object-based image analysis (obia) -...

37
Object-based image analysis (OBIA) Remote Sensing (GRS-20306)

Upload: lynga

Post on 03-Feb-2018

227 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Object-based image analysis (OBIA)

Remote Sensing (GRS-20306)

Page 2: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Objects in context

What is this?

Page 3: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Objects in context

Human cognition is able to identify objects within an image and classify them within a certain context Context is given by the presence of other objects and

their spatial/visual characteristics

Page 4: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Objects in context: examples

Looking for agricultural land and forest

A

B C

D

Page 5: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Objects in context: examples

Hierarchical cognition

Forest Water

River Lagoon

Page 6: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Pixel based classification

Pixels by themselves are not able to place objects in context Pixel based classification approaches do not deal well

with the high information content of RS data Much human interaction is required to clean up the

“salt and pepper” effect

Page 7: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

How many trees?

Original Pan Image

Pixel based classification

Object based classification

Page 8: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Object-based image analysis

Human cognition ability is translated into a computational language Image objects are classified

after the pixel info is compressed into a layer of homogeneous regions Regions are separated if they

are significantly different from adjacent features

Page 9: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Object-based image analysis

Beyond spectral info (mean, st. dev.), other object attributes are used to classify:

● Size ● Shape (smooth borders, compactness) ● Texture ● Location in relation to other objects (relative

border, surrounded by) ● Object statistics

Other RS and GIS sources can be combined to add more attributes to the objects (Lidar data, satellite images, vector layers, ............)

Page 10: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Steps

Segmentation: from pixels to vector objects

Definition of the object network (optional)

Classification using spectral and spatial attributes of objects

Export of the results to a vector/raster format for further analysis (i.e. shapefile)

Page 11: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Segmentation algorithms of eCognition

Chessboard segmentation Quadtree-based segmentation Contrast-split segmentation Multi-resolution segmentation Spectral difference segmentation Multi-threshold segmentation Contrast filter segmentation

Page 12: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Chessboard segmentation

Parameters ● Object size 10 (squares of 10x10 pixels)

Page 13: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Quadtree based segmentation

Parameters ● Mode: color ● Scale: 100

Page 14: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Contrast Split Segmentation

Parameters ● Band: Blue ● Min threshold: 120 (dark), Max threshold 253 (bright)

Single band in grey scale Dark and bright areas

Page 15: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Multiresolution segmentation

Parameters ● Scale: 15 ● Shape: 0.1 ● Compactness: 0.5

Page 16: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Multi-resolution concept flow diagram

Scale Parameter Defines the maximum St. Dev. of the homogeneity criteria. The larger the value, the larger the resulting objects

Homogeneity Composed of 4 criteria which define the total relative homogeneity for the resulting objects Digital values

Color = 1 - shape

Criteria work in pairs (equalized to a value of 1)

Shape

Optimizes the resulting objects in regards to smooth borders Smoothness = (1-b* compactness)*shape

Optimizes the resulting objects in regards to the overall compactness Compactness = (b* compactness*shape)

Color

Defines textural homogeneity Shape = Smoothness + Compactness

Compactness

Smoothness

Page 17: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Compactness and smoothness

Compactness = [object border-length] √(#pixels)

Smoothness = [object border-length] [border-length for the given boundary box]

Page 18: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Examples of smoothness / compactness

Smoothness Compactness 12/12=1 12/√9=4

12/12=1 12/ √5=5.45

12/10=1.2 12/ √5=5.45

20/20=1 20/ √9=6.6

Page 19: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Original colour image (example)

Page 20: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Scale 25 - shape 0.1, compactness 0.5

Page 21: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Scale 50 - shape 0.1, compactness 0.5

Page 22: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Scale 50 - shape 0.9, compactness 0.5

Page 23: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Scale 50 - shape 0.9, compactness 0.1

Page 24: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Scale 50 - shape 0.9, compactness 0.9

Page 25: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Each image object uses the homogeneity criterion to determine the best neighbour to merge with

If the first image object's best neighbour (red) does not recognize the first image object (grey) as best neighbour, the algorithm moves on (red arrow) with the second image object finding the best neighbour

This branch-to-branch hopping repeats until mutual best fitting partners are found

Page 26: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

If the homogeneity of the new image object does not exceed the scale parameter, the two partner image objects are merged

Page 27: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Hierarchical structure

Users can create various levels of data by grouping image objects in different ways Classifications can refer to the sub objects/super

objects of an image

Page 28: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Classification

Infinite possibilities: Spectral based classification Sophisticate GIS functions (topological relationships,

reshape algorithms, hierarchical analysis)

Page 29: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Summary

Spatial resolution improvements of Remote Sensing imagery has led into the need of extracting more information from these datasets Object-oriented analysis allows to get information of

features from an image considering contextual information Segmentation is the key process to go from pixels to

objects. It groups homogeneous pixels into meaningful objects using spectral and spatial criteria.

Page 30: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Summary

Object-oriented analysis allows to build up a hierarchical network of the different objects/entities present in an image. This methodology integrates GIS and remote sensing

tools. Sophisticated GIS functions as well as RS traditional functions can be used for image classification

Page 31: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Quickbird panchromatic band (60 cm pixel size) - eCognition software -

Page 32: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Quadtree-based segmentation (scale parameter = 20)

Page 33: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Apply threshold range for defining tree objects (in green)

Page 34: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Merge all unclassified pixels

Page 35: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Remove ‘holes’ in tree crowns and merge

Page 36: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Minimum crown size is 15 pixels, merge and NDVI threshold

What next ???

Page 37: Object-based image analysis (OBIA) - WURscomp5063.wur.nl/courses/grs20306/course/Schedule/Object-based... · Pixel based classification Pixels by themselves are not able to place

Thank you for your attention

Questions?