automated extraction of landforms from dem data

Automated Extraction of Landforms from DEM data

R. A. MacMillanLandMapper Environmental Solutions Inc.

Outline• Rationale for Automated Landform

Classification– Theoretical, methodological, cost and

efficiency arguments• Classification of Landform Elements

– Conceptual underpinnings – hill slope segments

– Implementation methods – various examples• Classification of Landform Patterns

– Conceptual underpinnings – size, shape, scale, context

– Implementation methods – various examples• Miscellaneous Bits and Pieces

– Thoughts and ideas that may or may not prove useful

• Discussion and Conclusions– What works and what doesn’t?– Future developments & challenges to be

addressed

Rationale for Automated Landform Classification

Scientific and theoretical arguments

Business case – costs and efficiency

Rationale: Scientific and Theoretical

• Why Delineate Landforms?– Landforms define boundary

conditions for processes operative in the fields of:• Geomorphology• Hydrology• Ecology• Pedology• Forestry ... others

– Landforms control or influence the distribution and redistribution of water, energy and matterSource: MacMillan and Shary, (2009)

Rationale: Costs and Efficiency

• Why Automate the Delineation of Landforms?– Speed and cost of production

• Never likely to ever again see investments in large groups of human interpreters to produce global maps

• Governments can’t afford and are unwilling to pay for manual interpretation and delineation of landforms

– Consistency and reproducibility• Manual human interpretation can never

be entirely consistent or reproducible• Automated methods can be constantly

improved and re-run to produce updated products.

Source: MacMillan and Shary, (2009)

Conceptual Hierarchy of Landforms

• Focus here is on two main levels– Landform elements (facets here)– Landform patterns (repeating

landform types)

Source: MacMillan, 2005

Conceptual Hierarchy of Landforms


Rationale: Process-Form Relationships

• Landform Elements related to hill slope processes– Forms are related to processes and also

control them

Source: Skidmore et al. (1991)

Rationale: Process-Form Relationships

Source: Ventura and Irwin. (2000)

Rationale: Recognizing Landform Patterns

• Landform Patterns Establish Context and Scale– Different landform patterns exhibit

differences in• Relief energy available to drive processes

such as runoff, erosion, mass movement, solar illumination, energy flows

• Size and scale of landform features such as slope lengths, slope gradients, surface texture, complexity of slopes, degree of incision of channel networks

• Contextual position in the larger landscape– Runoff producing or runoff receiving area– Sediment accumulation or removal area– Elevated water tables or artesian conditions versus

recharge areas


Classification of Landform Elements

Conceptual underpinningsImplementation examples

Landform Elements: Conceptual Underpinnings

• Many similar ideas on partitioning of hill slopes– Simplest and most basic

conceptualization– 2d not 3d partitioning of a hill slope into

elements


• Many similar ideas on partitioning of hill slopes

Ruhe and Walker (1968)

Reprinted from Ventura and Irvin (2000)


• 2D Concepts in 3D– Ventura & Irwin

(2000)• Ridge top• Shoulder• Backslope• Footslope• Toeslope• Floodplain

– Based solely on slope and curvature• No landform

position

Source: Ventura and Irvin (2000)


• 3D concepts more comprehensive than 2D– Erosion, deposition, and transit are

influenced by both profile and plan (across-slope) curvature

Source: Shary et al., (2005)Concepts: Gauss, 1828


• 3D concepts profile and plan curvature

Source: Shary et al., (2000)


• 3D conceptualization

• 3D classes

Source: Pennock et al., (1987)Source: Pennock et al., (1987)


• Complete system of classification by curvature

Source: Shary et al., (2005)


• 3D conceptualization– Initially based

solely on local surface form• Convex, concave,

planar– In theory surface

form should reflect• Landscape

position• Hillslope

processes– Surface shape not

always sufficient• Need better

context

• 3D classes

Source: Dikau et al., (1989)

PEAK CELL

DIVIDECELL

PIT CELL

05 14 5 8 7 6 24 3 2 1

Downslope length - cell to pit (cells)

63

3 m

063 20100100 1080 88 75 50 38 25 12

Relative slope position as % length upslope

1

2

3

4

5

Downslope drainage direction (DDIR)

