A novel approach to monitor peanut equivalent water thickness through modular training and transfer learning of an improved PROSAIL model using a Wasserstein generative adversarial network