Improving crop classification accuracy using Sentinel-1 C-SAR data and GAN-generated optical images

We improve the classification accuracy for cultivated crops using Sentinel-1 C- SAR data by combining the backscatter coefficients of vertical–horizontal and vertical–vertical polarizations and their polarization ratio with optical images (true-color and infrared-color images) generated from synthetic aperture radar (SAR) data using generative adversarial networks (GANs). Seven types of GANs—CycleGAN, feature-guided SAR-to-optical image translation, NiceGAN, pix2pix, pix2pixHD, query-selected attention for contrastive learning in I2I translation (QS-Attn), and continuous and diverse image-to-image translation via signed attribute vectors—are evaluated for generating optical images. For each, we observe significant improvements in accuracy compared with using only the original SAR data (p<0.05 based on Z-test), with overall accuracy increasing from 71.1% to 87.1%. However, we find no significant differences among the GAN types, except for pix2pixHD (p>0.05 based on Z-test). The generated images contribute to improved identification accuracy for grasslands and maize fields; however, the F-1 score for maize fields did not reach 0.5, indicating a substantial gap compared with using actual optical data from the Sentinel-2 multispectral imager (F-1 score of 0.627 for maize fields). Nevertheless, these results demonstrate the potential utility of SAR-to-optical conversion using GANs in the agricultural sector.