Combining orbital SAR and optical data for global classification

Abstract

The potential of combining microwave and optical data for land-cover classification is explored. The data sets used in this study consist of a JERS-1 scene, an ERS-1 scene and a multispectral-mode SPOT scene. The test site is located in Raco, northern Michigan. A hybrid unsupervised/supervised classifier was applied to the SAR data alone, SPOT data alone, and several combinations of both. The level I classifier discriminates five land-cover types from water to vegetation, and then these classes are further separated into eleven categories at level II. The results indicate that the JERS-1/ERS-1 composite dominates the level I classification performance with an overall accuracy of 90.2%, compared to 84.8% for SPOT data alone. An increase was achieved by the combination of SAR and SPOT data (94.8%). Where level II is concerned, the best combination is the pair of JERS-1/ERS-1 together with the near infrared band of SPOT data. It resulted in an accuracy of 83%, 7% higher than SAR data alone and 17% higher than SPOT alone. It is concluded that well-chosen optical data (cloud-free, appropriate band) can contribute significant classification improvement when used along with SAR data.