Optimal selection of remote sensing feature variables for land cover classification

With the abundance of spectral information and texture information of remote sensing images, the feature variables applicable to land cover classification also increase. But the high-dimensional features will deteriorate the classifier performance. In this paper, an improved selection method was proposed by considering separability and redundancy between feature variables. The study area - You County was divided into eight land cover types based on the SPOT-5 remote sensing image in 2009 combined with the continuous inventory data of forest resources during the same period. A total of 3,921 sample plots containing classification information were selected, of which 2/3 were used as training data, 1/3 as testing data. Waveband operations and texture extraction were performed on the images and a total of four bands information, eight texture factors, and nine vegetation indices were extracted as candidate feature variables subset. Calculate the weighted average of the Jeffries-Matusita (JM) distance between different classes of each feature variable, as the measure of class separability, and select the variable ordered first. Then calculate the PEARSON correlation coefficient between the remaining variables and the selected variables as the measure of variables redundancy. The ratio of separability and redundancy is referred to as the improved JM distance, which is successively substituted into the Bayesian classifier in descending order. The results show that (1) From the JM distance, the spectral features are more suitable for the land cover classification in the study area than the texture features. (2) The overall classification accuracy can be improved through the feature variables selected by improved JM distance. (3) The number of feature variables using the improved JM distance to achieve stable classification accuracy is far less than traditional JM distance, which effectively reduce feature dimensions. Therefore, the improved selection method proposed in this paper not only retains the class separability but also avoids the information redundancy, and can effectively select the feature variables.