Improved Crop Classification with Rotation
Knowledge using Sentinel-1 and -2 Time Series
Sébastien Giordano, Simon Bailly, Loic Landrieu, and Nesrine Chehata
Abstract
Leveraging the recent availability of accurate, frequent, and
multimodal (radar and optical) Sentinel-1 and -2 acquisitions,
this paper investigates the automation of land parcel identi-
fication system (
LPIS
) crop type classification. Our approach
allows for the automatic integration of temporal knowledge,
i.e., crop rotations using existing parcel-based land cover
databases and multi-modal Sentinel-1 and -2 time series. The
temporal evolution of crop types was modeled with a linear-
chain conditional random field, trained with time series of
multi-modal (radar and optical) satellite acquisitions and
associated
LPIS
. Our model was tested on two study areas in
France (
≥
1250 km
2
) which show different crop types, various
parcel sizes, and agricultural practices: .the Seine et Marne
and the Alpes de Haute-Provence classified accordingly
to a fine national 25-class nomenclature. We first trained
a Random Forest classifier without temporal structure to
achieve 89.0% overall accuracy in Seine et Marne (10 classes)
and 73% in Alpes de Haute-Provence (14 classes). We then
demonstrated experimentally that taking into account the
temporal structure of crop rotation with our model resulted
in an increase of 3% to +5% in accuracy. This increase was
especially important (+12%) for classes which were poorly
classified without using the temporal structure. A stark posi-
tive impact was also demonstrated on permanent crops, while
it was fairly limited or even detrimental for annual crops.
Introduction
The Sentinel-1 and -2 satellites provide open and free acquisi-
tions exhibiting unprecedented characteristics which are
well-suited to agriculture monitoring: high temporal frequen-
cy (5–6 days), the complementary C-band Sentinel-1 radar
images and multispectral Sentinel-2 images with relevant
spectral bands to crop mapping, and high spatial resolution
(10–20 m). In Europe, several cases of agricultural monitoring
using Sentinel images have been proposed (European Com-
mission 2016), such as observing crops (e.g., for crop area
estimates, crop map products, crop phenology indicators) and
controlling Common Agricultural Policy payments (e.g., for
permanent grasslands, greening measures). Sentinel Images
have also been used for updating and controlling the quality
of the land parcel identification system (
LPIS
), a geographical
information system on agricultural parcels, at a national scale,
updated annually (Boryan, Yang, Mueller
et al.
2011).
This paper introduces a tool for automated
LPIS
crop type classi-
fication frommulti-modal Sentinel time series which incorporates
knowledge from existing
LPIS
editions to improve its accuracy.
Multi-Temporal Satellite Images for Crop Mapping
Satellite time series are particularly well-suited for identify-
ing different crop types, as they allow for the monitoring of the
evolution of the plant’s phenology. This is particularly crucial
in the growing or harvest seasons. Synthetic aperture radar (
SAR
)
data are crucial as well, as they mitigate the effect of cloud cover.
Many studies have demonstrated the potential of multi-
temporal Sentinel and Landsat-8 images for crop type
mapping (Palchowdhuri, Valcarce-Di
ň
eiro, King
et al.
2018;
Veloso, Mermoz, Bouvet
et al.
2017; Vuolo, Neuwirth, Im-
mitzer
et al.
2018; Belgiu and Csillik 2018; Ottosen, Lommen,
SkjãÃÿth, 2019; Defourny, Bontemps, Bellemans
et al.
2019)
and the contribution of
SAR
time series for crop monitoring
(Whelen and Siqueira, 2017; Li, Zhang, Zhang
et al.
2019).
Inglada, Arias, and Tardy
et al.
(2015) assessed state-of-the-
art methods for automatic crop mapping with multi-temporal
and high resolution optical images. Five different classification
approaches using
SPOT4
and Landsat-8 images were compared
for six annual crops, over 12 different study areas, with the
best results (overall accuracy) (
OA
= 80%) obtained using the
Random Forest classifier. In Kussul, Lemoine, and Gallego
et
al.
(2016), Landsat-8 and Sentinel-1 time series were used on
a study area in Ukraine. A pixel-based classification com-
bined with a parcel-based regularization (majority voting) was
proposed using
LPIS
ancillary data. An
OA
of 89% was reached,
but only on a nomenclature comprised of six annual crops
and large parcels (
>
250 hectares (ha)). Wagner, Narasimhan,
and and Waske (2018) combined Sentinel-1 and -2 to improve
land cover mapping in cloud-prone regions. Veloso, Mermoz,
and Bouvet
et al.
(2017) showed the importance of radar data
for crop mapping. More recently, the Sentinel-2 Agriculture
Consortium (Sen2Agri) has led experiments at the country
level (Czech Republic) using Sentinel time series for crop
mapping (Sen2-Agri
GISAT
s.r.o. 2018). A multi-sensor (Sen-
tinel-1, Sentinel-2) pixel-based supervised classification was
performed. The
LPIS
was used for both learning and validation
steps. Monthly cropland maps were produced with an overall
accuracy greater than 80%, and each land cover type had a F-
score greater than 60%. In Defourny, Bontemps, Bellemans
et
al.
(2019), three entire countries (Ukraine, Mali, and South Af-
rica) and five local cities were mapped using Sen2Agri system.
Overall accuracy values were higher than 90%, and already as
high as 80% midseason. However, only the five major crops
were considered for each site. For the Sen2Agri framework,
the nomenclature was generally limited to 5–7 classes and did
not fully integrate temporal knowledge from existing data.
Crop Rotation Integration
Crop rotation knowledge can be used to improve agricultural
yields (Berzsenyi, Györffy, and Lap
et al.
2000) and soil qual-
ity (Karlen, Hurley, Andrews
et al.
2006). Crop type predic-
tion can also be improved using
prior
knowledge on crop
rotations per parcel since a crop type is strongly correlated
to past crop types. Modeling such temporal structures from
Sébastien Giordano, Loic Landrieu, and Nesrine Chehata are
with the University Paris-Est, LASTIG MATIS, IGN, ENSG,
F-94160 Saint-Mandé, France.
Simon Bailly is with Wanaka, Paris, France.
Nesrine Chehata is with EA 4592, G & E Lab, Bordeaux
INP/Bordeaux Montaigne University, France.
Photogrammetric Engineering & Remote Sensing
Vol. 86, No. 7, July 2020, pp. 431–441.
0099-1112/20/431–441
© 2020 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.86.7.431
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2020
431
