TomoSAR II: 3D Semantische Szeneninterpretation auf Grundlage tomographischer SAR-Daten

• Antragsteller: Olivier D' Hondt, Ph.D.
• Fachliche Zuordnung: Geodäsie, Photogrammetrie, Fernerkundung, Geoinformatik, Kartographie ; Bild- und Sprachverarbeitung, Computergraphik und Visualisierung, Human Computer Interaction, Ubiquitous und Wearable Computing
• Förderung: Förderung von 2015 bis 2022
• Projektkennung: Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 273687959

• Erstellungsjahr: 2023

Zusammenfassung der Projektergebnisse

The project’s primary objective was to advance the semantic interpretation of Tomographic Synthetic Aperture Radar (TomoSAR) data. It aimed to automate TomoSAR data processing efficiently and improve object detection, addressing open challenges. Initially, the project sought to extend object detection methods using a database of simulated objects, considering two approaches: ”perceptual grouping” and localizing 3D object pose. Simultaneously, the project aimed to develop algorithms for pre-processing TomoSAR data, addressing issues such as phase calibration, model order selection, DEM refinement, and phase ambiguity. Moreover, the project intended to propose enhancements to adaptive filtering of the covariance matrix and devise a method for fast TomoSAR focusing employing machine learning. To make the link with our previous project (TomoSAR), an in-depth evaluation of TomoSAR’s potential for land cover classification was conducted and published in the ”Remote Sensing” journal. This new in-depth evaluation introduced new 3D features and compared the approach to a baseline method. Results from this evaluation confirmed our initial results: TomoSAR data, when classified with a suitable classifier and hand crafted features, can outperform PolSAR classification, particularly benefiting classes with 3D structures like Forest, City, and Shrubland. Then an application of the method for Urban/Forest classification on more recent data was conducted in collaboration with DLR. The method effectivenes in separating challenging forest classes was demonstrated. The intial project also underwent a significant transformation by adopting a novel end-to-end strategy that leveraged convolutional neural networks (CNNs) for 3D semantic interpretation. This shift aimed to address challenges associated with the project’s multi-stage approach. However, the latter stages of the project were impeded by the COVID-19 pandemic, which, coupled with personal constraints, resulted in a slowdown of progress. Consequently, while the results displayed promise, they were considered preliminary due to these constraints and their impact on the project’s timeline. In a notable development, the project introduced an innovative approach for 3D reconstruction and semantic segmentation of TomoSAR data using deep learning. This approach eliminated the need for manual labeling by generating training data through simulations. It utilized CNNs to capture spatial information and achieve 3D reconstruction. Experimental results showed that the proposed method closely matched true values when tested on simulated data. Additionally, it yielded promising results when applied to real TomoSAR data, offering an effective alternative to traditional methods for urban structure extraction and 3D interpretation. In summary, the project made substantial progress in automating and enhancing TomoSAR data processing, offering improved object detection and advanced classification capabilities. It also demonstrated the potential of deep learning in 3-D interpretation, despite facing challenges related to the COVID-19 pandemic.

Projektbezogene Publikationen (Auswahl)

• Exploiting GAN-Based SAR to Optical Image Transcoding for Improved Classification via Deep Learning. EUSAR 2018; 12th European Conference on Synthetic Aperture Radar, 2018, pp. 1-6
  A. Ley; O. D’Hondt; S. Valade; R. Haensch & O. Hellwich
  
• Exploiting SAR Tomography for Supervised Land-Cover Classification. Remote Sensing, 10(11), 1742.
  D’Hondt, Olivier; Hänsch, Ronny; Wagener, Nicolas & Hellwich, Olaf
  
• Towards Global Volcano Monitoring Using Multisensor Sentinel Missions and Artificial Intelligence: The MOUNTS Monitoring System. Remote Sensing, 11(13), 1528.
  Valade, Sébastien; Ley, Andreas; Massimetti, Francesco; D’Hondt, Olivier; Laiolo, Marco; Coppola, Diego; Loibl, David; Hellwich, Olaf & Walter, Thomas R.
  
• Reference-Free Despeckling of Synthetic-Aperture Radar Images Using a Deep Convolutional Network. IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, 3908-3911.
  Davis, T.; Jain, V.; Ley, A.; D'Hondt, O.; Valade, S. & Hellwich, O.
  
• Deep Learning Based Joint Reconstruction and Extraction of Urban Structures from Tomographic SAR Data. 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, 3085-3088.
  D'Hondt, Olivier & Hellwich, Olaf
  
• Semantic Segmentation of High- Resolution Airborne SAR Images using Tomographic Information. EUSAR 2021, Virtual event
  O. D’Hondt; R. Haensch; V. Cazcarra-Bes & O. Hellwich