text
stringlengths
0
820
4. Experiments
We apply TSViT to two tasks using SITS records X∈
RT×H×W×Cas inputs: classification and semantic seg-
mentation. At the object level, classification models learn
a mapping f(X)∈RKfor the object occupying the center
of the H×Wregion. Semantic segmentation models learn
a mapping f(X)∈RH×W×K, predicting class probabili-
ties for each pixel over the spatial extent of the SITS record.
We use an ablation study on semantic segmentation to guide
model design and hyperparameter tuning and proceed with
presenting our main results on three publicly available SITS
semantic segmentation and classification datasets.
4.1. Training and evaluation
Datasets To evaluate the performance of our proposed
semantic segmentation model we are using three publicly
available S2land cover recognition datasets. The dataset
presented in [47] covers a densely cultivated area of interest
of102×42km2north of Munich, Germany and contains 17
distinct classes. Individual image samples cover a 240×240
m2area ( 24×24pixels) and contain 13 bands. The PASTIS
10423
dataset [14] contains images from four different regions in
France with diverse climate and crop distributions, span-
ning over 4000 km2and including 18 crop types. In total,
it includes 2.4k SITS samples of size 128×128, each con-
taining 33-61 acquisitions and 10 image bands. Because
the PASTIS sample size is too large for efficiently train-
ing TSViT with available hardware, we split each sample
into24×24patches and retain all acquisition times for a
total of 60k samples. To accommodate a large set of ex-
periments we only use fold-1 among the five folds provided
in PASTIS. Finally, we use the T31TFM-1618 dataset [60]
which covers a densely cultivated S2tile in France for years
2016-18 and includes 20 distinct classes. In total, it includes
140k samples of size 48×48, each containing 14-33 ac-
quisitions and 13 image bands. For the SITS classification
experiments, we construct the datasets from the respective
segmentation datasets. More specifically, for PASTIS we
use the provided object instance ids to extract 24×24pixel
regions whose center pixel falls inside each object and use
the class of this object as the sample class. The remain-
ing two datasets contain samples of smaller spatial extent,
making the above strategy not feasible in practice. Here, we
choose to retain the samples as they are and assign the class
of the center pixel as the global class. We note that this
strategy forces us to discard samples in which the center
pixels belongs to the background class. Additional details
are provided in the supplementary material.
Implementation details For all experiments presented
we train for the same number of epochs using the provided
data splits from the respective publications for a fair com-
parison. More specifically, we train on all datasets using
the provided training sets and report results on the valida-
tion sets for Germany and T31TFM-1618, and on the test
set for PASTIS. For training TSViT we use the AdamW op-
timizer [23] with a learning rate schedule which includes a
warmup period starting from zero to a maximum value 10−3
at epoch 10, followed by cosine learning rate decay [33]
down to 5∗10−6at the end of training. For Germany and
T31TFM-1618 we train with the above settings and report
the best performances between what we achieve and the
original studies. Since we split PASTIS, we are training
with both settings and report the best results. Overall, we
find that our settings improve model performance. We train
with a batch size of 16 or 32 and no regularization on ×2
Nvidia Titan Xp gpus in a data parallel fashion. All mod-
els are trained with a Masked Cross-Entropy loss, masking
the effect of the background class in both training loss and
evaluation metrics. We report overall accuracy (OA), aver-
aged over pixels, and mean intersection over union (mIoU)
averaged over classes. For SITS classification, in addition
to the 1D models presented in section 2 we modify the best
performing semantic segmentation models by aggregating
extracted features across space prior to the application of aAblation Settings mIoU
Factorization orderSpatial & Temporal 48.8
Temporal & Spatial 78.5
#clstokens1 78.5
K 83.6
Position encodingsStatic 80.8
Date lookup 83.6
Interactions between
clstokensTemporal Spatial
✓ ✓ 81.5
✓✓✓ 83.6
Patch size2×2 84.8
3×3 83.6
6×6 79.6
Table 1. Ablation on design choices for TSViT . All proposed
design choices are found to have a positive effect on performance.
classifier, thus, outputing a single prediction. Classification
models are trained with Focal Loss [30]. We report OA and
mean accuracy (mAcc) averaged over classes.
4.2. Ablation studies
We perform an ablation study on design parameters of
our framework using the Germany dataset [47]. Starting
with a baseline TSViT with L= 4 for both encoder net-
works, a single clstoken, h=w= 3, t= 1, d=
128we successively update our design after each ablation.
Here, we present the effect of the most important design