text
stringlengths
0
820
ers. For MIM objectives, training data with a substantially
lower entropy can make for an easier reconstruction task,
since masked regions may be more similar to their neigh-
bors. Therefore, the network does not have to work as hard
to fill in the blanks, limiting the learning potential. Overall,
these result indicate that the noticeably narrow scope of fea-Table 1. Dataset Analysis. To evaluate each method, we finetune
the pretrained model on seven different tasks, outlined in Section
4 and report the ARP metric defined in Equation 1. We also report
the training time in hours on a V100 GPU, as well as the car-
bon impact estimations2in kg CO 2equivalent [24]. Overall, our
collected GeoPile pretraining dataset significantly improves down-
stream performance. †indicates the vanilla continual pretrain-
ing approach of initializing the model with ImageNet-22k weights
prior to conducting MIM training on GeoPile. To further improve
the performance in an efficient manner, we introduce our continu-
ous pretraining paradigm GFM.
Method # Images Epochs ARP ↑Time↓CO2↓
ImageNet-22k Sup. 14M - 0.0 - -
Sentinel-2 [28] 1.3M 100 -5.83 155.6 22.2
GeoPile 600k 200 0.92 133.3 19.0
GeoPile†600k 200 1.24 133.3 19.0
GeoPile†600k 800 1.45 533.2 76.0
GFM 600k 100 3.31 93.3 13.3
Table 2. Breakdown of datasets in the GeoPile. We gather approxi-
mately 600k samples from a combination of labeled and unlabeled
satellite imagery with various ground sample distances and scenes.
Dataset # Images GSD # Classes
NAIP [31] 300,000 1m n/a
RSD46-WHU [27] 116,893 0.5m - 2m 46
MLRSNet [33] 109,161 0.1m - 10m 60
RESISC45 [8] 31,500 0.2m - 30m 45
PatternNet [45] 30,400 0.1m - 0.8m 38
tures and limited per-sample information in Sentinel-2 data
may be limiting the potential of the pretrained model.
Therefore, we set out to collect a diverse geospatial pre-
training dataset. Sourcing from both labeled and unlabelled
data, we form a new pretraining dataset which we term
GeoPile. The breakdown of GeoPile is shown in Table 2.
For textural detail, we ensure a variety of ground sample
distances (GSD), including images with much higher reso-
lution than Sentinel-2 (which has a GSD of 10m). Further-
more, the selected labeled datasets encompass a wide vari-
ety of classes from general remote sensing scenes, ensuring
visual diversity across samples. We calculate the average
entropy of our GeoPile dataset, and find it to be 4.6, much
higher than that of Sentinel-2. Furthermore, the textural and
visual diversity is qualitatively evident in Figure 2. In Table
1, the enhancing effect of the data selection is clearly shown
by the substantial performance increase.
3.2. Vanilla Continual Pretraining
Next, after establishing our pretraining data selection, we
investigate an alternate pretraining paradigm that bridges
the gap between the two common approaches mentioned
2CO2estimations were completed with mlco2.github.io from [24].
ℱ𝑇
ℱ𝑆𝒫…
…… 𝒟ℒ𝑓𝑒𝑎𝑡
GeoPile
𝑓𝐿𝑇𝑃(𝑓𝐿𝑆)
ℒ𝑀𝐼𝑀
……
Foundation
ModelModel Zoo
Large -scale
Dataset…
Figure 3. Our GFM continual pretraining pipeline, which leverages publicly-available large-scale models in concert with our compiled
geospatial dataset and pretraining objective. First, we select a concise set of data from various sources, which we term GeoPile (Section
3.1). Next, we train GFM with our multi-objective continual pretraining approach. Our GFM framework is constructed as a teacher-student
paradigm, with two parallel model branches. The teacher FTis initialized with ImageNet-22k weights (top) and frozen during training.
The student FSis initialized from random initialization (bottom), and is trained to serve as the final geospatial foundation model. In a
continual pretraining fashion, we leverage the intermediate features of an ImageNet-22k pretrained model to guide and quicken learning.
Furthermore, we build in an MIM objective on the student branch to learn valuable in-domain features directly from the geospatial data.
in Section 1. Specifically, we investigate the potential of
continual pretraining in the context of geospatial pretrained
models. To do so, we first employ the vanilla continual pre-
training approach; that is, using the ImageNet-22k weights
as initialization prior to beginning the pretraining step with
GeoPile. We find this to be helpful in improving perfor-
mance over starting from scratch. This validates the pos-
sibility of continual pretraining as a beneficial paradigm to
provide performance gain without additional resource costs.
Nonetheless, the improvement is still limited, with ∼0.3%
ARP increase over starting from scratch and ∼1.24% ARP
over the baseline.
To further improve the performance of our pretrained
model in comparison to the ImageNet-22k baseline, we in-
crease the number of pretraining epochs in the next row of
Table 1. While we are able to make improvements, this
comes at the cost of substantially more computational cost
and carbon footprint for marginal gain. Therefore, we ask
the question: how can we significantly improve the per-
formance further while maintaining minimal compute and
carbon footprint overhead? To this end, we propose a sim-
ple and efficient approach for building geospatial pretrained
models capable of strong downstream performance.
3.3. GFM Pretraining
A significant number of geospatial foundation model
studies disregard the existing large-scale model represen-
tations. This is far from ideal, particularly for large trans-
former models known to require a vast amount of data and