SynPolSAR: Augmenting PolSAR and optical image-mask pairs using diffusion models

Due to the high cost and complexity of annotating polarized SAR images, publicly labeled PolSAR datasets are limited, which presents a considerable challenge for training deep learning models. Meanwhile, optical images, which are relatively easier to obtain and annotate, can provide complementary information for remote sensing applications. This paper presents a framework for generating pairs of PolSAR and optical image masks through fine-tuning networks, with the aim of augmenting PolSAR and optical segmented datasets. To enhance the efficiency of generating statistical features from polarized SAR images, the Kullback-Leibler loss function is incorporated. In addition, a fine-tuning network is employed to simulate the characteristics of various polarized SAR and optical images. Subsequently, data augmentation experiments are conducted on the AIR-PolSAR-Seg dataset, and the augmented dataset is utilized to train segmentation models, including FPN, PSPNet, and U-Net. The experimental results demonstrate that the proposed method significantly improves the segmentation accuracy of these models, with IoU metrics increasing by 1 to 2 percentage points. Moreover, the integration of optical image features further enhances the robustness of the segmentation models.