ASF Student Intern From Karlsruhe, Germany, Improves PolSAR Classification

by Donald Atwood, Ph.D.

For the last 4 years, ASF has actively supported an internship program in cooperation with Karlsruhe University in Germany. For graduation, Karlsruhe engineering students must complete a 6-month internship to apply their research skills. For ASF, this program has provided an opportunity to hire highly-skilled students to perform innovative research using ASF-archive datasets. Past projects have led to research on improved detection of ships, topographic compensation of polarimetric data, and the mapping of radio-frequency interference (RFI) effects on Advanced Land Observing Satellite Phased Array L band SAR (ALOS PALSAR). The advantages of the program are numerous: The cultivation of a new generation of SAR researchers, increased community awareness of ASF datasets, the development of new SAR innovations, and the development of cross-institutional ties.

In the summer of 2010, Karlsruhe Senior, Moriz Wurth, traveled to Alaska to work on the development of Polarimetric SAR (PolSAR) applications for the Jet Propulsion Laboratory (JPL) Uninhabited Aerial Vehicle (UAVSAR). The choice of UAVSAR was a natural for an ASF-intern project. ASF has become the archive for this JPL-created dataset. In addition to being unrestricted and available for immediate download (making it somewhat unique for SAR data), UAVSAR-PolSAR data are a magnificent dataset for those seeking to perform PolSAR research. The quad-pol, L-band data have a -47 dB noise floor, are high resolution (<6 m), and have been terrain corrected.

Besides investigating oil-spill signatures for the Deepwater Horizon event and addressing dihedral rotation in urban settings, Moritz explored the way that SAR scattering mechanisms are impacted by local incidence angles. Using polarimetric decompositions, such as those developed by Freeman, VanZyl, and Yamaguchi, the interactions of SAR microwaves with land cover can be characterized as being double bounce, surface bounce, or volume scattering. Moritz’ project was to follow up on results in which UAVSAR-polarimetric scattering mechanisms were shown to be strongly dependent upon local incidence angle. Moritz wanted to find out whether he could optimize PolSAR classification accuracy by utilizing multiple view angles.

This inquiry was facilitated with multiple UAVSAR views of the Kilauea Volcano on the Big Island of Hawai‘i. The area is interesting for its diversity of land cover as well as its significant topography. The latter meant that each of the three contemporaneous view angles of the same region afforded significantly different local incidence angles for each pixel on the ground. The impact of different local incidence angles on scattering mechanism was explored by performing VanZyl decomposition on each of the three images. Figure 4 shows RGB images of the three decompositions, in which red represents double bounce, blue represents surface bounce, and green is volume scattering. While qualitative agreement exists between the three images,
the variability caused by differences of incidence angle is clearly evident. A relevant question becomes whether the three datasets can be merged so as to improve the classification results over any one scene.

For each given ground sample (pixel), the three VanZyl decompositions are analyzed so as to identify the one offering the strongest double-bounce component. The polarimetric properties of that pixel were then included in a new synthesized polarimetric image. Thus, pixel-by-pixel, a new polarimetric image that optimizes the double-bounce scattering mechanism was created.

Using an unsupervised polarimetric classifier in PolSARpro, the synthesized image was classified into 16 segments and using cluster-busting techniques, the 16 segments were merged into five land-cover classes: water, vegetation, lava, open area, and urban. A similar classification procedure was applied to the three original images as well; yielding a total of four land-cover classifications that can be compared for accuracy. Classification accuracy was determined by using 400 randomly sampled points whose land-cover types were determined by inspection of aerial imagery (Figure 5).

With the exceptions of water and forest classes in the User-Accuracy graph, the synthesized image yielded either the first or second highest accuracy for every other class. The improvement in the case of open land and urban is seen to be very significant, with the synthesized accuracy more than 2x greater than that of individual user accuracies. In a field where significant effort has been expended to improve classification accuracy by several percentage points, improvements on the order of 50% are indeed striking. Nevertheless, further research may yield better ways to generate synthetic images, with even higher classification accuracy.

These results will be presented at the 9th European Conference on SAR (EUSAR) 2012, in Nuremberg, Germany. The generation of significant scientific outcomes from the internship program is a win/win for the student interns and ASF. The student interns get real-world research opportunities in remote sensing and ASF gets capable young researchers who bring energy and a European perspective to their 6-month stay in Alaska.