![Assessment of the Land Cover Classification
Accuracy of Venµs and Sentinel-2 Image Time
Series with respect to Formosat-2
Jordi Inglada, Olivier Hagolle, G´erard Dedieu
CNES-CESBIO, UMR 5126, Toulouse, France - jordi.inglada@cesbio.cnes.fr
Image Time Series
High temporal resolution: monitoring, analysis, and modeling of land surface functioning under the in-
fluences of environmental factors as well as human activities
Coming sensors:
• Venµs: superspectral sensor, dedicated to vegetation monitoring, with a 2-day revisit; developed as a
cooperation between France and Israel and should be launched in 2012.
• Sentinel-2: will provide systematic global acquisitions of high-resolution multispectral imagery with a high
revisit frequency (5 days) for the needs of operational land monitoring and emergency services.
Available image time series: Formosat-2 is already providing image time series with short revisit cycle
and constant viewing angle. CESBIO has been collecting this kind of data since 2006 over several experimental
sites.
(a) March 14, 2006 (b) July 7, 2006 (c) November 2, 2006
Objectives
Temporal sampling: A short revisit period increases the chance of obtaining cloud-free images. Partially
cloudy images can still be used sometimes. What is the influence of temporal sampling in land-cover map
production?
Spectral resolution: Formosat-2 has a very limited spectral resolution with respect to Venµs and Sentinel-2.
What is the influence of spectral resolution in land-cover map production?
Classification as metric: We use the classification accuracy as a comparison metric between the studied
sensors.
Image time series simulation: In order to perform a fair comparison between the sensors, we simulate a
time series for each sensor using the same input data.
Temporal sampling
Time series: A series of 49 Formosat-2 images is used
for this study.
Cloud cover: Only 10 images out of 49 are cloud-free.
The figure on the right shows the cloud percentage
for each acquisition.
Temporal sampling influence: The final temporal
sampling will depend on the acceptance threshold ap-
plied on the cloud cover. Mar 2006
Apr 2006
May 2006
Jun 2006
Jul 2006
Aug 2006
Sep 2006
Oct 2006
Nov 2006
0.0
0.2
0.4
0.6
0.8
1.0
Cloud %
40 dates
30 dates
Spectral Bands
Spectral sensitivity: the num-
ber of available bands and their
width are different for the 3 sen-
sors of interest.
Formosat-2: has visible bands (B,
G, R, NIR).
Venµs: has 10 different bands,
which are narrow, with a fine sam-
pling of the red edge, but lacking
the SWIR.
Sentinel-2: is similar to Venµs,
but with a coarser sampling of the
red edge but adds the SWIR.
500 1000 1500 2000
wavelength
0.0
0.2
0.4
0.6
0.8
1.0
Formosat-2
Relative Spectral Responses
500 1000 1500 2000
wavelength
0.0
0.2
0.4
0.6
0.8
1.0
Venus
500 1000 1500 2000
wavelength
0.0
0.2
0.4
0.6
0.8
1.0
Sentinel-2
Simulations
Formosat-2
Input Series
LAI(t)
Land Cover Map
Cab
Car
N
PROSPECT+SAIL Full Spectra Venµs RSR
Formosat-2 RSR
Sentinel-2 RSR
Main steps
• The Formosat-2 time series is used to estimate the
LAI for each pixel of the classes of interest.
• Leaf pigments are obtained using the LOPEX’93
data base [3].
• A land-cover map [4] is used as input to select the
pigment values for each pixel.
• LAI and leaf pigments are used to feed the
PROSPECT+SAIL [2, 5] simulation code.
• The simulated full spectra are reduced using the
relative spectral responses of each sensor in order
to generate the time series [1].
Limits of the approach
• Leaf pigments are constant for a given class
– Gives more weight to the spectral characteristics
with respect to the temporal profile
• No atmospheric effect is taken into account by the
simulation
– The relative robustness of the SWIR band
(Sentinel-2) is not highlighted
• Pixel-based simulations
– Some intra-plot variablities which can help to dis-
criminate some classes are not taken into account
Results
0 10 20 30 40 50
Number of dates
0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
KappaIndex
FSAT-2
Venus
Sentinel-2
Classification results as a function of the number of images
With a few images: the spectral resolution makes the difference and the accuracy increases with the number
of bands.
