Quantifying Error in Training Data for Mapping and Monitoring the Earth System - A Workshop on “Quantifying Error in Training Data for Mapping and Monitoring the Earth System” was held on January 8-9, 2019 at Clark University, with support from Omidyar Network’s Property Rights Initiative, now PlaceFund.
Tata AIG General Insurance Company - Insurer Innovation Award 2024
Crowdsourcing Land Cover and Land Use Data: Experiences from IIASA
1. Crowdsourcing land cover and land
use data: Experiences from IIASA
Workshop on Quantifying Error in Training Data and its Implications for Land Cover Mapping,
Clark University, Worcester, MA, USA
Presenter:
Juan Carlos Laso Bayas
Research Scholar
Center for Earth Observation and Citizen Science (EOCS)
Ecosystems Services and Management (ESM)
International Institute for Applied Systems Analysis (IIASA)
2. • Citizen Science campaigns: Cropland
validation, Field Size, Picture Pile
• Mobile applications
• Directed: Fotoquest
• Opportunistic: AgroTutor, GROW
• Hybrid: LACO-Wiki mobile
• Alerts: FloodCitiSense
• Remote sensing: Night lights - Poverty
mapping, Oil palm mapping
Indonesia
Crowdsourcing data collection
4. Cropland validation: Geo-Wiki interface
Supporting tools,
e.g. NDVI
Background
layers
Submission of
classification
Additional
tools, e.g.
Google Earth
and examples
300 x 300 m
grid to classify
5. Laso Bayas et al., (2017). Nature Scientific Data
Mean cropland percentage per location
Cropland validation: Control and results
• Quality score: Every
20 images randomly
one control image
• No immediate
feedback
• Incentives: €750 -
€25, scientific
publication
6. Field size campaign: Design
130.000 locations worldwide – 4 weeks – 4000 control points
130 volunteers, 390.000 validations. Initial training sites, feedback.
7. Field size campaign: Interface
Supporting
tools, e.g.
Measuring tool
Background
layers
Skipping,
Google Earth,
Examples,
Ask experts
Additional
info, e.g.
Image date
Three grid
colours /
sizes
Size and
dominance
selection
8. Field size campaign: How to tell?
• Very small: Fields smaller than the yellow
cells (less than 80 x 80 m)
• Small: fields of a size between a yellow
cell and four yellow cells (2.56 ha);
• Medium: fields smaller than the red box
(16 ha) but bigger than four yellow cells
• Large: fields smaller than the blue box
(100 ha) but bigger than the red box
• Very large: fields larger than the blue box.
9. Field size campaign: Results
Lesiv M, Laso Bayas JC, See L, et al., (2017). Global Change Biology
Dominant field size map
10. 1. Rapid image assessment
2. Change detection
Designed to be generic and
flexible tool customizable to
different domains that requires
EO data as an input resource.
Picture Pile: Rapid change mapping
Picture Pile
11. Hurricane Matthew post-
disaster damage mapping
volunteers validations
179 249K
Do you see damaged buildings?
Picture Pile: Post disaster mapping
12. Picture Pile: Control and feedback
Control images
used to learn, then
1 check every 10
images. Min 4
evaluations per
image
Experts
produce
control images
13. • 250.000 images
• Duration: 3 weeks with half of
the images classified in 5 days
• Damaged = 4 or more volunteers
classified with damage
• No damage = 4 or more no
damage
• Likely damaged = 3
• Unknown = no majority
• Not usable = 4 or more cloud
cover
Picture Pile: Damage map
14. LACO-Wiki: Web and mobile land cover
validation platforms
1. Upload
2. Sampling
3. Validation
4. Reporting
15. LACO-Wiki: ESA CCI 20 m validation
Grid step: ~12 km
Each location: a 20m x 20m cell
Kenya validation done on site using LACO-Wiki online: 55% accuracy
Systematic samples for:
• Egypt
• Cote D’Ivoire
• Gabon
• Zambia
• Kenya
16. LACO-Wiki: Features being added
NDVI
Mobile component based on FotoQuest
Classifications can be made freely available for use as training data
17. • Definitions: e.g. cropland/pastures, percentage covered (50%)
• Perception of sizes and proportions
• Imagery: Cloud coverage, acquisition time, resolution
• Competition: Quantity vs Quality – power users
• Contributors: Several opinions for one location
• Skipping images, not use of tools – increase of speed vs low
quality?
• Error on control images: Who is an expert?
Cropland validation campaign:
Potential error sources
18. • Clearer definitions – general and border line examples
• Analysis: Skipped images (nr. of), use of tools, quality scoring
• Obtain users profiles (survey) and consistency (performance)
• Models that consider weighing users contributions according to
performance
• Expert points – secret controls?
• Potential use of newly available information: e.g. availability of very
High Resolution Imagery (spatial and temporal availability)
• Training of users: Before and during the campaign
• Quality score: penalizing for speeding and immediate feedback
Cropland validation campaign:
Potential strategies to correct errors
19. Thank you for your atention
Contact:
Center for Earth Observation and Citizen Science
Juan Carlos Laso Bayas
lasobaya@iiasa.ac.at
Steffen Fritz
fritz@iiasa.ac.at
Registrations for the 2019 program are being accepted from 1 Oct 2018 - 11 Jan 2019.
The YSSP Program, for PhD students: 3 summer months at IIASA
20.
21. Center for Earth Observation
and Citizen Science at IIASA: EOCS
• Explore earth observation (EO) and
crowdsourcing (CS) capabilities
• New technologies - social innovation
• CS and EO for SDGs
• Lower costs and extend in-situ data
collection
• Environmental monitoring by citizens
through apps, virtual campaigns and
open platforms
22. Comparison of the crowd with experts
Individual
Experts
Not
usable
Damage No
damage
%
Not usable 926 184 108 76.0
Damage 79 8602 280 96.0
No damage 163 466 6243 90.9
% 79.3 93.0 94.2 92.5
562 expert locations but seen many times by the volunteers
Overall agreement of 92.5%
Missed damage buildings 7% of the time
Saw damage 4% of the time when experts did not
23. Campaign from 2017:
Match up to 78% with
major classes
Feedback provided
to users in less than
24 hours
24. LACO-Wiki: Web and mobile land cover
validation platforms
https://laco-wiki.net
25. LACO-Wiki: Tools
KML files to open in Google Earth
desktop application:
- Historical imagery
- Pictures
Different
layers
NDVI tool linked to GEE