1. Use of dense time series of high resolution
images for change detection
and land use classification
J. Inglada, B. Beguet, J.-F. Dejoux, C. Marais-Sicre, D. Ducrot, M. Huc,
O. Hagolle, F. Baup, G. Dedieu
CESBIO, UMR 5126, Toulouse, France - jordi.inglada@cesbio.cnes.fr
Real Time Land Cover Map Production
CESBIO’s Sud-Ouest Project: the goal of this project is to contribute to the understanding and the modeling of the continental surfaces at the landscape and regional levels and to
increase knowledge and develop generic methods. Yearly monitoring and regular satellite image acquisitions since 2006, 3 permanent instrumented sites.
Land-cover map production: as a means (input for models) but also as a research goal in itself.
Real time land cover map production: know as soon as possible in the agricultural season which is the crop which is going to be grown.
Satellite Data
Formosat-2: 11 images in 2008.
• For soil work:
– 29th August to 12th November 2008
– 5 Formosat-2 images
Image pre-processing: all data are geometrically corrected and ra-
diometrically calibrated; cloud screening is performed [3]
(a) February 11, 2008 (b) July 10, 2008 (c) October 26, 2008
Ground Data
2008 campaign:
• 14 terrain surveys
• 650 plots revisited
• only 501 plots were kept for the land-cover classifications
• accuracy of the surveys allowing for diachronic studies
Soil work: 7 field surveys for 300 plots
• Each ground sample is associated
with a confidence index
• Some soil states are visible on the
ground before being detected on
the images
Supervised Classification
Yearly classification: using the
data for the whole season.
Advanced methods: described
in [2, 4]
class accuracy (%)
broad leaf forest 97.88
needle leaf forest 97.05
eucalyptus 74.53
rape 99.33
barley 99.06
maize 99.60
sunflower 99.12
sorghum 100.00
soybean 97.36
fallow 97.75
grassland 95.16
Soil Work
Problem Position
Main goal: improve real-time crop classification; soil work can give hints on the type of crop
Soil map: is also interesting in itself as a product
Classes of interest:
Inter-crop Stubble disking Deep ploughing Harrowing Sowing preparation Emergence
Crops (C): Sunflowers, which are mostly dry in September and harvested in September or October. Irrigated
soybean and maize, which are green in September and begin to dry in October (harvest in October and
November).
Inter-crop (IC): Begin after harvest. No recent or visible soil tillage. Stubble stands often right, crop
residues may be visible on top soil. Some green plants can grow, like volunteers (or regrow) and weeds, if
climatic conditions are favorable (rain, etc.).
Stubble disking (SD): Superficial (5 to 15 cm) soil tillage in order to mix crop residues and soil and to
destroy green vegetation (weeds). Soil surface is irregular, has some small clods and a small roughness.
Stubble and crop residues are partly visible.
Deep ploughing (DP): Mainly mouldboard ploughing between 20 to 45 cm deep. More than 95% bare
soil: no visible crop residues. Visible clods and strong roughness.
Harrowing (H): Secondary or superficial tillage. More than 95% bare soil. There are medium sized clods.
Improper for seedling. Various tillage operations are possible: rotary harrowing, chiseling, superficial plough-
ing (less than 20 cm deep).
Remark: Some green plants (volunteers, weeds) may be visible for the 4 previous categories, only if climatic
conditions are favorable and duration between each stage or soil tillage is sufficient. In the present poster,
it was sometimes the case only in inter-crops or after stubble disking.
Sowing preparation (SP): More than 95% bare soil. Soil ready for seedling. Regular surface. Small clods.
Emergence (E): Germination. Plants are visible from field borders and are at cotyledons or first leaves
development stages. Plant height lower than 5 cm.
Approach
Radiometry only: only the reflectances and combi-
nations of them (indexes) are used; no texture, statis-
tics, nor object-based features.
Statistical analysis: the temporal evolution of the
reflectances and the indexes – globally and per class
– are studied.
2 kinds of analysis:
1. Identification of the soil state: classification
2. Identification of the transitions between states:
change detection
SVM classification: Support Vector Machines [1]
are both used as separability measure and as clas-
sification tool.
