Remote Sensing Image Classification Based on Improved Fast Independent Component Analysis

The increasing requirement of classification categories is followed by the increasing probabilities of wrong classification and the decreasing classification speed. If we can separate certain types of pixels out in advance, and then classify the remaining pixels, we can reduce the probabilities of mistakes effectively. This paper proposed an improved Fast Independent Component Analysis (ICA) based remote sensing image classification algorithm. Firstly we analyzed the core iterative process of Fast-ICA algorithm, and adopted adaptive step size control in our search strategy, thus avoid large number of iterations caused by too small or too large step. Secondly, due to the initial value of Fast-ICA algorithm effects the results very much, a favorable initial matrix was selected before our iterative process. Next we use the improved algorithm to separate out certain types of pixels in advance, in such a manner to simplify the following classification. At last we compared the results of this algorithm with general Fast-ICA algorithm、principal component analysis (PCA) and ratio transformation. The experiment result shows the effectiveness of using this algorithm in image classification.