针对高空间分辨率的遥感影像,提出了一种基于多尺度分割的变化检测算法。采用Mean-Shift分割算法对影像进行多尺度分割,构建了不同尺度上的地理对象,以不同尺度上的地理对象灰度均值构建了变化检测的多尺度特征向量,采用变化矢量分析法获得最后的变化检测结果。以城镇区和农田区的QuickBird影像对本文算法进行了检验,从精度评价的效果来看,无论城镇区还是农田区,采用面向对象的变化检测方法精度都高于基于单像素的检测方法,且当尺度层数固定时,多尺度组合的变化检测结果优于单一尺度的变化检测结果,对城镇、农田区域的变化检测的精度分别达到87.57%和81.55%。本文算法既可以顾及大面积同质区域变化,又可以反映小的地物目标及边缘部分的变化,能够很好地满足城镇、农田等不同环境背景下的变化检测需求,在国土资源监测中具有一定的应用价值。

Change detection determines changes in multitemporal images, which are widely used in deforestation, land use, and urban expansion, among others. Nonetheless, traditional pixel-based change detection methods cause confusion when used on high-spatial images, and generat salt-and-pepper noises on the changed map because of the presence of heterogeneous objects at a pixel level. The useful spatial or contextual information regarding the values of proximate pixels is typically ignored in a pixel-based method; therefore, an object-based method is a new approach to solve these problems in high-spatial resolution images. This study proposes a multilevel object-based method to detect changes in such images. First, we utilize the mean-shift segmentation method to segment the image and consider the heterogeneity of ground objects. Geographic objects are acquired from the segment results at different levels through multilevel information. Then, we combine the gray data of each geographic object at different scale levels to build a feature vector. The change vector analysis method is used to construct an intensity difference map at the multitemporal phase. The change results are generated by applying the expectation maximization algorithm to automatically obtain the thresholds for changed and unchanged areas; moreover, the multiscale change detection algorithm is verified with QuickBird images of urban and rural areas. The pixel-based, single scale level object-based, and multiscale level object-based methods are also compared with one another based on these two datasets. Results show that the change detection accuracy of the object-based change detection method is higher than that of the pixel-based approach in both urban and rural areas. When the total scale level number considered in the change detection method is fixed, the multiscale, object-based method always performs better than the single level object-based method. The change detection accuracies in the urban and rural areas are 87.57% and 81.55%, respectively. Furthermore, the qualitative analysis findings related to the change-detection maps suggest that the proposed technique induces high fidelity in both homogenous and small or border regions. Thus, the proposed algorithm can satisfy the requirements of change detection in urban and rural areas, which benefits land and resource monitoring.