gremlin97/RemoteSensingCorpus · Datasets at Hugging Face

text stringlengths 0 820
can significantly improve model performance (Rao et al., 2020; Hengl et al., 2017). We therefore
pre-trained Presto on a diverse set of directly-sensed and derived Earth observation products which
we pre-processed and exported using Google Earth Engine (Gorelick et al., 2017).
A pre-training batch contained several pixel-timeseries samples, each of which is a concatenation of
dynamic-in-time datapoints with each timestep representing a month (yielding T= 12 timesteps in
total). The following dynamic-in-time data products were used, yielding 15channels: (i) Sentinel-2
(S2) multispectral data, (ii) Sentinel-1 (S1) radar data, (iii) ERA5 climate reanalysis data, (iv) NDVI
(Rouse et al., 1974) derived from Sentinel-2 data and (v) land cover classes Vfrom Dynamic World.
To every pixel-timeseries we appended two static-in-time products: (i) topography data from the
SRTM digital elevation model (90m Digital Elevation Data, 2003) and (ii) location coordinates of
each pixel. Hence, one pre-training sample x, comprising a pixel-timeseries t∈[RT×15;VT×1]and
static variables s∈R1×5, is summarized as follows:
x=h
tS1
i;tS2
i;tERA5
i;tNDVI
i;tDW
i\|i= 1, ...,12
;sTG;sLoci
(1)
From now on, we use “pixel-timeseries” to refer to both the dynamic and the static variables.
3.2 Encoding and tokenization
We transformed the pixel-timeseries xinto a number of tokens (each represented by an embedding e)
to be processed by the Presto transformer. Per timestep 0≤i < T , we split the input variables into
channel groups Caccording to their type of sensor or source: e.g., the S1 bands form one channel
group. We describe these groups in more detail in Appendix A.1.3. Each real-valued channel group
represents a different sensor, native spatial resolution or (in the case of Sentinel-2 channel-groups)
region of the electromagnetic spectrum. We projected each channel group to a common latent space
of dimension deby separate learned linear projections hC: e.g., eS1
i=hS1(tS1
i). The Dynamic World
classes are categorical, so we embedded them by indexing them into an embedding matrix.
Unlike natural images in which the data and its label are self-contained, remote sensing labels are
inherently associated to a place and time on Earth (i.e., a latitude/longitude and timestamp). In
addition, while natural images contain RGB channels from the same camera sensor, Presto’s pixel-
timeseries input contains channels from multiple remote sensing instruments and data products. We
therefore wanted to communicate to the model: (i) the location of the datapoint (already present in
4
Table 1: We evaluated Presto on a wide variety of downstream tasks , including segmentation
(seg.), multi-label (ml) scene classification (class.) and regression (reg.) tasks. There is diversity in
terms of data composition, geographic area and training set size. Input shape describes the shape of a
single sample, in terms of [Height, Width, Timesteps, Channels]. We bold the temporal dimension,
to highlight time-series versus single-timestep inputs.
Dataset Task RegionInput shape
[H, W, T, C]Train
samples
CropHarvest Seg.Kenya
[1, 1, 12, 18]1,345
Brazil 203
Togo 1,319
S2-Agri 100 Class. France [5, 5, 24, 10] 1,500
TreeSatML
Class.Germany[6, 6, 1, 2]45,337[6, 6, 1, 11]
EuroSat Class. Europe[64, 64, 1, 3]21,600[64, 64, 1, 11]
Fuel Moisture Reg. USA [1, 1, 3, 19] 1,578
Algae Blooms Reg. USA [1, 1, 12, 19] 777
the input as static variable through coordinates sLoc) and a variable’s (ii) timestamp and (iii) channel
group. We did this by adding encodings to the previously described embeddings e. The complete
encoding has dimension deand contains a concatenation of positional, month, and learned channel
encodings described below.
•Positional: We used the sinusoidal positional encoding originally used by Vaswani et al. (2017).
•Month: We added an encoding representing the month being captured by each token, because we
expect timesteps from similar months to have similar features even if they are from different years.
We assign an integer to each month ranging from 0to11, yielding:
pmonth,2i= sin ((2 π×month )/12) (2)
pmonth,2i+1= cos ((2 π×month )/12) (3)
For static-in-time variables, the positional and month encodings were set to zero.
•Channel Group: Each token is associated with a set of input channels. In multispectral SatMAE
(Cong et al., 2022), a fixed encoding was used to communicate input-band information with
different channels representing different wavelengths, which is possible because only input data
from one sensor (Sentinel-2) is used. However, since Presto’s input data includes multiple remote
sensing products, we applied a learnable encoding for each channel group from the set of possible
channel groups C={S1,S2 RGB , ...,ERA5 ,TG,Loc}.
The transformer input E∈R(T·\|Cdynamic\|+\|Cstatic\|)×de(for encoder dimension de) is a concatenation of:
•Dynamic variables, for timesteps i < T and channel groups c∈ C :ec
i=hc(tc
i) +
[pc
channel ;psin(i);pmonth(i) ]
• Topographical data: eTG=hTG(sTG) + [pTG
channel ; 0; 0]
• Coordinates: eLoc=hLoc(sLoc)
3.3 Pre-training via Structured Masking
A key requirement for Presto was to perform well even with incomplete inputs (i.e., when there are
missing timesteps, channels, or both). When masking out part of the input x, we therefore tailored
the masking strategies to encourage the model to learn representations that perform well when given
a subset of bands or timesteps for downstream tasks. For a T×Dinput of Ttimesteps and Dtotal
input channels, we used the following masking techniques (illustrated in Figure 1), where Presto
considers a token to be a 1×dinput (a single timestep of dgrouped channels). The coordinates were
never masked but the static topological tokens can be.
1.Random :(t×d)masked values, with t < T andd < D
5
Table 2: Mean F1 score across all CropHarvest tasks. Presto outpeforms TIML (Tseng et al., 2022)
and MOSAIKS-1D while requiring the adaptation of far fewer parameters. The TIML and
MOSAIKS-1D model did not receive Dynamic World as input, so we measured Presto’s performance
both with and without it.
#. parameters
Model Total Adapted Mean F1
Random Forest 0.441