Index Formula
NDVI NIR−R
NIR+R
Color R−B
R
Brightness
√
G2 + R2 + NIR2
Shape 2R−G−B
G−B
Redness R−V
R+V
Classification
Direct approach: each soil state is
considered a class and a supervised
classification is performed
Crop class: not so easy to classify,
since it corresponds to several crop
types
Errors:
• IC can be confused with MT, since
the amount of green vegetation
before tilling varies very much;
• many confusions between bare soil
classes;
• germination is correctly detected
Grouping soil classes improves the
classification
C IC SD DP H SP E
C 66.4 9.54 7.08 4.67 0.35 7.53 4.43
IC 6.54 64.67 14.14 0.95 4.71 3.89 5.1
S 4.08 6.6 63.5 1.5 13.6 6.94 3.78
DP 6.36 2.76 2.64 57.54 16.51 10.53 3.66
H 1.53 1.35 6.9 20.85 44.09 23.17 2.11
SP 3.6 0.0 6.17 23.1 13.12 41.52 12.49
E 1.28 5.85 1.67 0.08 1.72 1.6 87.8
Overall Accuracy = 0.6085
Kappa = 0.541
C IC SD Soil E
C 65.65 10.47 8.9 7.12 7.86
IC 6.19 65.22 16.16 6.23 6.2
SD 5.29 6.61 67.88 15.43 4.79
Soil 3.92 2.16 8.18 77.27 8.47
E 2.98 6.39 2.32 2.31 86.0
Overall Accuracy = 0.7235
Kappa = 0.6555
Change Detection
Classes are transitions: supervised classification is used in order to detect transitions between soil states.
C→ IC SD DP H SP
IC 73.49 16.23 0.03 1.07 9.18
SD 6.93 53.17 5.63 12.73 21.54
DP 0.65 2.14 83.07 3.54 10.6
H 2.47 5.93 10.7 74.97 5.93
SP 5.08 1.63 8.96 2.5 81.83
Overall Accuracy = 0.7354
Kappa = 0.667
IC→ SD DP H SP
SD 75.99 11.54 10.97 1.5
DP 3.58 89.67 6.75 0.0
H 14.32 15.87 62.15 7.66
SP 0.58 0.0 2.92 96.5
Overall Accuracy = 0.8125
Kappa = 0.7485
SD→ DP H SP E
DP 74.1 14.02 3.2 8.68
H 23.96 32.88 28.59 14.57
SP 7.51 13.69 65.91 12.89
E 6.82 5.69 7.25 80.24
Overall Accuracy = 0.633
Kappa = 0.5105
Transition D→H H→SP H→E SP→E
Accuracy (%) 97.0 88.74 87.91 96.76
• The number of transitions is very low for some
cases (between 12 and 50 plots; or between 1000
and 10000 pixels)
• Many transitions between states can’t be de-
tected accurately
• However, some changes are well detected (about
90% and more)
Conclusions
Soil work knowledge is needed to improve real-time land-cover map production; soil maps
are also useful in themselves
Soil states are difficult to identify using direct classification and optical radiometry only
Soil state changes can be detected in some cases, but many transitions seem difficult to identify
References
[1] C.J.C. Burges. A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery, 2(2):121–167, 1998.
[2] D. Ducrot, A. Masse, C. Marais-Sicre, J-F. Dejoux, and F. Baup. Multisensor and multitemporal image fusion methods to improve remote sensing image classification. In Recent Advances in Quantitative Remote
Sensing, September 2010.
[3] O. Hagolle, M. Huc, D. Villa Pascual, and G. Dedieu. A multi-temporal method for cloud detection, applied to FORMOSAT-2, VENµS, LANDSAT and SENTINEL-2 images. Remote Sensing of Environment,
114:1747–1755, August 2010.
[4] S. Idbraim and D. Ducrot. An unsupervised classification using a novel ICM method with constraints for land cover mapping from remote sensing imagery. International Review on Computers and Software (IRECOS),
March 2009.