Comparing supervised and unsupervised multiresolution segmentation approaches for extracting buildings from very high...

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Image Processing

Background:

• Multiresolution segmentation (MRS) is crucial for high-resolution imagery but can create inaccurate object geometries, impacting classification.
• Supervised and unsupervised MRS approaches exist, but comparative evaluations are lacking.
• Existing research suggests supervised MRS is preferable, though not definitively proven.

Purpose of the Study:

• To compare supervised and unsupervised approaches for multiresolution segmentation (MRS).
• To evaluate the impact of segmentation geometry on classification accuracy using satellite imagery.
• To assess the automation potential of object-based image analysis.

Main Methods:

• Tested one supervised and two unsupervised MRS methods on QuickBird and WorldView-2 satellite imagery.
• Assessed results using segmentation evaluation metrics and building classification accuracy.
• Analyzed differences in image object geometries and thematic accuracy.

Main Results:

• Both supervised and unsupervised MRS approaches produced similar classification accuracies (82%-86%).
• One unsupervised method closely matched the supervised method in optimal scale parameters and object geometry.
• A second unsupervised method yielded different object geometries but comparable classification accuracies.

Conclusions:

• Object-based image analysis can be automated without compromising classification accuracy.
• The study challenges the notion that classification accuracy is solely dependent on segmentation quality.
• Acceptable classification accuracy can be achieved even with imperfect segmentation, provided under-segmentation is controlled.