can significantly improve model performance (Rao et al., 2020; Hengl et al., 2017). We therefore

pre-trained Presto on a diverse set of directly-sensed and derived Earth observation products which

we pre-processed and exported using Google Earth Engine (Gorelick et al., 2017).

A pre-training batch contained several pixel-timeseries samples, each of which is a concatenation of

dynamic-in-time datapoints with each timestep representing a month (yielding T= 12 timesteps in

total). The following dynamic-in-time data products were used, yielding 15channels: (i) Sentinel-2

(S2) multispectral data, (ii) Sentinel-1 (S1) radar data, (iii) ERA5 climate reanalysis data, (iv) NDVI

(Rouse et al., 1974) derived from Sentinel-2 data and (v) land cover classes Vfrom Dynamic World.

To every pixel-timeseries we appended two static-in-time products: (i) topography data from the

SRTM digital elevation model (90m Digital Elevation Data, 2003) and (ii) location coordinates of

each pixel. Hence, one pre-training sample x, comprising a pixel-timeseries t∈[RT×15;VT×1]and

static variables s∈R1×5, is summarized as follows:

x=h

tS1

i;tS2

i;tERA5

i;tNDVI

i;tDW

i|i= 1, ...,12

;sTG;sLoci

(1)

From now on, we use “pixel-timeseries” to refer to both the dynamic and the static variables.

3.2 Encoding and tokenization

We transformed the pixel-timeseries xinto a number of tokens (each represented by an embedding e)

to be processed by the Presto transformer. Per timestep 0≤i < T , we split the input variables into

channel groups Caccording to their type of sensor or source: e.g., the S1 bands form one channel

group. We describe these groups in more detail in Appendix A.1.3. Each real-valued channel group

represents a different sensor, native spatial resolution or (in the case of Sentinel-2 channel-groups)

region of the electromagnetic spectrum. We projected each channel group to a common latent space

of dimension deby separate learned linear projections hC: e.g., eS1

i=hS1(tS1

i). The Dynamic World

classes are categorical, so we embedded them by indexing them into an embedding matrix.

