基于端元可变的喀斯特地形复杂区石漠化信息提取

喀斯特山区因地形复杂、地表破碎等特点使得遥感影像中阴影、混合像元及光谱变异现象普遍存在，传统基于多光谱遥感的像元二分法（DPM）在光谱变异和阴影显著的区域难以准确的对喀斯特石漠化（KRD）信息进行提取。采用高光谱遥感的混合像元分解技术可将复杂的混合像元分解为纯净的地物光谱与各地物光谱对应的混合比例，为复杂山区获取更高精度的石漠化信息提供可能。然而，由于光照、环境及大气等诸多因素的变化会引起端元发生不同程度变异，导致在混合像元分解过程中出现显著的误差，其次要从地形复杂、地表异质性强的山区影像上直接获取地物纯净光谱建立用于应对光谱变异的光谱库极其困难。因此，如何在这种情况下应对光谱变异和地形效应，获取有效、准确的对石漠化信息进行提取是当前研究的重点。针对以上问题，采用通过模拟由光照条件造成的地物反射率变化，并考虑每个波长间隔光谱变异情况的广义线性混合模型（GLMM），以减轻喀斯特地区石漠化信息提取过程中光谱变异与地形效应的影响。首先，从GF-5高光谱影像中提取喀斯特地区主要地物（植被、裸岩、裸土）的典型代表性光谱，然后基于提取的地物光谱模拟不同光照下每个像元光谱的变异情况，选择最适合的光谱组合对像元进行分解，得到最优的解混效果。为了验证方法的可靠性，利用高分辨率影像目视解译的结果作为参考对方法预测结果进行验证，同时选择未考虑端元变异的全限制最小二乘法（FCLSU）和DPM进行对比。结果表明，在地形高度复杂的喀斯特山区，考虑阴影、混合像元及光谱变异是必要的，GLMM在石漠化信息提取中总精度达到了84.89%，明显高于其他两种方法的59.68%和67.34%。通过对光照区和阴影区分别进行精度检验，发现GLMM在光照区与阴影区有着相似的精度表现，而另外两者则差异较大，阴影区明显低于光照区。这反映GLMM能较为有效地减轻地形效应的影响，对喀斯特石漠化信息提取的精度有一定提升。

Rocky Desertification Information Extraction in Karst Terrain Complex Area Based on Endmember Variable

Shadows, mixed pixels and spectral variations are common in remote sensing images in mountainous karst areas due to complex terrain and broken surface. Dimidiate pixel model (DPM) based on multispectral remote sensing is difficult to accurately extract rocky karst desertification (KRD) information in areas with significant spectral variations and shadows. The mixed pixel decomposition technology of hyperspectral remote sensing can decompose complex mixed pixels into the mixed ratio corresponding to the pure landmark spectrum and each landmark spectrum, which provides the possibility for obtaining higher precision rocky desertification information in complex mountainous areas. However, due to the changes in many factors such as illumination, environment and atmosphere, the end members will vary to varying degrees, which will result in significant errors in the process of mixed pixel decomposition. Secondly, it is difficult to directly obtain the pure landmark spectrum from mountain images with complex terrain and strong surface heterogeneity and establish a spectrum library to deal with spectral variation. Therefore, the focus of current studies is how to deal with spectral variation and terrain effect in this case and obtain effective and accurate information extraction of rocky desertification. In order to solve the above problems, the generalized linear mixed model (GLMM), which simulates the reflectivity change of ground objects caused by illumination conditions and considers the spectral variation at each wavelength interval, was adopted to reduce the influence of spectral variation and terrain effect in the process of information extraction of rocky desertification in karst areas. First of all, the typical representative spectra of main ground objects (vegetation, bare rock and bare soil) in the karst area were extracted from GF-5 hyperspectral images. Then the spectral variation of each pixel under different illumination was simulated based on the extracted landmark spectrum, and the most suitable spectral combination was selected to decompose the pixels to get the best unmixing effect. In order to verify the reliability of the method, the visual interpretation results of high-resolution images were used as a reference to verify the prediction results of the method, and the fully constrained least squares linear spectral unmixing (FCLSU) DPM without considering end-member variation were selected for comparison. The results showed that it was necessary to consider shadows, mixed pixels and spectral variation in karst mountainous areas with highly complex terrain. The total accuracy of GLMM in rocky desertification information extraction reached 84.89%, significantly higher than that of the other two methods (59.68% and 67.34%). The accuracy of GLMM in the illumination area and shadow area was similar to that of GLMM in the illumination area and shadow area. However, the other two were quite different, and the shadow area was lower than the illumination area, which reflects that GLMM can effectively reduce the influence of terrain effect and improve the accuracy of information extraction of rocky karst desertification.