives introduce various design choices. For
instance, SeCo [28] creates multiple feature subspaces dur-
ing pretraining, each one invariant to a separate form of aug-
mentation. [1] also employs temporal positives, but instead
chooses the sampling locations for the pretraining data to
ensure that image pairs contain primarily natural illumina-
tion and viewing angle variant, without major changes such
as new urban developments.
Continual Pretraining . Continual pretraining has been
primarily introduced in the natural language domain [16,
17, 26], in order to improve large language models (LLM).
[16] illustrates the viability of two additional stages of
pretraining, using in-domain data (domain-adaptive), and
then even further using task-specific data (task-adaptive).
[17] proposes a continual training paradigm for enabling
temporal reasoning abilities to pretrained language mod-
els. [26] focus on using continual pretraining to enable
mixed language neural machine translation. In the vision
domain, [22] employs a BYOL [15] style continual pre-
training paradigm for 2D medical image segmentation. [34]
explores a hierarchical pretraining approach for task adap-
tation. However, they primarily focus on adapting to a spe-
cific downstream task at a time, employing three training
stages on top of an existing pretrained model for each task
individually. In contrast, we employ one efficient in-domain
pretraining setting that can generalize to many downstream
tasks, as illustrated in Section 4. Furthermore, rather than
directly loading the pretrained weights from existing mod-
els as initialization, we find instead that leveraging the rep-
resentations as an auxiliary distillation objective during the
pretraining process enables learning stronger representa-
tions.
3. Methodology
In the following sections, we discuss the pretraining data
selection (Sec, 3.1), investigate vanilla continual pretraining
(Sec. 3.2), and present our GFM method (Sec. 3.3).
3.1. Pre-training Data Selection
A particularly common choice of source data among
geospatial contrastive pretraining works is Sentinel-2 im-
agery [28, 1, 37] due to its large corpus of available data
and ease of access. Therefore, to begin our study, we first
gather a pretraining dataset of 1.3 million Sentinel-2 im-
ages using the sampling technique from [28]. After gather-
ing the Sentinel-2 data, we employ it to pretrain a Swin-B
[25] model with the masked image modeling (MIM) objec-
tive from [41]. We then finetune and evaluate this model
on a wide variety of downstream datasets to get a broad un-
derstanding of its performance potential in many tasks (see
Section 4 for task details). For a comparison, we finetune
the ImageNet-22k pretrained Swin-B from the official Swin
Transformer repository [25] on all downstream tasks as a
Figure 2. We visualize some example images from the pretraining
datasets with Sentinel-2 (left) and GeoPile (right). Sentinel-2 has
noticeably much lower feature diversity within a single image and
across images than that of our GeoPile pretraining dataset.
baseline. In order to compare these models across all tasks,
we introduce an average relative performance metric (ARP)
in which we take the relative difference on each task with
respect to the ImageNet-22k baseline, and then average that
difference:
ARP (M) =1
NNX
i=1score (M,taski)−score (baseline ,taski)
score (baseline ,taski).
(1)
Here “baseline” is the Swin-B model pretrained on
ImageNet-22k, as mentioned above. Mdenotes the model
for performance evaluation, and N is the number of tasks.
There are 7tasks used in Section 4 covering important
geospatial applications such as classification, multi-label
classification, semantic segmentation, change detection,
and super-resolution. The reported ARP value is scaled by
100 to show as a percentage.
We compare these two models in Table 1. Interest-
ingly, we find that the Sentinel-2 model performs poorly
on downstream tasks compared to the ImageNet-22k base-
line. To investigate further, we visualize multiple samples
from Sentinel-2 in the left columns of Figure 2. Upon in-
spection, we note that the feature diversity within a single
image and across images of Sentinel-2 is perceivably low.
To further quantify this suspicion, we calculate the average
image entropy over a randomly sampled set of 3000 im-
ages from the collected Sentinel-2 data as well as the typ-
ical ImageNet dataset as a baseline. Overall, the Sentinel
images have an average entropy of 3.9 compared to 5.1 of
ImageNet. Such an evaluation provides insights into the
potential pitfalls of Sentinel-2 data in pretraining transform-