20

40

60

80

100

83 72 1 0 1 2 64 5 6 7

Upslope length - cell to peak (cells)

Upslope drainage direction (UDIR)

Relative slope position as % height above pit level

Elevation of each cell above pit elevation (m)

5 m

5 cells

3 cells

PEAK CELL

DIVIDECELL

PIT CELL

05 14 5 8 7 6 24 3 2 1


05 14 5 8 7 6 24 3 2 1


63

3 m

063 20100100 1080 88 75 50 38 25 12


063 20100100 1080 88 75 50 38 25 12


1

2

3

4

5

1

2

3

4

5

Downslope drainage direction (DDIR)Downslope drainage direction (DDIR)

20

40

60

80

100

83 72 1 0 1 2 64 5 6 7



83 72 1 0 1 2 64 5 6 7


83 72 1 0 1 2 64 5 6 7



Relative slope position as % height above pit level

Elevation of each cell above pit elevation (m)

5 m

5 cells

3 cells


• Adding landform position to 3D improves 3D.

Source: MacMillan, 2000, 2005

Measures of Absolute Landform Position

Computed by LandMapR• Flow Length N

to Peak• Vertical Distance Z

to Ridge

FLOW UP TO RIDGE FROM EVERY CELLFLOW UP TO PEAK FROM EVERY CELL

Image Data Copyright the Province of British Columbia, 2003

Source: MacMillan et al., 2007

Measures of Relative Relief (in Z) Computed by

LandMapR• Percent Z Pit to Peak • Percent Z Channel to

Divide

MEASURE OF LOCAL CONTEXTMEASURE OF REGIONAL CONTEXT



Measures of Relative Slope Length (L)

Computed by LandMapR• Percent L Pit to Peak • Percent L Channel to

Divide

MEASURE OF LOCAL CONTEXTMEASURE OF REGIONAL CONTEXT




Measures of Relative Slope Position Computed

by LandMapR• Percent Diffuse

Upslope Area• Percent Z Channel to

Divide

RELATIVE TO MAIN STREAM CHANNELSSENSITIVE TO HOLLOWS & DRAWS


Multiple Resolution Landform Position

Source: Geng et al., 2012

What you see depends upon how closely you look

Different results with different window sizes and grid resolutions

Relative position is always relative to something and varies across an area

MRVBF: Multi-resolution valley bottom flatness

• Valley bottom flatness from:– Flatness (inverse of slope)– Local lowness (ranking in a 6 cell

circular region)• Multi-resolution:

– Compute valley bottom flatness at different resolutions• Smooth and subsample the DEM

Source: Gallant, 2012

MRVBF: Generalise DEMSmooth and subsample

Original: 25 m Generalised: 75 m Generalised 675 mFlatness

Bottomness

Valley Bottom Flatness

Valley Bottom Flatness

Bottomness

Flatness


MRVBF: Multi-ResolutionFlatness and bottomness at multiple

resolutions

25 m

75 m

675 m

Flatness Bottomness Valley Bottom Flatness


Calculating MRVBF

45555

23333

12222

)1(4

)1(2

)1(1

MRVBFWVBFWMRVBF

MRVBFWVBFWMRVBF

VBFWVBFWMRVBF

Weight function Wn gives abrupt transition,

depends on n

2*W2

5*W5

VBF

W


Multiple Resolution Landform Position MRVBF

Example Outputs


Broader Scale 9” DEM

MRVBF for 25 m DEM

Landform Elements: Other Measures of Landform

Position• SAGA-RHSP:

relative hydrologic slope position

• SAGA-ABC: altitude above channel

Source: C. Bulmer, unpublishedCalculation based on: MacMillan, 2005

Source: C. Bulmer, unpublished


Position• SAGA-MRVBF:

valley bottom flatness index

• SAGA-Combined RHSP and MRVBF



Position• SAGA-Combined

RHSP and MRVBF vs Soil Map

• SAGA-Combined RHSP and MRVBF vs Soil Map

Source: C. Bulmer, unpublishedCalculation based on: MacMillan, 2005



Position• TOPHAT – Schmidt

and Hewitt (2004)• Slope Position –

Hatfield (1996)

