3D-SAR Tomography and Machine Learning for High-Resolution Tree Height Estimation

Aug 5, 2026

Accurate forest biomass estimation depends on reliable tree height measurements, and this work extracts forest height from Synthetic Aperture Radar (SAR) using machine learning. Models are built from the TomoSense dataset, combining SAR inputs from Single Look Complex (SLC) images and tomographic cubes with airborne LiDAR as ground truth. The study evaluates classical methods, a 3D U-Net, and Bayesian-optimized approaches across different SAR frequencies and polarimetries. Best results reach a mean absolute error of 2.82 m for canopies around 30 m, supporting future ESA Biomass Satellite height and biomass modelling.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages14

Reading Time35 min

AccessFree to download

Introduction and Methodology

TomoSense dataset and experimental setup

Model types and input representations

Evaluation metrics

Tabular models

3D-SAR Tomography and Machine Learning for High-Resolution Tree Height Estimation arXiv :2409 .05636v 1 [ cs .CV] 9 Sep 2024 Grace Colverd University of Cambridge [gb669@cam.ac.uk](gb669@cam.ac.uk) Jumpei Takami United Nations Office for Outer Space Affairs [jumpei.takami@un.org](jumpei.takami@un.org) Laura Schade UK Department for Energy Security and Net Zero [laura.schade@energysecurity.gov.uk](laura.schade@energysecurity.gov.uk) Karol Bot INSEC TEC Joseph A. Gallego-Mejia Drexel University [joseph.gallegomejia@drexel.edu](joseph.gallegomejia@drexel.edu) Abstract Accurately estimating forest biomass is crucial for global carbon cycle modelling and climate change mitigation. Tree height, a key factor in biomass calculations, can be measured using Synthetic Aperture Radar (SAR) technology. This study applies machine learning to extract forest height data from two SAR products: Single Look Complex (SLC) images and tomographic cubes, in preparation for the ESA Biomass Satellite mission. We use the TomoSense dataset, containing SAR and LiDAR data from Germany’s Eifel National Park, to develop and evaluate height estimation models. Our approach includes classical methods, deep learning with a 3D U-Net, and Bayesian-optimized techniques. By testing various SAR frequencies and polarimetries, we establish a baseline for future height and biomass modelling. Best-performing models predict forest height to be within 2.82m mean absolute error for canopies around 30m, advancing our ability to measure global carbon stocks and support climate action. 1 Introduction and Methodology We present work modelling forest height from three-dimensional tomographic SAR (TomoSAR) data, developing a robust method for estimating tree heights. The primary objective of this research is to extend our understanding of the benefits of tomography (multiple image acquisition) within SAR research, given the incoming launch of a tomographic SAR satellite (Biomass Satellite mission ESA [1]) . TomoSAR captures three-dimensional representations of forest structures, requiring multiple SLC images captured from different incidence angles and applied geometric processing including performing a Fourier transformation to create a three-dimensional representation (Ferro-Famil [2]) . This results in a volumetric scattering distribution that provides detailed information about the vertical structure of the forest, beyond traditional SLC images. TomoSAR datasets are inherently designed for tomographic applications, significantly reducing preprocessing time and complexity, and enabling more immediate and detailed forest structure analysis. We first present details of our models and the different experiments. We provide further background information and discussion of related works in Appendix A.1 . Preprint. Under review. Figure 1: Geographic splits of training/validation/test data. L-r: Swathe, square, quadrant. We use the TomoSense Dataset (Tebaldini et al. [3]) which provides a set of 30 calibrated twodimensional SLC images from different incidence angles captured by plane, a three-dimensional, processed tomographic cube (tomocube) representing forest scattering and an airborne LiDAR scan for the case study area. The forest canopy sits around 20-35m (see Appendix A.3 for 3D visualisation) . The 2 SAR bands used in this analysis are L and P (wavelengths of 22cm and 69cm respectively) . 2 versions of L-band data are included: monostatic and bistatic acquisition (captured with one plane in a fly-back vs. two planes in tandem) . The resolution of L-monostatic, L-bistatic and P bands:(3m, 0.55m, 1.3m),(3m, 0.55m, 2.3m),(5m, 1m, 3m) . The LiDAR scan (2D) matches the range and azimuth (x,y) dimensions. The processed tomocube’s all have pixel sizes of (321, 665, 36) . We test the ability of the tomographic cube as input to predict forest height, using LIDAR scans as ground truth. We compare a single-pixel approach with tabular machine learni

FAQs

What data sources are used to train and evaluate the tree height models?

The study uses the TomoSense dataset, including tomographic SAR products (SLC images and tomographic cubes) and airborne LiDAR scans as ground truth for the case study area.

Which modelling approaches are compared for forest height estimation?

The work compares classical machine learning, deep learning using a 3D U-Net, and Bayesian-optimized techniques, including tabular methods and CNN-based image approaches.

How is model performance measured against LiDAR canopy heights?

Mean Absolute Error (MAE) is used to quantify average absolute differences, alongside RMSE and the model R-squared (R²) value to assess predictive quality.

Aug 5, 2026

Upload by Valentina