STEREO RADARGRAMMETRY USING DEEP LEARNING FROM AIRBORNE SAR IMAGES

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.

Proposed Method Overview

We leverage the stereo radargrammetry framework of our previous work^*, which can measure the elevation from two SAR images by stereo vision. The internal and external parameters are obtained from the metadata in SAR image acquisition and the SAR projection model takes the earth ellipsoid into account. RoMa is expected to be effective for finding correspondence between SAR images because of its high accuracy for large deformation and poor-texture regions between camera images. As mentioned above, not only replacing the image correspondence part of our conventional method with RoMa, the proposed method divides the SAR image into patches for image correspondence, and fine-tunes RoMa using the SAR images. Fig. 1 illustrates the overview of the proposed method, which consists of (i) patch-wise processing, (ii) fine-tuning, and (iii) 3D measurement.
^* Karl Insfran et al., "ACCURATE 3D MEASUREMENT FROM TWO SAR IMAGES WITHOUT PRIOR KNOWLEDGE OF SCENE", 2021.

Figure 1: Overview of the proposed method consisting of (i) patch-wise processing, (ii) fine-tuning, and (iii) 3D measurement.

SAR Dataset Creation Process

We create the dataset to apply deep learning-based methods to stereo radargrammetry. The dataset consists of SAR image pairs and their corresponding elevation map pairs. SAR images are divided into patches of trainable size as shown in Fig. 2, since the size of the SAR image is approximately 8, 000×8, 000 pixels. First, the elevation map corresponding to the SAR image pair is obtained from DSM using the metadata and the projection model as shown in Fig. 2 (a). Next, patches are extracted from the SAR image pair as shown in Fig. 2 (b). A patch is defined on the reference (Ref.) image and extracted from the corresponding position on the source (Src.) image using the latitude and longitude in the metadata. Note that the patches on Ref. image should be selected so that 1/3 of the image overlaps between adjacent patches so that disparity can be calculated from a patch pair. A patch is also extracted from the elevation map based on the location of the patch on Ref. image.

Figure 2: Overview of creating the SAR image dataset for applying image correspondence methods using deep learning to stereo radargrammetry.

Qualitative Results

Figure 3: Elevation maps and error maps obtained from each method.

BibTeX


        @article{Sasayama-IGARSS-2025,
          author  = "Sasayama, Tatsuya and Ito, Shintaro and Ito, Koichi and Aoki, Takafumi",
          title   = "STEREO RADARGRAMMETRY USING DEEP LEARNING FROM AIRBORNE SAR IMAGES",
          journal = "Proceedings of IEEE International Geoscience and Remote Sensing Symposium",
          year    = "2025",
          month   = Aug,
        }