Integration of Multi-Frequency Data into the Periodic Remote Delineation of Wetlands

Synthetic Aperture Radar turned out to be a suitable tool for diverse monitoring purposes because images can be acquired independently of weather or illumination. Mainly in higher latitudes, this is the only reliable data source available throughout the year. Regarding the increasing number of SAR satellite sensors – each sensor with its own characteristic advantages – it is reasonable to integrate all available image acquisitions in one uniform image processing and evaluation environment. Such a mathematical framework was developed recently for the integration of multi-mode SAR data, i.e. multi-frequency (and therewith multi-sensor), multi-temporal, multi-polarized, and multi-scale. This technique bases on the description of SAR image data by the help of the Kennaugh elements. These layers are orthorectified, further calibrated, and finally normalized in order to restrict all measured variables to the uniform number range regardless of the data source. Thus, images of all current SAR sensors – be it TerraSAR-X, RADARSAT-2, or ALOS-PALSAR – are processed and evaluated together in the same geometric and radiometric frame. Multi-frequency data consequently can be used to fill temporal gaps in-between the acquisitions of one special sensor on the one hand, but it can also be used to increase the information content by contributing multi-spectral layers on the other hand. The SAR image pre-processing approach additionally includes a sophisticated automated image enhancement step: the multi-scale and multi-directional multi-looking. Instead of applying a uniform look number to the whole image, the look number and therewith the degree of smoothing is adapted locally to the image content both in scale and direction. This primarily results in a smooth, but detail-preserved intensity image. Furthermore, it provides a very valuable texture information layer indicating the appropriate shape, scale, and orientation of a set of pre-defined filter masks, known as “Schmittlets”. This even allows the evaluation of local image patterns. The final step of the processing chain is a very simple, but extremely effective classification technique based on the correlation of local histograms with sample histograms of predefined classes. Thanks to the closed and uniform number range of the normalized Kennaugh elements, the histograms can easily be compared. The occurrences of the single Schmittlets are aggregated to sets of relative angles (e.g. parallel, diagonal, rectangular) in order to describe local patterns regardless of their absolute orientation in the image. Hence, radiometric information (via the Kennaugh elements) and geometric information (via the Schmittlet indices) is utilized in conjunction. In summary, this contribution presents an up-to-date approach to the automated monitoring of land cover and land cover changes on a regional scale (at least) applied on TerraSAR-X and RADARSAT-2 acquisitions of wetlands in the Peace River – Athabasca Lake – Delta in Alberta, Canada