首页 > , Vol. , Issue () : 1993-2002
为了实现两个不同年份单时相遥感影像之间的土地覆盖变化检测，提出了一种基于土地覆盖类型特征自适应确定阈值的遥感影像变化检测方法。本文以2015年土地覆盖数据为基础，综合2013年和2015年Landsat8-OLI影像数据，首先，采用时相不变点群法（temporally invariant cluster, TIC）保证了两期影像辐射水平的一致性。其次，对两期影像进行多尺度分割，并在各级尺度下构建分割对象的变化向量。然后，采用最大类间方差的方法分别进行单一变化阈值变化检测以及基于土地覆盖类型的多阈值变化检测分析，并利用目视解译样点进行精度验证与评价。结果表明：（1）单一阈值变化检测结果的总体精度为79.6%，Kappa系数为0.601，多阈值变化检测结果的总体精度为87.2%，Kappa系数为0.741，多阈值变化检测具有更高的精度。（2）进一步逐土地覆盖类型精度评价可知，多阈值变化检测能在一定程度上减弱了物候期的影响，具有更高的稳定性。该研究以土地覆盖数据为底图，逐类别的选取变化检测阈值，提高了变化区域检测的精度，在大范围高效更新土地覆盖数据的应用中具有一定的参考价值。
Change detection with single phase remote sensing image between two different time is widely used in land cover, urban expansion, coral reef health, forest fire events and deforestation. And the most important step in change detection is to determine the change threshold value which is used to distinguish change area and no-change area. However, traditional change detection methods usually just determine one threshold. They neglect the difference of spectral value range between different land cover types. The same land cover type may have great difference. For example, the areas of farmland which have been harvest is totally different from the area which have not been harvest. So, this paper proposes an adaptive multi threshold values remote sensing image change detection method which is based on land cover type feature. This paper collects the land cove data for 2015 and two Landsat8-OLI images for 2013 and 2015. Firstly, it uses the method of temporally invariant cluster(TIC) to ensure the consistency of the radiometric level of the two images. Then, in order to avoid the salt-and-pepper noises, we segment the remote sensing image with multiscale segmentation algorithm. The segmentation spatial scales are 200,150,100, which are used for different land cover types. Change vectors of the image objects at different segmentation spatial scales are then constructed. The maximum inter-class variance method is used to determine the change detection single threshold value and multi threshold values which are based on land cover types. Finally, we collect 500 samples by visual interpretation and then we conduct accuracy assessment on the result of single threshold value change detection and multi threshold values change detection. The experiment outcomes show that the multi threshold values change detection method has higher accuracy and greater stability. The total accuracy of the multi threshold values change detection is 87.2%, whereas the total accuracy of the single threshold value change detection is 79.6%. The Kappa coefficient is 0.741 and 0.601, respectively. In order to further compare the proposed multi threshold values method with the traditional single threshold value method, we conduct the further accuracy assessment with each land cover type. It shows that the producer’s accuracy of the no-change area in farmland, water, developed land and barren improve obviously. The user’s accuracy of the change area in farmland and water also improve. The multi threshold values change detection method weakens the influence of phenology phase to some extent and has better applicability. The TIC relative radiometric normalization method overcome the shortcomings of traditional visual interpretation for selecting time-invariant pixels. It avoids the influence of subject factors and can normalization images accurately and efficiently. In addition, the multiscale segmentation algorithm can provide different spatial segmentation scales to avoid over segmentation and under segmentation problems. The proposed method in this study involves change vector analysis driven by different thresholds based on land cover type rather than a single threshold value. It has improved the accuracy of the change detection and can provide a reference for the application of efficiently updating of land cover data in the large scale area.