Quad Polarization PALSAR Radar, Radar Texture, and Optical Data for Land-Cover Classification

Faculty and students within the Department of Geography at George Mason University in Fairfax, Virginia, have conducted research with radar and the integration of radar and optical data for improved land-cover mapping for many years. Much of that work has been possible by acquiring RADARSAT-1, and more recently ALOS Phased Array L-band SAR (PALSAR), data from ASF.

The basic research included two related activities: 1) examining methods to improve digital classification of land cover by extraction of radar-derived data layers, and 2) comparing radar and optical data results with the option of integrating them to assess multisensory classification accuracy. These were important as most of the spaceborne systems were single wavelength and single polarization, thus greatly limiting digital-classification strategies.

The primary method to extract radar-derived values was by using various texture measures. Variance, mean Euclidean distance, kurtosis, and skewness were often evaluated at different window sizes ranging from 5x5 to 13x13. Very frequently, texture measures provide improved classification accuracies sometimes for individual classes and often for overall accuracies over the original radar. Merging texture and original radar measurements generally results in better accuracies. Other spatial methods that have been examined include the comparison of despeckle techniques with different window sizes, despeckling at a small window size, then extracting texture at a larger window size, and post-classification filtering.

The classification results from radar and radar-derived values were compared to those from registered optical [Landsat Thematic Mapper (TM), SPOT, ASTER] data using the same classification and accuracy assessment methods. Different methods to integrate the radar, radar-derived values, and optical data were examined to evaluate multisensor classification accuracies. These methods included relative weighting of bands from the respective sensors and use of principle components analysis

Figure 2a: PALSAR image (65 km by 35 km) for Bangladesh, acquired 14 March 2003. 2b: PALSAR image (65 km by 35 km) for Nairobi, Kenya, acquired 12 May 2007. Polarization for both images is HH, HV and VH; RGB.

Figure 2b: PALSAR image (65 km by 35 km) for Nairobi, Kenya, acquired 12 May 2007. Polarization for both images is HH, HV and VH; RGB.

Recently, the Department has used quad-pol PALSAR data provided by ASF for studies in Bangladesh (Figure 2a), Sudan, Kenya (Figure 2b), California, and Washington, D.C. The PALSAR data were obtained at a 12.5-m spatial resolution and ASF remote-sensing software, MapReady, was used for pre-processing the imagery.

The recent, more-widely available, quad-polarization radar, such as PALSAR, increases the usefulness for feature delineation and allows for comparison of the various polarizations. For example, for four land covers in the Kenya scene, the VV polarization provided an overall accuracy of 85 percent versus 62 to 71 percent for the other three bands.

The use of radar texture on PALSAR images was often a useful procedure for increasing the capability to distinguish among the different land covers. Generally, the Variance measure was best among those examined and the appropriate window size will vary by location. In Bangladesh, using a Variance 7x7 window, the producer’s capability to classify agriculture increased from 71-percent accuracy using four original bands to 90 percent using a combination of radar and texture. In Kenya, the producer’s accuracy for urban increased from 64 percent in the original radar to 87 percent using texture.

The question of whether classification accuracies can be increased by fusing multisensor data has been an important issue in the remote-sensing community. For these sites, the fusing of PALSAR with optical imagery, yielded better classification accuracies as opposed to taking either dataset individually. In Bangladesh, the overall accuracy for radar was 91 percent and increased to 98 percent with a merge of Landsat TM and radar texture. In Kenya, the original radar had an overall accuracy increase from 77 percent with radar to 86 percent with Landsat TM and radar texture.

The wider availability of multiple-polarization spaceborne radar has made available to the geospatial industry a great wealth of data. There is good evidence in these studies that radar-derived measures and the merger of radar and optical data can be useful strategies in classifying land cover in diverse regions around the world.