基于对象分类的遥感影像森林变化检测方法

Forest change detection method for remote sensing images based on object classification

针对湖北省宜昌市点军区森林变化检测应用需求,采用两期高分二号(GF-2)数据进行对比分析。定义了基于邻域差分绝对值与标准差比的多尺度分割评价函数,用来确定对遥感图像分割的分割尺度、形状因子以及紧凑度。通过试验,利用神经网络分类方法确定了基于对象分类的最优特征组合,并采用基于对象的最近邻(k NN)分类方法对遥感图像进行分类,最后对两期遥感影像分类结果中的森林类别进行变化检测。结果显示,在分类过程中,基于对象的分类总体精度为0.9866,Kappa系数为0.9752,高于神经网络和最大似然分类方法。在以森林地为主的丘陵地带变化检测应用中具有较好的适用性。

For the application of forest change detection in Yichang,Hubei Province,two-phase Gaofen-2(GF-2)remote sensing image data were used for comparison and analysis.Based on the absolute value and standard deviation ratio of neighborhood difference,a multi-scale segmentation parameter evaluation function was defined to determine the segmentation scale,shape factor,and compactness in image segmentation.The optimal feature combination based on object classification was then determined using the experimental neural network classification method,and the object-based k-nearest neighbor classification method was employed to classify remote sensing images.Finally,the forest category in the classification results of the two-phase GF-2 remote sensing images was detected.The results showed that the overall accuracy of object-oriented classification was 0.9866,and the Kappa coefficient was 0.9752,which were both higher than those of the neural network and maximum likelihood classification methods.Thus,the object-oriented classification method can be effectively used in forest change detection applications.