Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files Community
1
text
stringlengths
0
820
Designed
for
Men
by
Caroline
Criado
Perez
●
The
Kindness
of
the
Hangman
by
Henr y
Oster
●
So
Y ou
W ant
t o
T alk
About
Race
by
Ijeoma
Oluo
●
Ut opia
for
Realists
by
Rutger
Bregman
Sitemap
FOLLOW:
GITHUB
FEED
©
2024
Hannah
Kerner.
Powered
by
Jekyll
&
AcademicPages
,
a
fork
of
Minimal
Mistakes
.
Lightweight, Pre-trained Transformers for
Remote Sensing Timeseries
Gabriel Tseng1,2Ruben Cartuyvels1,3Ivan Zvonkov4Mirali Purohit5
David Rolnick1,2Hannah Kerner5
1Mila – Quebec AI Institute
2McGill University
3KU Leuven
4University of Maryland, College Park
5Arizona State University
Abstract
Machine learning methods for satellite data have a range of societally relevant
applications, but labels used to train models can be difficult or impossible to
acquire. Self-supervision is a natural solution in settings with limited labeled data,
but current self-supervised models for satellite data fail to take advantage of the
characteristics of that data, including the temporal dimension (which is critical for
many applications, such as monitoring crop growth) and availability of data from
many complementary sensors (which can significantly improve a model’s predictive
performance). We present Presto (the Pretrained Remote Sensing Transf ormer), a
model pre-trained on remote sensing pixel-timeseries data. By designing Presto
specifically for remote sensing data, we can create a significantly smaller but
performant model. Presto excels at a wide variety of globally distributed remote
sensing tasks and performs competitively with much larger models while requiring
far less compute. Presto can be used for transfer learning or as a feature extractor
for simple models, enabling efficient deployment at scale.
1 Introduction
Machine learning is increasingly being applied to the remote sensing domain, in particular to under-
stand the evolution of the Earth’s surface over time (Brown et al., 2022; V oosen, 2020; Abys et al.,
2024; Wang et al., 2020b). These applications can have important societally beneficial outcomes,
ranging from tracking progress on sustainable development goals (Ferreira et al., 2020) to improved
weather forecasting (English et al., 2013; V oosen, 2020) to disaster management (Kansakar and
Hossain, 2016). However, labeled datasets often contain labels that are few, sparse, and unreliable
(Bressan et al., 2022), especially for under-resourced geographies, leading to poor global gener-
alization (Yifang et al., 2015; Kerner et al., 2020; Nakalembe et al., 2021). This has spurred the
investigation of self-supervised learning algorithms for remote sensing data.
Current self-supervised approaches for remote sensing data have drawn from methods in computer
vision, yielding models that treat remote sensing data as single-timestep images (Jean et al., 2019;
Manas et al., 2021; Ayush et al., 2021). Such models (i) cannot benefit from patterns that emerge
when an area is monitored over time, which is especially important for agriculture and other seasonal
landcover, (ii) typically only consider a single satellite product (such as Sentinel-2 multispectral data),
despite there being hundreds of publicly available satellite data products (GEE), (iii) are typically
large and computationally expensive (Reed et al., 2022; Cong et al., 2022; Fuller et al., 2023), making
the deployment of these models at scale challenging, and (iv) cannot natively handle the labels for
Preprint.arXiv:2304.14065v4 [cs.CV] 5 Feb 2024