The research group “SiPEO” develops explorative algorithms to improve information retrieval from remote sensing data, in particular those from current and the next generation of Earth observation missions such as TerraSAR-X, TanDEM-X, TerraSAR-X follow-on, Tandem-L and EnMAP. Currently, the team is working on the following main areas:
The improved retrieval of geo-information from EO data can be used to better support cartographic applications, resource management, civil security, disaster management, planning and decision making.
Sparse signals are commonly expected in remote sensing. By exploiting the sparsity of signals, e.g. by using the Compressive Sensing theory, we can either achieve higher resolution compared to the Nyquist sampling theory or reduce the required number of samples with respect to a given resolution request. The team explores this idea for
- Tomographic SAR inversion
- SAR imaging
- Hyperspectral unmixing
- Hyperspectral resolution enhancement
- Water vapor modeling using GNSS and InSAR data
For instance, for Tomographic SAR inversion, a super-resolution factor of up to 25 can be achieved. Fig. 1 presents a point cloud from Tomographic SAR reconstruction of an area in Berlin using a TerraSAR-X high resolution spotlight image stack. For more examples, please click here.
Noise reduction is a standard step in EO data processing. Often classical local filters are used, e.g. look-processing for SAR and InSAR data, which always reduce the spatial resolution. This calls for non-local approaches that take advantage of the high degree of redundancy of natural images. The team explores the non-local concepts for
- InSAR filtering
- Hyperspectral image denoising
Often remote sensing techniques suffer from unmodeled noise contributions and a large amount of outliers that make the employment of robust estimators important. The team explores robust estimators for:
- Object reconstruction from TomoSAR point clouds
- Deformation monitoring of non-urban areas using interferometric SAR data stacks
- Hyperspectral nonlinear unmixing
- Dimensionality reduction in hyperspectral images
Figure 4 shows an example of the estimated linear deformation rate up to 20cm/year over an active volcanic area (Stromboli Volcano, Italy).
Deep learning has become one of the biggest breakthroughs in the machine learning field. In remote sensing, deep learning can make use of the nature of big remote sensing data, and provides an end-to-end fashion where powerful hierarchical feature representations can be automatically learned from the raw data. Our team explores deep learning for:
- Convolutional networks and recurrent networks for hyperspectral image analysis
- Novel and unconventional network architectures for fusion of SAR and hyperspectral data
- Fusion of social media and remote sensing satellite data
Figure 5 shows an example of the classification map of the hyperspectral image of Houston, US