Land cover classification by SAR

A 7.5 km×12.4 km test site located in northern Michigan was used to evaluate the land-cover classification accuracy attainable by automatic classifiers using various combinations of SAR polarizations and frequency bands. The scene was imaged by the JPL AirSAR at L- and C-bands. Using all available polarimetric data at both bands, a level-1 classifier was developed capable of correctly classifying each pixel in the scene as belonging to one of the following four categories: urban, tall vegetation (trees), short vegetation (grasses and crops), and bare surfaces (water, roads and bare soil) with an accuracy of 95% or better. The classification results provided by this L/C polarimetric combination are compared in this study with results based on L-band hh-polarization alone (representing the JERS-1 SAR), C-band vv-polarization alone (representing ERS-1 SAR), both of the preceding channels in combination, as well as other frequency/polarization combinations. The strengths and weaknesses of the different combinations are evaluated and related to the physical interactions taking place. Subsequently, a level-2 classifier was then developed to distinguish between three different tree types. Again, a comparison was made with the various combinations, and an analysis is given of the important physical interactions. Lastly, a study was made of the importance of resolution and number of looks in regard to the use of texture for classification