Object Features For Classification in Remote Sensing
Introduction
The classification of remote sensing imagery has traditionally relied on pixel-based analysis, which considers spectral values . However, object-based classification provides a more advanced approach by incorporating shape, texture, and contextual relationships.
This study explores the segmentation and classification process applied to a Quickbird image of Salzburg and pseudo airquality image, using eCognition software to extract meaningful features for classification.
Segmentation
Segmentation is a fundamental step in object-based image analysis. Two segmentation approaches were used:
Chessboard Segmentation: This method creates uniform grid-based objects, which is useful for initial feature assessment. This method is not suitable for distinguishing meaningful land cover features. A comparison of segmented objects, such as a ship and water for example, revealed:
The NDVI for water is negative due to its low near-infrared (NIR) reflectance and high visible reflectance and the ship exhibits higher reflectance across all bands, likely due to its surface color.
However, because all segments are of equal size, the segmentation does not align well with natural object boundaries, making it unsuitable for this classification task.
- Multiresolution Segmentation: This technique groups pixels into meaningful objects based on spectral and spatial properties. It is more suited for real-world classification as it captures object boundaries more accurately.
Feature Selection and Object Analysis
Object features play a crucial role in classification. The key categories of features examined in this study include:
Spectral Features: Mean and maximum pixel values per band.
Geometrical Features: Shape index and object area.
Textural Features: Haralick-based texture measures.
Hierarchical Features: Relationships between super-objects and sub-objects.
Class-Related Features: Proximity to specific classified objects
Classification
Step 1: NDVI Calculation
To classify vegetation, the Normalized Difference Vegetation Index (NDVI) was computed using:
NDVI = (NIR-Red)/(NIR+Red)
Objects with NDVI values above 0.25 were classified as vegetation, while those below -0.15 were classified as water bodies.
Step 2: Assigning Object Classes
Using the “Assign Class” algorithm:
Vegetation was classified using the NDVI threshold.
Water bodies were extracted based on low NDVI values.
A “Boat” class was assigned using the “Relative Border to Water” feature, identifying objects completely surrounded by water.
Step 3: Integrating Air Quality Data
The air quality raster layer was used to refine the vegetation class:
Vegetation was split into high-air-quality (in the image light blue color) and low-air-quality (in the image below green color) areas based on pixel values from the air quality dataset.
This created subclasses, allowing for a more detailed environmental analysis.
Hierarchical classification
To refine object classification, hierarchical segmentation was applied by creating a lower-level segmentation from an existing level using multiresolution segmentation with a smaller scale parameter. This approach enhances detail by generating smaller objects, capturing fine structures such as individual trees, buildings, or narrow water bodies. The multi-level segmentation enables a hierarchical classification process, allowing for the refinement of class boundaries and improving overall classification accuracy.
Conclusion
The object-based approach provided improved classification accuracy compared to pixel-based methods.
The multiresolution segmentation effectively delineated meaningful objects, making the classification more representative of real-world features.
Feature-based classification allowed for the differentiation of vegetation, water, and built-up areas, enhancing thematic mapping.
The use of multiple scale parameters enabled hierarchical classification, ensuring different levels of detail could be incorporated into the analysis.
The integration of air quality data demonstrated how multi-source information can refine classifications, making results more applicable to environmental studies.