基于E0-1 Hyperion影像地物识别与分类不同方法的效果比较

Comparison of Land Use Identification and Classification using Different Models for E0-1 Hyperion Images

以广州市大学城为研究区域裁剪出Hyperion影像,直接从影像提取地物端元,采用光谱角匹配、无约束和全约束的线性混合像元分解等3种方法进行地物识别与分类,分别生成地物分类图;选用Kappa系数和总精度两个参数,对比不同方法的地物识别效果.结果显示,分类图像的总精度均大于60%,Kappa系数在0.61—0.80,表明将高光谱影像用于地物识别与分类,可获得好的地物识别与分类结果.其中,光谱角匹配和全约束的线性混合分解法得到的两个参数明显更大.就地物类型而言,3种方法都对自然地物有较好的识别与分类效果,如林地和草地的用户精度高于80%,裸地和水体的值也高于50%.对人工地物的高反射率表面也有好的识别效果,用户精度达100%;对老建筑屋顶以及低反射率表面的识别效果不佳.尽管如此,光谱角匹配和全约束的线性混合像元分解法对于识别建筑屋顶的水泥混凝土表面表现出优势,用户精度大于71.4%.

Based on E0-1 Hyperion data,images which enclose the Higher Education Mega Center were cut,and endmembers for classification objects were picked up directly from images. Spectrum angle matching and mixed pixels linear decomposition method including non-constrained and condition restriction were applied for land cover identification and classification, and classification maps were created. Two parameters, Kappa coefficient and total accuracy, were chosen to compare the effects for object identifying. Total accuracy from the three classification maps was higher than 60%, while Kappa coefficients fell between 0. 61-0. 80. And the results indicated three classification methods were adoptive for E0-1 Hyperion images. However, spectrum angle matching and mixed pixels linear decomposition method with condition restriction were more precise than mixed pixels linear decomposition method with non-constrained. Two parameters were different among different land cover classifications. Three methods were fit for natural surface identification and classification, such as user accuracy, for forest and grassland surface, were higher than 80%, while that for bare land and water body were still higher than 50%. It was also found that artificial and high reflectance surface could be classified with accuracy 100% ,while it could not acquire higher value to cement concrete old roofs and low reflectance surface. However, spectrum angle matching and mixed pixels linear decomposition method with condition restriction have obvious advantage with accuracy higher than 71.4%.