Multi-Resolution Segmentation Explained

Algorithm Fundamentals

Multi-resolution segmentation employs a bottom-up, region-merging approach. Starting with individual pixels as initial objects, the algorithm iteratively merges adjacent objects based on a heterogeneity criterion. The process continues until no pair of objects can be merged without exceeding a user-defined threshold.

The algorithm's defining characteristic is its optimization of a composite heterogeneity measure that incorporates both spectral properties and object shape. This dual consideration produces segments that are not only spectrally coherent but also geometrically reasonable.

The Scale Parameter

The scale parameter is the most influential setting in multi-resolution segmentation. It defines the maximum allowed increase in heterogeneity when merging two objects. Higher scale values permit more heterogeneity, resulting in larger segments; lower values produce finer segmentation.

Scale and Object Size

The relationship between scale parameter and resulting object size is not linear and depends on image characteristics. Homogeneous areas (water bodies, bare soil) tend to produce larger objects at a given scale than heterogeneous areas (urban environments, mixed forests).

Selecting Optimal Scale

Optimal scale selection remains a challenge in GEOBIA practice. Several approaches have been developed:

• ESP (Estimation of Scale Parameter): Analyzes local variance across scales to identify natural breaks
• Reference-based optimization: Compares segmentation results against known object boundaries
• Application-driven selection: Matches segment size to target feature dimensions

Shape Criterion

The shape criterion determines the relative importance of spectral homogeneity versus object shape in the merging decision. It is expressed as a value between 0 and 1, where:

• 0: Pure spectral criterion—objects merge based only on spectral similarity
• 1: Pure shape criterion—objects merge to optimize geometric properties

In practice, values between 0.1 and 0.5 are most common. Higher shape values are useful when target features have characteristic geometric properties (buildings, agricultural fields), while lower values suit natural features with irregular boundaries.

Compactness Parameter

Within the shape criterion, the compactness parameter balances two geometric measures: compactness and smoothness. Like the shape criterion, it ranges from 0 to 1:

• Higher compactness: Favors circular, compact objects
• Lower compactness (higher smoothness): Favors objects with smooth, regular boundaries

For urban mapping with rectangular buildings, higher compactness values (0.5-0.8) often work well. For natural features like water bodies or vegetation patches, lower values may produce more appropriate boundaries.

Layer Weights

When working with multispectral imagery, layer weights determine the relative contribution of each band to the heterogeneity calculation. By default, all bands receive equal weight, but adjusting weights can improve segmentation for specific applications.

For vegetation mapping, emphasizing near-infrared bands can improve discrimination of vegetation types. For urban applications, visible bands may provide more meaningful boundaries between built structures.

Hierarchical Segmentation

One of multi-resolution segmentation's key strengths is its ability to create hierarchical object structures. By performing segmentation at multiple scale levels, analysts can establish parent-child relationships between objects.

This hierarchical approach enables feature extraction at multiple scales and supports contextual classification where an object's class is informed by its neighbors or super-objects.

Practical Hierarchy Design

• Start with a fine segmentation level for detailed features
• Create coarser levels for contextual information
• Ensure scale ratios allow meaningful object aggregation
• Consider application requirements when defining hierarchy depth

Computational Considerations

Multi-resolution segmentation is computationally intensive, particularly for large images or fine scale parameters. Processing time increases non-linearly with image size and inversely with scale parameter.

For large-area projects, consider tiled processing with careful attention to edge effects, or explore cloud-based processing solutions that can parallelize segmentation tasks.

Comparison with Other Methods

While MRS dominates commercial GEOBIA software, alternative segmentation techniques may be preferable in certain contexts:

• Mean-shift: Parameter-free mode detection, available in open-source tools
• Watershed: Effective for imagery with strong edge features
• Superpixels: Faster processing, consistent object sizes

The choice of segmentation algorithm should be guided by application requirements, available software, and computational constraints.