Multi-Pixel Simultaneous Classification of PolSAR Image Using Convolutional Neural Networks.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in optical image processing.
• CNNs are increasingly applied to polarimetric synthetic aperture radar (PolSAR) image classification.
• Current CNN methods often classify PolSAR pixels independently, ignoring spatial context and interrelations.

Purpose of the Study:

• To develop a novel CNN-based method for PolSAR image classification that considers pixel interrelations.
• To improve the efficiency and accuracy of PolSAR image classification.
• To introduce a fixed-feature-size CNN (FFS-CNN) capable of simultaneous patch classification.

Main Methods:

• A fixed-feature-size Convolutional Neural Network (FFS-CNN) was developed.
• FFS-CNN classifies all pixels within a patch simultaneously, unlike traditional pixel-wise methods.
• The FFS-CNN learns interrelations between different land covers within a patch during training.

Main Results:

• The FFS-CNN demonstrated faster classification speeds compared to common CNNs.
• The method effectively utilizes land cover interrelations to enhance classification accuracy.
• Experimental results on Gaofen-3 and other PolSAR datasets show comparable performance to state-of-the-art methods.

Conclusions:

• FFS-CNN offers an efficient and effective approach for PolSAR image classification.
• Simultaneous patch classification and learning interrelations significantly improve results.
• The proposed method shows strong potential for advanced PolSAR data analysis.