Deep learning-based vegetation index estimation

Abstract This chapter proposes a novel approach to estimate the Normalized Difference Vegetation Index (NDVI) using an unsupervised generative adversarial network architecture. This unsupervised model is based on an image-to-image translation applying a cycled generative adversarial network (CycleGAN). The model proposes a translation between a grayscale image to a near-infrared (NIR) unpaired image to obtain a synthetic NIR. Once this NIR image is generated, it is used to estimate the NDVI. The translation is obtained by means of a ResNet architecture and a multiple loss function. Experimental results obtained with the proposed scheme show the validity of the implemented model. It has been compared with the state-of-the-art approaches showing better results.