Feature Selection in Object-Based Image Analysis

Why Feature Engineering Matters

The shift from pixels to objects opens vast possibilities for feature extraction. A single object can yield dozens or hundreds of potential features. However, more features do not guarantee better classification—irrelevant or redundant features can decrease accuracy and increase computational costs.

Effective feature selection identifies the minimal set of features that maximize class separability while avoiding overfitting. This process is essential when working with GEOBIA because the feature space can quickly become high-dimensional.

Spectral Features

Spectral features describe the radiometric properties of image objects. They remain fundamental to most classification tasks, particularly when distinguishing land cover types with distinct spectral signatures.

Basic Spectral Statistics

• Mean: Average pixel value per band—the most commonly used spectral feature
• Standard deviation: Measures internal spectral variability
• Minimum/Maximum: Extreme values, useful for detecting outliers or mixed pixels
• Brightness: Average across all bands, general reflectance measure

Derived Spectral Indices

Spectral indices calculated at the object level often provide better discrimination than raw band values:

• NDVI: Normalized Difference Vegetation Index for vegetation assessment
• NDWI: Normalized Difference Water Index for water body detection
• NDBI: Normalized Difference Built-up Index for urban mapping
• Custom ratios: Band ratios tailored to specific discrimination tasks

Geometric Features

Geometric or shape features describe the spatial properties of objects—their size, form, and boundary characteristics. These features are particularly valuable when target classes have distinctive morphological properties.

Size Measures

• Area: Number of pixels or metric area
• Perimeter: Boundary length
• Length/Width: Dimensions of bounding rectangle

Shape Indices

• Compactness: Circularity measure (area relative to perimeter squared)
• Elongation: Ratio of length to width
• Rectangularity: How well the object fits its bounding rectangle
• Convexity: Ratio of convex hull area to actual area
• Fractal dimension: Boundary complexity measure

For urban mapping applications, shape features can distinguish buildings (compact, rectangular) from roads (elongated) or vegetation patches (irregular).

Textural Features

Texture describes spatial patterns of pixel values within objects. It captures structural information invisible to spectral features alone—the difference between smooth water and rough forest canopy, for instance.

GLCM-Based Texture

Gray-Level Co-occurrence Matrix (GLCM) features are the most widely used texture measures in remote sensing:

• Homogeneity: Measures local uniformity
• Contrast: Local intensity variation
• Entropy: Randomness or disorder
• Correlation: Linear dependencies between neighboring pixels
• Dissimilarity: Average difference between neighboring pixels

GLCM features depend on direction and distance parameters. Computing multiple directions and averaging (or selecting maximum) can provide rotation-invariant measures.

Contextual Features

Contextual features leverage relationships between objects—their neighbors, super-objects, or position within the scene. This is a distinctive advantage of object-based analysis over pixel-based methods.

Neighbor Relations

• Number of neighbors: Count of adjacent objects
• Border contrast: Spectral difference along shared boundaries
• Relative position: Direction or distance to specific feature types

Hierarchical Features

When using multi-resolution segmentation, features can be derived from parent or child objects:

• Super-object class: Classification of containing object
• Sub-object statistics: Aggregated properties of child objects
• Relative area: Proportion of super-object area

Feature Selection Methods

With potentially hundreds of available features, systematic selection is essential. Common approaches include:

Filter Methods

• Correlation analysis: Remove highly correlated (redundant) features
• Separability measures: Rank features by class discrimination (e.g., Jeffries-Matusita distance)
• Information gain: Measure entropy reduction for each feature

Wrapper Methods

Wrapper methods evaluate feature subsets using classifier performance:

• Forward selection: Add features incrementally based on accuracy improvement
• Backward elimination: Remove features that least impact accuracy
• Genetic algorithms: Optimize feature subsets through evolutionary search

Embedded Methods

Some classifiers provide built-in feature importance measures:

• Random Forest importance: Mean decrease in accuracy or Gini impurity
• Regularization: L1 regularization in linear models

Best Practices

• Start with domain knowledge—include features known to distinguish target classes
• Remove highly correlated features early to reduce dimensionality
• Use cross-validation when evaluating feature subsets
• Consider computational costs for production workflows
• Document selected features for reproducibility

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