CAJ | 학술논문

传统的高光谱分类通常仅考虑单一像元的光谱或纹理特征，分类后容易出现地物破碎的现象。鉴于此，本文提出了一种面向对象的混合分类方法，将面向对象的分割结果与传统的像元级分类结果进行有机融合，充分利用对象的光谱特征和空间结构特征。在此基础上，引入了2种具体的混合分类方法，即多尺度分割的SVM分类和多波段分水岭分割的SVM分类。前者将地物光谱的可变性进行弱化处理，转化为多尺度均质对象单元进行分类；后者融入了地物的空间信息和形态学特征，对分割得到的同质区域进行分类。将这2种分类方法应用于航空高光谱数据，实验结果表明：面向对象的混合分类方法的总体精度分别为92.63%和96.13%，与传统的像元级分类法相比，分别提高了10.14%和13.64%，有效地解决了分类后地物的破碎现象。
전통적고광보분류통상부고필단일상원적광보혹문리특정，분류후용역출현지물파쇄적현상。감우차，본문제출료일충면향대상적혼합분류방법，장면향대상적분할결과여전통적상원급분류결과진행유궤융합，충분이용대상적광보특정화공간결구특정。재차기출상，인입료2충구체적혼합분류방법，즉다척도분할적SVM분류화다파단분수령분할적SVM분류。전자장지물광보적가변성진행약화처리，전화위다척도균질대상단원진행분류；후자융입료지물적공간신식화형태학특정，대분할득도적동질구역진행분류。장저2충분류방법응용우항공고광보수거，실험결과표명：면향대상적혼합분류방법적총체정도분별위92.63%화96.13%，여전통적상원급분류법상비，분별제고료10.14%화13.64%，유효지해결료분류후지물적파쇄현상。
Hyperspectral imagery generally contains hundreds of contiguous narrow bands, which could provide detailed spectral information for target detection and image classification. Traditional hyperspectral classification fails to generate expected results since it simply considers spectral or textural properties at the pixel scale in the context of natural complexity. In this article, a hybrid classification method was proposed which takes full advan-tages of spectral and spatial features by fusing object-based segmentation results with traditional per-pixel classi-fication results. Based on this concept, two specific hybrid classification approaches were employed:(1) the mix-ture of multi-scale segmentation and SVM classification and (2) the mixture of multi-band watershed segmenta-tion and SVM classification. In the first proposed method, spectral variations were attenuated by converting them into homogenous image objects at multiple scales;while, the latter method aggregates spatial information and morphological profiles into the segmented objects to achieve the homogeneous classification. The two classi-fication algorithms were applied to airborne hyperspectral imagery and the results show that the overall accuracy based on traditional pixel-wise classification reaches about 82.49%, relatively lower compared with the hybrid object-based classification methods, which are 92.63%(Method 1) and 96.13%(Method 2) respectively. In addi-tion, Method 2 performs better than Method 1 since it produced a smoother boundary, partly because Method 2 needs less user-defined parameters, and the iterative“trial-and-error”of which may affect the classification re-sults. In conclusion, this study demonstrates that the hybrid of object-based classification is a significantly more robust approach than the traditional per-pixel classifier. The proposed method overcomes the spectral confusion, solves the problem of land fragmentation, and provides a solution to map complex environments accurately.