Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
achieving comparable results with SITS-Former despite having 6×fewer parameters (shown in Table
6). This shows that Presto can ingest timeseries at different temporal resolutions and at varying
intervals .
In addition, the S2-Agri dataset is missing pixel location metadata, which is always passed to Presto
during pre-training. S2-Agri was sampled from a single S2-tile, so we used the location of the central
pixel of this tile for all pixels in the dataset. Even with this much less accurate location metadata,
Presto remained performant.
10
Table 7: Structured masking strategies yield the best downstream performance . We measured
Presto R’s F1 score on the CropHarvest validation task. Combining structured strategies outperformed
the “Random” masking employed by (He et al., 2022).
Channel
GroupsRandom TimestepsContiguous
TimestepsF1
Score
✓ 0.646
✓ 0.653
✓ 0.664
✓ 0.649
✓ ✓ ✓ ✓ 0.665
5.4 Ablations
We conducted three ablations to better understand Presto’s performance:
•Structured masking strategies perform best : Table 7 shows results from ablating the masking
strategies. Unlike other masked autoencoder methods (He et al., 2022), we found that combining
structured masking with random masking outperforms random masking alone.
•Pre-training Presto is critical to achieve strong performance : In Tables 3, 5 and Table 6, we
compared the performance of a randomly -initialized Presto architecture with the pre-trained model.
Pre-training yielded a significant increase in performance (a 50% increase in accuracy on the
S2-Agri 100dataset). Even when the downstream training dataset size was large (EuroSat has
21,600 training samples), pre-training yielded a 14% increase in accuracy given RGB inputs and
up to 22% increase in accuracy at lower resolutions (Table 11). For TreeSatAI with S1 data (Table
15), a randomly initialized model slightly outperformed the pre-trained model. We hypothesize
that this is due to the difference in input relative to the pre-training data, since the TreetSatAI input
consists of a single image from only one timestep and one channel group.
•Presto’s performance scales with model size : To measure how different model sizes affect Presto’s
performance, we pre-trained two larger Presto variants: a deeper variant with 4 encoder layers
instead of 2, and a wider variant with a doubled encoder size (Table 8). Performance improved
as model size increased, suggesting that practitioners who can afford greater computational costs
could obtain better results by training a larger Presto model.
6 Discussion & Conclusion
Limitations Presto is designed to ingest 10m/px resolution imagery and is pre-trained on products
at this scale. This decision is motivated by the free, global availability over time of products at
this scale (such as Sentinel-1 and Sentinel-2). Presto does not natively process very-high resolution
imagery such as <1m/px imagery from commercial satellites or drones, which can be costly and
often lack complete coverage globally and temporally. In addition, Presto is a pixel-timeseries model.
While we demonstrated Presto’s flexibility on single-timestep image datasets, image-based models
may be preferred if a user’s goal is to process entire images to make a prediction. We observed that
Presto’s performance on the EuroSAT dataset plateaued as the input resolution increased (Table 5),
due to images from classes where the relevant pixels for the class are a minority of the pixels in the
image (e.g., highways). In such scene classification challenges, image-based models which can learn
the shape of the relevant pixels may be better suited. We discuss this further in Section A.6.
Conclusion We present Presto: a lightweight, pre-trained timeseries transformer for remote sensing.
By leveraging structure unique to remote sensing data—specifically, (i) an important temporal
dimension, (ii) associated metadata and (iii) a diversity of sensors, we are able to train an extremely
lightweight model which achieves state-of-the-art results in a wide variety of globally distributed
evaluation tasks. Computational efficiency is of paramount importance in remote sensing settings
and often determines which models ultimately get selected for deployment. We demonstrated that
strong performance can be achieved while meeting this constraint, and that self-supervised learning
can provide significant benefits even for small models.
11
Table 8: Effect of model size on validation performance . To understand the effect of model size
on performance, we pre-train two larger variants of Presto. As in Table 7, we measure Presto R’s
performance on the CropHarvest validation task. The number of parameters includes both the encoder
and decoder parameters. The FLOPS are computed for a “full” input (12 timesteps, with no missing
channels), when passed through the encoder and decoder.
Depth Width# params
(M)FLOPs
(M)F1
score
2 128 0.81 88.94 0.665
2 256 2.02 220.81 0.687
4 128 1.21 132.42 0.669
Impact statement
Machine learning applications to remote sensing have a wide range of 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).
Presto is designed to be accessible to a wide range of practitioners; we achieve this by only training
Presto on publicly available data and by keeping the model size small enough so it can be leveraged
in compute-constrained environments. In addition to increasing Presto’s accessibility, its small size
also lowers its carbon footprint (Strubell et al., 2019).
As described by Tuia et al. (2023), a natural concern when applying machine learning algorithms to
remote sensing data is its use to collect information about individuals who are unaware that data is
being collected, and therefore cannot consent to this practice. We therefore encourage deployment
of Presto in collaboration with local communities and stakeholders (Krafft; Kshirsagar et al., 2021;
Nakalembe and Kerner, 2023).
Acknowledgements
This work was supported by NASA under the NASA Harvest Consortium on Food Security and
Agriculture (Award #80NSSC18M0039). This research was enabled in part by compute resources
provided by Mila (mila.quebec); in addition, we acknowledge material support from NVIDIA
Corporation in the form of computational resources. We thank Esther Rolf and Caleb Robinson for
reviewing drafts of this manuscript.
References
Earth engine data catalogue. https://developers.google.com/earth-engine/datasets/
catalog . Accessed: 2023-01-31.
Tick tick bloom: Harmful algal bloom detection challenge.
https://www.drivendata.org/competitions/143/tick-tick-bloom/page/649/, 2023. Accessed:
2023-03-10.
S. 90m Digital Elevation Data. The CGIAR consortium for spatial information, 2003.
C. Abys, S. Skakun, and I. Becker-Reshef. Two decades of winter wheat expansion and intensification