The “Hyperspectral Remote Sensing and Traffic Monitoring” team develops concepts and products for imaging sensors on satellites, aircraft / helicopters and UAV.
Hyperspectral Remote Sensing
Operational processing systems for ground segments are developed and provided for satellite missions. Currently, processing chains are being developed for the German hyperspectral mission EnMAP (Environmental Mapping and Analysis Program) and the US-German hyperspectral mission DESIS (DLR Earth Sensing Imaging Spectrometer), which is installed on the ISS (International Space Station).
The processing of the resulting image information includes conversion of the data to physical measurements that are georeferenced and atmospherically corrected and form the basis for deriving value-added products. For this purpose the team develops new, basic methodologies like data classification using modern “deep learning” techniques, spectral demixing based on “sparse reconstruction” methodologies, innovative image optimisation methods for suppressing noise, removing clouds from images with the help of time series, and improving geometric resolution by fusing hyperspectral and multispectral data sets.
Figure: Classification of a scene of an urban area into 20 categories using a multimodal data set (IEEE GRSS Data Fusion Contest 2018: hyperspectral data with 48 bands in a wavelength range between 380 – 1050 nm and 1 m ground resolution; RGB data with 5 cm ground resolution; elevation model derived from lidar data; multispectral lidar data with 50 cm ground resolution). The classification precision is given as an Overall (OA) value for the individual classes.
Complete solutions – from hardware to software – are developed and operated for aircraft/helicopter/UAV-mounted sensor systems. Using aerial optical cameras with high-rate serial imagery makes it possible to promptly derive specific traffic data for large areas. This data source is an ideal complement to already-available traffic data sources in cases of disasters and major public events. It supports decision making by rescue teams and public agencies and organizations with responsibilities for assuring public safety.
Complex analytic methodologies are required to automatically extract the relevant traffic data. They can be categorized as methods for street detection, vehicle detection, or vehicle tracking. Since vehicles can be detected and tracked in image sequences with high updating rates it is possible to derive velocity as well as vehicle density.
To achieve this, various camera systems have been assembled at the institute (3K and 4k camera systems) and approved for mounting on various flight platforms like the Cessna Grand Caravan 208B, the Dornier Do228, and the Bo105 und EC135 helicopters of the Eurocopter company. The camera systems consist of three commercially available 20 MPix Canon cameras and a navigation unit and they can be used to record serial image data for large areas with frequencies up to 12 Hz (see link to the aerial camera systems). At a flight altitude of 1000m above ground, a surface area up to 2.5 x 1 km2 can be recorded on a single image, in other words in this case a surface of 2.5 x 10 km² in ca. 2 minutes.
At present the team is switching from classic machine learning algorithms to deep learning in order to improve the automated extraction and tracking of objects as well as the generation of high-definition (HD) road maps for automated driving and for monitoring the static and dynamic environment of self-driving vehicles.