1. The document describes a method to update the Copernicus High-Resolution Imperviousness layer using data from the Sentinel-2A satellite. It exploits Sentinel-2's repetitiveness to create cloud-free mosaics and uses a random forest classifier and data fusion techniques to combine information from multiple images.
2. Testing on areas in Prague and Rennes showed that uncertainty decreases with data fusion as more images are combined. Multi-temporal and multi-source fusion also helps limit errors and improve geometric accuracy.
3. An evaluation found good accuracy between the updated 2015 classifications and the original 2012 imperviousness layer. The method provides an efficient, operational, and automatic way to process Sentinel
1. Using European Sentinel-2A to update
the Copernicus High-Resolution Layer
Imperviousness
Antoine Lefebvre1
, Christophe Sannier2
(1) CNES, UMR IRISA, Rennes, France
(2) SIRS SAS, Villeneuve d’Ascq, France
2. CONTEXT – Copernicus High Resolution Layers
• 5 Layers produced for 2012:
– Forest
– Grassland
– Wetland
– Permanent water bodies
– Imperviousness
• HRL Imperviousness, main specifications
– Resolution : 20x20m ; 100x100m
– Soil sealing degree : 0 – 100%
• Seamless European-wide coverage over 39
EEA countries
3. CONTEXT – Sentinel-2
• Launched on 23 June 2015
• 13 multi-spectral bands
– High Spatial Resolution :
• up to 10m
• Spectral capabilities :
– 3 bands in the red-edge
– 2 bands in the SWIR
• Repetitiveness :
– up to 5 days (with constellation)
4. • None provides a cloud-free coverage
• Incomplete time series (gaps)
– How to get an homogeneous result over Europe ?
– How to do it in an automatic and operational way ?
t1 t2
t3 t4
KEY IDEA – Exploiting Sezntinel-2
Repetitiveness
5. • Sum of overlays provides a cloud-free coverage
t1 t2
t3 t4 Sum of overlays
KEY IDEA – Exploiting Sezntinel-2
Repetitiveness
6. • Sum of overlays provides a cloud-free coverage
• Combination of single scene classification with data fusion technique
t1 t2
t3 t4 Sum of overlays
KEY IDEA – Exploiting Sezntinel-2
Repetitiveness
7. • 2 Former Urban Areas
– Prague, Czech Republic
• 6,900 km2
– Rennes, France
• 2,500 km2
• Sentinel-2 and Landsat-8
images
METHOD – Study Areas
8. • Input dataset
– All spectral bands
– Pantex
• Based on blue band (Sentinel-2)
• Based on Panchromatic band
(Landsat-8)
• Cloud mask extraction
– Sentinel-2: semi-automatic
– Landsat-8: F-mask
S2 PANTEX
S2 Cloud mask
S2 image
S2 image
METHOD – Data Preparation
9. • Sampling
– Automatic sampling on Copernicus High Resolution Layer (2012)
• 20x20m
• Range from 1 to 100%
• Available on all 39 EEA countries
• Classifier
– Random Forest
• Uncertainty
– Kappa
METHOD – Single Scene Classification
10. METHOD – Data Fusion
• Dempster-Shafer Theory (DST):
– Dealing with imprecision and uncertainty
• Kappa computed on each individual scene
• Associative rule:
– Combination of numerous separate information
• Sentinel-2, Landsat-8, … and some more if available
17. RESULTS - Benefits of multi-temporal data
Limitation of
commision errors
Urban areas
Agricultural areas
18. RESULTS - Benefits of multi-source data
Geometric
accuracy
enhancement
Urban areas
Imprecise edges
19. • Comparison: 2015 classifications and 2012 HRL Imperviousness
• Validation strategy (Built-Up / Non Built-Up)
– 400 randomly selected pixels and stratified selection
• 200 points in the change areas
• 200 points in the unchanged areas
– To prevent problems due to geometrical accuracy
• selected sample is surrounded by the same thematic class in a 3x3 window.
– Interpretation of samples is performed manually
Update Copernicus HRL Imperviousness
24. CONCLUSION
• Simple method to process Sentinel-2 time-series
• Overcome missing data
• Provide cloud-free coverage
• Sentinel-2 information
• Spectral resolution: cloud extraction, classification
• Spatial resolution: texture features
• Temporal resolution: repetitiveness
• Efficiency & Operationality
• Limited interaction with the user
• Automatic sampling (can be applied on the all 39 EEA countries)
• No image selection
• Good accuracy
• More overlays, less uncertainty, better accuracy
• Multi-source ability
• Complementarity with other sensors (Landsat-8 here but also, possible with any others)