E Cognition User Summit2009 Pbunting University Wales Forestry

Advances in the use of eCognition for forest research and applications

Dr. Pete Bunting

Contents

• Individual tree analysis– High resolution forest mask

– Delineation Approach

• Fusing with other high resolution data

• Scaling to the landscape

• Future work…

Individual Tree Analysis

Individual Tree Analysis

Forest Mask

• To delineate crowns the non-tree areas need to be removed.– Otherwise, bright areas (e.g., bare soil) would

be delineated as if they were crowns.

• Unfortunately, there is no single solution to the classification of forest/non-forest from high resolution imagery. – But, there are methodologies which can help.

Indexes and Indices for Forest Discrimination

• Normalised Difference Vegetation Index (NDVI).

• Forest Discrimination Index (FDI)– Requires hyper-spectral data over the red

edge.€

NDVI =r 750 - r 680

r 750 + r 680=

NIR - RED

NIR + RED

€

FDI = r 838 - r 714 + r 446( ) = NIR - (RE + BLUE )

Forest Discrimination Methodology

• A common problem is the variability in image brightness across the scene. – North/South facing slops

– Sensor noise

– Contrast with other ground cover types.

• Using two levels where the discrimination threshold(s) is varied with respect to the brightness of the upper level.

Forest Discrimination Methodology

• Image processed in sections (large segments).– Do not need to be squares any segmentation

will do.

Individual Tree Analysis

Hill and Valley Model

• It is helpful to view the data with this model.

• Works with either brightness or height.

• High points the crown tops.

• Valleys crown edges.

Individual Tree Analysis

Splitting the Forest into Crowns

We locate the bright areas of the crown and grow to

the crown edge.

Using a Global VariableSimplify your process with a variable:

WithoutWith

Setup variable

Loop until reach the required value

Increment the variable

Individual Tree Analysis

Merging Small Objects

• During the splitting process small bits of crowns can ‘knocked off’.

• Following splitting a process which merges small objects (a few pixels in size) with their largest neighbor is executed.

Individual Tree Analysis

Classifying Tree Crowns• Objects representing whole crowns were

classified to prevent further splitting.

• Rules to identify crowns are mostly based on their shape properties, including– Elliptical fit,

– Roundness,

– Length/width ratio.

• Additionally, some spectral properties can be useful– For example, standard deviation.

Individual Tree Analysis

Examples of Merging CrownsBright point merging Including small objects

Before AfterBefore After

Relative Border Relative size

Before After Before After

Parco Nazionale d’Abruzzo, Lazio e Molise, Italy

www.definiens.com

Object Variables: Mean-lit Spectra• To associate delineated crowns with a

species type, we extract and use the reflectance spectra from the ‘brightest’ part of the crown.

• These ‘mean-lit’ spectra allow better discrimination between tree species.

• eCognition allows the extraction of values on a per object basis and their assignment as local variables (e.g., tree reflectance spectra).

• These can be used as object features in the subsequent classification of species.

Level 2

Level 1

Object Variables for Tree Species Classification

Object Mean Object Variables

An example of tree species classification in Australia

Eucalyptus populnea

Eucalyptus melanaphloia

Stereo Air-Photo

26

LiDAR Height

CASI reflectance

LiDAR HSCOI

CASI band ratio

CASI Tree Crowns

LiDAR Tree Crowns - Before auto-registration of CASI data

LiDAR Tree Crowns - After auto-registration of CASI data

Species Map of crowns from CASI data

Biomass Map

Stem Locations

Integration of CASI/LIDAR Data

Branch Locations

27

Automated delineation of forest communities

Landsat / AIRSAR Classification

• Using grids (at 25 m resolution) and the dominate and co-dominate species

• Landsat spectral data• Landsat FPC• AIRSAR LHH and LHV (Available on ALOS-PALSAR)

• Produce a rulebase object-oriented classification

28

Comparison to Landsat

29

CASISpeciesCrown Cover

Identifying thresholds

30

eCognition Process

31

Classification of Communities

• Integration of L-band (HH/HV) SAR and optical Landsat data.

• Rules identified using communities identified from the high resolution datasets.

32

33

Future Work…

Long-term change observed from LiDAR, Injune

August 2000 – Optech ALTM1020

April 2009 – Riegl LMS-Q560

0m 30m

HeightJorg Hacker, Ariborne Research Australia, Alex Lee/John Armston

LiDAR v TLS

Thank you for listening

[email protected]