Multiscale Supervised Classification of Point Clouds with Urban and Forest Applications.

Area of Science:

• Geospatial analysis
• Computer vision
• Machine learning

Background:

• Object detection from 3D point clouds is crucial for geospatial applications.
• Existing methods often rely on sensor-specific data, limiting broader applicability.
• Geometric information alone offers a system-independent approach.

Purpose of the Study:

• To evaluate multiscale supervised classification for object detection in laser scanning and photogrammetric point clouds.
• To assess the utility of geometric information independent of data acquisition systems.
• To compare the performance of different classification algorithms.

Main Methods:

• Utilized multiscale supervised classification algorithms.
• Employed only geometric point cloud data (coordinates).
• Applied Principal Component Analysis (PCA) at six scales with up to five features.
• Tested four multiclass classifiers: Linear Discriminant Analysis, Logistic Regression, Support Vector Machines, and Random Forest.

Main Results:

• Achieved high accuracy: over 80% for urban datasets and over 93% for forest datasets.
• Results are comparable to state-of-the-art methods.
• Random Forest demonstrated superior performance considering accuracy, computation time, and variable importance.

Conclusions:

• Multiscale supervised classification using geometric data is a robust method for object detection in 3D point clouds.
• The approach is independent of specific data acquisition systems.
• Random Forest is recommended for its balanced performance across key metrics.