IMPROVEMENT OF HYPERSPECTRAL IMAGE CLASSIFICATION USING GENETIC ALGORITHM FOR FEATURE SELECTION AND SVMs PARAMETERS OPTIMIZATION

In the recent decades hyperspectral remotely sensed data present the excellent potential for information extraction of the earth surfaces. These observations are normally sampled in several hundred narrow and continuous bands. Such high dimensionality of data provides more discrimination ability in classification task, but also imposes high computational cost and complexity in data modeling. In particular, because of inadequate number of training samples related to dimension of data, the proof of “curse of dimensionality” may occur. Therefore, as a necessary pre-processing step the data reduction techniques, are indispensable to perform by feature selection. Feature selection methods are divided into two categories: the wrapper methods and the filter methods. In the first methods, the “goodness” of selected feature is estimated based on classification accuracy, whereas the second methods use an independent criterion to find the most proper subset of features. This paper suggests a framework to combine filter and wrapper feature selection methods to find the optimal or the near optimal feature subset and optimize the Support Vector Machines (SVM) kernel parameters at the same time. The proposed approach includes two steps: in the first step the filter method selection is applied. Then, using four different filter feature selection methods, a feature pool is produced. In the second step, a single objective genetic algorithm, as a wrapper method, searches previous resulted feature pool to find the final feature subset and optimize SVM’s parameters. The Genetic Algorithm (GA) is a global optimizer which belongs to the randomized heuristic search techniques. Based on the Darwinian principle of ‘survival of the fittest’, the GA obtains the optimal solution after a series of iterative computations. The suggested algorithm is applied to the Hyperion images acquired over the Okavango Delta in Botswana. The results showed the great improvement in classification accuracy and the significant decrease in feature subset. The number of bands used for classification is reduced from 145 to 16, while the classification accuracy increased from 89.87% to 94.11%.