Stereo Radargrammetry Using Deep Learning from Airborne SAR Images

TL;DR

This work addresses the challenge of robust elevation retrieval from airborne SAR imagery despite inherent geometric modulation. It proposes a deep-learning–driven stereo radargrammetry pipeline that uses RoMa for dense image correspondence, trained and fine-tuned on a SAR-specific Mt. Aso dataset with patch-wise processing that avoids ground-projection distortion and preserves image quality. The approach is governed by a joint loss $\mathcal{L} = \mathcal{L}_D + \lambda \mathcal{L}_C$ to optimize disparity and confidence, enabling accurate 3D reconstruction across diverse terrains. Experiments show that RoMa-based fine-tuning yields higher elevation accuracy and wider coverage than POC-based and MegaDepth-finetuned baselines, demonstrating the method's practical viability for SAR–driven stereo radargrammetry. The work also provides a public SAR image dataset to accelerate future SAR-specific deep learning research in this domain.

Abstract

In this paper, we propose a stereo radargrammetry method using deep learning from airborne Synthetic Aperture Radar (SAR) images. Deep learning-based methods are considered to suffer less from geometric image modulation, while there is no public SAR image dataset used to train such methods. We create a SAR image dataset and perform fine-tuning of a deep learning-based image correspondence method. The proposed method suppresses the degradation of image quality by pixel interpolation without ground projection of the SAR image and divides the SAR image into patches for processing, which makes it possible to apply deep learning. Through a set of experiments, we demonstrate that the proposed method exhibits a wider range and more accurate elevation measurements compared to conventional methods. The project web page is available at: https://gsisaoki.github.io/IGARSS2025_sasayama/

Figures (5)

• Figure 1: Overview of creating the SAR image dataset for applying image correspondence methods using deep learning to stereo radargrammetry.
• Figure 2: Overview of the proposed method consisting of (i) patch-wise processing, (ii) fine-tuning, and (iii) 3D measurement.
• Figure 3: SAR images in the test dataset of Table \ref{['tbl:dataset_split']}, whose observation area is 2 km$\times$2 km and the intersection angle is about 43 degrees.
• Figure 4: Elevation maps and error maps obtained from each method for Area 1.
• Figure 5: Percentages of 3D points less than the thresholds for errors.