• Region = a spatially contiguous area with common attributes or uniform behaviour (internal ‘homogeneity’)


• Regionalisation = spatial classification, i.e. classification under spatial constraints (contiguity of an attribute in space)


• Why does it work? Principle of spatial auto-correlation (“1st law of Geography”, W Tobler, 1970)

not a trivial task...

• Challenges in generating regions

They need to be spatially coherent
They do not exist a priori


• Defining Homogeneity criterion H

Variance* of pixels within a region
Spectral distance (SD) of pixels
Vectors in n-dimensional feature space

Which image objects are we able to perceive?

• From 252 digital numbers (14 x 18 image matrix) to ‘Abraham Lincoln’ (one individual person)

What do images evoke in our visual context and our brain?

Local pixel neighborhood

• Spatial autocorrelation among pixels
• neighbouring pixels tend to have similar values
• But not everywhere!

Two interrelated methodological pillars (Lang 2008)

(1) segmentation / regionalisation for nested, scaled representations ('object scape')

• Segmentation of different types of continuous data
• Optimizing segmentation (algorithms and parameterization) in terms of usability, theoretical foundation, etc.
• Systematic comparison of segmentation results

(2) rule-based classifiers based on spectral and geometrical properties as well as spatial relationships

• Ontology driven classification, advanced sets of target classes
• Process of OBIA not necessarily sequential (i.e. as described here), rather sequential, iterative (class modelling)
• Smooth transition to methods to further characterise and analyse classification results (e.g. landscape metrics)

Segmentation in several hierarchical scales

scale-specific vs. scale-adapted segmentation

Representations in various nested levels
Each object 'knows' its neighbours and hierarchical level

We all have a concept of “cold”, “warm”, “hot”

• In remote sensing and image analysis a lot of expert knowledge exists
• Such knowledge is implicit or explicit
• The latter can be formalised and encoded in rules, (e.g. NDVI > 0.2 => vegetation)

Knowledge in image interpretation: visual skills are complemented by explicit knowledge, enriched by experience and training

Structural knowledge

• how concepts within a domain are interrelated
• links between image objects and real-world geographical features

Procedural knowledge

• specific computational functions, algorithms
• Encoding of structural knowledge, e.g. by a set of rules … or by a set of representative samples

• Physical laws and principles (e.g. Stefan Boltzmann‘s Law : The hotter a body, the more energy is emitted; Wien‘s Displacement Law: The hotter a body, the shorter the wavelength of the maximum radiation)
• Different land surface types show specific spectral profiles
• Spatial autocorrelation
• ...

Heuristics describing certain geographical phenomena are encoded in a set of rules. Rules can address all kinds of features.

Example soccer field = a special kind of grassland patch with certain size and shape.

• Grassland appears “red” in a false-colour band combination
• How does a football field look like on Sentinel-2?
• Object features (color, shape, size, etc.)
• Object-to-neighbour relationships
• Hierarchical relationships>

We all have a concept of “cold”, “warm”, “hot”

• We discretise a gradual phenomenon by categorizing it
• Boundaries between these concepts are not crisp
• The category under concern gets more stable with increasing ‘distance’ from boundaries
• Fuzzy sets: a way to model / operationalise vague knowledge

Classes are embedded in a hierarchical heritage scheme
(1) Feature-based inheritance

• child classes inherit all descriptive features from their parent classes
• Not confined to spectral values

(2) Semantic inheritance

• Classes can be grouped semantically
• Belong to the same logical parent class
• Semantic inheritance may turn into feature-based when additional explicit information is provided

Example: Monitoring of water quality in swimming lakes over three seasons
Image search criteria: [(0) sensor type], (1) time window , (2) geographical focus (AOI), (3) cloud cover , (4) thematic focus , (5) monitoring period.

• Use standard metadata (location) and cloud mask
• Thematic focus: requires a fundamental (i.e. low level semantic) concept of the existence and representation of (swimming) lakes.
• Temporal aspect: (constancy or change) implies the notion for changes in critical indicators, e.g. by indices in IR reflectance recorded by multi-temporal imagery

How to measure the information in an image?

• Histogram (distribution of values)
• Distribution of colours / colour levels
• Image content (based on classification)

Information measurement requires information units
(bins, classified pixels, objects, etc.)

• Image content can be measured based on classification
• Indices: Diversity H, Dominance, Evenness
• Key elements: categorised units, number and percentages of classes

where P = percentage, i = class 1 … i

Structural knowledge can be organized in knowledge organizing systems (KOS)

• Realised as graphic notations such as semantic nets or frames
• Semantic knowledge representation
• using inheritance concept, e.g. ‘is part of’, ‘is more specific than’, ‘is an instance of’

Semantic net

• Control over existing connections, once established
• Transparency and operability

Corine Land Cover (CLC)

• Standard EU mapping scheme (used by European Environmental Agency EEA) initiated in 1985 (reference year 1990)
• Updates produced in 2000, 2006, 2012, 2018 with the latest update (CLC+) under production - integrating the first time OBIA relevant segmentation techniques
• CLC+ technical document: https://land.copernicus.eu/user-corner/technical-library/clc-core-consultations-for-the-technical-specifications

Land Cover Classification System (FAO-LCCS)

• capable to record any land cover type, independent of specific applications and/or geographical areas
• overcome problems associated with the interpretation of different land cover class definitions (Mucher et al. 2014)
• uses a set of independent diagnostic criteria strictly based on vegetation physiognomy and structure rather than establishing land cover classes based on terminology (Kosmidou et al. 2014)
• comprehensive land cover characterization, regardless of mapping scale, data collection method or geographic location.

Spectral libraries represent spectral reflectance and physical properties of different land cover types

Level of landscape elements

• Spectrally homogenous, correspond to agricultural fields, road segments, blocks of houses etc.

• Contain elements (building blocks)
• Are spectrally heterogeneous, but functionally homogenous
• Geons (Lang et al. 2014)

Ambition: to delineate and classify composite objects in a complex real-world scene

• To represent the scene in at least in two hierarchical levels
• To delineate homogenous elementary units on lower level(s)
• To establish lateral and vertical relationships among objects (object-relationship modelling, ORM)
• To characterise the composition of target classes based on the arrangement of elementary units

A strategy to approach composite classes by applying multi-scale segmentation and spatial modelling

• The target classes represent functionally homogenous units – often spectrally heterogeneous –, as being composed by sub-units
• The delineation of target units may be ambiguous (but inter-subjectively agreeable – thus non-arbitrary)
• The spatial properties of composite classes (or objects) can be described and their structural arrangements of building blocks (elements) can be characterised

“Multiscale” is implicit in the term class modelling

• A complex class (composite object) consists of several functional parts in a specific arrangement, such as:
• ‘mixed arable land’ composed by a mix of agricultural and grassland fields
• ‘residential area’ composed by family houses, gardens, shopping malls, etc.
• Complex classes are hierarchically structured (‘body plan’)

Spectral signatures of elements

• Spectral profiles are created by charting the object mean of representative class samples
• Profiles can be translated in rules
• Fuzzy rules can be applied to translate uncertainty in class descriptions

Defining an object’s body-plan

• A higher-level objects consists of …
• x% of object type A
• y% of object type B
• z% of …

Critical reconditions

• Elementary units need to be classified (‘Lego blocks’)
• Outline of composite object needs to be delineated

Specify set of target classes

• Define class descriptors
• Including non-target classes
• Perform classification

Composite object (yellow outline) consists of several components (sub-objects)

• Create customized feature
• Specify relative area of classified sub-objects