Feature Selection Method Based on High-Resolution Remote Sensing Images and the Effect of Sensitive Features on...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• High spatial resolution remote sensing imagery offers abundant features for analysis.
• Effective feature selection is crucial for reducing redundancy and enhancing classification accuracy and efficiency.
• Existing methods may not optimally handle the high dimensionality of remote sensing features.

Purpose of the Study:

• To propose and evaluate a novel integrated feature selection approach (RFGASVM) for building extraction.
• To improve the efficiency and accuracy of building extraction from high-resolution remote sensing data.
• To demonstrate the method's effectiveness across diverse sensors including GF-2, BJ-2, and UAV imagery.

Main Methods:

• Integration of ReliefF for preliminary feature filtering, genetic algorithms for feature subset selection and parameter optimization (C and gamma), and support vector machines (SVM) for classification.
• Development of a fitness function balancing identification accuracy, feature subset size, and feature cost.
• Application and validation using high-resolution imagery from GF-2, BJ-2, and UAV platforms.

Main Results:

• The RFGASVM method achieved significant feature reduction and high building extraction accuracy (Overall Accuracy > 85%, Kappa > 0.80).
• Precision for each image exceeded 85%, demonstrating robust performance across different sensors.
• The proposed method exhibited a two-fold increase in time efficiency compared to using SVM with all features.

Conclusions:

• The RFGASVM approach offers substantial feature reduction and high extraction performance for building detection.
• This method is effective and efficient for feature selection in high-resolution remote sensing applications.
• The integrated RFGASVM technique provides a valuable tool for optimizing remote sensing data analysis.