Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
crop type recognition.
3.3. Tokenization of SITS inputs
A SITS record X∈RT×H×W×Cconsists of a series
ofTsatellite images of spatial dimensions H×Wwith
Cchannels. For the tokenization of our 3D inputs, we can
extend the tokenization-as-convolution approach to 3D data
and apply a 3D kernel with size (t×h×w)at stride (t, h, w )
across temporal and spatial dimensions. In this manner
N=⌊T
t⌋⌊H
h⌋⌊W
w⌋non-overlapping tokens xi∈RthwC
are extracted, and subsequently projected to ddimensions.
Using t >1, all extracted tokens contain spatio-temporal
information. For the special case of t= 1 each token
contains spatial-only information for each acquisition time
and temporal information is accounted for only through the
encoder layers. Since the computation cost of global self-
attention layers is quadratic w.r.t. the length of the token se-
quence O(N2), choosing larger values for t, h, w can lead
to significantly reduced number of FLOPS. In our experi-
ments, however, we have found small values for t, h, w to
work much better in practice. For all presented experiments
we use a value of t= 1motivated in part because this choice
simplifies the implementation of acquisition-time-specific
temporal position encodings, described in section 3.6. With
regards to the spatial dimensions of extracted patches we
have found small values to work best for semantic segmen-
tation, which is reasonable given that small patches retain
additional spatial granularity. In the end, our tokenization
scheme is similar to ViT’s applied in parallel for each acqui-
sition as shown in Fig.3, however, at this stage, instead of
unrolling feature dimensions, we retain the spatial structure
of the original input as reshape operations will be handled
by the TSViT encoder submodules.
3.4. Encoder architecture
In the previous section we presented a motivation for us-
ing small values t, h, w for the extracted patches. Unless
other measures are taken to reduce the model’s computa-
tional cost this choice would be prohibitive for process-
ing SITS with multiple acquisition times. To avoid such
problems, we choose to factorize our inputs across their
temporal and spatial dimensions, a practice commonly em-
10421
Figure 4. TSViT submodules. (a) Temporal encoder. We reshape tokenized inputs, retaining the spatio-temporal structure of SITS, into
a set of timeseries for each spatial location, add temporal position encodings PT[t,:]for acquisition times t, concatenate local clstokens
ZTcls (eq.5) and process in parallel with a Transformer. Only the first Koutput tokens are retained. (b)Spatial encoder. We reshape
the outputs of the temporal encoder into a set of spatial feature maps for each clstoken, add spatial position encodings PS, concatenate
global clstokens ZScls(eq.6) and process in parallel with a Transformer. (c)Segmentation head. Each local clstoken is projected into hw
values denoting class-specific evidence for every pixel in a patch. All patches are then reassembled into the original image dimensions. (d)
Classification head. Global clstokens are projected into scalar values, each denoting evidence for the presence of a specific class.
ployed for video processing [17,27,37,56,66,72]. We note
that all these works use a spatial-temporal factorization or-
der, which is reasonable when dealing with natural images,
given that it allows the extraction of higher level, semanti-
cally aware spatial features, whose relationship in time is
useful for scene understanding. However, we argue that in
the context of SITS, reversing the order of factorization is
a meaningful design choice for the following reasons: 1) in
contrast to natural images in which context can be useful for
recognising an object, in crop type recognition context can
provide little information, or can be misleading. This arises
from the fact that the shape of agricultural parcels, does not
need to follow its intended use, i.e. most crops can gener-
ally be cultivated independent of a field’s size or shape. Of
course there exist variations in the shapes and sizes of agri-
cultural fields [34], but these depend mostly on local agri-
cultural practices and are not expected to generalize over
unseen regions. Furthermore, agricultural parcels do not in-
herently contain sub-components or structure. Thus, know-
ing what is cultivated in a piece of land is not expected to
provide information about what grows nearby. This is in
contrast to other objects which clearly contain structure, e.g.
in human face parsing there are clear expectations about
the relative positions of various face parts. We test this hy-
pothesis in the supplementary material by enumerating over
all agricultural parcels belonging to the most popular crop
types in the T31TFM S2tile in France and taking crop-type-
conditional pixel counts over a 1km square region from their
centers. Then, we calculate the cosine similarity of thesevalues with unconditional pixel counts over the extent of
the T31TFM tile and find a high degree of similarity, sug-
gesting that there are no significant variations between these
distributions; 2) a small region in SITS is far more informa-
tive than its equivalent in natural images, as it contains more
channels than regular RGB images ( S2imagery contains 13
bands in total) whose intensities are averaged over a rela-
tively large area (highest resolution of S2images is 10×10
m2); 3) SITS for land cover recognition do not typically
contain moving objects. As a result, a timeseries of single
pixel values can be used for extracting features that are in-
formative of a specific object part found at that particular
location. Therefore, several objects can be recognised us-
ing only information found in a single location; plants, for
example, can be recognised by variations of their spectral
signatures during their growth cycle. Many works perform-
ing crop classification do so using only temporal informa-
tion in the form of timeseries of small patches [47], pixel
statistics over the extent of parcels [46] or even values from
single pixels [40, 48]. On the other hand, the spatial pat-