Application of Genetic Algorithm and Support ...

The use of multi-source remote sensing data for improved land cover classification has attracted the attention of many researchers. On the other hand, such an approach increases the data volume with more redundant information and increased levels of uncertainty within datasets, which may actually reduce the classification accuracy. It is therefore an essential, though challenging task to select appropriate features and combine datasets for classification. The combination of feature selection techniques using the Genetic Algorithm (GA) and Support Vector Machines (SVMs) classifiers has been used in various application fields in a number of studies on classification of hyperspectral data. However, the performance of this technique for classifying multi-source remote sensing data has not been well evaluated in the literature. In this study, the GA-SVM model was proposed and implemented to classify multiple combined datasets, consisting of Landsat 5 TM, multi-date dual polarization ALOS/PALSAR images and their multi-scale textural information. The performance of the proposed method was compared with that of the traditional stack-vector approach. A large number of different combined datasets were generated and classified. It is revealed that the proposed method is very efficient for handling multisource data. Results indicated that the GA-SVM approach clearly outperforms the stack-vector approach, with significantly higher classification accuracy and much fewer input features. The highest classification accuracy achieved was 96.47% with only 81 out of 189 features being selected. This study also demonstrated the advantages of using multi-source data over single source data.