Image Segmentation Techniques for GEOBIA

The Role of Segmentation in GEOBIA

Unlike pixel-based classification, which analyzes individual pixels in isolation, GEOBIA begins by grouping pixels into homogeneous regions called segments or objects. These segments serve as the fundamental units of analysis, enabling the extraction of spatial, spectral, and contextual features that would be impossible to derive from individual pixels.

Effective segmentation must balance two competing objectives: creating segments that are internally homogeneous while also being distinct from neighboring segments. Over-segmentation produces too many small fragments, while under-segmentation merges distinct land cover types into single objects.

Region-Growing Segmentation

Region-growing algorithms start from seed points and iteratively merge adjacent pixels based on similarity criteria. The process continues until no more pixels can be added without violating homogeneity thresholds.

Algorithm Process

1. Select initial seed points (randomly or systematically)
2. Examine neighboring pixels for similarity
3. Merge pixels that meet homogeneity criteria
4. Update region statistics after each merge
5. Repeat until convergence

Advantages and Limitations

Region-growing is conceptually simple and produces connected regions by design. However, results can be sensitive to seed point selection, and the algorithm may struggle with gradual transitions between land cover types.

Multi-Resolution Segmentation

Multi-resolution segmentation (MRS), developed by Baatz and Schäpe, is the most widely used algorithm in commercial GEOBIA software. It employs a bottom-up region-merging approach that optimizes for both spectral homogeneity and object shape.

Key Parameters

• Scale parameter: Controls the maximum allowed heterogeneity of segments
• Shape criterion: Balances spectral homogeneity against compactness
• Compactness: Influences whether segments favor compact or elongated shapes

The scale parameter is particularly critical—higher values produce larger segments, while lower values create finer partitions. Tools like ESP (Estimation of Scale Parameter) can help identify optimal scale values for specific applications.

Watershed Segmentation

Watershed segmentation treats the image as a topographic surface where pixel values represent elevation. The algorithm identifies catchment basins and watershed lines by simulating flooding from local minima.

This technique is particularly effective for imagery with clear boundaries, such as agricultural fields or urban structures. However, it tends toward over-segmentation and often requires preprocessing (gradient computation) and post-processing (region merging) steps.

Mean-Shift Segmentation

Mean-shift is a non-parametric clustering technique that identifies modes in the feature space. For image segmentation, it considers both spatial coordinates and spectral values, grouping pixels that converge to the same mode.

The algorithm requires two bandwidth parameters: one for spatial distance and one for spectral distance. Mean-shift is available in open-source tools like Orfeo ToolBox, making it accessible for open-source GEOBIA workflows.

Superpixel Methods

Superpixel algorithms, originally developed for computer vision, have gained traction in remote sensing applications. Methods like SLIC (Simple Linear Iterative Clustering) generate compact, approximately uniform regions efficiently.

Superpixels are particularly useful for integrating GEOBIA with machine learning, as they provide consistent, computationally manageable units for feature extraction and classification.

Selecting Appropriate Segmentation Parameters

Parameter selection remains one of the most challenging aspects of GEOBIA. No universal settings exist—optimal parameters depend on image resolution, land cover complexity, and analysis objectives.

General Guidelines

• Match segment size to the scale of target features
• Consider using hierarchical segmentation at multiple scales
• Evaluate segmentation quality before classification
• Document parameter choices for reproducibility

Quality Assessment

Evaluating segmentation quality is essential before proceeding to classification. Assessment methods include:

• Visual inspection: Overlay segments on imagery to check boundary alignment
• Supervised evaluation: Compare segments against reference objects
• Unsupervised metrics: Measure intra-segment homogeneity and inter-segment heterogeneity

For detailed guidance on evaluating GEOBIA results, see our article on accuracy assessment methods.