Generation of Lidar-Predicted Forest Biomass Maps from Radar Backscatter with Conditional Generative Adversarial Networks

This paper studies the generation of LiDAR-predicted aboveground biomass (AGB) maps from synthetic aperture radar (SAR) intensity images by use of conditional generative adversarial networks (cGANs). The purpose is to improve on traditional regression models based on SAR intensity, which are trained with a limited amount of AGB in situ measurements. Although they are costly to collect, data from airborne laser scanning (ALS) sensors are highly correlated with AGB and can replace in situ measurements as the regression target. Thus, the amount of training data increases dramatically, and we can learn an expressive two-stage regression model for SAR backscatter intensity. We propose to model the regression function between SAR intensity and ALS-predicted AGB with a Pix2Pix convolutional neural network for image translation that uses a ResNet-5-based cGAN architecture with the Wasserstein GAN gradient penalty (WGAN-GP) objective function. The synthesized ALS-predicted AGB maps are evaluated qualitatively and quantitatively against real ALS-predicted AGB maps. Our results show that the proposed architecture manages to capture characteristics of the real data, which suggests further use of the ResNet-5 for a SAR intensity regression model of AGB.