提出了一种高光谱遥感图像半监督分类算法DE-self-training。利用少量标记样本作为初始训练集,基于改进的Self-training算法构建初始分类器,对未标记样本进行预测;然后从分类结果中按一定比例随机选取部分样本,连同其类别标记一起加入训练集中,再用扩大的训练集重新训练分类器,并对剩余的未标记样本进行预测。如此迭代地进行训练-预测-挑选样本扩大训练集过程。同时,在迭代训练过程中,运用基于最近邻域规则的数据剪辑策略对扩大训练集时产生的误标记样本进行过滤,以保证训练集的质量,不断迭代地训练出更精确的分类器,最终使所有未标记样本都获得类别标记。以AVIRIS Indian Pines和Hyperion EO-1 Botswana作为实验数据对DE-self-training算法进行测试,并与基于支持向量机的分类结果作比对。实验表明,DE-self-training算法可以在标记样本数量有限条件下,充分挖掘未标记样本的有用信息,使总体分类精度和Kappa系数都有不同程度的提高。
This study presents an investigation of the scattering and backscattering properties of the particulates in three Chinese inland lakes(the Taihu Lake, the Chaohu Lake and the Dianchi Lake) based on in situ measurements taken at 119 sites. We modeled the particulate scattering spectra using a wavelength-dependent power-law function, finding that the power-law exponents in the Taihu Lake and the Chaohu Lake differ from those in the Dianchi Lake but are similar to the values in the U.S. coastal waters. In contrast to the open ocean, the backscattering properties in the three lakes can not be determined only from chlorophyll-a concentration. The backscattering ratio spectra exhibit a wavelength dependence feature in all three lakes, generally decreasing with the increasing wavelength. Analysis results of the correlations between the backscattering ratio and the individual water quality parameters clearly show that there are distinctive relations among the three lakes, attributed primarily to different compositions of optically active materials in the three lakes. Analysis of the mass-specific scattering and backscattering coefficients shows that the coefficients at wavelength 532 nm in the Taihu Lake and Chaohu Lake are similar, but they are apparently different from those in the Dianchi Lake. Lastly, Model I multiple linear regressions were adopted to partition the mass-specific cross-sections for scattering and backscattering into organic and inorganic cross-sections to further interpret the scattering and backscattering properties. The relative contribution of organic and inorganic particulates to scattering and backscattering is clearly different among the three lakes. The scattering and backscattering properties of the particulates in the three inland lakes vary significantly based on our collected data. The results indicated that the existing semi-analytical water quality retrieval models of the Taihu Lake can not be applied perfectly to the Chaohu Lake and the Dianchi Lake.
提出了一种融合光谱和空间结构信息的高光谱遥感影像增量分类算法INC_SPEC_MPext。通过主成分分析(PCA)提取高光谱影像的若干主成分,利用数学形态学提取各主分量影像对应的形态学剖面(MP),再将所有主分量影像的形态学剖面归并联结,组成扩展的形态学剖面(MPext)。将MPext与光谱信息相结合以增加知识,最大限度地挖掘未标记样本的有用信息,优化分类器的学习能力。不断从分类器对未标记样本的预测结果中甄选置信度高的样本加入训练集,并迭代地利用扩大的训练集进行分类器构建和样本预测。以不同地表覆盖类型的AVIRIS Indian Pines和Hyperion EO-1Botswana作为测试数据,分别与基于光谱、MPext、光谱和MPext融合的分类方法进行比对。试验结果表明,在训练样本数量有限情况下,INC_SPEC_MPext算法在降低分类成本的同时,分类精度和Kappa系数都有不同程度的提高。
