PolGAN: A deep-learning-based unsupervised forest height estimation based on the synergy of PolInSAR and LiDAR data

Abstract

This paper describes a deep-learning-based unsupervised forest height estimation method based on the synergy of the high-resolution L-band repeat-pass Polarimetric Synthetic Aperture Radar Interferometry (PolInSAR) and low-resolution large-footprint full-waveform Light Detection and Ranging (LiDAR) data. Unlike traditional PolInSAR-based methods, the proposed method reformulates the forest height inversion as a pan-sharpening process between the low-resolution LiDAR height and the high-resolution PolSAR and PolInSAR features. A tailored Generative Adversarial Network (GAN) called PolGAN with one generator and dual (coherence and spatial) discriminators is proposed to this end, where a progressive pan-sharpening strategy underpins the generator to overcome the significant difference between spatial resolutions of LiDAR and SAR-related inputs. Forest height estimates with high spatial resolution and vertical accuracy are generated through a continuous generative and adversarial process. UAVSAR PolInSAR and LVIS LiDAR data collected over tropical and boreal forest sites are used for experiments. Ablation study is conducted over the boreal site evidencing the superiority of the progressive generator with dual discriminators employed in PolGAN (RMSE: 1.21 m) in comparison with the standard generator with dual discriminators (RMSE: 2.43 m) and the progressive generator with a single coherence (RMSE: 2.74 m) or spatial discriminator (RMSE: 5.87 m). Besides that, by reducing the dependency on theoretical models and utilizing the shape, texture, and spatial information embedded in the high-spatial-resolution features, the PolGAN method achieves an RMSE of 2.37 m over the tropical forest site, which is much more accurate than the traditional PolInSAR-based Kapok method (RMSE: 8.02 m).