Machine learning classification based on k-Nearest Neighbors for PolSAR data.

Area of Science:

• Remote Sensing
• Machine Learning
• Image Processing

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) imagery presents unique challenges for automated analysis.
• Standard machine learning classifiers often struggle with the inherent speckle and complex scattering mechanisms in PolSAR data.

Purpose of the Study:

• To evaluate the performance of well-known machine learning techniques for PolSAR image classification.
• To assess the trade-off between classification performance and computational complexity for different methods.
• To identify suitable machine learning approaches for PolSAR data analysis.

Main Methods:

• Comparison of K-Nearest Neighbors (KNN), Support Vector Machine (SVM), randomized decision tree, and a stochastic distance-based classifier (Kullback-Leibler).
• Application and testing of these classifiers on real PolSAR datasets.
• Analysis of classification accuracy and computational efficiency.

Main Results:

• K-Nearest Neighbors (KNN) and Support Vector Machine (SVM) exhibited poor performance, likely due to their inability to handle PolSAR speckle and terrain properties.
• Adapted standard machine learning methods demonstrated a favorable performance-to-complexity ratio.
• The Kullback-Leibler stochastic distance method showed superior results compared to KNN and SVM.

Conclusions:

• Standard machine learning techniques, when appropriately adapted, can achieve excellent performance-to-complexity trade-offs for PolSAR image classification.
• The Kullback-Leibler stochastic distance method is a promising approach for PolSAR image classification.
• Further research into adapting machine learning for PolSAR data is warranted.