text
stringlengths
0
820
and semantic segmentation presented in Table 2 and Fig.1.
2. Related work
2.1. Crop type recognition
Crop type recognition is a subcategory of land use recog-
nition which involves assigning one of Kcrop categories
(classes) at a set of desired locations on a geospatial grid.
For successfully doing so modelling the temporal patterns
of growth during a time period of interest has been shown
to be critical [15, 44]. As a result, model inputs are time-
series of Tsatellite images of spatial dimensions H×W
withCchannels, X∈RT×H×W×Crather than single ac-
quisitions. There has been a significant body of work on
crop type identification found in the remote sensing liter-
ature [8, 9, 19, 39, 41, 55]. These works typically involve
1without any convolution operationsmultiple processing steps and domain expertise to guide
the extraction of features, e.g. NDVI [25], that can be
separated into crop types by learnt classifiers. More re-
cently, Deep Neural Networks (DNN) trained on raw op-
tical data [22, 26, 29, 46, 47, 62] have been shown to outper-
form these approaches. At the object level, (SITS classi-
fication) [24, 40, 48] use 1D data of single-pixel or parcel-
level aggregated feature timeseries, rather than the full SITS
record, learning a mapping f:RT×C→RK. Among
these works, TempCNN [40] employs a simple 1D convo-
lutional architecture, while [48] use the Transformer archi-
tecture [64]. DuPLo [24] consists of an ensemble of CNN
and RNN streams in an effort to exploit the complementar-
ity of extracted features. Finally, [16] view satellite images
as un-ordered sets of pixels and calculate feature statistics
at the parcel level, but, in contrast to previously mentioned
approaches, their implementation requires knowledge of the
object geometry. At the pixel level (SITS semantic seg-
mentation), models learn a mapping f(X)∈RH×W×K.
For this task, [47] show that convolutional RNN variants
(CLSTM, CGRU) [54] can automatically extract useful fea-
tures from raw optical data, including cloudy images, that
can be linearly separated into classes. The use of CNN
architectures is explored in [45] who employ two models:
a UNET2D feature extractor, followed by a CLSTM tem-
poral model (UNET2D-CLSTM); and a UNET3D fully-
convolutional model. Both are found to achieve equivalent
performances. In a similar spirit, [7] use a FPN [31] feature
extractor, coupled with a CLSTM temporal model (FPN-
CLSTM). The UNET3Df architecture [60] follows from
UNET3D but uses a different decoder head more suited to
contrastive learning. The U-TAE architecture [14] follows
a different approach, in that it employs the encoder part of
a UNET2D, applied on parallel on all images, and a sub-
sequent temporal attention mechanism which collapses the
temporal feature dimension. These spatial-only features are
further processed by the decoder part of a UNET2D to ob-
tain dense predictions.
2.2. Self-attention in vision
Convolutional [20, 28, 57] and fully-convolutional net-
works (FCN) [52, 53] have been the de-facto model of
choice for vision tasks over the past decade. The convo-
lution operation extracts translation-equivariant features via
application of a small square kernel over the spatial extent
of the learnt representation and grows the feature recep-
tive field linearly over the depth of the network. In con-
trast, the self-attention operation, introduced as the main
building block of the Transformer architecture [64], uses
self-similarity as a means for feature aggregation and can
have a global receptive field at every layer. Following the
adoption of Transformers as the dominant architecture in
natural language processing tasks [6, 12, 64], several works
10419
have attempted to exploit self-attention in vision architec-
tures. Because the time complexity of self-attention scales
quadratically with the size of the input, its naive implemen-
tation on image data, which typically contain more pixels
than text segments contain words, would be prohibitive. To
bypass this issue, early works focused on improving effi-
ciency by injecting self-attention layers only at specific lo-
cations within a CNN [5, 67] or by constraining their re-
ceptive field to a local region [38, 42, 65], however, in prac-
tice, these designs do not scale well with available hard-
ware leading to slow throughput rates, large memory re-
quirements and long training times. Following a different
approach, the Vision Transformer (ViT) [13], presented in
further detail in section 3.1, constitutes an effort to apply
a pure Transformer architecture on image data, by propos-
ing a simple, yet efficient image tokenization strategy. Sev-
eral works have drawn inspiration from ViT to develop
novel attention-based architectures for vision. For image
recognition, [32, 69] re-introduce some of the inductive bi-
ases that made CNNs successful in vision, leading to im-
proved performances without the need for long pre-training
schedules, [43, 59] employ Transformers for dense predic-
tion, [11,58,71] for object detection and [3,36,68] for video
processing. Among these works, our framework is more
closely related to [3] who use ViT for video processing.
However, we deviate significantly from their design by in-
troducing a dictionary of acquisition-time-specific position
encodings to accommodate an uneven distribution of im-
ages in time, by employing a different input factorization
strategy more suitable to SITS, and by being interested in
both global and dense predictions (section 3.4). Finally,
we are using a multi-token strategy that allows better han-