Feature Extraction and Classification for Remote Sensing Imagery Based on Orthogonal Frequency Division Method

Abstract

Remote sensing (RS) is the unique way to monitor whole world dynamically. Feature extraction and classification for RS image is an important research direction in this field. Image classification is based on feature extraction, and is also the premise of image processing. However, at present, image classification lacks theoretical basis and is still in the stage of visual observation and empirical interpretation. There are many reasons for poor image classification, e.g. lack of clarity or lack of classification basis. In this article, a new method for RS image feature extraction and classification called the “orthogonal frequency division” (OFD) method is proposed. The method combines filtering and stratification, and the experiments prove that it can effectively classify and extract the features of RS images. The traditional filtering methods regard image contents as a whole signal space, after each filtering, only the high frequency values can be separated in the form of edges, while the low frequency values still remains in the signal space, forming an image which pixel values are doubled. If we want to label a region, we have to superimpose these edges, so we need to do filtering for many times, which is a tedious process. Moreover, the scaling ratio is usually small and fixed, therefore, if the differences among regions are large, multiple filtering must be carried out, otherwise the filtering effect is not obvious. Then, although the image compression effect is achieved, the image will become too small and lose its ability to be recognized. The OFD method only needs to be filtered once, its filter group can classify both low and high frequency values and can increase the differences among regions, so that the image contents form layers, different classification results can be extracted by different thresholds. These classification results can be used as the material of neural network training set, at this point, the classification result regions can be labeled directly. By comparing the classification results with some prevailing image processing methods (such as PCA method or wavelet transform method), the classification effect is superior to the traditional methods in terms of classification accuracy.