PIDLF+: A Novel Adaptive PIDLF for Forestry Height Mapping Based on Single-Baseline PolInSAR Data

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 18; pp. 20054 - 20071
• Main Authors: Li, Dandan, Lu, Hailiang, Paoletti, Mercedes E., Haut, Juan M., Gu, Juntao, Li, Chao, Jing, Weipeng
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: <inline-formula xmlns:ali="http://www.niso.org/schemas/ali/1.0/" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"> <tex-math notation="LaTeX"> L</tex-math> </inline-formula>-band; Accuracy; Adaptation models; Decorrelation; Deep learning; Deep learning (DL); Estimation; Europe; forest height estimation; Forestry; Laser radar; polarimetric interferometric synthetic aperture radar (PolInSAR); Synthetic aperture radar; terrain slope; Training
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2025.3594201

Forest height inversion based on polarimetric interferometric synthetic aperture radar data has demonstrated significant potential for producing large-scale high-resolution forest height maps and has been the subject of extensive research in recent decades. Machine learning and deep learning (DL) techniques frequently utilize the height of the forest derived from light detection and ranging (LiDAR) as labels, exhibiting superior inversion accuracy. Nevertheless, due to the complex terrain of forest areas, it is difficult to ensure the inversion accuracy when training in a specific forested scene and then applying it to other areas. Meanwhile, the lack of training samples imposes additional constraints on the efficacy of supervised DL-based methods. Given this context, an unsupervised framework for forest height inversion that can be used with several existing DL models and does not require previous knowledge from LiDAR data is very appealing. In this article, we delve into this framework by introducing a novel terrain compensation strategy, which allows the framework to be adapted to different forest terrain scenarios. Furthermore, a more efficient network architecture called ResUnet++ is used to extensively extract forest height information from pseudo-labels, resulting in the novel adaptive polarimetric interferometric deep-learning-based framework ( PIDLF+ ). Three study sites with different terrain characteristics, i.e., Lopé, Pongara, and Rabi, were selected for validation. The experiments show that the proposed PIDLF+ works well in different forest terrains and improves accuracy. It achieved a root-mean-square error (RMSE) of 8.69 m and a coefficient of determination (<inline-formula><tex-math notation="LaTeX">R^{2}</tex-math></inline-formula>) of 0.94 in the Lopé site, an RMSE of 11.31 m and an <inline-formula><tex-math notation="LaTeX">R^{2}</tex-math></inline-formula> of 0.90 in the Pongara site, and an RMSE of 8.86 m and an <inline-formula><tex-math notation="LaTeX">R^{2}</tex-math></inline-formula> of 0.92 in the Rabi site.