基于多尺度分割和决策树算法的山区遥感影像变化检测方法——以四川攀西地区为例

Change detection of remote sensing images based on multiscale segmentation and decision tree algorithm over mountainous area: a case study in Panxi region,Sichuan Province

山区遥感影像变化检测面临地形效应明显、空间异质性高等不利因素的影响,构建适用于复杂地形条件的变化检测方法一直是遥感应用研究的难点。在对影像进行多尺度分割的基础上,构建对象的多种光谱、形状及地形特征,将地形阴影、物候等造成的虚假变化当作"未变化"训练样本,利用决策树算法自动提取检测规则,建立复杂地形条件下面向对象的遥感影像变化检测方法,并将该方法用于四川攀西山区1989年和2009年TM影像的检测试验,最后对方法的不足和改进措施进行了讨论。主要结论包括:(1)文中构建的方法能够有效减弱山区复杂地形条件对遥感影像变化检测的不利影响,采用地面调查数据和分层随机采样的总体验证精度为93.57%,Kappa系数为0.8706。(2)C5.0决策树算法对于只有"变化"和"未变化"两种类别且同类间训练样本高度异质化的影像分类仍能取得较好的效果,具有较好的鲁棒性和适应能力。通过将地形、物候等引起的虚假变化当作"未变化"训练样本可以有效提高检测精度。(3)光谱特征仍是TM影像遥感分析的主要信息源,影像间NDVI的差值对于植被覆盖区域土地覆盖格局变化的检测具有良好的表征作用。(4)攀西地区1989—2009年间土地覆盖格局变化明显且与人类活动关系密切,典型的驱动方式包括退耕还林(草)工程、水利工程建设和矿山开采等。共检测出变化面积740.2km2,占影像总面积的2.49%。

In mountainous area,many factors such as the high spatial heterogeneity in land cover conditions and the topographic effect on remote sensing data acquirement significantly constraint the change detection by using remote sensing method. It has become a great challenge to develop an effective method to accurately map the land cover changes for complexterrain areas. In the traditional change detection methods,the spectral features are the major information. Due to the lack of the consideration of other information from structure, topography and other parameters, many false changes will be introduced in the final detection results with a big uncertainty. To avoid the false change and overcome the influence from terrain shadow and phenology differences, an object-oriented change detection method by combining multi-scale segmentation and decision tree algorithm for complex terrain condition is proposed in this study. In the new method,firstly,a series of spectrum information（ i.e. reflectance,change vector intensity,NDVI）,shape information（ i. e. area,lengthwidth ratio,shape index） and terrain parameters（ i.e. elevation,slope,aspect） are calculated for change detection rules construction. Secondly,the unchanged surfaces which are usually detected as the false changes due to the influence of terrain shadow and phenology difference are extracted as training sample. Finally,the change detection rules are built automatically by C5.0 decision tree algorithm. the proposed method is applied to Landsat-TM images acquired in 1989 and2009 at Panxi region,the detection results indicate：（ 1） Over the mountainous area,the method can effectively reduce the influence from topographic effect,and has a big improvement on detection accuracy with the overall accuracy of 93.57% and Kappa coefcient of 0.8706 by validating ground reference samples and stratified random samples.（ 2） The C5.0 decision tree algorithm performs well in terms of robustness and adaptability,which serves well in land cover classification with types only labeled as ＂changed ＂and ＂unchanged ＂and highly heterogeneous training samples. Considering the false changes result from topographic and phenological as ＂unchanged ＂training sample,the accuracy of change detection is improved effectively.（ 3） Spectral features are the predominant information in analyzing TM images,and the NDVI differences between the reference image and detected image is an effective indicator for change detection of land cover pattern in vegetation coverage region.（ 4） The changes of land cover pattern in Panxi region from 1989 to 2009 were closely associated with human activities. The typical driving forces of pattern change include returning farmland to forest policy,water conservancy project,mining,urban sprawl and so on. The changed area amounted to 740.2 km2,accounting for 2.49% of the total detected area.