摘要:在喀斯特分布区,基岩、植被、裸地等多种地表覆盖交错分布,地物覆盖高度异质,并且呈现出短周期规律性变化和长期动态趋势变化,单一时相的影像进行土地覆盖分类精度非常有限。针对这一问题,本文提出一种顾及物候特征的多时相遥感影像分类策略,利用具有高时间分辨率的MODIS NDVI时间序列产品作为数据源,选择BFAST(Breaks For Additive Seasonal and Trend)方法进行NDVI时间序列的物候分解,采用动态阈值法对时序分解的物候轨迹进行标记,最后将物候标记特征与原始光谱时序综合特征进行组合,选择支持向量机(SVM)分类器进行土地利用覆盖分类,并且对比了不同特征空间下的分类结果。以云南省壮族苗族自治州丘北县和砚山县为研究区进行分类实验,结果表明,BFAST模型可以有效地分解出NDVI时序中的关键物候特征,相比基于单纯光谱特征的分类,物候驱动的喀斯特断陷盆地区土地覆盖分类精度有明显的提升,在NDVI、光谱和物候组合特征空间下,土地覆盖分类精度最高,总体精度和Kappa系数分别为88.94%和0.8693,尤其在灌木林、有林地、石旮旯地与稀疏植被的区分中,SOS、POS和GSG等物候特征具有较强的可分性,表明物候特征在地物识别中的有效性。
摘要:Time-series remote sensing images were previously employed to detect land use and land-cover changes and to analyze related trends. However,land-cover change mapping using time-series remote sensing data,especially medium-resolution imagery,was often constrained by a lack of high-quality training and validation data,especially for historical satellite images. In this study,we tested and evaluated a generalized classifier for time series Landsat Thematic Mapper( TM) imagery based on spectral signature extension. First,a new atmospheric correction procedure and a robust relative normalization method were performed on time-series images to eliminate the radiometric differences between them and to retrieve the surface reflectance. Second,we selected one surface reflectance image from the time series as a source image based on the availability of reliable ground truth data. The spectral signature was then extracted from the training data and the source image. Third,the spectral signature was extended to all the corrected time-series images to build a generalized classifier. This method was tested on a time series consisting of five Landsat TM images of the Tibetan Plateau,and the results showed that the corrected time-series images could be classified effectively from the reference image using the generalized classifier. The overall accuracy achieved was between 88. 35% and 94. 25%,which is comparable with the results obtained using traditional scene-by-scene supervised classification. Results also showed that the performance of the extension method was affected by the difference in acquisition times of the source image and target image.