Complex-Valued Fully Convolutional Network for PolSAR Image Classification with Noisy Labels

• Published in: IEEE International Geoscience and Remote Sensing Symposium proceedings pp. 5962 - 5965
• Main Authors: Wang, Ningwei, Bi, Haixia, Wang, Xiaotian, Chen, Zhao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.07.2023
• Subjects: Complex-Valued Convolutional Neural Network (CV-CNN); Convolutional neural networks; correcting noisy labels; Data models; Deep learning; Geoscience and remote sensing; Maintenance engineering; Noise measurement; Polarimetric SAR (PolSAR) image classification; rebalancing loss function; Training
• ISSN: 2153-7003
• DOI: 10.1109/IGARSS52108.2023.10282961

The process of annotating PolSAR data is highly intricate and demands a proficient understanding of the subject matter. Due to the complexity inherent in this task, it may result in imprecise or erroneous labels being incorporated. The presence of noisy labels inevitably impacts the performance of models in this context, making PolSAR classification a challenging task. This paper proposes a module for correcting noisy labels, which utilizes a CV-CNN with two convolutional layers as its backbone and presents two key contributions: (1) an effective label correction method that leverages the inherent similarities between training samples to repair imprecise or erroneous labels, and (2) a rebalancing loss function that adjusts the weights of different classes to enhance the accuracy of smaller classes. Experimental evaluations on the Flevoland dataset demonstrate the efficacy of our proposed approach.