With many images: cloudy images lower the accuracy and the huge amount of data makes the learning
convergence difficult.
Venµs and Sentinel-2: are equivalent starting with about 15 images.
Formosat-2: need at least 20 images to give similar results, but this is difficult to obtain with a 5-day revisit
cycle.
Conclusions
A simulation framework which
• allows to compare different sensors in a realistic setting;
• needs improvements:
– realistic time evolution for crop characteristics
– atmospheric effects
Temporal vs. spectral resolutions
• the trade-off between them has been analyzed
Further work:
• study the use of spectral indexes (vegetation, soil, etc.)
• use feature selection approaches in order to determine which are the most useful spectral bands
References
[1] Germain Forestier, Jordi Inglada, C´edric Wemmert, and Pierre Gancarski. Mining spectral libraries to study sensors’ discrimination ability. In SPIE Europe Remote Sensing, Berlin, Germany, September 2009.
[2] J.B. F´eret, C. Fran¸cois, G.P. Asner, A.A. Gitelson, R.E. Martin, L.P.R. Bidel, S.L. Ustin, G. le Maire, and S. Jacquemoud. PROSPECT-4 and 5: advances in the leaf optical properties model separating photosynthetic
pigments. Remote Sensing of Environment, 112:3030–3043, 2008.
[3] B. Hosgood, S. Jacquemoud, G. Andreoli, J. Verdebout, A. Pedrini, and G. Schmuck. The JRC Leaf Optical Properties Experiment (LOPEX’93). Technical report, EUROPEAN COMMISSION, Directorate - General
XIII, Telecommunications, Information Market and Exploitation of Research, L-2920 Luxembourg; CL-NA-16095-EN-C, 1994.
[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.
[5] S. Jacquemoud, W. Verhoef, F. Baret, C. Bacour, P.J. Zarco-Tejada, G.P. Asner, C. Fran¸cois, and S.L. Ustin. PROSPECT + SAIL models: a review of use for vegetation characterization. Remote Sensing of Environment,
113:S56–S66, 2009.](https://image.slidesharecdn.com/simu-poster-101108044333-phpapp02/85/Multi-temporal-series-simulations-1-320.jpg)
Multi-temporal series simulations
- 1. Assessment of the Land Cover Classification Accuracy of Venµs and Sentinel-2 Image Time Series with respect to Formosat-2 Jordi Inglada, Olivier Hagolle, G´erard Dedieu CNES-CESBIO, UMR 5126, Toulouse, France - jordi.inglada@cesbio.cnes.fr Image Time Series High temporal resolution: monitoring, analysis, and modeling of land surface functioning under the in- fluences of environmental factors as well as human activities Coming sensors: • Venµs: superspectral sensor, dedicated to vegetation monitoring, with a 2-day revisit; developed as a cooperation between France and Israel and should be launched in 2012. • Sentinel-2: will provide systematic global acquisitions of high-resolution multispectral imagery with a high revisit frequency (5 days) for the needs of operational land monitoring and emergency services. Available image time series: Formosat-2 is already providing image time series with short revisit cycle and constant viewing angle. CESBIO has been collecting this kind of data since 2006 over several experimental sites. (a) March 14, 2006 (b) July 7, 2006 (c) November 2, 2006 Objectives Temporal sampling: A short revisit period increases the chance of obtaining cloud-free images. Partially cloudy images can still be used sometimes. What is the influence of temporal sampling in land-cover map production? Spectral resolution: Formosat-2 has a very limited spectral resolution with respect to Venµs and Sentinel-2. What is the influence of spectral resolution in land-cover map production? Classification as metric: We use the classification accuracy as a comparison metric between the studied sensors. Image time series simulation: In order to perform a fair comparison between the sensors, we simulate a time series for each sensor using the same input data. Temporal sampling Time series: A series of 49 Formosat-2 images is used for this study. Cloud cover: Only 10 images out of 49 are cloud-free. The figure on the right shows the cloud percentage for each acquisition. Temporal sampling influence: The final temporal sampling will depend on the acceptance threshold ap- plied on the cloud cover. Mar 2006 Apr 2006 May 2006 Jun 2006 Jul 2006 Aug 2006 Sep 2006 Oct 2006 Nov 2006 0.0 0.2 0.4 0.6 0.8 1.0 Cloud % 40 dates 30 dates Spectral Bands Spectral sensitivity: the num- ber of available bands and their width are different for the 3 sen- sors of interest. Formosat-2: has visible bands (B, G, R, NIR). Venµs: has 10 different bands, which are narrow, with a fine sam- pling of the red edge, but lacking the SWIR. Sentinel-2: is similar to Venµs, but with a coarser sampling of the red edge but adds the SWIR. 500 1000 1500 2000 wavelength 0.0 0.2 0.4 0.6 0.8 1.0 Formosat-2 Relative Spectral Responses 500 1000 1500 2000 wavelength 0.0 0.2 0.4 0.6 0.8 1.0 Venus 500 1000 1500 2000 wavelength 0.0 0.2 0.4 0.6 0.8 1.0 Sentinel-2 Simulations Formosat-2 Input Series LAI(t) Land Cover Map Cab Car N PROSPECT+SAIL Full Spectra Venµs RSR Formosat-2 RSR Sentinel-2 RSR Main steps • The Formosat-2 time series is used to estimate the LAI for each pixel of the classes of interest. • Leaf pigments are obtained using the LOPEX’93 data base [3]. • A land-cover map [4] is used as input to select the pigment values for each pixel. • LAI and leaf pigments are used to feed the PROSPECT+SAIL [2, 5] simulation code. • The simulated full spectra are reduced using the relative spectral responses of each sensor in order to generate the time series [1]. Limits of the approach • Leaf pigments are constant for a given class – Gives more weight to the spectral characteristics with respect to the temporal profile • No atmospheric effect is taken into account by the simulation – The relative robustness of the SWIR band (Sentinel-2) is not highlighted • Pixel-based simulations – Some intra-plot variablities which can help to dis- criminate some classes are not taken into account Results 0 10 20 30 40 50 Number of dates 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 KappaIndex FSAT-2 Venus Sentinel-2 Classification results as a function of the number of images With a few images: the spectral resolution makes the difference and the accuracy increases with the number of bands. With many images: cloudy images lower the accuracy and the huge amount of data makes the learning convergence difficult. Venµs and Sentinel-2: are equivalent starting with about 15 images. Formosat-2: need at least 20 images to give similar results, but this is difficult to obtain with a 5-day revisit cycle. Conclusions A simulation framework which • allows to compare different sensors in a realistic setting; • needs improvements: – realistic time evolution for crop characteristics – atmospheric effects Temporal vs. spectral resolutions • the trade-off between them has been analyzed Further work: • study the use of spectral indexes (vegetation, soil, etc.) • use feature selection approaches in order to determine which are the most useful spectral bands References [1] Germain Forestier, Jordi Inglada, C´edric Wemmert, and Pierre Gancarski. Mining spectral libraries to study sensors’ discrimination ability. In SPIE Europe Remote Sensing, Berlin, Germany, September 2009. [2] J.B. F´eret, C. Fran¸cois, G.P. Asner, A.A. Gitelson, R.E. Martin, L.P.R. Bidel, S.L. Ustin, G. le Maire, and S. Jacquemoud. PROSPECT-4 and 5: advances in the leaf optical properties model separating photosynthetic pigments. Remote Sensing of Environment, 112:3030–3043, 2008. [3] B. Hosgood, S. Jacquemoud, G. Andreoli, J. Verdebout, A. Pedrini, and G. Schmuck. The JRC Leaf Optical Properties Experiment (LOPEX’93). Technical report, EUROPEAN COMMISSION, Directorate - General XIII, Telecommunications, Information Market and Exploitation of Research, L-2920 Luxembourg; CL-NA-16095-EN-C, 1994. [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. [5] S. Jacquemoud, W. Verhoef, F. Baret, C. Bacour, P.J. Zarco-Tejada, G.P. Asner, C. Fran¸cois, and S.L. Ustin. PROSPECT + SAIL models: a review of use for vegetation characterization. Remote Sensing of Environment, 113:S56–S66, 2009.