基于ReliefF和PSO混合特征选择的面向对象土地利用分类

Object based land-use classification based on hybrid feature selection method of combining Relief F and PSO

针对面向对象土地利用分类存在特征维数过高的问题,提出了一种结合Relief F和粒子群优化算法(particle swarm optimization,PSO)的混合特征选择方法,即首先利用Relief F作为特征预选器滤除相关性小的特征,然后以PSO作为搜索算法,以支持向量机(support vector machine,SVM)的分类精度作为评估函数在剩余特征中选择出最优特征子集。该文以吉林省长春市部分区域为研究区,采用Landsat8遥感影像为数据源,首先对其进行多尺度分割,然后提取影像对象的光谱、纹理、形状和空间关系特征,利用提出的混合特征选择方法选取最优特征子集,最后使用SVM分类器对研究区进行土地利用分类,总体分类精度和Kappa系数分别为85.88%和0.8036,与基于4种其他特征选择方法的土地利用分类结果进行比较,基于Relief F和PSO的混合特征选择方法利用最少的特征获得最高的分类精度,能够有效地用于面向对象土地利用分类。

In recent years, object-based methods have been increasingly used for the land-use classification of remote sensing data. However, the availability of numerous features with object-based image analysis renders the selection of optimal features. In this study, a hybrid feature selection method that combined filter approach and wrapper approach was proposed. In the filter approach, the Relief F algorithm was employed to select features that had the higher relevance with land-use classes. The wrapper approach used the particle swarm optimization （PSO） algorithm as a search method and the classification accuracy of support vector machine （SVM） as an evaluator to search for an optimal feature subset from the selected features. The objective of this research was to examine the effectiveness of the proposed feature selection method on object-based classification. The study site was located in the southeastern part of Changchun City, Jilin Province. A Landsat8 image acquired on July 15, 2014 was selected as data source for this classification. To begin with, image objects were delineated by implementing multi-scale segmentation on the Landsat8 image. Second, a total of 95 features were extracted from the Landsat8 image. Third, the proposed hybrid feature selection method was employed to search for an optimal feature subset. In the first stage of the feature selection, the Relief F algorithm was applied to select 50 features that had the higher relevance with land-use classes in Weka 3.6. In the second stage, the PSO algorithm was used to optimize the kernel parameters of SVM simultaneously with the feature selection in Matlab 2010b. As a result, an optimal feature subset of 22 features was obtained. Finally, based on the selected features, land-use classification was performed using SVM classifier embedded in Definiens Developer 9.0. Using the confusion matrix that was determined on the basis of the visual interpretation map of Google Earth high-resolution remote sensing images, we calculated 4 statistical items for validation： overall accuracy, Kappa value, producer＆#39;s accuracy and user＆#39;s accuracy. The overall accuracy was 85.88% and the Kappa value was 0.8036. Among producers, several classes （such as water, cultivated land and forest） achieved over 90% classification accuracies, while building and grassland yielded over 85% accuracies. The lowest classification accuracy was 66.09% for road due to its mixing with building easily in city. At last, land-use classification was performed using SVM classifier. Then, a series of comparisons between the proposed method and the 4 other feature selection methods （i.e., Relief F algorithm, wrapper approach of combining PSO and SVM, all features, feature space optimization （FSO） algorithm） were made to verify the effectiveness of the proposed method further. The results of the comparisons showed that the proposed method obtained the highest classification accuracy among all the feature selection methods. The overall accuracy and the Kappa value of the classification using the Relief F feature selection method were 78.90% and 0.7079, much lower than the proposed method. The accuracy of the classification using the wrapper approach of combining PSO and SVM was similar with the proposed method. However, the proposed method was more efficient to select features than the wrapper approach. The classification that used all features obtained the lowest overall accuracy and Kappa value of 48.25% and 0.3225. The overall accuracy and the Kappa value of the classification using the FSO feature selection method declined by 10.53% and 0.1445 respectively compared with the classification results using the proposed method. The comparisons indicate that the proposed hybrid feature selection method of combining Relief F and PSO can be effectively applied to object-based land-use classification.