Source: Hatfield (1996)Source: Schmidt & Hewitt, (2004)


Position - Scilands

Source: Rüdiger Köthe , 2012

Landform Elements: Implementation Example -

Scilands


Landform Elements: Landform Elements:

Implementation Example - Scilands


Landform Elements: Implementation Example:

LandMapR• LandMapR 15 Default Landform

Classes

Source: MacMillan et al, 2000

Landform Elements: Implementation Example:

LandMapR• LandMapR 15 Default Landform

Classes


LandMapR: Different Classes in Different Areas

Normal Mesic

Moist Foot Slope

Warm SW Slope

Shallow Crest

Organic Wetland

Wet Toe Slope

Cold Frosty Wet

Permanent Lake


Example of Application of Fuzzy K-means Unsupervised

Classification

From: Burrough et al., 2001, Landscsape Ecology

Note similarity of unsupervised classes to

conceptual classes

Supervised Classification Using Fuzzy Logic

• Shi et al., 2004– Used multiple cases of

reference sites– Each site was used to

establish fuzzy similarity of unclassified locations to reference sites

– Used Fuzzy-minimum function to compute fuzzy similarity

– Harden class using largest (Fuzzy-maximum) value

– Considered distance to each reference site in computing Fuzzy-similarity

Fuzzy likelihood of being a broad ridge

Source: Shi et al., 2004

Classification of Landform Patterns

Conceptual underpinnings

Rationale: Identify Landscapes of Different Size,

Scale and Context


Conceptualization of Landform Patterns

• Landform Patterns Tend to Repeat – Landform patterns are typically of

larger size and scale and display greater complexity and variation• Hills, mountains, plains, plateaus,

tablelands– Landform patterns usually, but not

always, exhibit full or partial cycles of repetition of forms• Hills and mountains exhibit a full range

of landform positions, slope gradients, curvatures (mostly positive)

• Valleys and plains can exhibit undulations or cyclic variations OR they may be asymmetric.

Considerations Used in Several Systems of

Classifying Landform Patterns• Hammond (Dikau,

1991)– Considerations

• Slope gradient; percentage of gentle slopes (4 classes) in search window

• Local relief within a search window of fixed dimensions (6 classes)

• Profile type; percentage of cells classed as gentle slope in lowland versus upland locations (4 classes)

• SOTER (van Engelen & Wen, 1995)– Considerations

• Dominant slope gradient• Relief intensity• Hypsometry (elevation asl)• Degree of dissection

• Iwahashi & Pike (2006)– Considerations

• Local slope gradient (3x3 window)

• Texture - Local relief intensity assessed as number of pits and peaks within a fixed window of 10 cells

• Curvature; calculated as percentage of convex cells in a 10 cell radius

• eSOTER (Dobos et al., 2005)– Considerations

• Majority slope gradient (of 7 classes in 900 m block, then smoothed)

• Relief intensity (max-min elevation within a radius of 5 cells, 990 m classified into 4 classes & smoothed)

• Hypsometry (elevation asl, 10 classes)

• Dissection (# channel cells in radius)

Rationale for Classifying Landform Patterns

• So, Why Consider These Attributes?– Slope Gradient

• Steepness relates to energy, erosion, deposition, context

– Relief Intensity (or texture or local relief)• Provides an indication of amplitude of

landscape, amount of energy available for erosion, slope lengths, size and scale of hill slopes

– Profile Type (or shape, curvature, hypsometry)• Helps to differentiate uplands (convex)

from lowlands• Helps to establish broader landscape

context

Conceptualization of Landform Patterns


Classification of Landform Patterns

Implementation examples

Landform Patterns: Implementation Example of

the Hammond System• Hammond system; as per Dikau et al., 1991




Source: Zawadzka et al., in prep

9600m square window 9600m circular window 900m circular window

Nor

mal

Rel

ief I

ndex

Mod

ified

Rel

ief I

ndex




18000 m circular window

9200m circular window

900m circular window

Hammond approach tends to produce concentric rings related to how the search window observes the data

Hammond approach is very sensitive to differences in window size and shape or grid resolution



Source: MacMillan, (unpublished)


the Hammond System• Hammond landform underlying 1:650k soil map

Source: Reuter, H.I. (unpublished)


