基于特征优选随机森林算法的GF-2影像分类

Research on GF-2 Image Classification Based on Feature Optimization Random Forest Algorithm

基于随机森林算法(RF,RandomForest)对“高分二号”(GF-2)卫星遥感数据进行面向对象地表信息提取时存在如下不足:1)有限的光谱波段导致随机森林可选特征变量受限,影响分类器性能;2)面向对象影像分割尺度以经验判别为主,缺少定量化的判定标准。为了克服上述问题,文章提出了一种优化特征空间的随机森林分类算法。首先根据面向对象分割的理论方法,引入方差变化率,获取研究区影像的最优分割尺度;然后利用随机森林–平均精度减少模型(RF-MDA,Random Forest-Mean Decrease in Accuracy)与K折交叉验证算法(K-CV,K-Cross Validation),进行特征重要性排序并优化特征空间;最后,基于不同特征组合的随机森林分类算法进行面向对象分类,并对分类结果进行对比分析。结果表明,改进的基于特征优选随机森林分类算法的总体精度和Kappa系数分别为93.44%和0.928,优于原始RF算法。该方法能够有效提高GF-2卫星遥感影像在土地利用分类方面的精度,可为国土监测和管理提供技术支持和理论指导。

To overcome the following boundaries of random forest object-based classification for high-resolution remote sensing images,that 1) limited spectral bands of high spatial resolution remotely sensed data has restricted the performance of random forest;2) Segmentation scale of object-oriented method is based on empirical discrimination,which lacks quantitative criteria.In this paper,a random forest classification algorithm with optimized feature space is proposed.Firstly,according to the theory and method of object-oriented segmentation,the variance change rate is introduced to obtain the optimal segmentation scale of the image in the study area.Then,the Random Forest-Mean Decrease in Accuracy(RF-MDA) model and K-Cross Validation(K-CV) are used to rank the feature importance and optimize the feature space.Finally,the random forest classification algorithm based on different feature combinations is used for object-oriented classification,and the classification results are compared and analyzed.The results show that the overall accuracy and kappa coefficient of the improved random forest classification algorithm based on feature optimization are 93.44% and 0.928 respectively,which are better than the original RF algorithm.This method can effectively improve the accuracy of GF-2 remote sensing image in land use classification,and can provide technical support and theoretical guidance for land monitoring and management.