Modified pix2pixHD for Enhancing Spatial Resolution of Image for Conversion from SAR Images to Optical Images in Application of Landslide Area Detection

A method for the conversion of SAR (Synthetic Aperture Radar) images to optical images can be useful for disaster area detection for the following two reasons: (1) it is easier to detect disaster areas with optical images rather than with SAR images; (2) disasters may occur at night and in rainy and cloudy conditions (SAR images can be acquired in daytime and nighttime as well as all weather conditions). Therefore, it becomes easier to detect disaster areas with optical images converted from SAR images. Using GANs (Generative Adversarial Networks), it is possible to convert SAR images to optical images. In particular, pix2pix and pix2pixHD are used for this purpose. The author proposed spatial resolution-maintained pix2pixHD previously. In this paper, a new method of modifying pix2pixHD with a spatial attention mechanism and an edge enhancement mechanism with a Canny filter in the loss function is proposed, and the proposed method is compared to the pix2pixHD with a spatial attention mechanism and pix2pixHD as well as pix2pix. All of these four methods are compared in terms of the spatial resolution (frequency components) of converted optical images. By experiment, the superiority of the modified pix2pixHD with spatial attention and edge enhancement mechanisms is confirmed for disaster area detection (landslide area detection).