Iwahashi & Pike (2006)• Implemented by Zawadzka et al., (in prep)


8 classes 12 classes 16 classes

Iwahashi & Pike classes need to be labelled and interpreted


Iwahashi & Pike (2006)• Iwahashi landform underlying 1:650k soil map

Source: Reuter, H.I. (unpublished)

steep gentle

Terr

ain

Ser

ies

Fine texture,High convexityFine texture,

Low convexityCoarse texture,High convexityCoarse texture,Low convexity

Terrain Classes

1

4

5

8

9

12

13

2 6 10 14

3 7 11 15

16


eSOTER (Dobos, 2005)• Implemented by Dobos et al., (in 2005)

Source: Dobos et al., 20058 classes

Manual - yellow eSOTER - red

Manual - yellow eSOTER - red


Peak Shed Approach• Implemented by Zawadzka et al., (in prep)


Peak shed entities classified by clustering algorithm.

Resulting entities need to be labelled and interpreted


Peak Shed Approach• Implemented by Zawadzka et al., (in prep)


Peak shed entities labelled according to Hammond


Slope Break Approach• Implemented by Zawadzka et al., (in prep)


Run 2 Run 3


Homogeneous Objects (eCognition)• Implemented by Zawadzka et al., (in

prep)


Landform Patterns: Implementation Example -

Scilands


Source: http://eusoils.jrc.ec.europa.eu/projects/landform/

Landform Patterns: Implementation Example of Homogeneous Objects vs

Meybeck 2001• Implemented by Dragut, (unpublished)

Source: Drãgut & Eisank, 2011

Method: Meybeck et al., 2001

Implemented by: Reuter and Nelson

See: ai-relief.org

http://eusoils.jrc.ec.europa.eu/projects/landform/


Meybeck 2001 vs Homogeneous Objects• Implemented by Dragut,

(unpublished)

Method: Meybeck et

al., 2001

Source: Reuter

and Nelson

Source: Drãgut , unpublished

See: ai-relief.org

Landform Patterns: Example of Multi-scale Nested Homogeneous Objects• Implemented by Dragut,

(unpublished)

Source: Drãgut, unpublished

Source: Reuter & Bock, 2012Scilands GMK ClassificationSee: ai-relief.org

Source: Reuter & Bock, 2012Hammond Classification (after Dikau, 1991)See: ai-relief.org

Source: Reuter & Bock, 2012Iwahashi & Pike Classification (16 classes)See: ai-relief.org

Miscellaneous Bits and Pieces

Some thoughts and ideas that may or may not prove useful

We are Really Looking for Discontinuities!

Source: Minar and Evans. (2008)


Source: MacMillan, unpublished


• The more I think about it the clearer it becomes– We are really looking to locate abrupt

boundaries where the slope, texture, relief and context change• If we are looking for boundaries it makes

sense to try to extract vector objects • It makes less sense to classify grid cells

then agglomerate them, then de-speckle them, then vectorize them

• This argues in favor of approaches like Object Extraction (Dragut) or perhaps Scilands (Kothe)

Source: Minar and Evans. (2008)

There are Special Cases that do not fit in a General

Classification (e.g River Valleys)

I like this!


There are Special Cases that do not fit in a General

Classification (e.g River Valleys)• Many General Purpose

Classifications Need to be Extended to Handle Special Cases– River Valleys are a case in point

• They have forms and patterns that are not cyclical

• They have special features that have special interpretation

– Active flood plain, levee, low terrace, high terrace, inter-terrace scarp, ox-bow lake, abandoned channel, dry islands

– Other Special Cases no doubt exist too• Think of mineral and organic wetlands,

deserts, playas

Multi-Scale and Multi-Resolution Calculations are Important but Problematic

No single fixed window size fits all landscapes – Need to be locally adaptive


Multi-Scale and Multi-Resolution are Important but

Problematic• All algorithms and systems that

compute attributes within a fixed window are flawed– No single window fits all landscapes– User’s frequently adjust window size

subjectively to fit local landscape features – no longer universal!

