A Generative Adversarial Network for Downscaling Vegetation Index

Satellite sensors have been widely used for vegetation monitoring, and a number of vegetation index (VI) products have been developed using observations from various satellite sensors. The current operational Visible Infrared Imaging Radiometer Suite (VIIRS) offers observation for the globe with three satellites (i.e., Suomi-NPP, NOAA-20, and NOAA-21) with 16 days of revisiting time for each satellite. However, accurate VI monitoring using VIIRS measurement is still challenging due to the limitations of its spatial resolution. In this research, we propose to down-scale the VIIRS VI products using the deep-learning-based super-resolution Generative Adversarial Network (GAN) to improve the spatial resolution of the VI products from VIIRS. Particularly, the proposed Vegetation GAN (VegGAN) is composed of a generator and a discriminator for adversarial training. To train the designed VegGAN, the low-resolution VI products from VIIRS are used as inputs to the VegGAN, and the high-resolution VI products from Landsat satellites are used as target labels to optimize the network parameters. The experimental results show that the proposed VegGAN can significantly enhance the spatial resolution of the VI product derived from VIIRS observations in the selected study domain. Moreover, this framework offers a synthetic VI product with both high temporal and spatial resolutions.