Decision Tree and Texture Analysis for Mapping Debris-Covered Glaciers in the Kangchenjunga Area, Eastern Himalaya
Decision Tree and Texture Analysis for Mapping Debris-Covered Glaciers in the Kangchenjunga Area, Eastern Himalaya
复制标题
DOI：
10.3390/rs4103078
复制
发表时间：
2012-10-01
期刊：
影响因子：
5
通讯作者：
Williams, Mark W.
中科院分区：
文献类型：
--
作者：
Racoviteanu, Adina;Williams, Mark W.
在这项研究中，我们使用可见光，短波红外线和热高级星载热发射和反射辐射计（ASTER）的数据验证高分辨率Quickbird（QB）和Worldview 2（WV 2）在喜马拉雅东部的碎片覆盖测绘使用两个独立的方法：（a）决策树算法，和（B）纹理分析。决策树算法基于多光谱和地形变量，如波段比、地表反射率、ASTER波段10和12的动力学温度、倾角和海拔。决策树算法得到64 km（2.）被归类为碎片覆盖的冰，占冰川化面积的11%。总体而言，在Kangchenjunga地区的十个冰川舌，ASTER和QB地区之间的面积差异为16.2 km（2）（25%），主要是由于云和阴影造成的映射误差。纹理分析技术包括共现测量，地质统计学和过滤在空间/频率域。与其他类别相比，碎片覆盖在所有地形类别中的方差最大，熵最高，均匀性最低，例如平均方差为15.27，而云的平均方差为0，净冰的平均方差为0.06。碎片覆盖区域的纹理图像结果与决策树算法的结果相当，两种技术之间的面积差异为8%。
In this study we use visible, short-wave infrared and thermal Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data validated with high-resolution Quickbird (QB) and Worldview2 (WV2) for mapping debris cover in the eastern Himalaya using two independent approaches: (a) a decision tree algorithm, and (b) texture analysis. The decision tree algorithm was based on multi-spectral and topographic variables, such as band ratios, surface reflectance, kinetic temperature from ASTER bands 10 and 12, slope angle, and elevation. The decision tree algorithm resulted in 64 km(2.) classified as debris-covered ice, which represents 11% of the glacierized area. Overall, for ten glacier tongues in the Kangchenjunga area, there was an area difference of 16.2 km(2) (25%) between the ASTER and the QB areas, with mapping errors mainly due to clouds and shadows. Texture analysis techniques included co-occurrence measures, geostatistics and filtering in spatial/frequency domain. Debris cover had the highest variance of all terrain classes, highest entropy and lowest homogeneity compared to the other classes, for example a mean variance of 15.27 compared to 0 for clouds and 0.06 for clean ice. Results of the texture image for debris-covered areas were comparable with those from the decision tree algorithm, with 8% area difference between the two techniques.