High-Resolution Forest Mapping from L-Band Interferometric SAR Time Series using Deep Learning over Northern Spain Journal: ISPRS Journal of Photogrammetry and Remote Sensing
Chiara Telli, Oleg Antropov, Anne Lönnqvist, Marco Lavalle · 2025
Read High-Resolution Forest Mapping from L-Band Interferometric SAR Time Series using Deep Learning over Northern Spain Journal: ISPRS Journal of Photogrammetry and Remote Sensing by Chiara Telli, Oleg Antropov, Anne Lönnqvist, Marco Lavalle (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-04362.txt.
Introduction
Chiara Telli, Oleg Antropov, Anne Lonnqvist, and Marco Lavalle map forest structure at high resolution over Asturias in northern Spain using L-band interferometric radar time series and deep learning. Nine dual-polarization satellite radar images at near-zero spatial baseline form the input stack, with airborne laser scanning supplying reference data.
Several U-Net family models were tested with different polarimetric and interferometric feature mixes. Adding model-based inverted interferometric features or coherence layers beat backscatter intensity alone. Attention and nested-fusion variants beat a vanilla U-Net and traditional machine learning. Forest height errors landed in the low single metres at twenty-metre resolution in the reported ranges, supporting radar time series as a forest monitoring source.
Catalog record
| CREATOR | Chiara Telli, Oleg Antropov, Anne Lönnqvist, Marco Lavalle |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-04362
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
High-Resolution Forest Mapping from L-Band Interferometric SAR Time Series using Deep Learning over Northern Spain Journal: ISPRS Journal of Photogrammetry and Remote Sensing Chiara Telli, Oleg Antropov, Anne Lönnqvist, Marco Lavalle Abstract In this study, we examine the potential of high-resolution forest mapping using L-band interferometric time series datasets and deep learning modeling. Our SAR data are represented by a time series of nine ALOS-2 PALSAR-2 dual-pol SAR images 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 candidate deep learning models from UNet-family with various combinations of input polarimetric and interferometric features. In addition to basic Vanilla UNet, attention reinforced UNet model with squeeze-excitation blocks (SeU-Net) and advanced UNet model with nested structure and skip pathways are used. Studied features include dual pol interferometric observables additionally incorporating model-based derived measures. Results show that adding model-based inverted 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 UNet or traditional machine learning methods. Forest height retrieval accuracies range between 3.1-3.8 m ( R 2 R^{2} = 0.45–0.55) at 20 m resolution when only intensity data are used, and improve to less than 2.8 m when both intensity and interferometric coherence features are included. At 40 m and 60 m resolution, retrieval performance further improves, primarily due to higher SNR in both the intensity and interferometric layers. When using intensity at 60 m resolution, best achieved RMSE is 2.2 m, while when using all suitable input features the achieved error is 1.95 m. We recommend this hybrid approach for L-band SAR retrievals also suitable for NISAR and future ROSE-L missions. † † graphicalabstract: † † highlights: Demonstrated the ability of joint multi-temporal interferometric and backscatter observations from PALSAR-2 to capture structural variations in forest canopies. Developed a hybrid physics–deep learning framework that combines InSAR coherence model–based inversion, polarimetric and auxiliary observables with UNet-family convolutional neural networks for forest height retrieval. Achieved RMSEs between \sim 2.8 m at 20 m resolution and 2.0 m at 60 m strongly outperform traditional machine learning or physics-based inversion approaches. Attention mechanisms and nested feature fusion yield more spatially coherent forest height maps consistent with airborne laser scanning measurements. Moderate spatial aggregation (40–60 m) provided an effective noise suppression enabling scalable monitoring of forest structure from satellite L-band InSAR imagery. 