SPIE Journal Paper | 14 February 2025
KEYWORDS: Feature extraction, Image classification, Convolution, Polarimetry, Scattering, Matrices, Synthetic aperture radar, RGB color model, Performance modeling, Machine learning
Convolutional neural networks (CNNs) have been widely applied in the field of polarimetric synthetic aperture radar (PolSAR) image classification and have achieved good classification results. However, due to the complexity of polarimetric information, traditional CNN structures are difficult to fully extract and integrate information from multiple polarimetric channels, resulting in insufficient feature diversity, which limits further improvement in classification accuracy. To this problem, this article proposes a multi-scale convolutional fusion network (MCF-Net). MCF-Net employs convolutional kernels of various sizes to perform parallel convolutions, extracting multi-scale features and optimally fusing them, thereby greatly enhancing the diversity and completeness of feature representation. In addition, to enhance the network’s focus on important features and suppress irrelevant information, this article proposes an improved squeeze-and-excitation (SE) module, guided by mean, entropy, and Z-score (MEZ-SE), to weigh the importance of feature maps. By embedding the MEZ-SE module into MCF-Net, MEZ-MCF-Net significantly improves the completeness of feature representation and enhances the ability to distinguish different land cover categories. On the Flevoland, San Francisco, and Oberfaffenhofen PolSAR datasets, the overall accuracy of MEZ-MCF-Net exceeds that of the second-best method by 3.04%, 2.09%, and 1.99%, respectively. There is no doubt that MEZ-MCF-Net is an effective method for PolSAR image classification.