– Windows that don’t fit the landscape produce artifacts and unrealistic classes or values

– Need to use multiple windows and average (like MRVBF) or make windows self-adjusting

I Like Top-Down, Divisive, Multi-scale Fully Nested

Hierarchical Objects• Multi-scale Objects of Dragut, (unpublished)


I Like Top-Down, Divisive, Multi-scale Fully Nested

Hierarchical Objects• Advantages of multi-scale,

hierarchical, nested vector objects– They nest, or fit, within higher level

objects exactly– There is less arbitrary sliver removal,

filtering, speckle removal, smoothing and manipulation

– They seem to produce fewer artifacts and outright errors

– They produce consistent and comparable results for all similar terrains


The World is Divided into Things that Stick Up and Things that Stick Down

As a first step we should always strive to separate erosional uplands from lowlands



Extracting nested peaks may be a way to separate uplands from lowlands


Might work even better if applied to DEM of inverted Height Above Channel (Z2St)


• In the First Instances Many Landform Pattern Classifications are Binary (upland vs lowlands)– Systems of Iwahashi and Pike, eSOTER,

Hammond Scilands all recognize this in their own way

– Maybe we should be making a point of finding ways to explicitly separate erosional uplands from aggrading lowlands as a first step in any classification

– I have fooled around with the idea of extracting nested pits from an inverted DEM as a way to extract uplands


Are Landform Patterns and Landform Elements Really

Different Things??Maybe the only real difference is one of scale?


Many classifications of Landform Patterns look a lot like Hillslope Elements on a large scale


Are Landform Patterns and Landform Elements Really

Different Things?• The More I look, the more that

landform patterns begin to look like landform elements computed over larger areas and at a coarser scale– Maybe we need to look at approaches

like MRVBF that compute values at multiple scales then average them to produce a final value or class• Similarities to the work of Jo Wood.• We still want to first separate hills from

valleys and uplands from lowlands, then landform elements within these larger scale features.


I Have Personally Found Hierarchical Classification

Useful to Set ContextI first classified areas into 3-4 relief classes

Then I developed and applied different classification rules for each relief class


Discussion and Conclusions

What works and what doesn’t?

How can we tell what works? Challenges to be addressed

Future developments

What Works and What Doesn’t?

• All things being equal apply Ockham’s Razor– If you need to decide between several

competing methods and none is clearly superior to others• Pick the one that is simplest, fastest and

easiest to implement– Fewest input variables– Fewest processing steps– Fewest tuneable parameters– Fewest subjective decisions

• This points towards selection of one of the following

– Iwahashi and Pike, Dragut or Scilands


How Can We Tell What Works?

• How can we evaluate “Truth” for subjective classifications?– Hard to decide objectively which

classification method to use when all classifications appear partly useful and partly incorrect• Need objective criteria and methods of

computing them to assess different classifications and identify the most useful

• Should be based on the ability of the classification to predict ancillary environmental properties or conditions of interest


Challenges to be Addressed

• A diversity of methods and absence of standards– Classes and results need to be

comparable between different areas• This argues for selecting and applying one

method universally and not applying different methods in different regions

• Need to objectively compare methods and then select one to use widely (everywhere?).

• Method almost certainly has to be multi-scale, hierarchical and locally adaptive

• Method needs to be parsimonious and easy to apply


Future Developments

• Global standards– We need global standards to compare

results• Free and open-source data and tools

on-line– I see both data & tools increasingly

available on-line• Incorporation of ancillary (remotely

sensed) data to infer parent material attributes for landforms– Once delineated, objects need to be

attributed for pm• Innovations in multi-scale hierarchical

analysis– Way forward will undoubtedly be multi-

scale


Thank You

Extra Slides Follow

Classify Landforms by Size and Scale

Image Data Copyright the Province of British Columbia, 2003Source: MacMillan, 2005

Quesnel PEM Landform Classification


I Have Personally Found Hierarchical Classification

Useful to Set ContextI first classified areas into 3-4 relief classes


Then I developed and applied different classification rules for each relief class


Quesnel PEM Landform Classification



As a first step we should always strive to separate erosional uplands from lowlands


automated extraction of landforms from dem data

Technology

landform position source

landform types source

scale of landform features

d classes source

d conceptualization

d partitioning

d erosion

d concepts profile