Deep Learning based Enhancement of TomoSAR Stacks

Abstract

Synthetic Aperture Radar Tomography (TomoSAR) exploits the use of several co-registered SAR images, acquired with different tracks (later called TomoSAR stack), to synthesize an aperture in elevation. Spectral analysis methods recover the vertical backscattered Power Spectrum Pattern (PSP) allowing the 3D imaging of an illuminated scene [1]. The resulting tomograms present ambiguities located at a distance inversely proportional to the separation between baselines [2]. Therefore, the larger the TomoSAR stack, the better is the ambiguity rejection. In practical scenarios, the size of the TomoSAR stack is constrained to a revisit time, since temporal decorrelation problems may occur [2]. An additional limitation to the number of passes in the stack is the feasibility of individual missions to perform them. In order to ease this issue, we use a deep neural network to synthesize Single Look Complex (SLC) SAR images from an “artificial” baseline, i.e. a sensor path with a line-of-sight not acquired in a specific TomoSAR stack. Deep Learning methods use a cascade of nonlinear processing units to obtain features of a dataset as humans do, i.e., learn by example. This knowledge is later used to map an input into a desired output. A U-net encoder-decoder network architecture [3] is employed with five encoder blocks and five decoder blocks. The contractive path (encoders) half the number of features at each step, while the expansive path (decoders) double them. In our case, the network takes six of seven baselines as an input and estimates the remaining baseline. We envision a scenario where multiple TomoSAR stacks are required to cover a region of interest. The goal is to acquire a large stack only for a part of the scene. This data can subsequently be leveraged to train the network to synthesize a subset of the stack. For the remaining parts of the scene, only a smaller stack has to be acquired as it can be extended by SLCs synthesized by the network. As a proof of concept, we limit ourselves to a single tomographic stack that is divided into train and test regions by a split in azimuth. Experiments are performed using the 2015 UAVSAR mission from NASA/JPL over Munich. The training and testing subsets have a size of 14000x8000 pixels. Further experiments are done using FSAR data from DLR over Froschham in Germany. Preliminary results show the correct estimation of both the phase and the amplitude in the targeted SLC. The amplitude is evaluated with a scatter plot between the original and the estimated data. The phase is assessed by performing interferometry between adjacent baselines; in one case using both original SLCs and in the other using one original SLC and the artificial SLC. [1] A. Moreira, P. Prats-Iraola, M. Younis, G. Krieger, I. Hajnsek and K. P. Papathanassiou, "A tutorial on synthetic aperture radar," in IEEE Geoscience and Remote Sensing Magazine, vol. 1, no. 1, pp. 6-43, March 2013, doi: 10.1109/MGRS.2013.2248301. [2] G. D. Martín-del-Campo-Becerra, S. A. Serafín-García, A. Reigber and S. Ortega-Cisneros, “Parameter selection criteria for Tomo-SAR focusing,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 1580–1602, Jan. 2021. [3] O. Ronneberger, P. Fischer and T. Brox, "U-Net: Convolutional networks for biomedical image segmentation", MICCAI, vol. 9351, pp. 234-241, November 2015.