1 Introduction Accurate methods for assessing forest carbon using Earth Observation (EO) sensors are increasingly needed to support climate change monitoring, carbon stock accounting, and to address the operational needs of diverse forest management stakeholders ( Latterini et al. 2025 ) . Forests are essential in maintaining healthy ecosystem interactions and biodiversity on Earth, and they play a critical role in tackling climate change due to their capacity to rapidly restore carbon stocks. This ecological and climatic importance has placed forests at the center of international environmental policies, which call for accurate, timely, and reliable methods for large-scale forest monitoring ( Berger et al. 2025 ; Moersberger et al. 2024 ) . At present, forest carbon stock and its changes are typically estimated through manual inventories of structural forest variables, a process that remains highly labor-intensive and limited in spatial and temporal coverage ( Neumann et al. 2016 ; GFOI 2014 ; Penman et al. 2003 ) . EO sensors offer opportunities to overcome these limitations by providing consistent, wall-to-wall coverage with regular temporal sampling. The integration of satellite EO data with ground measurements has therefore become an established practice in forest mensuration and monitoring, enabling the estimation of forest variables across multiple spatial scales and mapping units ( McRoberts and Tomppo 2007 ) , and improving national forest inventories ( Hunka et al. 2025 ) . Such estimates can be obtained either by inverting physics-based models, which aim to describe the interaction between transmitted signals and forest structure, or by applying data-driven approaches, which directly learn the underlying relationships from large sets of reference samples ( GFOI 2014 ) . Within this context, multiparametric satellite optical and synthetic aperture radar (SAR) imagery represent critical sources of information for large-scale forest mapping. Optical sensors provide valuable insights into vegetation phenology and health, but their applicability is often constrained by cloud cover and atmospheric conditions. In contrast, SAR observations, being independent of daylight and cloud cover, enable consistent monitoring across regions and seasons. The interaction between the radar signal and the observed scatterers is strongly influenced by the scatterers’ geometric and dielectric properties, making SAR particularly sensitive to key vegetation characteristics, such as forest height, growing stock volume, and above-ground biomass ( Ulaby and Long 2014 ; Le Toan et al. 2004 ) To date, extensive studies demonstrated significant correlations between both SAR backscatter coefficient and SAR interferometric measurements, and forest variables ( Schmullius et al. 2015 ) . This sensitivity is influenced by several factors, including mission configuration, radar frequency, polarization, and the structural characteristics of the forest. Due to the enhanced penetration capability to vegetation structure compared with shorter wavelengths, L-band time series have proven particularly effective in estimating forest variables and are considered presently most suitable for SAR based forest variable retrieval ( GFOI 2014 ) . Currently reported retrieval accuracies, however, indicate that further improvements are needed to achieved more accurate and reliable estimates. The integration of deep learning techniques into EO-based forestry, together with the advent of new remote sensing missions offering advanced imaging capabilities, such as NISAR, holds strong potential for methodological advances when improved modeling approaches are combined with novel imaging systems. Research in this field remains active, with continued efforts to refine physically based retrieval models using L-band time series and to develop deep-learning frameworks for data-driven forest estimation parameters. The following sections summarize recent progress in both approaches, highlighting key trends and remaining challenges. 1.1 L-band SAR time series in forest mapping SAR intensity based approaches infer forest structural variables from SAR backscatter time series by leveraging the sensitivity of low-frequency radar to forest structure using physics-based models and, more recently, machine learning techniques. Early research on L-band time series for forest mapping focused primarily on estimating growing stock volume (GSV) in boreal forests using L-band SAR imagery, particularly from the JERS ( Kurvonen et al. 1999 ; Pulliainen et al. 1999 ; Rauste 2005 ; Santoro et al. 2006 ) . Later studies utilized ALOS PALSAR and ALOS-2 PALSAR-2 sensors to map not only boreal forests ( Cartus et al. 2012 ; Antropov et al. 2013 ; Antropov et al. 2017 ; Santoro et al. 2024a ) but also temperate and tropical forests ( Englhart et al. 2012 ; Berninger et al. 2018 ; Rodríguez-Veiga et al. 2019 ) , and even on global scales ( Santoro et al. 2024b ) . While models based approaches based on Water Cloud Model ( Santoro et al. 2024a ) provide a simplified representation of these processes, forward modeling of electromagnetic wave propagation within forested environments remains inherently complex due to multiple scattering interactions and the challenge of isolating contributions from distinct scattering components. In forested areas, L-band backscatter generally arises from three