building a validation database for land cover products...
TRANSCRIPT
Building a validation database for land cover products from high spatial resolution images
The Land Cover project of the ESA Climate Change Initiative
Bontemps Sophie1, Achard Frédéric2, Lamarche Céline1, Mayaux Philippe2, Seifert Frank-Martin3, Arino Olivier3, Defourny Pierre1 1: UCLouvain, Belgium; 2: Joint Research Center, Italy; 3: ESA, ESRIN, Italy
Land cover as an operational product
What about products quality?
Validation
• “The process of assessing, by independent means, the quality of
the data products derived from the system outputs” (CEOS-WGCV
definition)
• CEOS 4-level validation approach according to the temporal and spatial coverage of available reference data
CEOS validation
stage Characteristics
Stage 1 Product accuracy is assessed from a small (typically < 30) set of locations and time periods by comparison
with in-situ or other suitable reference data.
Stage 2
Product accuracy is estimated over a significant set of locations and time periods by comparison with
reference in situ or other suitable reference data. Spatial and temporal consistency of the product and with
similar products has been evaluated over globally representative locations and time periods. Results are
published in the peer-reviewed literature.
Stage 3
Uncertainties in the product and its associated structure are well quantified from comparison with
reference in situ or other suitable reference data. Uncertainties are characterized in a statistically
robust way over multiple locations and time periods representing global conditions. Spatial and temporal
consistency of the product and with similar products has been evaluated over globally representative
locations and periods. Results are published in the peer-reviewed literature.
Stage 4 Validation results for stage 3 are systematically updated when new product versions are released and as
the time-series expands.
The CCI Land Cover project
• ESA Climate Change Initiative to provide a comprehensive and timely response to the need for long-term satellite-based products in the climate domain
• Land Cover project to generate global LC maps consistent over time, to match the needs of key users of the climate modelling community.
• Phase 1 (2011-2013) – Phase 2 (2014-2016)
http://www.esa-landcover-cci.org/
The CCI Land Cover project
Envisat MERIS Full & Reduced Resolution 2002-2012 300-1000m
SPOT-Vegetation 1 & 2 1998-2012 1000m
Envisat ASAR 2005-2012 Mainly Wide Swath Mode (150m)
• 3 consistent global LC maps
CCI-LC validation strategy
CCI-LC quantitative accuracy
assessment
1. Stratified random sampling of sites
2. Collection of high resolution imagery
3. Pre-processing of high resolution imagery
4. Development of a validation dedicated tool
5. Generation of the reference dataset (image interpretation)
6. Comparison between reference dataset and CCI land cover products
Stratified random sampling
1. Stratified random sampling of sites – Sampling design : systematic + 2-stage stratification + random
Systematic sampling of the FAO FRA RSS / JRC TREES + diminution of sample intensity with higher latitudes (for equal-probability of samples selection) + stratification for lower intensity in very
homogenous desert areas
1. Stratified random sampling of high-resolution sites – Sampling design : systematic + 2-stage stratification + random
Final random selection => 2600 sampling units distributed globally in a systematic stratified random manner
Stratified random sampling 2
0 k
m
8 km
8 k
m
Collection of high resolution imagery
1. Stratified random sampling of sites
2. Collection of high resolution imagery To allow the validation of the 3 epochs (2010, 2005, 2000)
Relying on Google Earth facilities
Collecting Landsat imagery for years 2000, 2005 and 2010 (GLS)
2010
2005
2000
Imagery pre-processing
1. Stratified random sampling of sites
2. Collection of high resolution imagery
3. Pre-processing of high resolution imagery Per-object approach: selection of segmentation parameters (object size, compactness, etc.) suitable for a wide variety of landscapes to avoid under-and over-segmentation
Under segmentation
Imagery segmentation
Over segmentation
Imagery segmentation
Correct segmentation
Imagery segmentation
Validation interface
1. Stratified random sampling of sites
2. Collection of high resolution imagery
3. Pre-processing of high resolution imagery
4. Development of a validation dedicated tool
On-line interpretation interface derived
from the GlobCover tool
Validation interface
Validation interface
1. Layer box 2. Zooms
3. Tools (navigation, NDVI, select object, paint class)
4. Legend 5. Comment
Step 1: use all information for assign class in the 2010 period Step 2: indicate the level of certainty for the entire interpretation
Validation interface
Step 3: evaluate the change between the 3 epochs, by « back-dating » the 2010 interpretation Step 4: indicate the level of certainty for the entire interpretation
Validation interface
Insular SE Asia: Forest Class
Thailand: mixed land use
Image interpretation
1. Stratified random sampling of sites
2. Collection of high resolution imagery
3. Pre-processing of high resolution imagery
4. Development of a validation dedicated tool
5. Generation of the reference dataset
High-resolution image interpretation
by experts in a standardized manner
on the developed tool
Already a great success … with
great possibilities for future
• All experts involved in validation quite happy with the tool: friendly and easy to use, clear interpretation framework, on-line
• Resulting database: – Plenty information that need to be fully exploited
• 2010 interpretation + level of certainty
• 2010 - 2005 - 2000 change information + level of certainty
• Each sample as a set of objects
– Homogenity of the sample
– Interpretation of all LC classes in the unit in a single LC class, in a mosaic, in LC fractions, etc.
– Used in the CCI-LC project but the use of VHR images coupled with a per-object should allow using it for other maps with lower-higher spatial resolution
Sentinel-2 opportunities
• Use of S2 imagery instead of Google Earth – Less variability in the images to segment (sensor, dates, pre-processing
performance, etc.) => better segments delineation
– Deriving NDVI time profiles from high resolution dataset (S2 being compatible with other HR missions – Landsat 8), to use temporal information inside the sample unit
Broadleaf or needleleaf? Rainfed or irrigated?
Perspectives
• Validation tool based on GLC2000, GlobCover, FAO-FRA RSS / JRC-TREES previous projects lessons learned
• Precursor for operational LC validation activities in collaboration with international scientific community – As a concept:
• Per-object approach inside a given area ensure the scalability of the resulting database => could be used to validate Sentinel-2 LC maps
• Relies on the availability of a large amount of good quality reference imagery at high spatial resolution => boosted by the unique amount of Sentinel-2 data.
– As a tool: could be of use in many other types of activities (validation, in-situ data collection, crowd-sourcing, etc.), all of them being in link with high spatial resolution imagery
• Sentinel-2 to make a significant difference as data source and as application
Thank you for your attention