Unlike natural images in which the data and its label are self-contained, remote sensing labels are

inherently associated to a place and time on Earth (i.e., a latitude/longitude and timestamp). In

addition, while natural images contain RGB channels from the same camera sensor, Presto’s pixel-

timeseries input contains channels from multiple remote sensing instruments and data products. We

therefore wanted to communicate to the model: (i) the location of the datapoint (already present in

4

Table 1: We evaluated Presto on a wide variety of downstream tasks , including segmentation

(seg.), multi-label (ml) scene classification (class.) and regression (reg.) tasks. There is diversity in

terms of data composition, geographic area and training set size. Input shape describes the shape of a

single sample, in terms of [Height, Width, Timesteps, Channels]. We bold the temporal dimension,

to highlight time-series versus single-timestep inputs.

Dataset Task RegionInput shape

[H, W, T, C]Train

samples

CropHarvest Seg.Kenya

[1, 1, 12, 18]1,345

Brazil 203

Togo 1,319

S2-Agri 100 Class. France [5, 5, 24, 10] 1,500

TreeSatML

Class.Germany[6, 6, 1, 2]45,337[6, 6, 1, 11]

EuroSat Class. Europe[64, 64, 1, 3]21,600[64, 64, 1, 11]

Fuel Moisture Reg. USA [1, 1, 3, 19] 1,578

Algae Blooms Reg. USA [1, 1, 12, 19] 777

the input as static variable through coordinates sLoc) and a variable’s (ii) timestamp and (iii) channel

group. We did this by adding encodings to the previously described embeddings e. The complete

encoding has dimension deand contains a concatenation of positional, month, and learned channel

encodings described below.

•Positional: We used the sinusoidal positional encoding originally used by Vaswani et al. (2017).

•Month: We added an encoding representing the month being captured by each token, because we

expect timesteps from similar months to have similar features even if they are from different years.

We assign an integer to each month ranging from 0to11, yielding:

pmonth,2i= sin ((2 π×month )/12) (2)

pmonth,2i+1= cos ((2 π×month )/12) (3)

For static-in-time variables, the positional and month encodings were set to zero.

•Channel Group: Each token is associated with a set of input channels. In multispectral SatMAE

(Cong et al., 2022), a fixed encoding was used to communicate input-band information with

different channels representing different wavelengths, which is possible because only input data

from one sensor (Sentinel-2) is used. However, since Presto’s input data includes multiple remote

sensing products, we applied a learnable encoding for each channel group from the set of possible

channel groups C={S1,S2 RGB , ...,ERA5 ,TG,Loc}.

The transformer input E∈R(T·|Cdynamic|+|Cstatic|)×de(for encoder dimension de) is a concatenation of:

•Dynamic variables, for timesteps i < T and channel groups c∈ C :ec

i=hc(tc

i) +

[pc

channel ;psin(i);pmonth(i) ]

• Topographical data: eTG=hTG(sTG) + [pTG

channel ; 0; 0]

• Coordinates: eLoc=hLoc(sLoc)

3.3 Pre-training via Structured Masking

A key requirement for Presto was to perform well even with incomplete inputs (i.e., when there are

missing timesteps, channels, or both). When masking out part of the input x, we therefore tailored

the masking strategies to encourage the model to learn representations that perform well when given

a subset of bands or timesteps for downstream tasks. For a T×Dinput of Ttimesteps and Dtotal

input channels, we used the following masking techniques (illustrated in Figure 1), where Presto

considers a token to be a 1×dinput (a single timestep of dgrouped channels). The coordinates were

never masked but the static topological tokens can be.

1.Random :(t×d)masked values, with t < T andd < D

5

Table 2: Mean F1 score across all CropHarvest tasks. Presto outpeforms TIML (Tseng et al., 2022)

and MOSAIKS-1D while requiring the adaptation of far fewer parameters. The TIML and

MOSAIKS-1D model did not receive Dynamic World as input, so we measured Presto’s performance

both with and without it.

#. parameters

Model Total Adapted Mean F1

Random Forest 0.441