基于最优尺度的生产建设扰动图斑识别

On the identification of construction disturbance patches based on optimal segmentation scale

遥感识别生产建设扰动图斑对水土保持监测具有重要的意义。本文尝试提出一套基于高分辨率影像自动获取生产建设扰动图斑的方法,重点探讨扰动图斑最优分割尺度的自动获取方法,以解决当前过度依赖分类者的多次尝试和主观判断问题。以高分1号影像为例,利用局部方差方法和目标函数方法计算生产建设扰动图斑的最优分割尺度,结果显示,与目视判读的最优分割尺度对比,目标函数方法的结果精确,局部方差方法的结果略微偏大。在此基础上,利用面向对象的监督分类方法,实现了对扰动图斑的识别。利用现场调查的结果验证,本文中的方法较为可靠,制图精度和用户精度分别达到86.3%和84.2%,高于其他分割尺度的识别精度。

[Background] Identifying construction disturbance patches is significant to soil and water conservation. With high spatial-resolution imagery prevalent,object-based image analysis（ OBIA） is proposed because it can offer substantial information to overcome the infamous ＂salt-and-pepper＂speckle and phenomenon of ＂different objects with same spectrums＂or ＂different spectrums with same objects＂associating with pixel-based classification. Image segmentation and classification are two main steps in OBIA,wherein the former is foundation. Yet there has challenges of estimating appropriate scale parameters,known as optimal scale parameters（ OSP）,for an interest land cover. Although dozens of algorithms have been proposed,none of them were used in construction project recognition. [Methods]Two widespread algorithms of OSP were analyzed and evaluated. One is the local variance（ LV） methodwith the conception that the rate of change of LV（ Roc_ LV） can capture the variation of object heterogeneity within a scene. The other is an objective function method with taking into account both of object＇s internal homogeneity and external distinguishability. The calculated OSP from LV method and objective function method were validated by two items,i. e.,artificial sketched OSP according to field survey,and the identifying accuracy of construction disturbance patches was conducted by an objectoriented supervised classification method. [Results]Image segmentation was implemented in e Cognition software by the scale of 30 to 500 with an interval of 10. According to the UAV image and ground survey,300 was manually judged as the OSP that unambiguously distinguished boundary between construction disturbance patches and other land use type. This artificially determined OSP was used as‘real＇value to evaluate LV method and objective function method. The OSP calculated by LV method and objective function method were 310 and 300,which were slightly higher than and equal to the ‘real ＇value,respectively. Then,construction disturbance patches of each segmented image from scales of 150 to 390 were identified by an object-oriented supervised classification method and interpretation keys,and their identifying accuracies were calculated by the identification according to the UAV image,statistic data and ground survey. The result indicated that scale of 300（ i. e.,the OSP） had the highest accuracy with the producing accuracy and user accuracy of 86. 3% and 84. 2%,respectively,while other ＇s producing accuracy was less than 85% and user accuracy was lower than 82%. This result illustrated the consistency between OSP and construction disturbance patches identifying accuracy. [Conclusions]Two conclusions are reached. 1） Compared to LV method,the objective function method is recommended to calculate the OSP of construction project in given study area and GF-1 image. 2） The identification process based on objective function method and object-oriented supervised classification method is proposed because by which the better accuracies than other segmental scales was obtained.