High-resolution forest height retrieval from L-band interferometric SAR time series using deep learning over Northern Spain

In this study, we investigate high-resolution forest height retrieval using L-band interferometric SAR time series and physics-informed feature-level deep learning framework that combines physical model-based parameter inversion with subsequent deep learning regression modeling. A time series of nine ALOS-2 PALSAR-2 dual-polarization SAR images were acquired at near-zero spatial baseline over a study site in Asturias, Northern Spain. Reference data are collected using airborne laser scanning. We examine the performance of several U-Net-family models with different combinations of input polarimetric and interferometric features. Studied models include Vanilla U-Net, attention reinforced U-Net with squeeze-excitation blocks (SeU-Net) and a nested U-Net with dense skip pathways. Studied features include dual-pol interferometric observables and semi-empirical model-based derived measures. Results show that adding model-based InSAR features or InSAR coherence layers improves retrieval accuracy compared to using backscatter intensity only. Use of attention mechanisms and nested connection fusion provides better predictions than using Vanilla U-Net or traditional machine learning methods. Forest height retrieval accuracies range between 3.1-3.6 m ( = 0.43–0.58) at 20 m resolution when only intensity data are used, and improve to 2.6 m ( = 0.70) when interferometric coherence and geometric features are included. At 40 m and 60 m resolution, retrieval performance further improves, primarily due to higher signal-to-noise ratio in both the intensity and interferometric layers. When using all suitable input features at 40 m resolution, the achieved error is 1.96 m. We recommend this approach for L-band SAR forest height retrievals also suitable for NISAR and future ROSE-L missions.