Trusted Polarimetric Feature Fusion for Polsar Image Classification

Fangzhou Han,Lamei Zhang,Hongwei Dong

DOI: https://doi.org/10.1109/bigsardata59007.2023.10294730

2023-01-01

Abstract:Benefiting from the rapid advancements in Deep Learning (DL), data-driven features exhibit remarkable separability in PolSAR image classification. However, data-driven features are challenging to interpret and susceptible to resolution and noise. On the contrary, polarimetric features based on the target decomposition not only have high robustness, but also has physical interpretability. Consequently, researchers have begun to explore ways to integrate both types of features to achieve enhanced performance, which typically concat the data-driven features and polarimetric features directly. However, polarimetric features are not stable in separability, which is because the inherent defects of the target decomposition algorithm. To mitigate the adverse effects of low-quality features, this paper proposes a trusted polarimetric feature fusion (TPFF). The new paradigm calculates the confidence of features and employs a novel feature fusion method. The proposed method has been tested on the EMISAR Foulum dataset, with experimental results demonstrating improved classification performance and robustness.