Land Cover and the REVIGIS project.

1
Land Cover and the REVIGIS project.
Activating Metadata Case Study 1

• Richard Wadsworth1, Lex Comber2
• Peter Fisher3
• 1. CEH Monks Wood, UK, 2. Leicester University,
  3. University of London

2
What was REVIGIS?

• Revision of Uncertain Geographic Information
• EU Project (under IST-FET 2000 to 2004)
• Develop methods to process information that was
  uncertain in what and uncertain in where
• Case studies included
• Cadastre strong what, strong where
• Sand Dunes strong what, weak where
• Land Cover weak what, weak where

3
Underlying Problem

• Historically maps supported the description of
  the phenomenon.
• Now the documentation supports the map.
• The user has less information on the origins and
  meanings of the data.
• The user has less motivation for understanding
  the origins and meanings of the data.
• A lecture is the process by which information
  goes from the notes of the lecturer to the notes
  of the student without passing through the
  consciousness of either.

4
How did we get there?

• Technology, Science and Public Policy are always
  changing.
• Repeated natural resource inventories often use
  different methods and record different categories
  from earlier studies
• It is difficult to distinguish changes in the
  phenomenon from changes in the method, classes,
  technology, objectives etc.

5
Problem Domain

• Two land cover maps (LCMGB, LCM2000)
• Produced by the same organisation
• Funded from same sources (mostly)
• Produced from similar data (Landsat ground
  survey)
• But issued with the caveat that they should not
  be used to estimate change.
• Why the caveat?

6
Reasons for the Caveat

• Objectives
• scientific v. policy support
• Conceptualisations
• Target classes v. Broad Habitats
• Representation
• pixel v. parcel
• Technology
• GIScience and GISystems developments
• Metadata
• aspatial v. object level meta-data

7
Commissioning context LCM1990
8
Commissioning context LCM2000
9
Classes, labels meanings

• We all have prototypes in our heads
• We match a class label with that prototype
• There may be a mis-match
• Many examples in land cover (and every other type
  of geographic information), eg
• What is unimproved grassland?
• What is a forest?
• What is a bog?

10
Uncertainty in what
Improved Grassland Improved by reseeding and / or
fertiliser Includes Fertile pastures with Juncus
effusus May be confused with semi-natural swards
where abandoned or little-managed
Acid Grassland Generally not reseeded or
fertiliser-treated Pastures with Juncus effusus
are included Management may obscure distinctions
from Improved grassland.
11
What is a forest?
12
What is a Bog?

• In LCMGB (1990) Bog was defined as
• permanent waterlogging,
• permanent or temporary standing water
• Myrica gale and Eriophorum spp.
• water-logging, perhaps with surface water
• In LCM (2000) Bog was defined as
• areas with peat gt0.5 m deep
• Consequences in one 100 x 100km square
• 1990 12 pixels of bog (lt1 ha)
• 2000 120728 pixels of bog (75 km2)

13
Land Cover Maps of GB
LCMGB 1990
LCM 2000
14
Land Cover Maps of GB
LCMGB 1990
LCM 2000
But the fragments are obviously of the same area
15
Confusion matrix is confused
16
The traditional approach to inconsistency

• Aggregated to a reduced number of common super
  classes (thematic simplification).
• An extreme case (Wulder et al 2004) land cover
  maps of Canada are aggregated to just 2 classes,
  forest and not forest!
• (If necessary data sets are aggregated spatially)

17
Problem with traditional approach

• Reduce what can be said about the phenomenon.
• Effectiveness of the process is rarely tested.
• Sensitivity is rarely discussed.
• Can increase variability.
• (aggregation is subjective)

18
Semantic-Statistical Approach

• Extend many-to-one aggregation to a
  many-to-many look-up-table.
• Extend belongs/doesnt belong (1/0) to
  expected, uncertain, unexpected (1/0/-1).
• Identifies inconsistencies between
  representations (and one form of inconsistency is
  change).

19
Semantic relationships (Fragment of a LUT)
Swansea Rows IGBP Columns Evergreen Needle leaf Forest Evergreen Broadleaf Forest Deciduous Needle leaf Forest Savannas Grassland
Crop / Forest / Grass Complex 0 0 0 0 0
Light Coniferous (Larix / larch) -1 -1 1 -1 -1
Dark Coniferous 1 0 -1 -1 -1
Soft Deciduous -1 -1 0 -1 -1
Grassland -1 -1 -1 1 1
20
A hypothetical segment
For class A expected score 18, uncertain
score 7 (4 class B pixels 3 class C pixels)
unexpected score 1 (the single pixel of class
D).
21
Hypothetical segment second classifications
For class A expected score 19 (class X),
uncertain score 5 (class Z) unexpected score
2 (class Y).
22
Combining Scores
Scores are treated as if they were probabilities
then using Dempster-Shafer Belief (Bel1.Bel2
Unc1.Bel2 Unc2.Bel1) / ß where ß (1
Bel1.Dis2 Bel2.Dis1) Bel1 Bel2 the
beliefs (expected), Unc1 Unc2
uncertainties (uncertain), Dis1 Dis2
disbeliefs (unexpected). For class A. Bel1
18/26 0.692, Unc1 7/26 0.269, Dis1 1/26
0.038 Bel2 19/26 0.731, Unc2 2/26
0.077, Dis2 5/26 0.192 Therefore ß 1
0.6920.192 0.7310.038 0.839 Belief
(0.6920.731 0.6930.077 0.7310.269) / 0.839
0.901 The belief has increased therefore we
consider that the segment is consistent for A
23
An actual segment
24
An actual segment
25
Object level metadata!
26
Intersection with LCMGB
27
Expected, Unexpected and Uncertain

• With the LCMGB and LCM2000 maps we need two LUT
  to calculate the expected, uncertain and
  unexpected scores twice
• once from perpixlist (object level metadata)
• secondly by intersecting the LCMGB pixels.

28
Semantic relations expressed in LUT
Relating Broad Habitats to the spectral variants
in perpixlist
29
Semantic relations expressed in LUT
Expert opinion of relationship between Broad
Habitats (LCM2000) and Target Classes (LCMGB)
30
Good things about the Statistical-Semantic
approach

• Doesnt destroy or throw away any data
• Few (2) of consistent segments appear to have
  changed.
• Most (80) inconsistencies are due to error not
  change.

31
Problems with the Statistical-Semantic approach

• Semantic relations (LUT) between LCMGB and
  LCM2000 based on Expert opinion.
• Might not have an Expert
• Experts need to make lots of decisions.
• The allocation to 1/0/-1 is not very subtle.
• Decisions are opaque (why do A and X have
  an expected relationship?)
• Knowledge based corrections what to trust?

32
Unanswered questions

• How to record the commissioning context?
  (especially the compromise between what is
  desirable and what is feasible)
• How to describe the data conceptualisations (ie
  more than the class labels)?
• Who is responsible for specifying the
  relationships between conceptualisations?
• How is that information to be communicated to the
  user?
• Different Experts have different views how to
  capture these?

33
Acknowledgements and Thanks
We wish to thank all our collaborators on REVIGIS
as well as our colleagues in our host Institutes.
REVIGIS European Commission IST-1999-14189,
Coordinator Robert Jeansoulin, Universite de
Provence, France.