E Cognition User Summit2009 Pregesbauer Geo Info Li Dar Basic Landcover

Michael PregesbauerFolie 1

Basic Landcover Classification by LiDAR and Optical Data

eCognition User Summit 2009Munich

November 2009

Michael PregesbauerFolie 2

Contents

• Overview – data resources• Object Generation• Class Definition and Classification• Classification results• remarks on data accuracy and data precision

Michael PregesbauerFolie 3

Why basic landcover classification for the public sector?

• legislative duties (e.g. land use planning, building regulation, sound

wave propagation models)

• planning purposes (e.g. infrastructure networks, soil sealing)

• change detection (development of settlements)

Michael PregesbauerFolie 4

Which Data for Classification

• Digital Orthofotos, 4 Channels (Red, Green, Blue, near Infrared), 12.5cm Ground Sampling Distance

• Digital Terrain Model, 1m Grid width

• Digital Surface Model, 1m Grid width

DOP

DSM

DTM

Michael PregesbauerFolie 5

Aim

Classification of

• Sealed areas– Buildings– Road Network

• Vegetation (Forest Areas)

• comprehensive data set for the whole state [20.000 km²]• stable classified classes (reliability ~ 95%)

Michael PregesbauerFolie 6

Data processing

• Dataset Tiling (2000x2000 Pixel)

• Creation of Initial Objects

• Stitching to Object Primitives

• Classification

Michael PregesbauerFolie 7

Object Primitives byImage Object Fusion

• Object fusion based on a condition:– spectral difference– height difference– border condition

Seed

CandidateCandidate

Michael PregesbauerFolie 8

Creation of Object Primitives

• The class filter allows restricting the potential candidates by their

classification.

• Fitting function threshold allows to select a feature and a condition you want to optimize the fusion.

• Depending on the fitting mode, one or more candidates will be merged with the seed image object.

Michael PregesbauerFolie 9

Classification

Normalized Differenced Vegetation Index

Mean Height Buidlings

Normalized Differenced Vegetation Index (NDVI)

Mean Height Vegetation

Property Class

Mean High as a definite threshold

NDVI as fuzzy function

Class A Class B

Michael PregesbauerFolie 10

Improvement of Buildings

Michael PregesbauerFolie 11

Appraisal of results

Buildings

• ~ 88 % accurately classified

• ~ 9 % classified as elevated objects

• ~ 3 % not classified

Vegetation

• ~ 91 % accurately classified

• ~ 5 % false classified

• ~ 4% not classifiedElevated Objects

Buildings

Vegetation

Michael PregesbauerFolie 12

Misclassification - Example

• Building Objects within Vegetation Areas

• Elevated Objects next to Buildings

Elevated Objects

Buildings

Vegetation

Michael PregesbauerFolie 13

Misclassification - Example

• Shadows

• dark roofs

• edges of buildings

lead to misclassifications at borders of buildings

Michael PregesbauerFolie 14

Misclassification - Example

Object Properties• Mean NDVI > 0• Means nDS 4.7 m• Shadow Index > 0.07

⇒ Object Class: Building

enhanced approach: usage of masks

Michael PregesbauerFolie 15

Additional Mask Layer

RGBi Image Layer

NDVI Layer

non Vegetation Mask

Elevated Objects Mask

Layer arithmetic‘s([Mean nir]-[Mean red])/([Mean nir]+[Mean red])

Layer arithmetic’s1: (NDVI ≥ 0); 0: (NDVI < 0)

Layer arithmetic's 1: (NDVI ≥ 0) and (nDS > 1); 0: (NDVI < 0)

Michael PregesbauerFolie 16

Classification improvement

Michael PregesbauerFolie 17

Results

Buildings

• ~ 94.3 % accurately classified

• ~ 5.2 % classified as elevated objects

• ~ 0.5 % not classified

Vegetation

• ~ 96.1% accurately classified

• ~ 1.1 % false classified

• ~ 2.8 % not classified

Buildings

Vegetation

[Abbildung]

Michael PregesbauerFolie 18

Results

Michael PregesbauerFolie 19

Building Generalization

Michael PregesbauerFolie 20

Results – Building Generalization

• Building Objects are generalized by a bounding box

• export as shape with attribute mean height

Michael PregesbauerFolie 21

Performance Tests

Hardware• 1 Server HP DL380G4 , 4 CPUs• 1,8 Tb working storage

Software• 3 Definiens Server v7.0• 1 Definiens Developer v7.0

Processing Time• 1 processing unit (1000x1250m) = ~ 11min• 1 processing unit (1000x1250m) building generalization (v8.0 beta)

= ~ 25min

Michael PregesbauerFolie 22

Lessons learned

Classification quality depends essentially on the data quality:

• accuracy of the geo-referencation

• spectral quality of the optical data

• filter quality of the LiDAR data

• time lag between data acquisition (LiDAR and optical data)

Classification quality can be enhanced by the

• usage of true orthofotos

• usage of DTM, First- and Last Pulse data

Michael PregesbauerFolie 23

Contact:Michael PregesbauerState Government of Lower AustriaLandhausplatz 1, A-3109 St.PoeltenTel.: ++43(0)2742/9005/13404Mail.: [email protected]

Thanks toChristian Weise [Definiens AG]Gregor Willhauck [Definiens AG]