Generation of NAIP-like earth surface imagery using Generative Adversarial Networks

High-resolution earth surface imagery is significant for surface object identification and edge detection, supporting field-level object-oriented classification and classification refinement. The U.S. National Agriculture Imagery Program (NAIP) provides high-definition images covering the Conterminous United States during the agricultural growing season (June-August). Such an annual drone imagery product possesses spatial resolution as high as 0.6m, covering cloud-free red, green, blue, and near-infrared channels, acting as a premium data source to promote applications of field-level crop identification. However, the NAIP imageries are unavailable in other nations, leading to a challenge to global precise crop mapping.. This study explores a near-real high-resolution generation workflow to produce NAIP-like images with high spatial resolution by leveraging the capabilities of Generative Adversarial Networks (GANs). By employing the Sentinel-2 imagery as source data and NAIP data as destination data, we utilized the Pix2Pix GAN to produce high-resolution imagery with high-definition object context and edge like NAIP. We used pixel-to-pixel differences and correlations between processed and actual NAIP images in three bands to evaluate the model performance. This progress showcases the potential of GANs in translating satellite imagery to NAIP-like imagery. Moreover, it offers a clue to refine field-level crop-type mapping and boundary extraction in other nations without NAIP.