Multitemporal SAR-to-Optical Image Translation Using Pix2Pix With Application to Vegetation Monitoring

Information representation enhancement in synthetic aperture radar images is highly debated in remote sensing literature. In the past, the most credited solutions were based on traditional image processing integrating microwave scattering principles. More recently, neural network-based solutions conquered the scene, allowing for the introduction of representations obtained via domain translation. In this context, generative adversarial networks rule the roost. However, after an initial improvement, the performance of the newly developed solutions reached a plateau. This paper proposes a change of perspective, moving the translation problem from the architecture to input data and training modality. It will be shown that state-of-the-art multitemporal synthetic aperture radar processing provides products that, due to their enhanced texture and colorimetric attributes, appear more similar to their optical correspondent. Moreover, training constrained by spatial and temporal features is beneficial to increase the phenomenological correspondence between the radar reflectivity function and the terrain reflectance. As a result, the obtained translated products show a significant increase of the standard image quality parameters. Finally, the exploitability of multimodal products is demonstrated with an application concerning the estimation of the normalized difference vegetation index, which shows the comparability of the synthetic index with that calculated from native optical data.