Meta-Learning Classification Network for Few-Shot Polarimetric SAR Images

Polarimetric synthetic aperture radar (PolSAR) images hold significant importance in applications such as crop classification. To achieve better classification performance, deep learning generally requires a large number of samples, but compared to optical remote sensing images, PolSAR image samples are relatively limited. Moreover, due to the substantial differences in scattering characteristics across different PolSAR image bands, it is challenging to train a model on samples from one band and effectively apply it to another. To address these two challenges, we propose a meta-learning classification network (MLCN) consisting of a feature learning module (FLM) and a relation learning module (RLM). The FLM employs 3-D convolution (3DConv) to fully extract spatial and scattering information from PolSAR images, thereby enhancing the model’s capability to learn correlations across different bands. The RLM integrates coordinate attention (CA) with the parametric rectified linear unit (PReLU) activation function to further enhance the model’s feature representation ability. We conduct three types of experiments on PolSAR images in the L-and C-band, including same-band same-image, same-band different-image, and different-band different-image scenarios. The results demonstrate that MLCN achieves overall accuracies (OAs) of over 96% and 90% on the Flevoland-15 and Indian Head datasets, respectively, even with only 0.2% of the sample data.